Add Batch 3c2b4e72-1172-49ce-a541-ca9aba3324a4 data

.gitattributes

@@ -6233,3 +6233,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/Narrative[[:space:]]Style[[:space:]]and[[:space:]]the[[:space:]]Spread[[:space:]]of[[:space:]]Health[[:space:]]Misinformation[[:space:]]on[[:space:]]Twitter/a8c8cc18-8e7c-4dd9-9691-b66d08e7c797_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/NarrativeXL_[[:space:]]a[[:space:]]Large-scale[[:space:]]Dataset[[:space:]]for[[:space:]]Long-Term[[:space:]]Memory[[:space:]]Models/b1dd1ac9-8661-49b2-8615-29d8439200cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Natural[[:space:]]Language[[:space:]]Annotations[[:space:]]for[[:space:]]Reasoning[[:space:]]about[[:space:]]Program[[:space:]]Semantics/bf035ad6-88af-4241-97db-5a095465958d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Natural[[:space:]]Response[[:space:]]Generation[[:space:]]for[[:space:]]Chinese[[:space:]]Reading[[:space:]]Comprehension/deacd081-229d-47a6-82a1-c3842e05b39b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Neuro-Symbolic[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]with[[:space:]]Dynamic[[:space:]]Word[[:space:]]Sense[[:space:]]Disambiguation/f2ea77f7-c2a2-49e4-b922-17fa39103dc4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/New[[:space:]]Datasets[[:space:]]and[[:space:]]Controllable[[:space:]]Iterative[[:space:]]Data[[:space:]]Augmentation[[:space:]]Method[[:space:]]for[[:space:]]Code-switching[[:space:]]ASR[[:space:]]Error[[:space:]]Correction/704f5812-1da2-4355-bf1f-7c6e5a7b4a97_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/No[[:space:]]offence,[[:space:]]Bert[[:space:]]-[[:space:]]I[[:space:]]insult[[:space:]]only[[:space:]]humans![[:space:]]Multilingual[[:space:]]sentence-level[[:space:]]attack[[:space:]]on[[:space:]]toxicity[[:space:]]detection[[:space:]]networks/6570fd92-ec0d-40f4-8099-434dc1b0509c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Noise-Robust[[:space:]]Fine-Tuning[[:space:]]of[[:space:]]Pretrained[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]External[[:space:]]Guidance/569fc499-7adf-4636-bebe-f243ca93393e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Noise-Robust[[:space:]]Semi-Supervised[[:space:]]Learning[[:space:]]for[[:space:]]Distantly[[:space:]]Supervised[[:space:]]Relation[[:space:]]Extraction/fc076ec4-b157-42d2-a4a8-4199f3a5df0a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Noisy[[:space:]]Pair[[:space:]]Corrector[[:space:]]for[[:space:]]Dense[[:space:]]Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Noisy[[:space:]]Self-Training[[:space:]]with[[:space:]]Synthetic[[:space:]]Queries[[:space:]]for[[:space:]]Dense[[:space:]]Retrieval/260bb510-850d-4d78-a651-3f79c3ed0b9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Non-Autoregressive[[:space:]]Document-Level[[:space:]]Machine[[:space:]]Translation/6f06de2c-1532-40db-88ad-1021607cb22a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Non-Autoregressive[[:space:]]Sentence[[:space:]]Ordering/7f62cb94-7ce6-4950-afee-ba58d63ed721_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Non-Compositionality[[:space:]]in[[:space:]]Sentiment_[[:space:]]New[[:space:]]Data[[:space:]]and[[:space:]]Analyses/3184a1e0-ef47-41f1-9868-1f2f8500adaa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Non-compositional[[:space:]]Expression[[:space:]]Generation[[:space:]]Based[[:space:]]on[[:space:]]Curriculum[[:space:]]Learning[[:space:]]and[[:space:]]Continual[[:space:]]Learning/f2f88f8e-59fe-468f-ad81-14ffa69defaa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Non-parallel[[:space:]]Accent[[:space:]]Transfer[[:space:]]based[[:space:]]on[[:space:]]Fine-grained[[:space:]]Controllable[[:space:]]Accent[[:space:]]Modelling/29dfb883-4c9c-41fd-bdd7-1f7fc4e13373_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Normal-Abnormal[[:space:]]Decoupling[[:space:]]Memory[[:space:]]for[[:space:]]Medical[[:space:]]Report[[:space:]]Generation/b90e4221-98d5-45c3-a7ce-3e01b0ad4ff9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Not[[:space:]]All[[:space:]]Demonstration[[:space:]]Examples[[:space:]]are[[:space:]]Equally[[:space:]]Beneficial_[[:space:]]Reweighting[[:space:]]Demonstration[[:space:]]Examples[[:space:]]for[[:space:]]In-Context[[:space:]]Learning/e97af7b5-da38-43ac-9cde-d3ffa03f20c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Not[[:space:]]All[[:space:]]Languages[[:space:]]Are[[:space:]]Created[[:space:]]Equal[[:space:]]in[[:space:]]LLMs_[[:space:]]Improving[[:space:]]Multilingual[[:space:]]Capability[[:space:]]by[[:space:]]Cross-Lingual-Thought[[:space:]]Prompting/5e6cc189-8023-4887-a5d7-ec59f21ea497_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/NovaCOMET_[[:space:]]Open[[:space:]]Commonsense[[:space:]]Foundation[[:space:]]Models[[:space:]]with[[:space:]]Symbolic[[:space:]]Knowledge[[:space:]]Distillation/f2bb36f5-68ab-4034-8f93-8bd45f069dbe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Novel[[:space:]]Relation[[:space:]]Detection_[[:space:]]Discovering[[:space:]]Unknown[[:space:]]Relation[[:space:]]Types[[:space:]]via[[:space:]]Multi-Strategy[[:space:]]Self-Supervised[[:space:]]Learning/145539c6-2bb9-47af-abc0-f8ccafa2913b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Novel[[:space:]]Slot[[:space:]]Detection[[:space:]]With[[:space:]]an[[:space:]]Incremental[[:space:]]Setting/6eaba6e2-7e04-4fdf-9bac-6e740367034c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]Event[[:space:]]Individuation[[:space:]]for[[:space:]]Document-Level[[:space:]]Information[[:space:]]Extraction/b3e55c88-5789-48bb-8447-324998f4ba31_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]General[[:space:]]Language[[:space:]]Understanding/0c1e898b-df01-4993-8e52-ab20ff82a4e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]Robustness[[:space:]]of[[:space:]]Finetuned[[:space:]]Transformer-based[[:space:]]NLP[[:space:]]Models/1aa0a00a-57a5-4b97-9940-f6c4372ecdb4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]Surgical[[:space:]]Fine-tuning[[:space:]]for[[:space:]]Language[[:space:]]Encoders/683c3d63-8780-4980-997c-78842446d9f8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]Task-personalized[[:space:]]Multimodal[[:space:]]Few-shot[[:space:]]Learning[[:space:]]for[[:space:]]Visually-rich[[:space:]]Document[[:space:]]Entity[[:space:]]Retrieval/651da9af-63a8-4f44-baaf-4d6439bf9bfb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]Uncertainty[[:space:]]Calibration[[:space:]]and[[:space:]]Selective[[:space:]]Generation[[:space:]]in[[:space:]]Probabilistic[[:space:]]Neural[[:space:]]Summarization_[[:space:]]A[[:space:]]Benchmark[[:space:]]Study/ce833aa2-3742-40ab-a6fd-203011ba65db_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]the[[:space:]]Calibration[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]and[[:space:]]Alignment/6075dddf-f372-4c32-9dae-381129a16631_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]the[[:space:]]Dimensionality[[:space:]]of[[:space:]]Sentence[[:space:]]Embeddings/ee5aec9d-b452-471f-8441-096a6bdafdee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]the[[:space:]]Impact[[:space:]]of[[:space:]]Cross-Domain[[:space:]]Data[[:space:]]on[[:space:]]German[[:space:]]Language[[:space:]]Models/8b41b321-4686-40d5-8868-72a373ea3020_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]the[[:space:]]Relation[[:space:]]between[[:space:]]Sensitivity[[:space:]]and[[:space:]]Accuracy[[:space:]]in[[:space:]]In-Context[[:space:]]Learning/8720e86c-0aa6-4cab-8825-c5acef4719d1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]the[[:space:]]Risk[[:space:]]of[[:space:]]Misinformation[[:space:]]Pollution[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/88587d5b-576f-4456-8c6c-4303076e2c80_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]the[[:space:]]Transferability[[:space:]]of[[:space:]]Visually[[:space:]]Grounded[[:space:]]PCFGs/c1a2d272-d49b-4073-9ecd-6dd9a775c813_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/On[[:space:]]the[[:space:]]Zero-Shot[[:space:]]Generalization[[:space:]]of[[:space:]]Machine-Generated[[:space:]]Text[[:space:]]Detectors/dd132655-0477-4d19-be93-a85eb2350cdf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/One[[:space:]]For[[:space:]]All[[:space:]]&[[:space:]]All[[:space:]]For[[:space:]]One_[[:space:]]Bypassing[[:space:]]Hyperparameter[[:space:]]Tuning[[:space:]]with[[:space:]]Model[[:space:]]Averaging[[:space:]]for[[:space:]]Cross-Lingual[[:space:]]Transfer/87703e50-59f3-4a45-a99a-9f2af7594f09_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/One-Model-Connects-All_[[:space:]]A[[:space:]]Unified[[:space:]]Graph[[:space:]]Pre-Training[[:space:]]Model[[:space:]]for[[:space:]]Online[[:space:]]Community[[:space:]]Modeling/cfe3b1af-89a5-4c6c-855b-74c74f9926ca_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Open[[:space:]]Domain[[:space:]]Multi-document[[:space:]]Summarization_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Study[[:space:]]of[[:space:]]Model[[:space:]]Brittleness[[:space:]]under[[:space:]]Retrieval/a53a7dde-9f55-4097-879d-8d0f47c9f695_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Open-Ended[[:space:]]Instructable[[:space:]]Embodied[[:space:]]Agents[[:space:]]with[[:space:]]Memory-Augmented[[:space:]]Large[[:space:]]Language[[:space:]]Models/2d8aafee-214b-4bd6-badd-e87f1acc2f5b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Open-ended[[:space:]]Commonsense[[:space:]]Reasoning[[:space:]]with[[:space:]]Unrestricted[[:space:]]Answer[[:space:]]Candidates/29b3f86c-b009-4927-b2f7-fc283fd73359_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Open-source[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]Strong[[:space:]]Zero-shot[[:space:]]Query[[:space:]]Likelihood[[:space:]]Models[[:space:]]for[[:space:]]Document[[:space:]]Ranking/3dfd33bf-8767-48fa-9101-9000558996d1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Orca_[[:space:]]A[[:space:]]Few-shot[[:space:]]Benchmark[[:space:]]for[[:space:]]Chinese[[:space:]]Conversational[[:space:]]Machine[[:space:]]Reading[[:space:]]Comprehension/2a96f6ed-9c5b-4f78-9a3e-32eec0900efb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Orthogonal[[:space:]]Subspace[[:space:]]Learning[[:space:]]for[[:space:]]Language[[:space:]]Model[[:space:]]Continual[[:space:]]Learning/d8acce6d-e884-48f7-8bfe-19dd0d322521_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PARROT_[[:space:]]Zero-Shot[[:space:]]Narrative[[:space:]]Reading[[:space:]]Comprehension[[:space:]]via[[:space:]]Parallel[[:space:]]Reading/574f0e5d-980c-4a84-9181-4ac1599ad1ed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PAXQA_[[:space:]]Generating[[:space:]]Cross-lingual[[:space:]]Question[[:space:]]Answering[[:space:]]Examples[[:space:]]at[[:space:]]Training[[:space:]]Scale/41d7db00-7b39-4eaa-8f42-f0e43adacd55_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PCMID_[[:space:]]Multi-Intent[[:space:]]Detection[[:space:]]through[[:space:]]Supervised[[:space:]]Prototypical[[:space:]]Contrastive[[:space:]]Learning/c3853808-510c-4970-85fa-c7de96201ee0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PIVOINE_[[:space:]]Instruction[[:space:]]Tuning[[:space:]]for[[:space:]]Open-world[[:space:]]Entity[[:space:]]Profiling/a351f98d-20a7-48c9-9f31-2534c52a2afb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PMIndiaSum_[[:space:]]Multilingual[[:space:]]and[[:space:]]Cross-lingual[[:space:]]Headline[[:space:]]Summarization[[:space:]]for[[:space:]]Languages[[:space:]]in[[:space:]]India/29603321-be17-44fe-822c-c3854eaa848d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/POSQA_[[:space:]]Probe[[:space:]]the[[:space:]]World[[:space:]]Models[[:space:]]of[[:space:]]LLMs[[:space:]]with[[:space:]]Size[[:space:]]Comparisons/aa6de3c0-e648-4323-8a39-5231a7425401_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PR-MCS_[[:space:]]Perturbation[[:space:]]Robust[[:space:]]Metric[[:space:]]for[[:space:]]MultiLingual[[:space:]]Image[[:space:]]Captioning/6e4e1df8-bddd-49bd-a95e-d0ff2c65ba12_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PROTEGE_[[:space:]]Prompt-based[[:space:]]Diverse[[:space:]]Question[[:space:]]Generation[[:space:]]from[[:space:]]Web[[:space:]]Articles/e6f561f5-83c4-4528-89f3-c844f40f2ece_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PUNR_[[:space:]]Pre-training[[:space:]]with[[:space:]]User[[:space:]]Behavior[[:space:]]Modeling[[:space:]]for[[:space:]]News[[:space:]]Recommendation/7839b0e8-4c65-49dc-bc9a-993093cfeed1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PaRaDe_[[:space:]]Passage[[:space:]]Ranking[[:space:]]using[[:space:]]Demonstrations[[:space:]]with[[:space:]]LLMs/b4457758-3317-44d1-a7a5-2d3ab010104b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Parameter[[:space:]]Efficient[[:space:]]Multi-task[[:space:]]Fine-tuning[[:space:]]by[[:space:]]Learning[[:space:]]to[[:space:]]Transfer[[:space:]]Token-wise[[:space:]]Prompts/a194612c-07a7-4f0e-8134-8623472c3b04_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Parameter-Efficient[[:space:]]Cross-lingual[[:space:]]Transfer[[:space:]]of[[:space:]]Vision[[:space:]]and[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Translation-based[[:space:]]Alignment/26d9d9a3-7a9f-474f-91b0-40379ce51f35_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Parameter-Efficient[[:space:]]Prompt[[:space:]]Tuning[[:space:]]Makes[[:space:]]Generalized[[:space:]]and[[:space:]]Calibrated[[:space:]]Neural[[:space:]]Text[[:space:]]Retrievers/a5b5efa0-10f1-46fa-8e37-dde7f71d29e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ParroT_[[:space:]]Translating[[:space:]]during[[:space:]]Chat[[:space:]]using[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]tuned[[:space:]]with[[:space:]]Human[[:space:]]Translation[[:space:]]and[[:space:]]Feedback/c95bd735-348c-490e-a94a-47ffe248b75c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Pass-Tuning_[[:space:]]Towards[[:space:]]Structure-Aware[[:space:]]Parameter-Efficient[[:space:]]Tuning[[:space:]]for[[:space:]]Code[[:space:]]Representation[[:space:]]Learning/8286ea35-5e0c-4493-ae9b-59a2896a756a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Perceptual[[:space:]]Structure[[:space:]]in[[:space:]]the[[:space:]]absence[[:space:]]of[[:space:]]grounding_[[:space:]]the[[:space:]]impact[[:space:]]of[[:space:]]abstractedness[[:space:]]and[[:space:]]subjectivity[[:space:]]in[[:space:]]color[[:space:]]language[[:space:]]for[[:space:]]LLMs/d7bd4224-fa65-4388-8ce8-3c9d6cbed3db_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PersonaLM_[[:space:]]Language[[:space:]]Model[[:space:]]Personalization[[:space:]]via[[:space:]]Domain-distributed[[:space:]]Span[[:space:]]Aggregated[[:space:]]K-Nearest[[:space:]]N-gram[[:space:]]Retrieval[[:space:]]Augmentation/38f2675d-a219-4d39-81c8-04fede72fa30_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PerturbScore_[[:space:]]Connecting[[:space:]]Discrete[[:space:]]and[[:space:]]Continuous[[:space:]]Perturbations[[:space:]]in[[:space:]]NLP/97e8067e-bdd6-46fa-85ba-546c32b7dbe3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Pit[[:space:]]One[[:space:]]Against[[:space:]]Many_[[:space:]]Leveraging[[:space:]]Attention-head[[:space:]]Embeddings[[:space:]]for[[:space:]]Parameter-efficient[[:space:]]Multi-head[[:space:]]Attention/5702c6b5-dcdc-45a8-b363-53e194749193_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PivotFEC_[[:space:]]Enhancing[[:space:]]Few-shot[[:space:]]Factual[[:space:]]Error[[:space:]]Correction[[:space:]]with[[:space:]]a[[:space:]]Pivot[[:space:]]Task[[:space:]]Approach[[:space:]]using[[:space:]]Large[[:space:]]Language[[:space:]]Models/b736fba6-2126-498e-abe8-157b8884fc6f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Plausibility[[:space:]]Processing[[:space:]]in[[:space:]]Transformer[[:space:]]Language[[:space:]]Models_[[:space:]]Focusing[[:space:]]on[[:space:]]the[[:space:]]Role[[:space:]]of[[:space:]]Attention[[:space:]]Heads[[:space:]]in[[:space:]]GPT/1f2cea45-b6d9-47cd-8837-32016db5c4d2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/PlugMed_[[:space:]]Improving[[:space:]]Specificity[[:space:]]in[[:space:]]Patient-Centered[[:space:]]Medical[[:space:]]Dialogue[[:space:]]Generation[[:space:]]using[[:space:]]In-Context[[:space:]]Learning/ca6ff9d9-6a9e-4180-b3da-945a582cff5a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Pragmatics[[:space:]]in[[:space:]]Language[[:space:]]Grounding_[[:space:]]Phenomena,[[:space:]]Tasks,[[:space:]]and[[:space:]]Modeling[[:space:]]Approaches/2ffd9f81-9a22-4f30-9148-303416c975f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Pre-trained[[:space:]]Speech[[:space:]]Processing[[:space:]]Models[[:space:]]Contain[[:space:]]Human-Like[[:space:]]Biases[[:space:]]that[[:space:]]Propagate[[:space:]]to[[:space:]]Speech[[:space:]]Emotion[[:space:]]Recognition/ea8be79b-f55e-4ced-81fb-8b01940cc317_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Natural Response Generation for Chinese Reading Comprehension/deacd081-229d-47a6-82a1-c3842e05b39b_content_list.json

@@ -0,0 +1,1915 @@
+[
+    {
+        "type": "text",
+        "text": "Natural Response Generation for Chinese Reading Comprehension",
+        "text_level": 1,
+        "bbox": [
+            151,
+            89,
+            847,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Nuo Chen†,* Hongguang Li‡, Yinan Bao‡",
+        "bbox": [
+            326,
+            123,
+            673,
+            141
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Baoyuan Wang\\* and Jia Li",
+        "bbox": [
+            371,
+            142,
+            630,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{\\dagger}$ Hong Kong University of Science and Technology (Guangzhou),",
+        "bbox": [
+            233,
+            159,
+            766,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hong Kong University of Science and Technology",
+        "bbox": [
+            295,
+            175,
+            705,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "‡ Xiaobing.AI",
+        "bbox": [
+            443,
+            192,
+            559,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "†chennuo26@gmail.com, jialee@ust.hk",
+        "bbox": [
+            294,
+            208,
+            707,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Machine reading comprehension (MRC) is an important area of conversation agents and draws a lot of attention. However, there is a notable limitation to current MRC benchmarks: The labeled answers are mostly either spans extracted from the target corpus or the choices of the given candidates, ignoring the natural aspect of high-quality responses. As a result, MRC models trained on these datasets can not generate human-like responses in real QA scenarios. To this end, we construct a new dataset called Penguin to promote the research of MRC, providing a training and test bed for natural response generation to real scenarios. Concretely, Penguin consists of 200k training data with high-quality fluent, and well-informed responses. Penguin is the first benchmark towards natural response generation in Chinese MRC on a relatively large scale. To address the challenges in Penguin, we develop two strong baselines: end-to-end and two-stage frameworks. Following that, we further design Prompt-BART: fine-tuning the pre-trained generative language models with a mixture of prefix prompts in Penguin. Extensive experiments validated the effectiveness of this design. Our benchmark and codes are available at https://github.com/nuochenpku/Penguin.",
+        "bbox": [
+            144,
+            282,
+            460,
+            695
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            711,
+            258,
+            726
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Machine Reading comprehension (MRC) aims to empower the machine to correctly answer queries based on the given passages or paragraphs, which is an active field in NLP. Much of this popularity can be attributed to the release of many annotated and publicly available datasets (Rajpurkar et al., 2016; Trischler et al., 2016; Chen et al., 2022a; You et al., 2022; Chen et al., 2023a). Formally, these MRC efforts can be classified into two most",
+        "bbox": [
+            112,
+            738,
+            489,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "popular streams<sup>1</sup> from the answer type perspective: span-extraction (Rajpurkar et al., 2016; Trischler et al., 2016; Cui et al., 2019; Chen et al., 2022b; You et al., 2021a) and multiple choices (Lai et al., 2017; Zellers et al., 2018; Wang et al., 2020). The former requires the model to locate the text span in the given passage as the answer, e.g., SQuAD (Rajpurkar et al., 2016) and NewQA (Trischler et al., 2016). SQuAD, for example, introduces extracted span answers. The latter needs to choose the answer from a list of candidates like RACE (Lai et al., 2017). Considering most of the existing works are built in English, several efforts have been proposed to further advance MRC in Chinese such as DuReader (He et al., 2017), CMRC 2018 (Cui et al., 2019), ReCo (Wang et al., 2020).",
+        "bbox": [
+            507,
+            252,
+            884,
+            508
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Nonetheless, despite significant achievements in the various types and on a relatively large scale, a significant barrier that has rarely been discussed in previous MRC efforts remains: Annotated answers among these datasets are almost limited to spans in the source document or the given candidates. While spans directly copied from the source document and simple words like \"yes/no\" can be seen as ground-truth answers, they don't sound like natural responses in real-life scenarios. In reality, the responses of humans are always generated with flexibility and information rather than being forced to select certain type spans (e.g., entities and person) or one of the candidates.",
+        "bbox": [
+            507,
+            511,
+            884,
+            734
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This work devotes itself to natural response generation for Chinese MRC, catering to real-world QA scenarios. Intuitively, we can try two existing technical routes to generate natural responses towards MRC at the moment: 1) Optimizing generative language models (GLMs) such as BART and T5 with predetermined and inflexible answers on the above benchmarks. 2) Use extremely large-",
+        "bbox": [
+            507,
+            738,
+            884,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Work done when interned at Xiaobing.AI. § Indicates Corresponding authors.",
+        "bbox": [
+            112,
+            892,
+            487,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>Some answer free-form datasets with extracted span answers and yes/no/unknown candidates classification also can be seen as the span-extraction and multiple choice settings.",
+        "bbox": [
+            507,
+            881,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "11068",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 11068-11081",
+        "bbox": [
+            208,
+            945,
+            784,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            277,
+            958,
+            719,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/3995d783969ffe7705391ffa2af0ef3c65d6b89ae984350721722a74fac5d410.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Passage</td><td>...Bream, which is traditionally steamed or braised, is a popular fish. Soup should of course be cooked, but soles is not suitable for this dish because of its fleshy nature...\n...鳊鱼一直备受人们喜爱，鳊鱼的传统做法是清蒸或者红烧。当然不排除可以烧汤，但是鉴于鳊鱼的肉质特色，不是很适合烧汤的...</td></tr><tr><td>Question</td><td>Can bream be stewed in soup?\n鳊鱼可以炖汤吗?</td></tr><tr><td>Candidates</td><td>Of Course | No, you can&#x27;t | Uncertain\n可以|不可以|不确定</td></tr><tr><td>BART</td><td>Of Course\n可以</td></tr><tr><td>GPT-3</td><td>Yes! If you want to stew soup, in think-ing, then we still choose to have.\n是可以的！想要炖汤的话，在思考，那么我们还是选择有。</td></tr><tr><td>Expected\nResponse</td><td>Bream can be used in soups but is generally not recommended.\n鳊鱼可以用来煮汤，但是一般不推荐这么做。</td></tr></table>",
+        "bbox": [
+            122,
+            80,
+            480,
+            404
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1: An example from ReCo (Wang et al., 2020). The text in bold refers to the answer in the candidates or the expected response. English translation is also given for comparison.",
+        "bbox": [
+            112,
+            412,
+            489,
+            470
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "scale pre-trained GLMs like GPT-3 (more than 100 billion parameters) (Brown et al., 2020) with few-shot in-context learning.",
+        "bbox": [
+            112,
+            479,
+            489,
+            527
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We showcase a multi-choice MRC example from ReCo (Wang et al., 2020) in Table 1, where the query is a typical yes-or-no question and the answer is required to be one of the candidates. We fine-tune BART-large (Lewis et al., 2020) on ReCo and find that the resulting model can accurately predict the answer: \"of course\" in this case. However, the prediction is accurate but stiff and lacking in detail. On the other hand, we use GPT-3 under few-shot from ReCo and validate its generated response via the same case in Table 1. Obviously, the produced result \"Yes! If you want to... we still choose to have\" is fluent yet incoherent and semantically incorrect. Ideally, the qualified response of humans should be informative and fluent, like \"Bream can be used .... not recommended\". In other words, although these models are proficient, they can't produce fluent and accurate responses for reading comprehension.",
+        "bbox": [
+            112,
+            529,
+            489,
+            818
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Based on the above analysis, we can draw the conclusion that it is impossible to build an effective MRC model for natural $^2$  response generation in the absence of a benchmark with well-labeled",
+        "bbox": [
+            112,
+            819,
+            489,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "responses. To this end, we move forward to push the boundary of MRC: Firstly, we collect the first Chinese generative reading comprehension (GRC) dataset named Penguin to facilitate research of natural response generation for Chinese MRC. Formally, Penguin poses a new challenge to current MRC models: Not only should the responses be accurate, but they also need to be natural (as fluent and informative as possible). Concretely, Penguin contains 200k training and a 15K test corpus with high-quality fluent and informative responses. Table 2 shows the overview of current MRC datasets and Penguin.",
+        "bbox": [
+            505,
+            84,
+            884,
+            292
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Secondly, we design two GRC pipelines based on this non-trivial dataset: i) end-to-end; ii) two-stage training. For the former, we directly take the question and corresponding passage as input, generating a sequence of words as responses. The latter consists of two modules: Answerer and Responser. Answerer is responsible for generating initial answers based on the question and passage. Then Responser rewrites and expands the generated answers into human-like responses. Experimentally, two pipelines achieve promising results in Penguin.",
+        "bbox": [
+            505,
+            294,
+            884,
+            470
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Finally, we propose Prompt-BART, an approach for fine-tuning pre-trained generative language models using a wide variety of prefix prompts in Penguin. The resulting model, as stated in the prompt that accompanied it, boosted the overall performance.",
+        "bbox": [
+            507,
+            473,
+            882,
+            569
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            583,
+            665,
+            599
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Machine Reading Comprehension During the past few years, Machine Reading Comprehension (MRC) (Rajpurkar et al., 2016; Trischler et al., 2016; You et al., 2021b; Chen et al., 2021; Song et al., 2023; Chen et al., 2023b) receives lots of attention from research communities and various related benchmarks have been proposed: SQuAD (Rajpurkar et al., 2016), MS-MarCO (Nguyen et al., 2016), NewQA (Trischler et al., 2016) and RACE (Lai et al., 2017), etc. Following this line of research, some works aim to accelerate the development of Chinese MRC on language diversity, such as DuReader (He et al., 2017), CMRC 2018 (Cui et al., 2019), ReCo (Wang et al., 2020).",
+        "bbox": [
+            505,
+            612,
+            884,
+            837
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "These well-designed and large-scale benchmarks facilitate the research of various types of MRC tasks, but they ignore the fact that answers are primarily confined to text span extraction from the source document or choosing candidates. In com",
+        "bbox": [
+            507,
+            838,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "2In this work, we argue that natural response are fluent and informative.",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "11069",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "parison with natural human responses, these answers are less informative and less fluid. In other words, models trained on the above datasets do not have strong ability to generate fluent, natural as well as informative responses. Although recent works (Bi et al., 2019; Li et al., 2021) develop sequence-to-sequence frameworks to address the challenge in MS-MarCO (Nguyen et al., 2016), they still regard some abstractive answers as the optimized objective, which consist of some uninformative extracted spans from the passage, leading to responses that are far away from fluent and fluid text. To this end, we propose a new dataset Penguin to encourage further progress in natural response generation of Chinese MRC.",
+        "bbox": [
+            112,
+            84,
+            492,
+            326
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Natural Generative Question Answering Similar to our work, natural generative question answering (NGQA) (Yin et al., 2016; Huang and Zhong, 2018; Bi et al., 2019; Jiang et al., 2022) also aims at generating natural responses to questions. Although both tasks share a similar target, the research question and task settings differ: Firstly, NGQA produces responses to simple factoid questions based on the naive entity triplets that are given as external knowledge. In contrast, we focus on yielding fluent responses based on a deep understanding of the question and passage. Such a deep understanding of interaction between question and passage is more challenging than the simple entity triplets. Moreover, GRC closely resembles humans asking questions about a certain topic and generating corresponding replies. Last, previous NGQA efforts were limited to the English corpus, whereas we are expanding it to Chinese.",
+        "bbox": [
+            112,
+            348,
+            490,
+            653
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 DataSet",
+        "text_level": 1,
+        "bbox": [
+            112,
+            678,
+            218,
+            694
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Task Definition",
+        "text_level": 1,
+        "bbox": [
+            112,
+            713,
+            280,
+            728
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Generally, machine reading comprehension task can be illustrated as  $\\langle \\mathcal{P}, \\mathcal{Q}, \\mathcal{A} \\rangle$ , where  $\\mathcal{P}$  is the paragraph or the whole passage,  $\\mathcal{Q}$  represents questions,  $\\mathcal{A}$  refers to answers. Previous works tend to model MRC as a classification task with the objective of selecting  $\\mathcal{A}$  from several candidates or finding the text span of  $\\mathcal{A}$  in  $\\mathcal{P}$ . In contrast, we formalize MRC as a sequence-to-sequence task, where the objective is to generate fluent and factual responses:  $\\mathcal{R}$ . In this work, we include  $\\langle \\mathcal{P}, \\mathcal{Q}, \\mathcal{A}, \\mathcal{R} \\rangle$  in Penguin.",
+        "bbox": [
+            112,
+            741,
+            490,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/0de47f22c2549fcd759900c4b5e10bdaf4a5fd425c21c4d4ddc92bedf8b5e373.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Lang.</td><td>Answer Type</td><td>Size</td></tr><tr><td>SQuAD (2016)</td><td>EN</td><td>Span</td><td>100K</td></tr><tr><td>RACE (2017)</td><td>EN</td><td>Multi-Choices</td><td>87K</td></tr><tr><td>MarCO (2016)</td><td>EN</td><td>Free-Form</td><td>100K</td></tr><tr><td>NewsQA (2016)</td><td>EN</td><td>Span</td><td>100K</td></tr><tr><td>DuReader (2017)</td><td>CN</td><td>Free-Form</td><td>200k</td></tr><tr><td>CMRC (2019)</td><td>CN</td><td>Span</td><td>20K</td></tr><tr><td>ReCo (2020)</td><td>CN</td><td>Multi-Choices</td><td>300K</td></tr><tr><td>Penguin</td><td>CN</td><td>Natural Response</td><td>200K</td></tr></table>",
+        "bbox": [
+            510,
+            80,
+            887,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Datasets Comparison. EN and CN refer to English and Chinese, separately.",
+        "bbox": [
+            507,
+            222,
+            882,
+            250
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Dataset Collection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            260,
+            700,
+            274
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As aforementioned, a high-quality dataset is a prerequisite for building effective GRC models. Unfortunately, current datasets don't contain well-informed and natural responses. To facilitate the study of MRC, we collect a new dataset: Penguin, in the hopes of creating sophisticated GRC models that can generate natural responses for practical QA situations. Considering constructing such a dataset on a large scale is non-trivial, we initialize our dataset from the current Chinese MRC dataset corpus to get raw passage-question-answer  $(<\\mathcal{P}, \\mathcal{Q}, \\mathcal{A})$  triplets, including CMRC 2018, DuReader, and ReCo.",
+        "bbox": [
+            507,
+            281,
+            884,
+            488
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "It is extremely difficult and expensive to ask the annotator to start from scratch and write a response  $\\mathcal{R}$  that makes sense and sounds natural for each  $<\\mathcal{P}, \\mathcal{Q}, \\mathcal{A}>$  triplet. Therefore, we generate informative and fluent responses via the following steps: (1) We first utilize a specific response generative model (short for Responser) to generate initial responses for the above triplets; (2) Then, we use state-of-the-art semantic matching models alongside certain manually-created criteria to exclude samples of incoherence and semantic shift; (3) Last, we employ three professional annotators to rewrite and recheck undesired cases, therefore regulating data quality. As a result, we collect a sequence of 4-tuples:  $<\\mathcal{P}, \\mathcal{Q}, \\mathcal{A}, \\mathcal{R}>$  in Penguin, where  $\\mathcal{R}$  is the labeled response. In the following, we will introduce the detailed procedures of these steps, sequentially. The conceptual process of data collection is presented in Figure 1.",
+        "bbox": [
+            507,
+            491,
+            884,
+            796
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Responder Prior to going further, the main concern is how to get an effective sequence-to-sequence responder model. Furthermore, more pressing than any of these other issues is figuring out where to collect the high-quality Chinese corpus needed to train the responder. Inspired by NGQA works, We first get the English <question,",
+        "bbox": [
+            507,
+            806,
+            885,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "11070",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/0d340609ffe895c643a7c0f4c5381ab5b681f4ad5d6d3f89a882bb0b01994b18.jpg",
+        "image_caption": [
+            "Figure 1: Data collection pipeline"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            157,
+            84,
+            467,
+            429
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "answer, response> triplets in (Baheti et al., 2020). Then we utilize machine translation systems to translate these data from English to Chinese. Considering translated corpus will inevitably have erroneous examples, we utilize perplexity (Brown et al., 1992) to identify and weed out some bad cases. Following the line of Yin et al. (2016), Bart-large is employed as the backbone of responder to fine-tune on these translated corpus, which takes the question as well as answer as input and then outputs response during training. Subsequently, the resulting responder models are used to inference on each  $\\langle \\mathcal{P}, \\mathcal{Q}, \\mathcal{A} \\rangle$  triplet from CMRC 2018, DuReader and ReCo, generating initial responses. Concretely, responder takes the passage and question as input, then outputs the response during inference.",
+        "bbox": [
+            112,
+            479,
+            489,
+            736
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Data Filtering Intuitively, the collected initial responses might be stilted, evasive, and semantically challenging to connect with its corresponding questions and answers. Therefore, we employ state-of-the-art Chinese semantic matching models on LCQMC leadboard<sup>3</sup> to compute the sentence similarity of the generated response and answer. Roughly, we will throw out an instance if the simi",
+        "bbox": [
+            112,
+            750,
+            489,
+            879
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/3bd857cc6ca00886ac0ee63135e2db8b8ec24021d78527bc9239c1cbaacabbf2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Statistics</td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td colspan=\"2\">&lt;P, Q, A, R&gt;</td><td>200k</td><td>10k</td><td>15k</td></tr><tr><td rowspan=\"3\">Passage tokens</td><td>Max.</td><td>492</td><td>486</td><td>490</td></tr><tr><td>Min.</td><td>24</td><td>19</td><td>31</td></tr><tr><td>Ave.</td><td>84.3</td><td>82.45</td><td>84.85</td></tr><tr><td rowspan=\"3\">Question tokens</td><td>Max.</td><td>61</td><td>37</td><td>17</td></tr><tr><td>Min.</td><td>4</td><td>5</td><td>5</td></tr><tr><td>Ave.</td><td>11.5</td><td>11.6</td><td>11.5</td></tr><tr><td rowspan=\"3\">Answer tokens</td><td>Max.</td><td>20</td><td>11</td><td>20</td></tr><tr><td>Min.</td><td>1</td><td>1</td><td>1</td></tr><tr><td>Ave.</td><td>2.8</td><td>2.6</td><td>2.1</td></tr><tr><td rowspan=\"3\">Response tokens</td><td>Max.</td><td>42</td><td>26</td><td>33</td></tr><tr><td>Min.</td><td>6</td><td>6</td><td>6</td></tr><tr><td>Ave.</td><td>10.1</td><td>9.6</td><td>9.7</td></tr></table>",
+        "bbox": [
+            522,
+            82,
+            870,
+            292
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Overall data Statistics in Penguin.",
+        "bbox": [
+            547,
+            302,
+            840,
+            317
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "larity score between the response and the answer is below the pre-defined threshold. To be more specific, the generated response just can be seen as replacing a few words in the answer in some cases, resulting in the extremely high similarity between the answer and the response. e.g., given the question \"How tall is Yao Ming\", and the answer \"He is  $2.29\\mathrm{m}$ \". the generated response could be \"Yao Ming is  $2.29\\mathrm{m}$ .\" For these cases, the threshold used to filter incoherent samples will be manually larger (more than 0.95). On the other hand, in certain instances, where the generated response provides the final piece of missing information based on the answer, the threshold will be manually lowered (lower than  $0.3)^4$ . Besides, we employ additional scoring and filtering rules to identify the most grammatically correct and contextually relevant response (if any) to each  $\\langle \\mathcal{P}, \\mathcal{Q}, \\mathcal{A} \\rangle$  triplet. For instance, we remove the samples where the length of the response is less than the length of the answer plus 5.",
+        "bbox": [
+            507,
+            326,
+            884,
+            646
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Human Annotation In the above, we automatically generate the response via the effective responder, and then try to filter out some meaningless examples. To further control the training data quality, we recruit another three professional annotators to re-check each example and revise undesirable cases. In detail, each piece of data is verified by two different annotators. These annotators review the generated responses after reading the relevant question, passage, and answer. If the generated response is unnatural, they try rephrasing the statement to obtain a fair response. Thereafter, another annotator rates their annotations' quality, avoiding bias and getting high-agreement data. As a result,",
+        "bbox": [
+            507,
+            658,
+            884,
+            883
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "4We manually select these thresholds based on experiments.",
+        "bbox": [
+            507,
+            892,
+            882,
+            916
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{3}$ https://paperswithcode.com/sota/chinese-sentence-pair-classification-on-lcqmc",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "11071",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/cec546d28a321e596f13e046d20fa46b16adc0640b1cf60e9e604155d4e6173c.jpg",
+        "image_caption": [
+            "Figure 2: Question types in Penguin. Empty colored blocks indicate suffixes that are too rare to show."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            84,
+            460,
+            329
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "we collect high quality 4-tuples:  $< \\mathcal{P}, \\mathcal{Q}, \\mathcal{A}, \\mathcal{R} >$ .",
+        "bbox": [
+            112,
+            388,
+            473,
+            406
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Test set Annotation We additionally hire eight expert annotators to choose and create questions and their answers from social media and themselves, bringing the acquired data closer to real-world events and increasing variety. Concretely, eight annotators are split into four smaller groups, and each of them consists of one questioner and one answerer. Questioners first select questions or question the news they are interested in from current social media platforms like weibo and zhihu. Then, answerers get related passages that contain corresponding knowledge about each question from Bing search. Following that, they must select the excerpt from the retrieved ones that is the most relevant to the annotated question at hand. After reading the question and its accompanying passage, they then annotate responses at last. Similarly, we recruit two annotators to recheck the data twice, further guaranteeing data quality. This way, we collect 15k data in our test set.",
+        "bbox": [
+            112,
+            416,
+            489,
+            736
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The general statistics in Penguin are presented in Table 3. In total, we collect train, development, and test set, which consists of  $200\\mathrm{k}$ ,  $10\\mathrm{k}$ , and  $15\\mathrm{k} < \\mathcal{P}$ ,  $\\mathcal{Q}, \\mathcal{A}, \\mathcal{R}$  data, separately. The collected responses are noticeably more informative than the answers given, as seen by their length.",
+        "bbox": [
+            112,
+            739,
+            489,
+            835
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Data Statics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            848,
+            257,
+            862
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Question Analysis The pilot investigations (He et al., 2017) assisted us in reaching an agreement on the following question type taxonomy: Yes/No,",
+        "bbox": [
+            112,
+            871,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/c0cd72329d832bbba2bd4d618a4a6f7f8004ce45960c877b2389ed8cc335b55f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>Example</td><td>Percentage</td></tr><tr><td>Extension</td><td>Q: The design life of Toyota engine?A: 0.24 million kmR: Toyota engine has a de-sign life of 0.24 million km.</td><td>62.4%</td></tr><tr><td>Rewritten</td><td>Q: Can we carry tea in trains?A: YesR: Of course, we can bring tea in trains.</td><td>26.3%</td></tr><tr><td>Others</td><td>Q: Can you fix the Apple 5S if it&#x27;s flooded?A: It can be fixed.R: In general, if the iPhone 5s gets water, it can defi-nitely be repaired.</td><td>11.3%</td></tr></table>",
+        "bbox": [
+            514,
+            70,
+            880,
+            302
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Comparison between answer and response in Penguin. For ease of reading, English translation data is only provided here.",
+        "bbox": [
+            507,
+            311,
+            882,
+            355
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Fact, and Opinion. Yes/No questions are often asked for confirmation, and Fact questions are mostly about specific entities, timing, or description of some facts. Opinion questions always contain why/what/how questions, inquiring about the reason or advice of things. We draw the distribution of these three question types from our collected Penguin in Figure 2, showing Penguin contains various questions, catering to more realistic scenarios.",
+        "bbox": [
+            507,
+            369,
+            884,
+            514
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Response Analysis After deep analysis of responses and answers in Penguin, we classify the connection between them into three broad categories: Extension, Rewritten, Others. Extension refers to the response containing exactly the content of the answer and going into more detail about it. Of note, although these answers are semantically right, we argue that the responses that extend to them could be more informative and cater to real human responses. Rewritten denotes the response totally rewritten the answer to make it more informative and natural while retaining the original meaning. That is to say, the response almost has no overlap with the answers in this situation. Others indicates the response may contain part of the answer, as well as expands it with lexical, coreferential, or casual inferential knowledge. We showcase of each category in Table 4, respectively.",
+        "bbox": [
+            507,
+            541,
+            884,
+            831
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Model",
+        "text_level": 1,
+        "bbox": [
+            507,
+            844,
+            601,
+            859
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose two frameworks to address the challenge in Penguin: 1) end-to-end; 2) two stage.",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "11072",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ee5b3a037fe0bfeafa0c37f89802315da3989d9ff181bc7dd19d694efb226bdd.jpg",
+        "image_caption": [
+            "Figure 3: Overview of our proposed frameworks: (a) End-to-end; (b) Two-stage. Answerer denotes a answer generator module and Responser refers to a response generator module. English translation is only given for reading."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            156,
+            72,
+            842,
+            250
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1847cd1e6506db16962f8226db0d2dce35867f76846ac405af9171df549834f9.jpg",
+        "image_caption": [
+            "Figure 4: Prompt templates."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            139,
+            303,
+            460,
+            588
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "End-to-end Intuitively, we can directly concat question and passage to construct the input  $\\mathbf{X}_{\\mathrm{e}}$ :",
+        "bbox": [
+            112,
+            626,
+            487,
+            658
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {\\mathrm {e}} = \\left\\{\\left[ \\text {C L S} \\right] \\text {Q u e s t i o n [ S E P ] P a s s a g e [ S E P ]} \\right\\} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            124,
+            665,
+            487,
+            680
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where [CLS] and [SEP] refer to the beginning and separate symbols. Then a pre-trained generative model is utilized to take  $\\mathbf{X}_{\\mathrm{e}}$  as input, and then generates a response after deeply understanding of the interaction between question and passage, as shown in Figure 3 (a). Concretely, we optimize the model with the widely used negative log-likelihood loss during training:",
+        "bbox": [
+            112,
+            689,
+            487,
+            818
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {e} = - \\log \\left(p _ {\\theta} \\left(\\mathcal {R} \\mid \\mathbf {X} _ {\\mathbf {e}}\\right)\\right) \\\\ = - \\sum_ {i = 1} ^ {| \\mathcal {R} |} \\log \\left(p _ {\\theta} \\left(r _ {i} \\mid \\mathbf {X} _ {\\mathbf {e}}, \\mathcal {R} _ {<   i}\\right)\\right). \\tag {2} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            171,
+            826,
+            485,
+            889
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Two stage Besides, we also design a two-stage framework to deal with the challenge in Penguin",
+        "bbox": [
+            112,
+            888,
+            485,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "which consists of a answer generator module (short for Answerer) and a response generator module (short for Responser), seen in Figure 3 (b). In two stage frameworks, Answerer can be thought of as a reading comprehension module that is in charge of thoroughly comprehending the passage and the question before generating a answer; then Responser rewrites and extends the generated answer to a more human-like response.",
+        "bbox": [
+            507,
+            303,
+            884,
+            447
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To be more specific, we first take  $\\mathbf{X}_{\\mathrm{e}}$  into Answerer, which yields initial answer:",
+        "bbox": [
+            507,
+            448,
+            882,
+            480
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {a} = - \\sum_ {j = 1} ^ {| \\mathcal {A} |} \\log \\left(p _ {\\alpha} \\left(a _ {j} \\mid \\mathbf {X} _ {\\mathbf {e}}, \\mathcal {A} _ {<   j}\\right)\\right). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            564,
+            492,
+            882,
+            539
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Then we construct a new input  $\\mathbf{X}_{\\mathbf{r}}$  with generated answer and question:",
+        "bbox": [
+            507,
+            550,
+            880,
+            581
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {X} _ {\\mathbf {r}} = \\left\\{\\left[ \\text {C L S} \\right] \\text {Q u e s t i o n [ S E P ] A n s w e r [ S E P ]} \\right\\} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            593,
+            882,
+            608
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Subsequently, Responser is responsible for reading the  $\\mathbf{X}_{\\mathbf{r}}$  to predict the target response  $\\mathcal{R}$ :",
+        "bbox": [
+            507,
+            621,
+            882,
+            653
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {r} = - \\sum_ {i = 1} ^ {| \\mathcal {R} |} \\log \\left(p _ {\\theta} \\left(r _ {i} \\mid \\mathbf {X} _ {\\mathbf {r}}, \\mathcal {R} _ {<   i}\\right)\\right). \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            665,
+            880,
+            709
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Notice that, Answerer and Responser are both the same generative models but not sharing parameters in our experiments.",
+        "bbox": [
+            507,
+            715,
+            880,
+            764
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Prompt-Tuning Inspired by recent works (Wei et al., 2022; Brown et al., 2020), we fine-tune in Penguin with a variety of prefix prompt with different template types. Concretely, we manually design five templates for each generative module from end-to-end and two-stage frameworks, as several examples shown in Figure 4. We use these prompts to formulate different inputs in training, obtaining the final checkpoint. Experimentally, we choose",
+        "bbox": [
+            507,
+            774,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "11073",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/4f8b9f520f859f19e5c1062cb2218b623e3530d315afc6bf8ccbb0af53c51972.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Framework</td><td>BackBone</td><td>Dist-1</td><td>Dist-2</td><td>BLEU-1</td><td>BLEU-2</td><td>R.-L</td><td>EM</td><td>Relv.P</td><td>Flu.</td><td>Info.</td></tr><tr><td rowspan=\"3\">End-to-End</td><td>T5-base</td><td>2.7</td><td>33.5</td><td>80.8</td><td>75.4</td><td>79.1</td><td>31.4</td><td>3.41</td><td>3.79</td><td>3.84</td></tr><tr><td>BART-Large</td><td>2.7</td><td>34.4</td><td>82.0</td><td>77.1</td><td>80.8</td><td>35.1</td><td>3.50</td><td>3.88</td><td>3.84</td></tr><tr><td>Prompt-BART</td><td>2.8</td><td>34.7</td><td>82.4</td><td>77.6</td><td>81.0</td><td>35.5</td><td>3.56</td><td>3.87</td><td>3.90</td></tr><tr><td rowspan=\"3\">Two-Stage</td><td>T5-base</td><td>2.7</td><td>33.8</td><td>80.4</td><td>75.0</td><td>79.0</td><td>31.0</td><td>3.40</td><td>3.84</td><td>3.87</td></tr><tr><td>BART-Large</td><td>2.7</td><td>33.7</td><td>82.5</td><td>77.3</td><td>81.2</td><td>36.7</td><td>3.67</td><td>3.87</td><td>3.90</td></tr><tr><td>Prompt-BART</td><td>2.9</td><td>33.8</td><td>83.2</td><td>78.2</td><td>81.8</td><td>37.2</td><td>3.76</td><td>3.87</td><td>3.93</td></tr></table>",
+        "bbox": [
+            134,
+            70,
+            863,
+            203
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5: End-to-end and two-stage frameworks' performances with different backbone models. EN and CN refer to English and Chinese, separately. We highlight the best performances in the table. For smaller versions' the results of each backbone model can be seen in Appendix D. R.-L and EM refer to ROUGH-L and Exact Match, separately. We run each model three times and report their average results.",
+        "bbox": [
+            112,
+            212,
+            884,
+            269
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "BART-large as our baseline, the resulting model is named as Prompt-BART after prompt-tuning. All prompts can be seen in Appendix G, Table 13.",
+        "bbox": [
+            112,
+            294,
+            487,
+            344
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            357,
+            260,
+            375
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Evaluation Metric",
+        "text_level": 1,
+        "bbox": [
+            112,
+            386,
+            305,
+            401
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Different from previous span-extraction and multiple-choice MRC tasks, we use automatic and human metrics to evaluate the model performances.",
+        "bbox": [
+            112,
+            409,
+            489,
+            456
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Automatic Metrics We utilize following commonly-used metrics to evaluate the quality of generated responses: Dist-1, Dist-2 (Li et al., 2016), BLEU-1,BLEU-2 (Papineni et al., 2002), ROUGH-L and Exact Match. We give detailed explanations of these metrics in Appendix A. For the above metrics, the larger the value, the better language modeling performance.",
+        "bbox": [
+            112,
+            468,
+            489,
+            596
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Human Metrics To further get a fuller picture of how well the generation-based model works, we asked 4 annotators to look at the quality and consistency of the generated responses on our test set. These annotators are from a third-party and skilled with language tasks but have no knowledge of the models. Concretely, we randomly sample 500 examples in our test set for model's evaluation. Human annotators are tasked with rating dialogue quality using three principle measures: Fluency (Flu.), Informativeness (Info.) and Relevance with Passage (Relv.P). Each criterion is given a rating from 1 to 5, with 1, 3, 5 representing poor, mediocre and flawless performances.",
+        "bbox": [
+            112,
+            609,
+            489,
+            834
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Backbone",
+        "text_level": 1,
+        "bbox": [
+            112,
+            848,
+            240,
+            862
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To investigate the performances of current state-of-the-art generative models, we choose two popular pre-trained generative models: T5 (Raffel et al.,",
+        "bbox": [
+            112,
+            871,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c6076eb4fbbeb6b04e68e194db77d4201b8800091711ee6ef4f68e4c2418ec3d.jpg",
+        "image_caption": [
+            "Figure 5: Model performances on different question types. Here, we choose BART-large as the backbone."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            294,
+            870,
+            445
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "2020) and BART (Lewis et al., 2020) as our backbones. In the Appendix B, we briefly introduce each model. Detailed experimental setup of training these models can be seen in Appendix C.",
+        "bbox": [
+            507,
+            500,
+            885,
+            565
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            577,
+            613,
+            592
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Automatic Evaluations Table 5 shows the main results of all models. From the table, we can conclude the following: 1) The designed two frameworks with different backbone models all achieve competitive results in each automatic metric, proving their effectiveness. 2) Importantly, the results of each base model under the two different frameworks are comparable. More interestingly, an end-to-end framework is a more advanced and promising system when you take into account that the parameters and inference cost of a two-stage framework are almost two times higher. 3) Prompt-tuning can further boost the model's performances in all evaluation metrics, and the improvement is more obvious in two-stage training (e.g., 81.8 vs. 81.2 in ROUGH-L).",
+        "bbox": [
+            505,
+            601,
+            885,
+            858
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Human Evaluations We also draw the following conclusions from the Table 5: 1) All models achieve good results in fluency, informativeness",
+        "bbox": [
+            507,
+            871,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "11074",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5821d99558c9879fde8ef2918c02d4a5f4f12e690318b57cd42f7de755b92469.jpg",
+        "image_caption": [
+            "Figure 6: Win/Lose human evaluation results  $(\\%)$  between end-to-end and Answer in two-stage pipeline with p-value  $< 0.05$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            76,
+            472,
+            242
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and relevance, showing their generated responses are close to human-like. 2) Prompt-BART in two-stage training performs better in relevance than other models, indicating a better understanding of the passage. In general, human evaluations of these models basically mirror automated evaluations.",
+        "bbox": [
+            112,
+            313,
+            487,
+            409
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4 Discussion and Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            423,
+            349,
+            437
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Aside from the high performance of our proposed models, we are concerned about the following questions:  $\\mathbf{Q}_1$ : What does the model perform when it answers different question types?  $\\mathbf{Q}_2$ : Is the generated response more natural to the generated answer?  $\\mathbf{Q}_3$ : Can extremely large model like GPT-3 almost address the challenge in Penguin? (Appendix E.1, Table 8)  $\\mathbf{Q}_4$ : What is the majority of the improvement brought by the designed prompts? (Appendix E.2, Table 9).",
+        "bbox": [
+            112,
+            445,
+            489,
+            606
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Answer to  $\\mathbf{Q}_1$ : We conduct fine-grained analysis to validate model performances on different question types. As seen in Figure 5, we can observe: 1) each model performs best on Yes/No questions and worst on Fact questions, indicating generating natural responses about specific facts needs deeper understanding of questions and passages; 2) prompt-tuning brings more benefits when answering Fact questions (76.5 vs. 74.3). We further present each case on three question types in Table 11 and Table 12.",
+        "bbox": [
+            112,
+            618,
+            489,
+            793
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Answer to  $\\mathbf{Q}_2$ : In order to test whether the generated responses are more natural than the fixed answers. Three professional workers are asked to evaluate the generated responses from end-to-end framework and generated answers from Answerer in two-stage pipeline. Concretely, they need to choose which is the best one (Win/Lose) for",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a45db3a8f52187f070a482eaab47763578458fc935bcbe1d1f1e4a9c2401563f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Passage</td><td>...If you already have a loan, the bank will definitely consider the debt ratio, depending on the assets and repayment ability....</td></tr><tr><td>Question</td><td>Will Credit Loan Affect Mortgage?</td></tr><tr><td>Response</td><td>It could affect your Mortgage, but it depends on your income level and assets.</td></tr><tr><td colspan=\"2\">Generated Response-One stage</td></tr><tr><td>T5</td><td>Credit loan must affect Mortgage.</td></tr><tr><td>BART</td><td>It will affect your Mortgage.</td></tr><tr><td>Prompt-BART</td><td>Credit loan could affect mortgage, but depends on assets.</td></tr><tr><td colspan=\"2\">Generated Response-Two stage</td></tr><tr><td>T5</td><td>Credit loan must affect Mortgage.</td></tr><tr><td>BART</td><td>Credit loan will affect mortgage</td></tr><tr><td>Prompt-BART</td><td>Credit loan may affect your Mortgage.</td></tr></table>",
+        "bbox": [
+            510,
+            80,
+            887,
+            362
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6: A case study in our collected Penguin, and generated responses from the large version of baselines are provided. The text in bold refers to the expected response from human annotation.",
+        "bbox": [
+            507,
+            370,
+            882,
+            429
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "each example in randomly selected 200 test samples from three views, as shown in Figure 6. We can observe that more than  $85\\%$  of the generated responses are superior to their counterpart answers across all evaluation perspectives, indicating they are closer to human-like responses than the generated answers. These Win/Lose results also prove the value of Penguin, that is, providing a training and testing bed for natural response generation to real-world scenarios.",
+        "bbox": [
+            507,
+            437,
+            882,
+            596
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Case Study In addition, we showcase an example and its generated responses from Penguin in Table 6. We can see that almost all of the responses from the different models do a pretty good job of roughly capturing the intended meaning. Moreover, Prompt-BART could generate more complete and accurate responses, showing its efficiency.",
+        "bbox": [
+            507,
+            607,
+            882,
+            720
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            732,
+            640,
+            746
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we summarize the limitation of current MRC efforts, and propose a new dataset named Penguin for natural response generation in Chinese MRC. Penguin requires the predicted responses not only be accurate but also natural-sounding, catering to real scenarios. Thereafter, we propose end-to-end and two-stage GRC pipelines to address the challenge in Penguin, achieving promising results. Moreover, we design prompt-tuning, which involves fine-tuning the pre-trained genera",
+        "bbox": [
+            507,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "11075",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tive language models with a mix of different prefix prompts in Penguin, making the system perform better and more efficiently. As an extension of future work, we will explore other techs to build more human-like GRC models.",
+        "bbox": [
+            112,
+            84,
+            487,
+            162
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            175,
+            220,
+            191
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The main target of this paper is to generate natural responses for Chinese Reading Comprehension. We present a new benchmark named Penguin in the hope of creating an effective GRC model. More generally, we expect the proposed Penguin can facilitate the research of MRC, catering to more realistic QA scenarios. Our collected corpus will be soon made public for the whole research communities. Admittedly, one limitation of this work is that the proposed dataset is restricted to Chinese. Another limitation is that our labeled responses are not  $100\\%$  from human annotation due to the high annotation cost. These concerns warrant further research and consideration when utilizing this work to build effective GRC models.",
+        "bbox": [
+            112,
+            200,
+            489,
+            441
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            468,
+            213,
+            482
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Prithviraj Ammanabrolu, Renee Jia, and Mark O. Riedl. 2022. Situated dialogue learning through procedural environment generation. In ACL (1), pages 8099-8116. Association for Computational Linguistics.",
+            "Ashutosh Baheti, Alan Ritter, and Kevin Small. 2020. Fluent response generation for conversational question answering. In ACL, pages 191-207. Association for Computational Linguistics.",
+            "Bin Bi, Chen Wu, Ming Yan, Wei Wang, Jiangnan Xia, and Chenliang Li. 2019. Incorporating external knowledge into machine reading for generative question answering. In EMNLP/IJCNLP (1), pages 2521-2530. Association for Computational Linguistics.",
+            "Peter F. Brown, Stephen Della Pietra, Vincent J. Della Pietra, Jennifer C. Lai, and Robert L. Mercer. 1992. An estimate of an upper bound for the entropy of english. Comput. Linguistics, 18(1):31-40.",
+            "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In NeurIPS."
+        ],
+        "bbox": [
+            115,
+            489,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Nuo Chen, Hongguang Li, Yinan Bao, Junqing He, Xinshi Lin, Qi Yang, Jianfeng Liu, Ruyi Gan, Ji-axing Zhang, Baoyuan Wang, et al. 2023a. Orca: A few-shot benchmark for chinese conversational machine reading comprehension. arXiv preprint arXiv:2302.13619.",
+            "Nuo Chen, Fenglin Liu, Chenyu You, Peilin Zhou, and Yuexian Zou. 2021. Adaptive bi-directional attention: Exploring multi-granularity representations for machine reading comprehension. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7833-7837. IEEE.",
+            "Nuo Chen, Linjun Shou, Ming Gong, and Jian Pei. 2022a. From good to best: Two-stage training for cross-lingual machine reading comprehension. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10501-10508.",
+            "Nuo Chen, Linjun Shou, Ming Gong, Jian Pei, and Daxin Jiang. 2022b. Bridging the gap between language models and cross-lingual sequence labeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1909–1923, Seattle, United States. Association for Computational Linguistics.",
+            "Nuo Chen, Linjun Shou, Tengtao Song, Ming Gong, Jian Pei, Jianhui Chang, Daxin Jiang, and Jia Li. 2023b. Structural contrastive pretraining for crosslingual comprehension. In Findings of the Association for Computational Linguistics: ACL 2023, pages 2042-2057, Toronto, Canada. Association for Computational Linguistics.",
+            "Nuo Chen, Yan Wang, Haiyun Jiang, Deng Cai, Ziyang Chen, and Jia Li. 2022c. What would harry say? building dialogue agents for characters in a story. arXiv preprint arXiv:2211.06869.",
+            "Yiming Cui, Ting Liu, Wanxiang Che, Li Xiao, Zhipeng Chen, Wentao Ma, Shijin Wang, and Guoping Hu. 2019. A span-extraction dataset for chinese machine reading comprehension. In EMNLP/IJCNLP (1), pages 5882-5888. Association for Computational Linguistics.",
+            "Wei He, Kai Liu, Yajuan Lyu, Shiqi Zhao, Xinyan Xiao, Yuan Liu, Yizhong Wang, Hua Wu, Qiaoqiao She, Xuan Liu, et al. 2017. Dureader: a chinese machine reading comprehension dataset from real-world applications. arXiv preprint arXiv:1711.05073.",
+            "Yanzhou Huang and Tao Zhong. 2018. Multitask learning for neural generative question answering. Mach. Vis. Appl., 29(6):1009-1017.",
+            "Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham Neubig. 2022. Understanding and improving zero-shot multi-hop reasoning in generative question answering. In COLING, pages 1765-1775. International Committee on Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "11076",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. Race: Large-scale reading comprehension dataset from examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 796-805. Association for Computational Linguistics.",
+            "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7871-7880. Association for Computational Linguistics.",
+            "Chenliang Li, Bin Bi, Ming Yan, Wei Wang, and Songfang Huang. 2021. Addressing semantic drift in generative question answering with auxiliary extraction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 942-947, Online. Association for Computational Linguistics.",
+            "Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110-119, San Diego, California. Association for Computational Linguistics.",
+            "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.",
+            "Benjamin Muller, Luca Soldaini, Rik Koncel-Kedziorski, Eric Lind, and Alessandro Moschitti. 2021. Cross-lingual genqa: A language-agnostic generative question answering approach for opendomain question answering. CoRR, abs/2110.07150.",
+            "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268.",
+            "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad:  $100,000+$  questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392. Association for Computational Linguistics.",
+            "Tengtao Song, Nuo Chen, Ji Jiang, Zhihong Zhu, and Yuexian Zou. 2023. Improving retrieval-based dialogue system via syntax-informed attention. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE.",
+            "Adam Trischler, Tong Wang, Xingdi Yuan, Justin Harris, Alessandro Sordoni, Philip Bachman, and Kaheer Suleman. 2016. Newsqa: A machine comprehension dataset. arXiv preprint arXiv:1611.09830.",
+            "Bingning Wang, Ting Yao, Qi Zhang, Jingfang Xu, and Xiaochuan Wang. 2020. Reco: A large scale chinese reading comprehension dataset on opinion. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 9146-9153. AAAI Press.",
+            "Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners. In ICLR. OpenReview.net.",
+            "Jun Yin, Xin Jiang, Zhengdong Lu, Lifeng Shang, Hang Li, and Xiaoming Li. 2016. Neural generative question answering. In Proceedings of the Workshop on Human-Computer Question Answering, pages 36-42.",
+            "Chenyu You, Nuo Chen, Fenglin Liu, Shen Ge, Xian Wu, and Yuexian Zou. 2022. End-to-end spoken conversational question answering: Task, dataset and model. In *NAACL-HLT* (Findings), pages 1219–1232. Association for Computational Linguistics.",
+            "Chenyu You, Nuo Chen, and Yuexian Zou. 2021a. Knowledge distillation for improved accuracy in spoken question answering. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7793-7797. IEEE.",
+            "Chenyu You, Nuo Chen, and Yuexian Zou. 2021b. Self-supervised contrastive cross-modality representation learning for spoken question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 28-39, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Rowan Zellers, Yonatan Bisk, Roy Schwartz, and Yejin Choi. 2018. SWAG: A large-scale adversarial dataset for grounded commonsense inference. In EMNLP, pages 93-104. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            910
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "11077",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/41f4309ffe3e683d930cd66e54376491ab6145029b12b039ec9760d7c2f9d45a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Parameter</td><td>T5</td><td>BART</td><td>Prompt</td></tr><tr><td>Batch size</td><td>8</td><td>10</td><td>10</td></tr><tr><td>Learning Rate</td><td>2e-4</td><td>2e-5</td><td>2e-5</td></tr><tr><td>Epoch</td><td>50</td><td>20</td><td>20</td></tr><tr><td>Gradient accumulation steps</td><td>1</td><td>10</td><td>10</td></tr><tr><td>Encoder Max Length</td><td>512</td><td>512</td><td>512</td></tr><tr><td>Decoder Max Length</td><td>40</td><td>128</td><td>128</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            485,
+            181
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/f11da67cb9956734a8df6928a447fbb37a9e7460c4ffb02b06760c3131ee8d76.jpg",
+        "table_caption": [
+            "Table 7: Hyper-parameters setup during fine-tuning BART and T5."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Module</td><td>BLEU-1</td><td>BLEU-2</td><td>ROUGH-L</td></tr><tr><td>GPT-3 (2-shot)</td><td>40.8</td><td>35.5</td><td>46.2</td></tr><tr><td>GPT-3 (5-shot)</td><td>48.7</td><td>42.9</td><td>58.0</td></tr><tr><td>BART-Large</td><td>84.8</td><td>80.0</td><td>84.3</td></tr></table>",
+        "bbox": [
+            124,
+            218,
+            475,
+            282
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Automatic Metrics",
+        "text_level": 1,
+        "bbox": [
+            114,
+            346,
+            314,
+            360
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Dist.: One of the most commonly-used metric that evaluate the diversity and informativeness of generated text. In this paper, we utilize Dist-1 and Dist-2 (Li et al., 2016) to validate the model performance.",
+        "bbox": [
+            112,
+            370,
+            485,
+            450
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "BLEU: Another famous metric is used to evaluate the difference between the generated text and its reference. Similarly, we use BLEU-1 and BLEU-2 (Papineni et al., 2002) in this work.",
+        "bbox": [
+            112,
+            460,
+            485,
+            523
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "ROUGH-L: An effective metric that measures the longest common subsequence (LCS) between model output and reference (Lin, 2004).",
+        "bbox": [
+            112,
+            533,
+            485,
+            580
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Exact Match: This metric measures the percentage of model prediction that match the reference exactly. The evaluated score is 1 if the prediction is exactly the same as the reference, otherwise 0.",
+        "bbox": [
+            112,
+            590,
+            485,
+            653
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Model Architecture",
+        "text_level": 1,
+        "bbox": [
+            112,
+            665,
+            319,
+            680
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "T5 is also a strong pre-trained encoder-decoder Transformer-based model which have achieved strong results on various generative benchmarks (Chen et al., 2022c; Muller et al., 2021; Ammanabrolu et al., 2022). Likewise, we use Chinese T5<sup>5</sup> with small and base versions in our experiments.",
+        "bbox": [
+            112,
+            690,
+            485,
+            801
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "BART is a widely used denoising auto encoder for pre-training sequence-to-sequence models, which is also a standard Transformer-based PLMs. In this work, we utilize Chinese BART with base and large version.",
+        "bbox": [
+            112,
+            803,
+            485,
+            883
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/538bc47ff2803a7de3bca8ecfe244d5868794d857dda43e51572aaf0362f7f6c.jpg",
+        "table_caption": [
+            "Table 8: GPT-3 performances with different few-shot examples."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Module</td><td>BLEU-1</td><td>BLEU-2</td><td>ROUGH-L</td></tr><tr><td>Answer</td><td>79.3</td><td>77.6</td><td>83.6</td></tr><tr><td>+Prompt</td><td>82.7</td><td>80.9</td><td>86.3</td></tr><tr><td>Responder</td><td>82.0</td><td>77.0</td><td>80.9</td></tr><tr><td>+Prompt</td><td>82.9</td><td>78.1</td><td>81.7</td></tr></table>",
+        "bbox": [
+            536,
+            80,
+            855,
+            165
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 9: Model performances with our designed prompts. Here we choose BART-Large as our backbone model. Notice that, we use ground-truth answers rather than responses to evaluate the performances of Answer.",
+        "bbox": [
+            507,
+            174,
+            880,
+            244
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            271,
+            719,
+            288
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 7 shows experimental setup in our experiments.",
+        "bbox": [
+            507,
+            298,
+            880,
+            328
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "D More Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            342,
+            663,
+            357
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We present T5-small, BART-small results in Table 10.",
+        "bbox": [
+            507,
+            369,
+            880,
+            398
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "E More results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            413,
+            655,
+            428
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "E.1 Answer to  $Q_{3}$",
+        "text_level": 1,
+        "bbox": [
+            507,
+            439,
+            665,
+            455
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In order to validate whether GPT-3 achieves promising results in Penguin. We train GPT-3 with few-shot examples, and construct a mini test set via randomly sampling 200 examples in Penguin's test set. Table 8 illustrates their corresponding results, we can observe there is a significant performance gap between GPT-3 and BART-Large. The results show current large language models with more 100 billion-level parameters can not address the challenge in our dataset, proving the difficulty of our Penguin.",
+        "bbox": [
+            507,
+            461,
+            882,
+            637
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "E.2 Answer to  $Q_{4}$",
+        "text_level": 1,
+        "bbox": [
+            507,
+            650,
+            665,
+            665
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we are devoted to investigating where the majority of the improvement that designed prompts bring to the two-stage framework stems from. We show more detailed results of Answerer and Responser in Table 9. We can easily draw the conclusions that 1) a well-performed Answerer is the main prerequisite for a well-performed Responser. 2) The benefits of the designed prompts are most seen in the Answerer module, which enhance Answerer to generate more accurate answers. For instance, Answerer + prompt obtains more than 3 points in BLEU scores.",
+        "bbox": [
+            507,
+            671,
+            882,
+            864
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "F Generated Cases of Question Types",
+        "text_level": 1,
+        "bbox": [
+            507,
+            876,
+            850,
+            892
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We show typical cases in Table 11 and Table 12.",
+        "bbox": [
+            507,
+            903,
+            865,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>5</sup>https://github.com/ZhuiyiTechnology/t5-pegasus",
+        "bbox": [
+            134,
+            890,
+            438,
+            904
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{6}$ https://github.com/fastnlp/CPT",
+        "bbox": [
+            136,
+            904,
+            334,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11078",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/d32d2b349c2bed6d35ae1eb0515cd5f89c186635902ba21160699b57f12cad87.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Framework</td><td>BackBone</td><td>Dist-1</td><td>Dist-2</td><td>BLEU-1</td><td>BLEU-2</td><td>ROUGH-L</td><td>EM</td><td>Average</td></tr><tr><td rowspan=\"2\">End-to-End</td><td>T5-small</td><td>2.7</td><td>34.5</td><td>80.1</td><td>74.8</td><td>78.9</td><td>30.5</td><td>50.3</td></tr><tr><td>BART-base</td><td>2.7</td><td>34.4</td><td>82.1</td><td>77.2</td><td>81.0</td><td>35.6</td><td>52.1</td></tr><tr><td rowspan=\"2\">Two-Stage</td><td>T5-small</td><td>2.7</td><td>34.0</td><td>80.2</td><td>75.0</td><td>79.0</td><td>30.8</td><td>50.3</td></tr><tr><td>BART-base</td><td>2.8</td><td>34.4</td><td>82.1</td><td>77.3</td><td>81.1</td><td>36.2</td><td>52.3</td></tr></table>",
+        "bbox": [
+            166,
+            70,
+            831,
+            167
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/9e797ac4b64adbf7ec8db062485a2ebca39e9a3fae79d11c7d473f1e9fad3584.jpg",
+        "table_caption": [
+            "Table 10: End-to-end and two-stage frameworks' performances with different backbone models. EN and CN refer to English and Chinese, separately. We highlight the best performances in the table."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Question Type</td><td>YES/NO (是否类)</td></tr><tr><td>Passage</td><td>...Three-year-old baby eat two or three at a time on the line, the child&#x27;s gastrointestinal function is relatively weak, eating more will cause gastrointestinal discomfort... ...三周岁宝宝每次吃两三个就行了,孩子的胃肠道功能比较弱,吃多了会引起肠胃不适...</td></tr><tr><td>Question</td><td>Is it good for three-year-old baby to litchi? 三周岁宝宝吃荔枝好吗?</td></tr><tr><td>Answer</td><td>Yes. 好</td></tr><tr><td>Expected Response</td><td>Three-year-old babies can eat two or three litchis, but not too much. 三周岁的宝宝可以吃两三个荔枝,但是不能吃太多。</td></tr><tr><td></td><td>Generated Response-One-Stage</td></tr><tr><td>T5</td><td>They can not eat litchi. 他们不可以吃荔枝</td></tr><tr><td>BART</td><td>They can eat litchi. 他们可以吃荔枝</td></tr><tr><td>Prompt-BART</td><td>They can eat litchi. 他们可以吃荔枝</td></tr><tr><td></td><td>Generated Response-Two-Stage</td></tr><tr><td>T5</td><td>Three-year-old baby can not eat litchi. 三周岁的宝宝不可以吃荔枝</td></tr><tr><td>BART</td><td>They can eat litchi. 他们可以吃荔枝</td></tr><tr><td>Prompt-BART</td><td>Three-year-old baby can eat litchi. 三周岁的宝宝可以吃荔枝</td></tr><tr><td></td><td>(a) A Generated Response example of Yes/No questions.</td></tr><tr><td>Question Type</td><td>Opinion (观点类)</td></tr><tr><td>Passage</td><td>...It&#x27;s not a matter of how fast you move, it&#x27;s a matter of understanding. I have been learning the piano for 10 years and I have to practice these czernies and so on over and over again this year. It is recommended to learn Bayer first, which greatly improves the basic skills.... ...进度快不快不是问题,主要看你的理解能力。我学了10年琴,车尔尼这些等等今年都要反复的练习。推荐还是先学拜厄,对基本功有很大的提高。...</td></tr><tr><td>Question</td><td>I&#x27;ve already learned Faber, so do I still need to practice Baier? 我已经学习了菲伯尔,那还需要练习拜厄吗?</td></tr><tr><td>Answer</td><td>Of course. 需要</td></tr><tr><td>Expected Response</td><td>You&#x27;d better practice Baier. It can improve your basic skills. 推荐还是练习拜厄,这对你的基本功有好处。</td></tr><tr><td></td><td>Generated Response-One-Stage</td></tr><tr><td>T5</td><td>You need to practice Baier. 还是要练习拜厄</td></tr><tr><td>BART</td><td>You still need to practice Baier. 仍然要练习拜厄</td></tr><tr><td>Prompt-BART</td><td>You still need to practice Baier. 仍然要练习拜厄</td></tr><tr><td></td><td>Generated Response-Two-Stage</td></tr><tr><td>T5</td><td>You still need to practice Baier. 仍然要练习拜厄</td></tr><tr><td>BART</td><td>You still need to practice Baier. 仍然要练习拜厄</td></tr><tr><td>Prompt-BART</td><td>He still needs to practice Baier. 他仍然要练习拜厄</td></tr><tr><td></td><td>(b) A Generated Response example of Opinion questions.</td></tr></table>",
+        "bbox": [
+            137,
+            217,
+            858,
+            847
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 11: Typical generated examples from each model in Penguin. English translation is also given for comparison.",
+        "bbox": [
+            112,
+            857,
+            880,
+            872
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "11079",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/5a88946cf8c50c27c7de13cec996a6a933bd7fff04e937bd61552d94019e35b2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Question Type</td><td>Fact (事实类)</td></tr><tr><td>Passage</td><td>...Zhang Lei is a Chinese football player and goalkeeper. Zhang Lei was selected to Guang-dong Hongyuan Youth Team at the age of 13, then played as the main goalkeeper for Dongguan Southern City before moving to Shenzhen in 2006. He was selected to the 2008 National Olympic Team and the 2005 World Youth Championships squad. In 2009, Beijing Guoan exchanged Cheng Yuelei, a backup goalkeeper, plus 400,000 yuan to Shenzhen in exchange for Zhang Lei, which was also the last domestic import transaction of Beijing Guoan in the first transfer market of the 2009 season. Zhang Lei did not get a chance to play because Beijing Guoan had Yang Zhi as their keeper. He was listed by Beijing Guoan and eventually moved to Changsha Jinde in 2010 after losing their main goalkeeper Song Zhenyu. In 2011, Zhang moved to Chongqing Lifan and became the team&#x27;s main goalkeeper. Served for five years, the first three years to play the main force, the last two years gradually as a substitute. In 2016, he was transferred to Hangzhou Greentown. Zhang Lei joined the Lijiang Flying Tigers on loan from Hangzhou Greentown in 2017....张磊，中国足球运动员��司职守门员。张磊13岁时入选广东宏远少年队，随后担任东莞南城的主力门将，2006年转会到深圳队。他入选了08之星国奥队，并进入2005年世青赛的大名单。2009年北京国安用替补门将程月磊加上40万元，与深圳队交换张磊，这也是2009赛季中超第一次转会市场上北京国安队最后一笔内援引进交易。因为北京国安有杨智把关，所以张磊没有得到出场机会，结果被北京国安挂牌，最终在2010年转会到失去主力门将宋振瑜的长沙金德，成为球队新的主力门将。2011年，张磊转会到重庆力帆成为球队的主力门将。效力了五年期间，前三年打主力，后两年渐为替补。2016年转会到杭州绿城。2017年，张磊从杭州绿城租借到丽江飞虎。...</td></tr><tr><td>Question</td><td>In what year did Zhang Lei transfer to Shenzhen?张磊哪一年转会去的深圳?</td></tr><tr><td>Answer</td><td>2006. 2006</td></tr><tr><td>Expected Response</td><td>He transfers to Shenzhen in 2006. 他于2006年转会到深圳。</td></tr><tr><td colspan=\"2\">Generated Response-One-Stage</td></tr><tr><td>T5BARTPrompt-BART</td><td>He transferred to Shenzhen in two years. 他于2年转会到深圳队He transferred to Shenzhen in two years. 他于2年转会到深圳队He transfers to Shenzhen in 2006. 他于2006年转会到深圳</td></tr><tr><td colspan=\"2\">Generated Response-Two-Stage</td></tr><tr><td>T5BARTPrompt-BART</td><td>He transfers to Shenzhen in 2006. 他于2006年转会到深圳He transfers to Shenzhen in 2006. 他于2006年转会到深圳Zhang Lei transfers to Shenzhen in 2006. 张磊于2006年转会到深圳</td></tr></table>",
+        "bbox": [
+            137,
+            80,
+            858,
+            580
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 12: A Generated Response example of Fact questions in Penguin. English translation is also given for comparison.",
+        "bbox": [
+            112,
+            589,
+            882,
+            619
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "G All Prompts",
+        "text_level": 1,
+        "bbox": [
+            112,
+            643,
+            260,
+            659
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We design five prompts for each module in end-to-end and two-stage frameworks, which are presented in Table 13.",
+        "bbox": [
+            112,
+            668,
+            489,
+            715
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "11080",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/97e722c508ea6716f37dcd96b2b4f2d4f074c7d38667317d9b024539937bcc36.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Framework</td><td>Prompts</td></tr><tr><td>One-Stage</td><td>1. Please generate a reasonable response based on the following question and passage. Question: Q.Passage: P.\n2. The question is Q, give the reasonable response based on the following paragraph: P.\n3. This is a reading comprehension problem, please generate the reasonable response based on the following question and passage. Question is Q. Passage is P.\n4. Please answer the question according to descriptions of the question and paragraph. Question is Q.Passage is P.\n5. The question is Q, Conduct question understanding and response generation, knowledge question and response task. Relevant paragraphs are as follows: P.</td></tr><tr><td>Answerer</td><td>1. Please generate the reasonable answer based on the following question and passage. Question: Q.Passage: P.\n2. The question is Q, give the reasonable answer based on the following paragraph: P.\n3. This is a reading comprehension problem, please generate the reasonable answer based on the following question and passage. Question is Q. Passage is P.\n4. Please answer the question according to descriptions of the question and paragraph. Question is Q.Passage is P.\n5. The question is Q, Conduct question understanding and answer generation, knowledge question and answer task. Relevant paragraphs are as follows: P.</td></tr><tr><td>Responder</td><td>1. Please generate the reasonable response based on the following question and answer. Question: Q.Answer: A.\n2. The question is Q, give the reasonable response based on the following answer: A.\n3. This is a answer rewriting problem, please generate the reasonable response based on the following question and answer. Question is Q. Answer is A.\n4. Please give the response according to descriptions of the question and answer. Question is Q. Answer is A.\n5. The question is Q, Conduct question understanding and response generation, answer rewriting task.Relevant answer is as follows: A.</td></tr></table>",
+        "bbox": [
+            126,
+            312,
+            872,
+            675
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 13: All used prompts during training. English translation is only given for reading.",
+        "bbox": [
+            196,
+            684,
+            798,
+            701
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "11081",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/Natural Response Generation for Chinese Reading Comprehension/deacd081-229d-47a6-82a1-c3842e05b39b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5bba15ed0c5a6d9111c973dfbc6d1595c30d0abc3161a1ee1af6fcccfb0a3c2
+size 982101

2023/Natural Response Generation for Chinese Reading Comprehension/full.md

@@ -0,0 +1,346 @@
+# Natural Response Generation for Chinese Reading Comprehension
+
+Nuo Chen†,* Hongguang Li‡, Yinan Bao‡
+
+Baoyuan Wang\* and Jia Li
+
+$^{\dagger}$ Hong Kong University of Science and Technology (Guangzhou),
+
+Hong Kong University of Science and Technology
+
+‡ Xiaobing.AI
+
+†chennuo26@gmail.com, jialee@ust.hk
+
+# Abstract
+
+Machine reading comprehension (MRC) is an important area of conversation agents and draws a lot of attention. However, there is a notable limitation to current MRC benchmarks: The labeled answers are mostly either spans extracted from the target corpus or the choices of the given candidates, ignoring the natural aspect of high-quality responses. As a result, MRC models trained on these datasets can not generate human-like responses in real QA scenarios. To this end, we construct a new dataset called Penguin to promote the research of MRC, providing a training and test bed for natural response generation to real scenarios. Concretely, Penguin consists of 200k training data with high-quality fluent, and well-informed responses. Penguin is the first benchmark towards natural response generation in Chinese MRC on a relatively large scale. To address the challenges in Penguin, we develop two strong baselines: end-to-end and two-stage frameworks. Following that, we further design Prompt-BART: fine-tuning the pre-trained generative language models with a mixture of prefix prompts in Penguin. Extensive experiments validated the effectiveness of this design. Our benchmark and codes are available at https://github.com/nuochenpku/Penguin.
+
+# 1 Introduction
+
+Machine Reading comprehension (MRC) aims to empower the machine to correctly answer queries based on the given passages or paragraphs, which is an active field in NLP. Much of this popularity can be attributed to the release of many annotated and publicly available datasets (Rajpurkar et al., 2016; Trischler et al., 2016; Chen et al., 2022a; You et al., 2022; Chen et al., 2023a). Formally, these MRC efforts can be classified into two most
+
+popular streams<sup>1</sup> from the answer type perspective: span-extraction (Rajpurkar et al., 2016; Trischler et al., 2016; Cui et al., 2019; Chen et al., 2022b; You et al., 2021a) and multiple choices (Lai et al., 2017; Zellers et al., 2018; Wang et al., 2020). The former requires the model to locate the text span in the given passage as the answer, e.g., SQuAD (Rajpurkar et al., 2016) and NewQA (Trischler et al., 2016). SQuAD, for example, introduces extracted span answers. The latter needs to choose the answer from a list of candidates like RACE (Lai et al., 2017). Considering most of the existing works are built in English, several efforts have been proposed to further advance MRC in Chinese such as DuReader (He et al., 2017), CMRC 2018 (Cui et al., 2019), ReCo (Wang et al., 2020).
+
+Nonetheless, despite significant achievements in the various types and on a relatively large scale, a significant barrier that has rarely been discussed in previous MRC efforts remains: Annotated answers among these datasets are almost limited to spans in the source document or the given candidates. While spans directly copied from the source document and simple words like "yes/no" can be seen as ground-truth answers, they don't sound like natural responses in real-life scenarios. In reality, the responses of humans are always generated with flexibility and information rather than being forced to select certain type spans (e.g., entities and person) or one of the candidates.
+
+This work devotes itself to natural response generation for Chinese MRC, catering to real-world QA scenarios. Intuitively, we can try two existing technical routes to generate natural responses towards MRC at the moment: 1) Optimizing generative language models (GLMs) such as BART and T5 with predetermined and inflexible answers on the above benchmarks. 2) Use extremely large-
+
+<table><tr><td>Passage</td><td>...Bream, which is traditionally steamed or braised, is a popular fish. Soup should of course be cooked, but soles is not suitable for this dish because of its fleshy nature...
+...鳊鱼一直备受人们喜爱，鳊鱼的传统做法是清蒸或者红烧。当然不排除可以烧汤，但是鉴于鳊鱼的肉质特色，不是很适合烧汤的...</td></tr><tr><td>Question</td><td>Can bream be stewed in soup?
+鳊鱼可以炖汤吗?</td></tr><tr><td>Candidates</td><td>Of Course | No, you can&#x27;t | Uncertain
+可以|不可以|不确定</td></tr><tr><td>BART</td><td>Of Course
+可以</td></tr><tr><td>GPT-3</td><td>Yes! If you want to stew soup, in think-ing, then we still choose to have.
+是可以的！想要炖汤的话，在思考，那么我们还是选择有。</td></tr><tr><td>Expected
+Response</td><td>Bream can be used in soups but is generally not recommended.
+鳊鱼可以用来煮汤，但是一般不推荐这么做。</td></tr></table>
+
+Table 1: An example from ReCo (Wang et al., 2020). The text in bold refers to the answer in the candidates or the expected response. English translation is also given for comparison.
+
+scale pre-trained GLMs like GPT-3 (more than 100 billion parameters) (Brown et al., 2020) with few-shot in-context learning.
+
+We showcase a multi-choice MRC example from ReCo (Wang et al., 2020) in Table 1, where the query is a typical yes-or-no question and the answer is required to be one of the candidates. We fine-tune BART-large (Lewis et al., 2020) on ReCo and find that the resulting model can accurately predict the answer: "of course" in this case. However, the prediction is accurate but stiff and lacking in detail. On the other hand, we use GPT-3 under few-shot from ReCo and validate its generated response via the same case in Table 1. Obviously, the produced result "Yes! If you want to... we still choose to have" is fluent yet incoherent and semantically incorrect. Ideally, the qualified response of humans should be informative and fluent, like "Bream can be used .... not recommended". In other words, although these models are proficient, they can't produce fluent and accurate responses for reading comprehension.
+
+Based on the above analysis, we can draw the conclusion that it is impossible to build an effective MRC model for natural $^2$  response generation in the absence of a benchmark with well-labeled
+
+responses. To this end, we move forward to push the boundary of MRC: Firstly, we collect the first Chinese generative reading comprehension (GRC) dataset named Penguin to facilitate research of natural response generation for Chinese MRC. Formally, Penguin poses a new challenge to current MRC models: Not only should the responses be accurate, but they also need to be natural (as fluent and informative as possible). Concretely, Penguin contains 200k training and a 15K test corpus with high-quality fluent and informative responses. Table 2 shows the overview of current MRC datasets and Penguin.
+
+Secondly, we design two GRC pipelines based on this non-trivial dataset: i) end-to-end; ii) two-stage training. For the former, we directly take the question and corresponding passage as input, generating a sequence of words as responses. The latter consists of two modules: Answerer and Responser. Answerer is responsible for generating initial answers based on the question and passage. Then Responser rewrites and expands the generated answers into human-like responses. Experimentally, two pipelines achieve promising results in Penguin.
+
+Finally, we propose Prompt-BART, an approach for fine-tuning pre-trained generative language models using a wide variety of prefix prompts in Penguin. The resulting model, as stated in the prompt that accompanied it, boosted the overall performance.
+
+# 2 Related Work
+
+Machine Reading Comprehension During the past few years, Machine Reading Comprehension (MRC) (Rajpurkar et al., 2016; Trischler et al., 2016; You et al., 2021b; Chen et al., 2021; Song et al., 2023; Chen et al., 2023b) receives lots of attention from research communities and various related benchmarks have been proposed: SQuAD (Rajpurkar et al., 2016), MS-MarCO (Nguyen et al., 2016), NewQA (Trischler et al., 2016) and RACE (Lai et al., 2017), etc. Following this line of research, some works aim to accelerate the development of Chinese MRC on language diversity, such as DuReader (He et al., 2017), CMRC 2018 (Cui et al., 2019), ReCo (Wang et al., 2020).
+
+These well-designed and large-scale benchmarks facilitate the research of various types of MRC tasks, but they ignore the fact that answers are primarily confined to text span extraction from the source document or choosing candidates. In com
+
+parison with natural human responses, these answers are less informative and less fluid. In other words, models trained on the above datasets do not have strong ability to generate fluent, natural as well as informative responses. Although recent works (Bi et al., 2019; Li et al., 2021) develop sequence-to-sequence frameworks to address the challenge in MS-MarCO (Nguyen et al., 2016), they still regard some abstractive answers as the optimized objective, which consist of some uninformative extracted spans from the passage, leading to responses that are far away from fluent and fluid text. To this end, we propose a new dataset Penguin to encourage further progress in natural response generation of Chinese MRC.
+
+Natural Generative Question Answering Similar to our work, natural generative question answering (NGQA) (Yin et al., 2016; Huang and Zhong, 2018; Bi et al., 2019; Jiang et al., 2022) also aims at generating natural responses to questions. Although both tasks share a similar target, the research question and task settings differ: Firstly, NGQA produces responses to simple factoid questions based on the naive entity triplets that are given as external knowledge. In contrast, we focus on yielding fluent responses based on a deep understanding of the question and passage. Such a deep understanding of interaction between question and passage is more challenging than the simple entity triplets. Moreover, GRC closely resembles humans asking questions about a certain topic and generating corresponding replies. Last, previous NGQA efforts were limited to the English corpus, whereas we are expanding it to Chinese.
+
+# 3 DataSet
+
+# 3.1 Task Definition
+
+Generally, machine reading comprehension task can be illustrated as  $\langle \mathcal{P}, \mathcal{Q}, \mathcal{A} \rangle$ , where  $\mathcal{P}$  is the paragraph or the whole passage,  $\mathcal{Q}$  represents questions,  $\mathcal{A}$  refers to answers. Previous works tend to model MRC as a classification task with the objective of selecting  $\mathcal{A}$  from several candidates or finding the text span of  $\mathcal{A}$  in  $\mathcal{P}$ . In contrast, we formalize MRC as a sequence-to-sequence task, where the objective is to generate fluent and factual responses:  $\mathcal{R}$ . In this work, we include  $\langle \mathcal{P}, \mathcal{Q}, \mathcal{A}, \mathcal{R} \rangle$  in Penguin.
+
+<table><tr><td>Dataset</td><td>Lang.</td><td>Answer Type</td><td>Size</td></tr><tr><td>SQuAD (2016)</td><td>EN</td><td>Span</td><td>100K</td></tr><tr><td>RACE (2017)</td><td>EN</td><td>Multi-Choices</td><td>87K</td></tr><tr><td>MarCO (2016)</td><td>EN</td><td>Free-Form</td><td>100K</td></tr><tr><td>NewsQA (2016)</td><td>EN</td><td>Span</td><td>100K</td></tr><tr><td>DuReader (2017)</td><td>CN</td><td>Free-Form</td><td>200k</td></tr><tr><td>CMRC (2019)</td><td>CN</td><td>Span</td><td>20K</td></tr><tr><td>ReCo (2020)</td><td>CN</td><td>Multi-Choices</td><td>300K</td></tr><tr><td>Penguin</td><td>CN</td><td>Natural Response</td><td>200K</td></tr></table>
+
+Table 2: Datasets Comparison. EN and CN refer to English and Chinese, separately.
+
+# 3.2 Dataset Collection
+
+As aforementioned, a high-quality dataset is a prerequisite for building effective GRC models. Unfortunately, current datasets don't contain well-informed and natural responses. To facilitate the study of MRC, we collect a new dataset: Penguin, in the hopes of creating sophisticated GRC models that can generate natural responses for practical QA situations. Considering constructing such a dataset on a large scale is non-trivial, we initialize our dataset from the current Chinese MRC dataset corpus to get raw passage-question-answer  $(<\mathcal{P}, \mathcal{Q}, \mathcal{A})$  triplets, including CMRC 2018, DuReader, and ReCo.
+
+It is extremely difficult and expensive to ask the annotator to start from scratch and write a response  $\mathcal{R}$  that makes sense and sounds natural for each  $<\mathcal{P}, \mathcal{Q}, \mathcal{A}>$  triplet. Therefore, we generate informative and fluent responses via the following steps: (1) We first utilize a specific response generative model (short for Responser) to generate initial responses for the above triplets; (2) Then, we use state-of-the-art semantic matching models alongside certain manually-created criteria to exclude samples of incoherence and semantic shift; (3) Last, we employ three professional annotators to rewrite and recheck undesired cases, therefore regulating data quality. As a result, we collect a sequence of 4-tuples:  $<\mathcal{P}, \mathcal{Q}, \mathcal{A}, \mathcal{R}>$  in Penguin, where  $\mathcal{R}$  is the labeled response. In the following, we will introduce the detailed procedures of these steps, sequentially. The conceptual process of data collection is presented in Figure 1.
+
+Responder Prior to going further, the main concern is how to get an effective sequence-to-sequence responder model. Furthermore, more pressing than any of these other issues is figuring out where to collect the high-quality Chinese corpus needed to train the responder. Inspired by NGQA works, We first get the English <question,
+
+![](images/0d340609ffe895c643a7c0f4c5381ab5b681f4ad5d6d3f89a882bb0b01994b18.jpg)  
+Figure 1: Data collection pipeline
+
+answer, response> triplets in (Baheti et al., 2020). Then we utilize machine translation systems to translate these data from English to Chinese. Considering translated corpus will inevitably have erroneous examples, we utilize perplexity (Brown et al., 1992) to identify and weed out some bad cases. Following the line of Yin et al. (2016), Bart-large is employed as the backbone of responder to fine-tune on these translated corpus, which takes the question as well as answer as input and then outputs response during training. Subsequently, the resulting responder models are used to inference on each  $\langle \mathcal{P}, \mathcal{Q}, \mathcal{A} \rangle$  triplet from CMRC 2018, DuReader and ReCo, generating initial responses. Concretely, responder takes the passage and question as input, then outputs the response during inference.
+
+Data Filtering Intuitively, the collected initial responses might be stilted, evasive, and semantically challenging to connect with its corresponding questions and answers. Therefore, we employ state-of-the-art Chinese semantic matching models on LCQMC leadboard<sup>3</sup> to compute the sentence similarity of the generated response and answer. Roughly, we will throw out an instance if the simi
+
+<table><tr><td colspan="2">Statistics</td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td colspan="2">&lt;P, Q, A, R&gt;</td><td>200k</td><td>10k</td><td>15k</td></tr><tr><td rowspan="3">Passage tokens</td><td>Max.</td><td>492</td><td>486</td><td>490</td></tr><tr><td>Min.</td><td>24</td><td>19</td><td>31</td></tr><tr><td>Ave.</td><td>84.3</td><td>82.45</td><td>84.85</td></tr><tr><td rowspan="3">Question tokens</td><td>Max.</td><td>61</td><td>37</td><td>17</td></tr><tr><td>Min.</td><td>4</td><td>5</td><td>5</td></tr><tr><td>Ave.</td><td>11.5</td><td>11.6</td><td>11.5</td></tr><tr><td rowspan="3">Answer tokens</td><td>Max.</td><td>20</td><td>11</td><td>20</td></tr><tr><td>Min.</td><td>1</td><td>1</td><td>1</td></tr><tr><td>Ave.</td><td>2.8</td><td>2.6</td><td>2.1</td></tr><tr><td rowspan="3">Response tokens</td><td>Max.</td><td>42</td><td>26</td><td>33</td></tr><tr><td>Min.</td><td>6</td><td>6</td><td>6</td></tr><tr><td>Ave.</td><td>10.1</td><td>9.6</td><td>9.7</td></tr></table>
+
+Table 3: Overall data Statistics in Penguin.
+
+larity score between the response and the answer is below the pre-defined threshold. To be more specific, the generated response just can be seen as replacing a few words in the answer in some cases, resulting in the extremely high similarity between the answer and the response. e.g., given the question "How tall is Yao Ming", and the answer "He is  $2.29\mathrm{m}$ ". the generated response could be "Yao Ming is  $2.29\mathrm{m}$ ." For these cases, the threshold used to filter incoherent samples will be manually larger (more than 0.95). On the other hand, in certain instances, where the generated response provides the final piece of missing information based on the answer, the threshold will be manually lowered (lower than  $0.3)^4$ . Besides, we employ additional scoring and filtering rules to identify the most grammatically correct and contextually relevant response (if any) to each  $\langle \mathcal{P}, \mathcal{Q}, \mathcal{A} \rangle$  triplet. For instance, we remove the samples where the length of the response is less than the length of the answer plus 5.
+
+Human Annotation In the above, we automatically generate the response via the effective responder, and then try to filter out some meaningless examples. To further control the training data quality, we recruit another three professional annotators to re-check each example and revise undesirable cases. In detail, each piece of data is verified by two different annotators. These annotators review the generated responses after reading the relevant question, passage, and answer. If the generated response is unnatural, they try rephrasing the statement to obtain a fair response. Thereafter, another annotator rates their annotations' quality, avoiding bias and getting high-agreement data. As a result,
+
+![](images/cec546d28a321e596f13e046d20fa46b16adc0640b1cf60e9e604155d4e6173c.jpg)  
+Figure 2: Question types in Penguin. Empty colored blocks indicate suffixes that are too rare to show.
+
+we collect high quality 4-tuples:  $< \mathcal{P}, \mathcal{Q}, \mathcal{A}, \mathcal{R} >$ .
+
+Test set Annotation We additionally hire eight expert annotators to choose and create questions and their answers from social media and themselves, bringing the acquired data closer to real-world events and increasing variety. Concretely, eight annotators are split into four smaller groups, and each of them consists of one questioner and one answerer. Questioners first select questions or question the news they are interested in from current social media platforms like weibo and zhihu. Then, answerers get related passages that contain corresponding knowledge about each question from Bing search. Following that, they must select the excerpt from the retrieved ones that is the most relevant to the annotated question at hand. After reading the question and its accompanying passage, they then annotate responses at last. Similarly, we recruit two annotators to recheck the data twice, further guaranteeing data quality. This way, we collect 15k data in our test set.
+
+The general statistics in Penguin are presented in Table 3. In total, we collect train, development, and test set, which consists of  $200\mathrm{k}$ ,  $10\mathrm{k}$ , and  $15\mathrm{k} < \mathcal{P}$ ,  $\mathcal{Q}, \mathcal{A}, \mathcal{R}$  data, separately. The collected responses are noticeably more informative than the answers given, as seen by their length.
+
+# 3.3 Data Statics
+
+Question Analysis The pilot investigations (He et al., 2017) assisted us in reaching an agreement on the following question type taxonomy: Yes/No,
+
+<table><tr><td>Type</td><td>Example</td><td>Percentage</td></tr><tr><td>Extension</td><td>Q: The design life of Toyota engine?A: 0.24 million kmR: Toyota engine has a de-sign life of 0.24 million km.</td><td>62.4%</td></tr><tr><td>Rewritten</td><td>Q: Can we carry tea in trains?A: YesR: Of course, we can bring tea in trains.</td><td>26.3%</td></tr><tr><td>Others</td><td>Q: Can you fix the Apple 5S if it&#x27;s flooded?A: It can be fixed.R: In general, if the iPhone 5s gets water, it can defi-nitely be repaired.</td><td>11.3%</td></tr></table>
+
+Table 4: Comparison between answer and response in Penguin. For ease of reading, English translation data is only provided here.
+
+Fact, and Opinion. Yes/No questions are often asked for confirmation, and Fact questions are mostly about specific entities, timing, or description of some facts. Opinion questions always contain why/what/how questions, inquiring about the reason or advice of things. We draw the distribution of these three question types from our collected Penguin in Figure 2, showing Penguin contains various questions, catering to more realistic scenarios.
+
+Response Analysis After deep analysis of responses and answers in Penguin, we classify the connection between them into three broad categories: Extension, Rewritten, Others. Extension refers to the response containing exactly the content of the answer and going into more detail about it. Of note, although these answers are semantically right, we argue that the responses that extend to them could be more informative and cater to real human responses. Rewritten denotes the response totally rewritten the answer to make it more informative and natural while retaining the original meaning. That is to say, the response almost has no overlap with the answers in this situation. Others indicates the response may contain part of the answer, as well as expands it with lexical, coreferential, or casual inferential knowledge. We showcase of each category in Table 4, respectively.
+
+# 4 Model
+
+In this paper, we propose two frameworks to address the challenge in Penguin: 1) end-to-end; 2) two stage.
+
+![](images/ee5b3a037fe0bfeafa0c37f89802315da3989d9ff181bc7dd19d694efb226bdd.jpg)  
+Figure 3: Overview of our proposed frameworks: (a) End-to-end; (b) Two-stage. Answerer denotes a answer generator module and Responser refers to a response generator module. English translation is only given for reading.
+
+![](images/1847cd1e6506db16962f8226db0d2dce35867f76846ac405af9171df549834f9.jpg)  
+Figure 4: Prompt templates.
+
+End-to-end Intuitively, we can directly concat question and passage to construct the input  $\mathbf{X}_{\mathrm{e}}$ :
+
+$$
+\mathbf {X} _ {\mathrm {e}} = \left\{\left[ \text {C L S} \right] \text {Q u e s t i o n [ S E P ] P a s s a g e [ S E P ]} \right\} \tag {1}
+$$
+
+where [CLS] and [SEP] refer to the beginning and separate symbols. Then a pre-trained generative model is utilized to take  $\mathbf{X}_{\mathrm{e}}$  as input, and then generates a response after deeply understanding of the interaction between question and passage, as shown in Figure 3 (a). Concretely, we optimize the model with the widely used negative log-likelihood loss during training:
+
+$$
+\begin{array}{l} \mathcal {L} _ {e} = - \log \left(p _ {\theta} \left(\mathcal {R} \mid \mathbf {X} _ {\mathbf {e}}\right)\right) \\ = - \sum_ {i = 1} ^ {| \mathcal {R} |} \log \left(p _ {\theta} \left(r _ {i} \mid \mathbf {X} _ {\mathbf {e}}, \mathcal {R} _ {<   i}\right)\right). \tag {2} \\ \end{array}
+$$
+
+Two stage Besides, we also design a two-stage framework to deal with the challenge in Penguin
+
+which consists of a answer generator module (short for Answerer) and a response generator module (short for Responser), seen in Figure 3 (b). In two stage frameworks, Answerer can be thought of as a reading comprehension module that is in charge of thoroughly comprehending the passage and the question before generating a answer; then Responser rewrites and extends the generated answer to a more human-like response.
+
+To be more specific, we first take  $\mathbf{X}_{\mathrm{e}}$  into Answerer, which yields initial answer:
+
+$$
+\mathcal {L} _ {a} = - \sum_ {j = 1} ^ {| \mathcal {A} |} \log \left(p _ {\alpha} \left(a _ {j} \mid \mathbf {X} _ {\mathbf {e}}, \mathcal {A} _ {<   j}\right)\right). \tag {3}
+$$
+
+Then we construct a new input  $\mathbf{X}_{\mathbf{r}}$  with generated answer and question:
+
+$$
+\mathbf {X} _ {\mathbf {r}} = \left\{\left[ \text {C L S} \right] \text {Q u e s t i o n [ S E P ] A n s w e r [ S E P ]} \right\} \tag {4}
+$$
+
+Subsequently, Responser is responsible for reading the  $\mathbf{X}_{\mathbf{r}}$  to predict the target response  $\mathcal{R}$ :
+
+$$
+\mathcal {L} _ {r} = - \sum_ {i = 1} ^ {| \mathcal {R} |} \log \left(p _ {\theta} \left(r _ {i} \mid \mathbf {X} _ {\mathbf {r}}, \mathcal {R} _ {<   i}\right)\right). \tag {5}
+$$
+
+Notice that, Answerer and Responser are both the same generative models but not sharing parameters in our experiments.
+
+Prompt-Tuning Inspired by recent works (Wei et al., 2022; Brown et al., 2020), we fine-tune in Penguin with a variety of prefix prompt with different template types. Concretely, we manually design five templates for each generative module from end-to-end and two-stage frameworks, as several examples shown in Figure 4. We use these prompts to formulate different inputs in training, obtaining the final checkpoint. Experimentally, we choose
+
+<table><tr><td>Framework</td><td>BackBone</td><td>Dist-1</td><td>Dist-2</td><td>BLEU-1</td><td>BLEU-2</td><td>R.-L</td><td>EM</td><td>Relv.P</td><td>Flu.</td><td>Info.</td></tr><tr><td rowspan="3">End-to-End</td><td>T5-base</td><td>2.7</td><td>33.5</td><td>80.8</td><td>75.4</td><td>79.1</td><td>31.4</td><td>3.41</td><td>3.79</td><td>3.84</td></tr><tr><td>BART-Large</td><td>2.7</td><td>34.4</td><td>82.0</td><td>77.1</td><td>80.8</td><td>35.1</td><td>3.50</td><td>3.88</td><td>3.84</td></tr><tr><td>Prompt-BART</td><td>2.8</td><td>34.7</td><td>82.4</td><td>77.6</td><td>81.0</td><td>35.5</td><td>3.56</td><td>3.87</td><td>3.90</td></tr><tr><td rowspan="3">Two-Stage</td><td>T5-base</td><td>2.7</td><td>33.8</td><td>80.4</td><td>75.0</td><td>79.0</td><td>31.0</td><td>3.40</td><td>3.84</td><td>3.87</td></tr><tr><td>BART-Large</td><td>2.7</td><td>33.7</td><td>82.5</td><td>77.3</td><td>81.2</td><td>36.7</td><td>3.67</td><td>3.87</td><td>3.90</td></tr><tr><td>Prompt-BART</td><td>2.9</td><td>33.8</td><td>83.2</td><td>78.2</td><td>81.8</td><td>37.2</td><td>3.76</td><td>3.87</td><td>3.93</td></tr></table>
+
+Table 5: End-to-end and two-stage frameworks' performances with different backbone models. EN and CN refer to English and Chinese, separately. We highlight the best performances in the table. For smaller versions' the results of each backbone model can be seen in Appendix D. R.-L and EM refer to ROUGH-L and Exact Match, separately. We run each model three times and report their average results.
+
+BART-large as our baseline, the resulting model is named as Prompt-BART after prompt-tuning. All prompts can be seen in Appendix G, Table 13.
+
+# 5 Experiments
+
+# 5.1 Evaluation Metric
+
+Different from previous span-extraction and multiple-choice MRC tasks, we use automatic and human metrics to evaluate the model performances.
+
+Automatic Metrics We utilize following commonly-used metrics to evaluate the quality of generated responses: Dist-1, Dist-2 (Li et al., 2016), BLEU-1,BLEU-2 (Papineni et al., 2002), ROUGH-L and Exact Match. We give detailed explanations of these metrics in Appendix A. For the above metrics, the larger the value, the better language modeling performance.
+
+Human Metrics To further get a fuller picture of how well the generation-based model works, we asked 4 annotators to look at the quality and consistency of the generated responses on our test set. These annotators are from a third-party and skilled with language tasks but have no knowledge of the models. Concretely, we randomly sample 500 examples in our test set for model's evaluation. Human annotators are tasked with rating dialogue quality using three principle measures: Fluency (Flu.), Informativeness (Info.) and Relevance with Passage (Relv.P). Each criterion is given a rating from 1 to 5, with 1, 3, 5 representing poor, mediocre and flawless performances.
+
+# 5.2 Backbone
+
+To investigate the performances of current state-of-the-art generative models, we choose two popular pre-trained generative models: T5 (Raffel et al.,
+
+![](images/c6076eb4fbbeb6b04e68e194db77d4201b8800091711ee6ef4f68e4c2418ec3d.jpg)  
+Figure 5: Model performances on different question types. Here, we choose BART-large as the backbone.
+
+2020) and BART (Lewis et al., 2020) as our backbones. In the Appendix B, we briefly introduce each model. Detailed experimental setup of training these models can be seen in Appendix C.
+
+# 5.3 Results
+
+Automatic Evaluations Table 5 shows the main results of all models. From the table, we can conclude the following: 1) The designed two frameworks with different backbone models all achieve competitive results in each automatic metric, proving their effectiveness. 2) Importantly, the results of each base model under the two different frameworks are comparable. More interestingly, an end-to-end framework is a more advanced and promising system when you take into account that the parameters and inference cost of a two-stage framework are almost two times higher. 3) Prompt-tuning can further boost the model's performances in all evaluation metrics, and the improvement is more obvious in two-stage training (e.g., 81.8 vs. 81.2 in ROUGH-L).
+
+Human Evaluations We also draw the following conclusions from the Table 5: 1) All models achieve good results in fluency, informativeness
+
+![](images/5821d99558c9879fde8ef2918c02d4a5f4f12e690318b57cd42f7de755b92469.jpg)  
+Figure 6: Win/Lose human evaluation results  $(\%)$  between end-to-end and Answer in two-stage pipeline with p-value  $< 0.05$
+
+and relevance, showing their generated responses are close to human-like. 2) Prompt-BART in two-stage training performs better in relevance than other models, indicating a better understanding of the passage. In general, human evaluations of these models basically mirror automated evaluations.
+
+# 5.4 Discussion and Analysis
+
+Aside from the high performance of our proposed models, we are concerned about the following questions:  $\mathbf{Q}_1$ : What does the model perform when it answers different question types?  $\mathbf{Q}_2$ : Is the generated response more natural to the generated answer?  $\mathbf{Q}_3$ : Can extremely large model like GPT-3 almost address the challenge in Penguin? (Appendix E.1, Table 8)  $\mathbf{Q}_4$ : What is the majority of the improvement brought by the designed prompts? (Appendix E.2, Table 9).
+
+Answer to  $\mathbf{Q}_1$ : We conduct fine-grained analysis to validate model performances on different question types. As seen in Figure 5, we can observe: 1) each model performs best on Yes/No questions and worst on Fact questions, indicating generating natural responses about specific facts needs deeper understanding of questions and passages; 2) prompt-tuning brings more benefits when answering Fact questions (76.5 vs. 74.3). We further present each case on three question types in Table 11 and Table 12.
+
+Answer to  $\mathbf{Q}_2$ : In order to test whether the generated responses are more natural than the fixed answers. Three professional workers are asked to evaluate the generated responses from end-to-end framework and generated answers from Answerer in two-stage pipeline. Concretely, they need to choose which is the best one (Win/Lose) for
+
+<table><tr><td>Passage</td><td>...If you already have a loan, the bank will definitely consider the debt ratio, depending on the assets and repayment ability....</td></tr><tr><td>Question</td><td>Will Credit Loan Affect Mortgage?</td></tr><tr><td>Response</td><td>It could affect your Mortgage, but it depends on your income level and assets.</td></tr><tr><td colspan="2">Generated Response-One stage</td></tr><tr><td>T5</td><td>Credit loan must affect Mortgage.</td></tr><tr><td>BART</td><td>It will affect your Mortgage.</td></tr><tr><td>Prompt-BART</td><td>Credit loan could affect mortgage, but depends on assets.</td></tr><tr><td colspan="2">Generated Response-Two stage</td></tr><tr><td>T5</td><td>Credit loan must affect Mortgage.</td></tr><tr><td>BART</td><td>Credit loan will affect mortgage</td></tr><tr><td>Prompt-BART</td><td>Credit loan may affect your Mortgage.</td></tr></table>
+
+Table 6: A case study in our collected Penguin, and generated responses from the large version of baselines are provided. The text in bold refers to the expected response from human annotation.
+
+each example in randomly selected 200 test samples from three views, as shown in Figure 6. We can observe that more than  $85\%$  of the generated responses are superior to their counterpart answers across all evaluation perspectives, indicating they are closer to human-like responses than the generated answers. These Win/Lose results also prove the value of Penguin, that is, providing a training and testing bed for natural response generation to real-world scenarios.
+
+Case Study In addition, we showcase an example and its generated responses from Penguin in Table 6. We can see that almost all of the responses from the different models do a pretty good job of roughly capturing the intended meaning. Moreover, Prompt-BART could generate more complete and accurate responses, showing its efficiency.
+
+# 6 Conclusion
+
+In this paper, we summarize the limitation of current MRC efforts, and propose a new dataset named Penguin for natural response generation in Chinese MRC. Penguin requires the predicted responses not only be accurate but also natural-sounding, catering to real scenarios. Thereafter, we propose end-to-end and two-stage GRC pipelines to address the challenge in Penguin, achieving promising results. Moreover, we design prompt-tuning, which involves fine-tuning the pre-trained genera
+
+tive language models with a mix of different prefix prompts in Penguin, making the system perform better and more efficiently. As an extension of future work, we will explore other techs to build more human-like GRC models.
+
+# Limitations
+
+The main target of this paper is to generate natural responses for Chinese Reading Comprehension. We present a new benchmark named Penguin in the hope of creating an effective GRC model. More generally, we expect the proposed Penguin can facilitate the research of MRC, catering to more realistic QA scenarios. Our collected corpus will be soon made public for the whole research communities. Admittedly, one limitation of this work is that the proposed dataset is restricted to Chinese. Another limitation is that our labeled responses are not  $100\%$  from human annotation due to the high annotation cost. These concerns warrant further research and consideration when utilizing this work to build effective GRC models.
+
+# References
+
+Prithviraj Ammanabrolu, Renee Jia, and Mark O. Riedl. 2022. Situated dialogue learning through procedural environment generation. In ACL (1), pages 8099-8116. Association for Computational Linguistics.  
+Ashutosh Baheti, Alan Ritter, and Kevin Small. 2020. Fluent response generation for conversational question answering. In ACL, pages 191-207. Association for Computational Linguistics.  
+Bin Bi, Chen Wu, Ming Yan, Wei Wang, Jiangnan Xia, and Chenliang Li. 2019. Incorporating external knowledge into machine reading for generative question answering. In EMNLP/IJCNLP (1), pages 2521-2530. Association for Computational Linguistics.  
+Peter F. Brown, Stephen Della Pietra, Vincent J. Della Pietra, Jennifer C. Lai, and Robert L. Mercer. 1992. An estimate of an upper bound for the entropy of english. Comput. Linguistics, 18(1):31-40.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In NeurIPS.
+
+Nuo Chen, Hongguang Li, Yinan Bao, Junqing He, Xinshi Lin, Qi Yang, Jianfeng Liu, Ruyi Gan, Ji-axing Zhang, Baoyuan Wang, et al. 2023a. Orca: A few-shot benchmark for chinese conversational machine reading comprehension. arXiv preprint arXiv:2302.13619.  
+Nuo Chen, Fenglin Liu, Chenyu You, Peilin Zhou, and Yuexian Zou. 2021. Adaptive bi-directional attention: Exploring multi-granularity representations for machine reading comprehension. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7833-7837. IEEE.  
+Nuo Chen, Linjun Shou, Ming Gong, and Jian Pei. 2022a. From good to best: Two-stage training for cross-lingual machine reading comprehension. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10501-10508.  
+Nuo Chen, Linjun Shou, Ming Gong, Jian Pei, and Daxin Jiang. 2022b. Bridging the gap between language models and cross-lingual sequence labeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1909–1923, Seattle, United States. Association for Computational Linguistics.  
+Nuo Chen, Linjun Shou, Tengtao Song, Ming Gong, Jian Pei, Jianhui Chang, Daxin Jiang, and Jia Li. 2023b. Structural contrastive pretraining for crosslingual comprehension. In Findings of the Association for Computational Linguistics: ACL 2023, pages 2042-2057, Toronto, Canada. Association for Computational Linguistics.  
+Nuo Chen, Yan Wang, Haiyun Jiang, Deng Cai, Ziyang Chen, and Jia Li. 2022c. What would harry say? building dialogue agents for characters in a story. arXiv preprint arXiv:2211.06869.  
+Yiming Cui, Ting Liu, Wanxiang Che, Li Xiao, Zhipeng Chen, Wentao Ma, Shijin Wang, and Guoping Hu. 2019. A span-extraction dataset for chinese machine reading comprehension. In EMNLP/IJCNLP (1), pages 5882-5888. Association for Computational Linguistics.  
+Wei He, Kai Liu, Yajuan Lyu, Shiqi Zhao, Xinyan Xiao, Yuan Liu, Yizhong Wang, Hua Wu, Qiaoqiao She, Xuan Liu, et al. 2017. Dureader: a chinese machine reading comprehension dataset from real-world applications. arXiv preprint arXiv:1711.05073.  
+Yanzhou Huang and Tao Zhong. 2018. Multitask learning for neural generative question answering. Mach. Vis. Appl., 29(6):1009-1017.  
+Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham Neubig. 2022. Understanding and improving zero-shot multi-hop reasoning in generative question answering. In COLING, pages 1765-1775. International Committee on Computational Linguistics.
+
+Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. Race: Large-scale reading comprehension dataset from examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 796-805. Association for Computational Linguistics.  
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7871-7880. Association for Computational Linguistics.  
+Chenliang Li, Bin Bi, Ming Yan, Wei Wang, and Songfang Huang. 2021. Addressing semantic drift in generative question answering with auxiliary extraction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 942-947, Online. Association for Computational Linguistics.  
+Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110-119, San Diego, California. Association for Computational Linguistics.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+Benjamin Muller, Luca Soldaini, Rik Koncel-Kedziorski, Eric Lind, and Alessandro Moschitti. 2021. Cross-lingual genqa: A language-agnostic generative question answering approach for opendomain question answering. CoRR, abs/2110.07150.  
+Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.
+
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad:  $100,000+$  questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392. Association for Computational Linguistics.  
+Tengtao Song, Nuo Chen, Ji Jiang, Zhihong Zhu, and Yuexian Zou. 2023. Improving retrieval-based dialogue system via syntax-informed attention. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE.  
+Adam Trischler, Tong Wang, Xingdi Yuan, Justin Harris, Alessandro Sordoni, Philip Bachman, and Kaheer Suleman. 2016. Newsqa: A machine comprehension dataset. arXiv preprint arXiv:1611.09830.  
+Bingning Wang, Ting Yao, Qi Zhang, Jingfang Xu, and Xiaochuan Wang. 2020. Reco: A large scale chinese reading comprehension dataset on opinion. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 9146-9153. AAAI Press.  
+Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners. In ICLR. OpenReview.net.  
+Jun Yin, Xin Jiang, Zhengdong Lu, Lifeng Shang, Hang Li, and Xiaoming Li. 2016. Neural generative question answering. In Proceedings of the Workshop on Human-Computer Question Answering, pages 36-42.  
+Chenyu You, Nuo Chen, Fenglin Liu, Shen Ge, Xian Wu, and Yuexian Zou. 2022. End-to-end spoken conversational question answering: Task, dataset and model. In *NAACL-HLT* (Findings), pages 1219–1232. Association for Computational Linguistics.  
+Chenyu You, Nuo Chen, and Yuexian Zou. 2021a. Knowledge distillation for improved accuracy in spoken question answering. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7793-7797. IEEE.  
+Chenyu You, Nuo Chen, and Yuexian Zou. 2021b. Self-supervised contrastive cross-modality representation learning for spoken question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 28-39, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Rowan Zellers, Yonatan Bisk, Roy Schwartz, and Yejin Choi. 2018. SWAG: A large-scale adversarial dataset for grounded commonsense inference. In EMNLP, pages 93-104. Association for Computational Linguistics.
+
+<table><tr><td>Parameter</td><td>T5</td><td>BART</td><td>Prompt</td></tr><tr><td>Batch size</td><td>8</td><td>10</td><td>10</td></tr><tr><td>Learning Rate</td><td>2e-4</td><td>2e-5</td><td>2e-5</td></tr><tr><td>Epoch</td><td>50</td><td>20</td><td>20</td></tr><tr><td>Gradient accumulation steps</td><td>1</td><td>10</td><td>10</td></tr><tr><td>Encoder Max Length</td><td>512</td><td>512</td><td>512</td></tr><tr><td>Decoder Max Length</td><td>40</td><td>128</td><td>128</td></tr></table>
+
+Table 7: Hyper-parameters setup during fine-tuning BART and T5.  
+
+<table><tr><td>Module</td><td>BLEU-1</td><td>BLEU-2</td><td>ROUGH-L</td></tr><tr><td>GPT-3 (2-shot)</td><td>40.8</td><td>35.5</td><td>46.2</td></tr><tr><td>GPT-3 (5-shot)</td><td>48.7</td><td>42.9</td><td>58.0</td></tr><tr><td>BART-Large</td><td>84.8</td><td>80.0</td><td>84.3</td></tr></table>
+
+# A Automatic Metrics
+
+Dist.: One of the most commonly-used metric that evaluate the diversity and informativeness of generated text. In this paper, we utilize Dist-1 and Dist-2 (Li et al., 2016) to validate the model performance.
+
+BLEU: Another famous metric is used to evaluate the difference between the generated text and its reference. Similarly, we use BLEU-1 and BLEU-2 (Papineni et al., 2002) in this work.
+
+ROUGH-L: An effective metric that measures the longest common subsequence (LCS) between model output and reference (Lin, 2004).
+
+Exact Match: This metric measures the percentage of model prediction that match the reference exactly. The evaluated score is 1 if the prediction is exactly the same as the reference, otherwise 0.
+
+# B Model Architecture
+
+T5 is also a strong pre-trained encoder-decoder Transformer-based model which have achieved strong results on various generative benchmarks (Chen et al., 2022c; Muller et al., 2021; Ammanabrolu et al., 2022). Likewise, we use Chinese T5<sup>5</sup> with small and base versions in our experiments.
+
+BART is a widely used denoising auto encoder for pre-training sequence-to-sequence models, which is also a standard Transformer-based PLMs. In this work, we utilize Chinese BART with base and large version.
+
+Table 8: GPT-3 performances with different few-shot examples.  
+
+<table><tr><td>Module</td><td>BLEU-1</td><td>BLEU-2</td><td>ROUGH-L</td></tr><tr><td>Answer</td><td>79.3</td><td>77.6</td><td>83.6</td></tr><tr><td>+Prompt</td><td>82.7</td><td>80.9</td><td>86.3</td></tr><tr><td>Responder</td><td>82.0</td><td>77.0</td><td>80.9</td></tr><tr><td>+Prompt</td><td>82.9</td><td>78.1</td><td>81.7</td></tr></table>
+
+Table 9: Model performances with our designed prompts. Here we choose BART-Large as our backbone model. Notice that, we use ground-truth answers rather than responses to evaluate the performances of Answer.
+
+# C Experimental Setup
+
+Table 7 shows experimental setup in our experiments.
+
+# D More Results
+
+We present T5-small, BART-small results in Table 10.
+
+# E More results
+
+# E.1 Answer to  $Q_{3}$
+
+In order to validate whether GPT-3 achieves promising results in Penguin. We train GPT-3 with few-shot examples, and construct a mini test set via randomly sampling 200 examples in Penguin's test set. Table 8 illustrates their corresponding results, we can observe there is a significant performance gap between GPT-3 and BART-Large. The results show current large language models with more 100 billion-level parameters can not address the challenge in our dataset, proving the difficulty of our Penguin.
+
+# E.2 Answer to  $Q_{4}$
+
+In this subsection, we are devoted to investigating where the majority of the improvement that designed prompts bring to the two-stage framework stems from. We show more detailed results of Answerer and Responser in Table 9. We can easily draw the conclusions that 1) a well-performed Answerer is the main prerequisite for a well-performed Responser. 2) The benefits of the designed prompts are most seen in the Answerer module, which enhance Answerer to generate more accurate answers. For instance, Answerer + prompt obtains more than 3 points in BLEU scores.
+
+# F Generated Cases of Question Types
+
+We show typical cases in Table 11 and Table 12.
+
+<table><tr><td>Framework</td><td>BackBone</td><td>Dist-1</td><td>Dist-2</td><td>BLEU-1</td><td>BLEU-2</td><td>ROUGH-L</td><td>EM</td><td>Average</td></tr><tr><td rowspan="2">End-to-End</td><td>T5-small</td><td>2.7</td><td>34.5</td><td>80.1</td><td>74.8</td><td>78.9</td><td>30.5</td><td>50.3</td></tr><tr><td>BART-base</td><td>2.7</td><td>34.4</td><td>82.1</td><td>77.2</td><td>81.0</td><td>35.6</td><td>52.1</td></tr><tr><td rowspan="2">Two-Stage</td><td>T5-small</td><td>2.7</td><td>34.0</td><td>80.2</td><td>75.0</td><td>79.0</td><td>30.8</td><td>50.3</td></tr><tr><td>BART-base</td><td>2.8</td><td>34.4</td><td>82.1</td><td>77.3</td><td>81.1</td><td>36.2</td><td>52.3</td></tr></table>
+
+Table 10: End-to-end and two-stage frameworks' performances with different backbone models. EN and CN refer to English and Chinese, separately. We highlight the best performances in the table.  
+
+<table><tr><td>Question Type</td><td>YES/NO (是否类)</td></tr><tr><td>Passage</td><td>...Three-year-old baby eat two or three at a time on the line, the child&#x27;s gastrointestinal function is relatively weak, eating more will cause gastrointestinal discomfort... ...三周岁宝宝每次吃两三个就行了,孩子的胃肠道功能比较弱,吃多了会引起肠胃不适...</td></tr><tr><td>Question</td><td>Is it good for three-year-old baby to litchi? 三周岁宝宝吃荔枝好吗?</td></tr><tr><td>Answer</td><td>Yes. 好</td></tr><tr><td>Expected Response</td><td>Three-year-old babies can eat two or three litchis, but not too much. 三周岁的宝宝可以吃两三个荔枝,但是不能吃太多。</td></tr><tr><td></td><td>Generated Response-One-Stage</td></tr><tr><td>T5</td><td>They can not eat litchi. 他们不可以吃荔枝</td></tr><tr><td>BART</td><td>They can eat litchi. 他们可以吃荔枝</td></tr><tr><td>Prompt-BART</td><td>They can eat litchi. 他们可以吃荔枝</td></tr><tr><td></td><td>Generated Response-Two-Stage</td></tr><tr><td>T5</td><td>Three-year-old baby can not eat litchi. 三周岁的宝宝不可以吃荔枝</td></tr><tr><td>BART</td><td>They can eat litchi. 他们可以吃荔枝</td></tr><tr><td>Prompt-BART</td><td>Three-year-old baby can eat litchi. 三周岁的宝宝可以吃荔枝</td></tr><tr><td></td><td>(a) A Generated Response example of Yes/No questions.</td></tr><tr><td>Question Type</td><td>Opinion (观点类)</td></tr><tr><td>Passage</td><td>...It&#x27;s not a matter of how fast you move, it&#x27;s a matter of understanding. I have been learning the piano for 10 years and I have to practice these czernies and so on over and over again this year. It is recommended to learn Bayer first, which greatly improves the basic skills.... ...进度快不快不是问题,主要看你的理解能力。我学了10年琴,车尔尼这些等等今年都要反复的练习。推荐还是先学拜厄,对基本功有很大的提高。...</td></tr><tr><td>Question</td><td>I&#x27;ve already learned Faber, so do I still need to practice Baier? 我已经学习了菲伯尔,那还需要练习拜厄吗?</td></tr><tr><td>Answer</td><td>Of course. 需要</td></tr><tr><td>Expected Response</td><td>You&#x27;d better practice Baier. It can improve your basic skills. 推荐还是练习拜厄,这对你的基本功有好处。</td></tr><tr><td></td><td>Generated Response-One-Stage</td></tr><tr><td>T5</td><td>You need to practice Baier. 还是要练习拜厄</td></tr><tr><td>BART</td><td>You still need to practice Baier. 仍然要练习拜厄</td></tr><tr><td>Prompt-BART</td><td>You still need to practice Baier. 仍然要练习拜厄</td></tr><tr><td></td><td>Generated Response-Two-Stage</td></tr><tr><td>T5</td><td>You still need to practice Baier. 仍然要练习拜厄</td></tr><tr><td>BART</td><td>You still need to practice Baier. 仍然要练习拜厄</td></tr><tr><td>Prompt-BART</td><td>He still needs to practice Baier. 他仍然要练习拜厄</td></tr><tr><td></td><td>(b) A Generated Response example of Opinion questions.</td></tr></table>
+
+Table 11: Typical generated examples from each model in Penguin. English translation is also given for comparison.
+
+<table><tr><td>Question Type</td><td>Fact (事实类)</td></tr><tr><td>Passage</td><td>...Zhang Lei is a Chinese football player and goalkeeper. Zhang Lei was selected to Guang-dong Hongyuan Youth Team at the age of 13, then played as the main goalkeeper for Dongguan Southern City before moving to Shenzhen in 2006. He was selected to the 2008 National Olympic Team and the 2005 World Youth Championships squad. In 2009, Beijing Guoan exchanged Cheng Yuelei, a backup goalkeeper, plus 400,000 yuan to Shenzhen in exchange for Zhang Lei, which was also the last domestic import transaction of Beijing Guoan in the first transfer market of the 2009 season. Zhang Lei did not get a chance to play because Beijing Guoan had Yang Zhi as their keeper. He was listed by Beijing Guoan and eventually moved to Changsha Jinde in 2010 after losing their main goalkeeper Song Zhenyu. In 2011, Zhang moved to Chongqing Lifan and became the team&#x27;s main goalkeeper. Served for five years, the first three years to play the main force, the last two years gradually as a substitute. In 2016, he was transferred to Hangzhou Greentown. Zhang Lei joined the Lijiang Flying Tigers on loan from Hangzhou Greentown in 2017....张磊，中国足球运动员，司职守门员。张磊13岁时入选广东宏远少年队，随后担任东莞南城的主力门将，2006年转会到深圳队。他入选了08之星国奥队，并进入2005年世青赛的大名单。2009年北京国安用替补门将程月磊加上40万元，与深圳队交换张磊，这也是2009赛季中超第一次转会市场上北京国安队最后一笔内援引进交易。因为北京国安有杨智把关，所以张磊没有得到出场机会，结果被北京国安挂牌，最终在2010年转会到失去主力门将宋振瑜的长沙金德，成为球队新的主力门将。2011年，张磊转会到重庆力帆成为球队的主力门将。效力了五年期间，前三年打主力，后两年渐为替补。2016年转会到杭州绿城。2017年，张磊从杭州绿城租借到丽江飞虎。...</td></tr><tr><td>Question</td><td>In what year did Zhang Lei transfer to Shenzhen?张磊哪一年转会去的深圳?</td></tr><tr><td>Answer</td><td>2006. 2006</td></tr><tr><td>Expected Response</td><td>He transfers to Shenzhen in 2006. 他于2006年转会到深圳。</td></tr><tr><td colspan="2">Generated Response-One-Stage</td></tr><tr><td>T5BARTPrompt-BART</td><td>He transferred to Shenzhen in two years. 他于2年转会到深圳队He transferred to Shenzhen in two years. 他于2年转会到深圳队He transfers to Shenzhen in 2006. 他于2006年转会到深圳</td></tr><tr><td colspan="2">Generated Response-Two-Stage</td></tr><tr><td>T5BARTPrompt-BART</td><td>He transfers to Shenzhen in 2006. 他于2006年转会到深圳He transfers to Shenzhen in 2006. 他于2006年转会到深圳Zhang Lei transfers to Shenzhen in 2006. 张磊于2006年转会到深圳</td></tr></table>
+
+Table 12: A Generated Response example of Fact questions in Penguin. English translation is also given for comparison.
+
+# G All Prompts
+
+We design five prompts for each module in end-to-end and two-stage frameworks, which are presented in Table 13.
+
+<table><tr><td>Framework</td><td>Prompts</td></tr><tr><td>One-Stage</td><td>1. Please generate a reasonable response based on the following question and passage. Question: Q.Passage: P.
+2. The question is Q, give the reasonable response based on the following paragraph: P.
+3. This is a reading comprehension problem, please generate the reasonable response based on the following question and passage. Question is Q. Passage is P.
+4. Please answer the question according to descriptions of the question and paragraph. Question is Q.Passage is P.
+5. The question is Q, Conduct question understanding and response generation, knowledge question and response task. Relevant paragraphs are as follows: P.</td></tr><tr><td>Answerer</td><td>1. Please generate the reasonable answer based on the following question and passage. Question: Q.Passage: P.
+2. The question is Q, give the reasonable answer based on the following paragraph: P.
+3. This is a reading comprehension problem, please generate the reasonable answer based on the following question and passage. Question is Q. Passage is P.
+4. Please answer the question according to descriptions of the question and paragraph. Question is Q.Passage is P.
+5. The question is Q, Conduct question understanding and answer generation, knowledge question and answer task. Relevant paragraphs are as follows: P.</td></tr><tr><td>Responder</td><td>1. Please generate the reasonable response based on the following question and answer. Question: Q.Answer: A.
+2. The question is Q, give the reasonable response based on the following answer: A.
+3. This is a answer rewriting problem, please generate the reasonable response based on the following question and answer. Question is Q. Answer is A.
+4. Please give the response according to descriptions of the question and answer. Question is Q. Answer is A.
+5. The question is Q, Conduct question understanding and response generation, answer rewriting task.Relevant answer is as follows: A.</td></tr></table>
+
+Table 13: All used prompts during training. English translation is only given for reading.

2023/Natural Response Generation for Chinese Reading Comprehension/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3dc75066e0af0603118e08c7d4ac8e5b3921c387c02efbffd6ccac1cb9643730
+size 1340829

2023/Neuro-Symbolic Sentiment Analysis with Dynamic Word Sense Disambiguation/f2ea77f7-c2a2-49e4-b922-17fa39103dc4_content_list.json

@@ -0,0 +1,1580 @@
+[
+    {
+        "type": "text",
+        "text": "Neurosymbolic Sentiment Analysis with Dynamic Word Sense Disambiguation",
+        "text_level": 1,
+        "bbox": [
+            275,
+            90,
+            722,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xulang Zhang\\*, Rui Mao\\*, Kai He\\* and Erik Cambria\\*",
+        "bbox": [
+            248,
+            147,
+            749,
+            165
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$\\spadesuit$ Nanyang Technological University, Singapore",
+        "bbox": [
+            309,
+            166,
+            692,
+            183
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "\\*National University of Singapore, Singapore",
+        "bbox": [
+            314,
+            183,
+            687,
+            200
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "xulang001@e.ntu.edu.sg; {rui.mao, cambria}@ntu.edu.sg; kai_he@nus.edu.sg",
+        "bbox": [
+            141,
+            200,
+            858,
+            216
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sentiment analysis is a task that highly depends on the understanding of word senses. Traditional neural network models are black boxes that represent word senses as vectors that are uninterpretable for humans. On the other hand, the application of Word Sense Disambiguation (WSD) systems in downstream tasks poses challenges regarding i) which words need to be disambiguated, and ii) how to model explicit word senses into easily understandable terms for a downstream model. This work proposes a neurosymbolic framework that incorporates WSD by identifying and paraphrasing ambiguous words to improve the accuracy of sentiment predictions. The framework allows us to understand which words are paraphrased into which semantically unequivocal words, thus enabling a downstream task model to gain both accuracy and interpretability. To better fine-tune a lexical substitution model for WSD on a downstream task without ground-truth word sense labels, we leverage dynamic rewarding to jointly train sentiment analysis and lexical substitution models. Our framework proves to effectively improve the performance of sentiment analysis on corpora from different domains.",
+        "bbox": [
+            144,
+            282,
+            460,
+            652
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            665,
+            258,
+            682
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To improve the accuracy of sentiment analysis, it is crucial to disambiguate polysemous words according to the context (Cambria et al., 2023). A word may have senses that carry opposite sentiments, e.g., \"fine\" has the positive meaning of being delicate and refined, and the negative meaning of monetary penalty. Additionally, words may appear in positive, negative, or neutral senses in different domains. For instance, the word \"bearish\" is often used in a negative sense to describe the stock market in the financial domain, while more likely to be used neutrally to describe hulking figures in other domains.",
+        "bbox": [
+            110,
+            694,
+            489,
+            903
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In traditional black-box sentiment analysis, the representations of word meanings are generated by neural networks (Mao and Li, 2021). As such, the disambiguation process is implicit and opaque. It is challenging to investigate what implicit senses are assigned to ambiguous words by a neural network model for it to make accurate predictions. Coupling symbolism and sub-symbolism, a neurosymbolic approach to sentiment analysis is transparent and interpretable, regarding the extraction and distinction of word meaning as a sub-problem (Cambria et al., 2017). However, research that explicitly deals with ambiguous words in sentiment analysis is limited (Xia et al., 2014, 2015; Cambria et al., 2015). Sense-annotated resources such as SentiWordNet (Baccianella et al., 2010) have been utilized, but only to detect and score sentimental words in the input (Hung and Lin, 2013; Nassirtoussi et al., 2015; Hung and Chen, 2016). The disambiguation of polysemous words in sentiment analysis inputs remains under-explored. Thus, we are motivated to develop a neurosymbolic framework for sentiment analysis with explicit Word Sense Disambiguation (WSD).",
+        "bbox": [
+            507,
+            253,
+            884,
+            638
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Traditional WSD aims to classify which sense of the target word is appropriate according to its context (Huang et al., 2019; Blevins and Zettlemoyer, 2020; Bevilacqua and Navigli, 2020; Song et al., 2021; Barba et al., 2021a,b). However, the limitations  $(\\ell)$  of their task setups are that  $(\\ell 1)$  the output of those WSD systems, whether sense label (Bevilacqua and Navigli, 2020) or gloss definition (Blevins and Zettlemoyer, 2020; Barba et al., 2021a), is difficult to integrate into downstream input;  $(\\ell 2)$  the target words must be pre-defined for WSD systems to disambiguate, making it difficult to apply them directly to downstream tasks; and  $(\\ell 3)$  the reliability of sense distinction of the benchmark has been questioned (Ramsey, 2017; Maru et al., 2022), as word senses are not discrete.",
+        "bbox": [
+            507,
+            642,
+            884,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8772",
+        "bbox": [
+            478,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8772-8783  \nDecember 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            779,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Attempts have been made to incorporate the disambiguation of word senses into downstream applications. Nonetheless, they are either impractical in real-world settings (Farooq et al., 2015; Pu et al., 2018), or only applicable to a specific downstream task, i.e., machine translation (Hangya et al., 2021; Campolungo et al., 2022). To address these issues, we leverage dynamic WSD in our neurosymbolic framework. We regard WSD as a lexical substitution task (for  $\\ell 1$ ), where ambiguous words in sentiment analysis input are paraphrased into ones that cause less sentiment confusion using a lexical substitution model with dynamic WSD.",
+        "bbox": [
+            112,
+            84,
+            492,
+            294
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our dynamic WSD method comprises two distinct aspects: the dynamic selection of the target word for disambiguation, based on downstream contextual information (for  $\\ell 2$ ); and the dynamic learning of the most suitable lexical substitutions, based on the prediction accuracy of the downstream model (for  $\\ell 3$ ). To identify target words that need to be substituted, we first adopt an attention-based explainable encoder to rank input words by the level of influence they have on the output, motivated by the assumption that not all polysemous words affect sentiment predictions (Cao et al., 2015). Then, the top-ranking words with the most diverse senses are selected to be targets, based on our hypothesis that an ambiguous word may coincide with significant disparities in the semantic space between the word itself and its potential substitution candidates.",
+        "bbox": [
+            115,
+            300,
+            490,
+            575
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As such, we can trace back which words are selected as targets, and which substitute words are used to disambiguate them, making our method interpretable and transparent. Furthermore, we fine-tune the lexical substitution model whilst training the sentiment analysis model by adopting a dynamic rewarding mechanism that does not require gold standard substitution labels. As such, the substitution model can be adjusted to provide more appropriate output for the downstream task as well as the target domain. Our model achieves more accurate sentiment analysis with explicit lexical-substitution-based WSD, despite not specifically trained on WSD and sentiment parallel datasets. Since the disambiguated word senses are represented as the lexical substitutions of the original words and an explainable encoder is used for retrieving the most contributing words, the neurosymbolic sentiment inference process is more interpretable for humans compared to black-box models.",
+        "bbox": [
+            112,
+            581,
+            490,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our model is examined with three publicly available datasets from movie review, finance, and Twitter domains, outperforming strong baselines by  $1.23\\%$  on average. The contributions of our work are twofold: (1) We propose a neurosymbolic sentiment analysis framework that leverages WSD as an auxiliary lexical substitution task, achieving state-of-the-art performance on our examined sentiment analysis datasets from different domains; (2) We design a transparent mechanism to select WSD target words, based on a) the importance of words to the learning and inference of a downstream task, and b) word sense diversity.",
+        "bbox": [
+            507,
+            84,
+            885,
+            294
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            309,
+            665,
+            325
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Word Sense Disambiguation is a task that assigns the appropriate meaning to an ambiguous word according to its context. Recent WSD systems attained F1 scores higher than  $80\\%$  on the benchmark datasets (Song et al., 2021; Bevilacqua and Navigli, 2020; Barba et al., 2021a,b), which is regarded as the estimated human performance since it is the highest inter-annotator agreement (Edmonds and Kilgarriff, 2002; Palmer et al., 2007). However, research on the application of WSD systems on downstream tasks is limited. Pu et al. (2018) utilized clustering algorithms to obtain sense embeddings by inferring from the similarity of the context vectors of a target word. The heavy requirement for disambiguating every homographs prior to downstream applications makes their method less feasible in many real-world situations. Campolungo et al. (2022) achieved the disambiguation of word sense in neural machine translation by leveraging a multilingual WSD system to establish a sense-tagged parallel corpus. Nonetheless, their approach relied on cross-lingual word alignments for reliable annotations, meaning that it cannot be adopted by other types of downstream tasks.",
+        "bbox": [
+            507,
+            338,
+            885,
+            724
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Sentiment Analysis can be generally defined as the task that determines if a piece of text expresses positive, negative, or neutral opinion (Mao et al., 2023c). Word sense information has been applied in sentiment analysis in the form of sentiment-annotated lexicons (Devitt and Ahmad, 2007; Ohana and Tierney, 2009; Hung and Lin, 2013; Nassirtoussi et al., 2015), such as SentiWordNet. However, these methods incorporate the sentiment information of a word without disambiguating its word sense according to the context, but resorting to the emotional score of the most frequent",
+        "bbox": [
+            507,
+            726,
+            885,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8773",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/8339731eaaa1ea3228dd3efe4658383c01c8f2bb2183ac4ccba471934e643f1c.jpg",
+        "image_caption": [
+            "Figure 1: Architecture of the proposed disambiguation framework for sentiment analysis. A dotted line means backpropagation. Green block means that the original word is replaced by a word provided by the lexical substitution model. The darker the green, the higher probability it is assigned to be the model. Orange block indicates the words with top ranking attention scores produced by HAN."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            80,
+            885,
+            325
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "sense, or the average score of all senses. Additionally, the main objective of this approach is to identify sentimental words in a sentence, instead of disambiguating polysemous words that might affect model predictions. Hung and Chen (2016); Farooq et al. (2015); Liu et al. (2018) attempted to disambiguate according to target domain and context, but either costed extensive human labour, or fail to incorporate word senses explicitly and transparently.",
+        "bbox": [
+            112,
+            420,
+            487,
+            580
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To address these limitations, our neurosymbolic framework aims to identify ambiguous target words in the context and replace them with clarifying ones automatically, making them interpretable. We also leverage a dynamic rewarding mechanism so that the substitutions can be fine-tuned to better serve the downstream task and the text domain.",
+        "bbox": [
+            112,
+            582,
+            487,
+            694
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Task Definition",
+        "text_level": 1,
+        "bbox": [
+            112,
+            712,
+            280,
+            728
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this work, we regard word sense ambiguity as the potential cause of errors in a downstream task, e.g., sentiment analysis. Inspired by the work of Mao et al. (2023b), we reframe WSD as a lexical substitution task, aiming to replace an ambiguous word with a synonym that not only represents the true sense of the original word in the context (Arthur et al., 2015), but also reduces misclassifications in the downstream task. Given an input sentence  $w = (w_{1},w_{2},\\dots ,w_{L})$ , we first aim to identify the target word  $t$  that is likely troublesome for",
+        "bbox": [
+            112,
+            741,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the downstream prediction from  $w$ . To resolve the potential ambiguity, we extract  $M$  of its synonyms from WordNet (Fellbaum, 1998) as candidates  $c = (c_{1},\\ldots ,c_{M})$ , forming the input  $(s)$  as",
+        "bbox": [
+            507,
+            420,
+            882,
+            483
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ns = \\langle s \\rangle , w _ {1}, w _ {2}, \\dots , t, \\dots , w _ {L}, \\langle / s \\rangle , \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            574,
+            493,
+            882,
+            514
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "$< \\mathrm{s}>$  and  $</ \\mathrm{s}>$  are special tokens that were defined by our employed pre-trained language model. The lexical substitution model should identify the best candidate from  $c$  as the substitute  $\\hat{c}$  which retains the original meaning of  $t$  in the context by using contrast learning. Thus,  $\\hat{c}$  is the symbolic representation of  $t$  in context  $w$ . Then, the paraphrased input  $w^{f} = (w_{1},\\dots,\\hat{c},\\dots,w_{L})$  is fed into a neural network classifier to predict a sentiment label. The objective is that the paraphrased input  $w^{f}$  increases the probability of correct sentiment prediction.",
+        "bbox": [
+            507,
+            529,
+            882,
+            706
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Methodology",
+        "text_level": 1,
+        "bbox": [
+            507,
+            717,
+            658,
+            734
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our proposed neurosymbolic framework can be viewed in Figure 1. There are two models in our framework, namely sentiment analysis and lexical substitution models. The sentiment analysis model consists of an encoder, an interpretable attention module, and multiple feedforward neural networks. First, through Figure 1 (1) and (2), the original input is fed into the encoder and the interpretable attention module to obtain the top  $I$  tokens with the highest attention weights, which contribute the most to sentiment prediction.",
+        "bbox": [
+            507,
+            741,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8774",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In (3), the sense diversity of each token is computed as the average distance of the token's hidden state to those of its substitution candidates. Then, through (4), top  $J$  tokens with the most diverse senses are selected as disambiguation targets, because these target tokens may have very different senses in different contexts. Subsequently, the pretrained lexical substitution model is applied to provide the best substitution to replace each target token, as shown in (5) and (6)a. The new input sentence is passed onto the sentiment analysis model for final prediction ((7)a and (8)a), which is used for the sentiment model backpropagation ((9)a).",
+        "bbox": [
+            112,
+            84,
+            489,
+            294
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to fine-tune the lexical substitution model to provide better substitutions that improve the accuracy of sentiment analysis, we adapt a dynamic rewarding mechanism (Ge et al., 2022) that was proposed for multi-task learning. As shown in Figure 1 (5) and (6)b, for each target word, we choose the top  $N$  candidates. Each of them is seen as a substitute candidate to calculate the probability of correct sentiment prediction after the replacement ((7)b and (8)b). Then, each probability is used to dynamically compute the loss weight when the corresponding candidate is learned by the lexical substitution model as a ground truth ((9)b and (10)b). If the replacement by a candidate leads to a higher chance of predicting the right sentiment, the candidate is more likely given a higher rank by the lexical substitution model. Otherwise, the model would be less likely to select the candidate as the best replacement. The important notations used in this paper are summarized in Table 1. The detailed training process can be seen in Algorithm 1.",
+        "bbox": [
+            115,
+            294,
+            489,
+            632
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Sentiment Analysis with Interpretability",
+        "text_level": 1,
+        "bbox": [
+            112,
+            642,
+            477,
+            657
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given input sentence  $w = (w_{1},\\dots ,w_{L})$ , the goal is to predict the correct sentiment label  $\\tilde{y}$ . The input is first fed into a pre-trained encoder:",
+        "bbox": [
+            112,
+            662,
+            487,
+            709
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nV = \\operatorname {E n c o d e r} (w), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            226,
+            721,
+            487,
+            738
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $V$  is hidden states. Next, we aim to find the tokens that contribute the most to sentiment inference. We adopt an interpretable attention module called Hierarchical Attention Network (HAN, Han et al., 2022), which effectively encodes hidden states with multiple non-linear projections and ranks the most influential tokens based on attention. Following the setup of Han et al. (2022), we stack two blocks of HAN to form our attention module.",
+        "bbox": [
+            112,
+            747,
+            489,
+            891
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nq, a = H A N (H A N (V)), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            200,
+            902,
+            487,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/3f931976f478537e282899b10e81e0924de510f5086401e4ef0f825eba60b565.jpg",
+        "table_caption": [
+            "Table 1: Notation table."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Description</td></tr><tr><td>L</td><td>Input sequence length</td></tr><tr><td>I</td><td>Number of words with top ranking attention scores to be potential target words</td></tr><tr><td>J</td><td>Number of words with top ranking diversity scores to be potential target words</td></tr><tr><td>N</td><td>Number of candidates used for finetuning lexical substitution via dynamic rewarding</td></tr><tr><td>K</td><td>Number of candidates for training lexical substitution model</td></tr><tr><td>w</td><td>The original input sentence</td></tr><tr><td>V</td><td>The hidden states of w</td></tr><tr><td>y</td><td>The ground-truth sentiment label of w</td></tr><tr><td>wf</td><td>The final input sentence where ambiguous words in w are substituted</td></tr><tr><td>ŷf</td><td>The predicted sentiment label with wf as input</td></tr><tr><td>wl</td><td>The l-th word in w</td></tr><tr><td>al</td><td>The attention weight of the l-th word in w</td></tr><tr><td>wiatt</td><td>The word in w with the i-th highest attention weight</td></tr><tr><td>viatt</td><td>The hidden state of wiatt</td></tr><tr><td>M</td><td>The number of synonyms of wiatt with the same part-of-speech that are provided by the WordNet</td></tr><tr><td>G</td><td>The hidden states of the M WordNet synonyms</td></tr><tr><td>gm</td><td>The hidden state of the m-th WordNet synonyms of wiatt</td></tr><tr><td>di</td><td>The average Euclidean distance between viatt and g1, . . . , gm, . . . , gM</td></tr><tr><td>tj</td><td>The j-th target word in w to be substituted</td></tr><tr><td>c</td><td>A set of candidate words from WordNet.</td></tr><tr><td>ˆc</td><td>The top N candidates ranked by the lexical substitution model</td></tr><tr><td>ˆcj,n</td><td>The top n-th candidate for replacing tj</td></tr><tr><td>wj,n</td><td>The resulting sentence when replacing tj in w withˆcj,n</td></tr><tr><td>ˆθj,n</td><td>The loss weight whenˆcj,n is treated as the gold standard substitution</td></tr><tr><td>s</td><td>The formatted input to lexical substitution model</td></tr><tr><td>U</td><td>The hidden state matrix of the input sentence in s</td></tr><tr><td>u1</td><td>The hidden state of the l-th input word in s</td></tr><tr><td>R</td><td>The hidden state matrix of the candidates in s</td></tr><tr><td>rk</td><td>The hidden state of the k-th candidate in s</td></tr><tr><td>τ</td><td>The temperature hyper-parameter for training the lexical substitution model</td></tr><tr><td>β</td><td>The hyper-parameter for balancing the sentiment analysis loss and fine-tuning loss</td></tr><tr><td>L(sa)</td><td>The sentiment analysis loss</td></tr><tr><td>L(tune)</td><td>The loss for finetuning the pretrained language model into a lexical substitution model</td></tr><tr><td>L(ls)</td><td>The lexical substitution loss whenˆcj,n is treated as the gold standard substitution</td></tr><tr><td>L(ls)</td><td>The final weighted lexical substitution loss for backpropagation</td></tr></table>",
+        "bbox": [
+            510,
+            107,
+            878,
+            743
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where vector  $q$  is the yielded hidden state.  $a$  is the attention weights, indicating the contribution of each token to the final sentiment prediction. To obtain the sentiment prediction,  $q$  is passed on to two layers of feed-forward neural networks (FNN) to obtain the probability distribution of sentiment prediction, with the first one being activated by ReLU (Agarap, 2018), and the second by softmax.",
+        "bbox": [
+            507,
+            790,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8775",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh = R e L U \\left(F N N _ {1} (q)\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            112,
+            487,
+            129
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {y} = \\operatorname {s o f t m a x} \\left(F N N _ {2} (h)\\right) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            200,
+            131,
+            487,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We denote the prediction of the sentiment analysis model as  $\\hat{y}^f$  when the input is  $w^f$ , which denotes a paraphrased  $w$  where all selected target tokens are replaced by sense-clarifying substitutions. Thus, the sentiment analysis loss is computed as:",
+        "bbox": [
+            112,
+            162,
+            487,
+            242
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} ^ {(s a)} = \\operatorname {C r o s s E n t r o p y} (\\hat {y} ^ {f}, \\tilde {y}) \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            181,
+            253,
+            487,
+            272
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Disambiguation by Lexical Substitution",
+        "text_level": 1,
+        "bbox": [
+            112,
+            294,
+            470,
+            311
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To obtain a lexical substitution model, we fine-tune the pre-trained language model ALM (Mao et al., 2023a)<sup>1</sup> with benchmark datasets from McCarthy and Navigli (2007); Kremer et al. (2014). We use the candidates from the datasets to formulate our input  $s$  as in Formula 1. The candidate with the highest score is set as the ground truth substitution  $\\tilde{c}$ . Given the input  $s$ , the model encodes it as:",
+        "bbox": [
+            112,
+            315,
+            487,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nU, R = A L M (s), \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            231,
+            458,
+            487,
+            474
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $U = [u_{1},\\ldots ,u_{L}]$  is the hidden states of the input sentence, and  $R = [r_1,\\dots,r_K]$  is the hidden states of the candidates. We denote the representation of the target word as  $u_{t}$ $(t\\in \\{1,\\ldots ,L\\})$",
+        "bbox": [
+            112,
+            486,
+            487,
+            551
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our training objective is to a) close the distance between the representation of the ground truth candidate  $r_k$ ,  $(k \\in \\{1, \\dots, K\\})$  with  $u_t$ , and b) push away incorrect candidates representations  $r_j$ $(j \\in \\{1, \\dots, K | j \\neq k\\})$  from  $u_t$ . Namely,  $(r_k, u_t)$  will be regarded as a positive pair, while  $(r_j, u_t)$  will be regarded as a negative pair. We follow the InfoNCE loss (Oord et al., 2018) to achieve these goals, which can be formulated as:",
+        "bbox": [
+            112,
+            552,
+            487,
+            695
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} ^ {(t u n e)} = - \\sum_ {i} \\log \\frac {e x p (d (u _ {t} , r _ {k}) / \\tau)}{\\sum_ {j} e x p (d (u _ {t} , r _ {i}) / \\tau)}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            131,
+            706,
+            487,
+            744
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $i\\in \\{1,\\dots ,K\\}$ $\\tau$  is a temperature hyperparameter, and  $d(\\cdot)$  is Euclidean distance.",
+        "bbox": [
+            112,
+            756,
+            487,
+            789
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1: Sentiment analysis with word sense disambiguation using a dynamic rewarding mechanism.",
+        "bbox": [
+            515,
+            87,
+            838,
+            134
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [],
+        "code_body": "1 Initialize sentiment analysis model as  $\\Phi$  , pre-trained   \n2 lexical substitution model as  $\\Psi$  .   \n3 Initialize hyperparameters  $\\beta$ $I,J,N$    \nwhile not done do   \n4 Sample a sentence  $w = w_{1},w_{2},\\ldots ,w_{L}$  .   \nfor  $l = 1:L$  do   \n6 Compute the attention weight  $a_{l}$  of token   \n $w_{l}$  .   \n7 end   \n8  $w^{att}\\gets$  Top I of  $w$  ordered by attention weights   \n $a = a_1,a_2,\\dots ,a_L$  .   \nfor  $i = 1:I$  do   \n10  $c\\gets$  all possible candidates with the same   \npart-of-speech type as  $w_{i}^{att}$  , provided by   \nWordNet;   \n11 Compute the average Eulidean distance  $d_{i}$    \nbetween the hidden states of  $w_{i}^{att}$  and  $c$  produced by the encoder in  $\\Phi$  .   \nend   \n13  $t = (t_1,t_2,\\dots ,t_J)\\gets$  Top J of  $w^{att}$  ordered   \nby  $d = d_{1},d_{2},\\ldots ,d_{I}$ $w^{f}\\gets w$  .   \nfor  $j = 1:J$  do   \n16 Input w into  $\\Psi$  to produce top N candidates   \n $\\hat{c}_j = (\\hat{c}_{j,1},\\dots ,\\hat{c}_{j,N})$  for target word  $t_j$  ordered by probability;   \n17 Replace  $t_j$  in  $w^f$  with  $\\hat{c}_{j,1}$    \ns  $\\leftarrow (< s > ,w_{1},\\dots ,t_{i},\\dots ,w_{L}, <   / s > ,\\right.$ $\\hat{c}_{j,1}, <   / s > ,\\dots ,\\hat{c}_{j,N}, <   / s > )$  .   \nfor  $n = 1:N$  do   \n21  $w^{j,n}\\gets$  Replace  $t_j$  in  $w$  with  $\\hat{c}_{j,n}$  .   \n22 Input  $w^{j,n}$  into  $\\Phi$  to obtain the probability of correct sentiment prediction  $P(\\hat{y} = \\tilde{y})_{j,n}$  ..   \n23  $\\theta_{j,n}\\gets \\beta P(\\hat{y} = \\tilde{y})_{j,n}^2$  .   \n24 Compute  $\\mathcal{L}_{j,n}^{(ls)}$  by feeding s with  $\\hat{c}_n$  labeled as true substitute into  $\\Psi$  ..   \n25  $\\mathcal{L}_{j,n}^{(ls)}\\gets \\theta_{j,n}\\mathcal{L}_{j,n}^{(ls)};$    \nend   \nend   \n28  $\\mathcal{L}^{(ls)}\\gets \\mathcal{L}_{1,1}^{(ls)} + \\dots +\\mathcal{L}_{1,N}^{(ls)} + \\mathcal{L}_{J,1}^{(ls)} + \\dots +\\mathcal{L}_{J,N}^{(ls)}$  .   \n29 Compute sentiment analysis loss  $\\mathcal{L}^{(sa)}$  using  $w^{f}$  as input;   \n30  $L\\gets \\mathcal{L}^{(sa)} + \\mathcal{L}^{(ls)}$  .",
+        "bbox": [
+            510,
+            137,
+            852,
+            728
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During the inference stage, we choose the candidate  $\\hat{c}$  whose corresponding hidden state  $r_k$  is the most similar to  $u_t$ , measured by Euclidean distance:",
+        "bbox": [
+            507,
+            760,
+            884,
+            822
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {c} = \\arg \\min  \\left(d \\left(u _ {t}, r _ {k}\\right)\\right) \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            606,
+            827,
+            882,
+            844
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We then use the resulting lexical substitution model to find replacements for the selected target words in sentiment analysis input. To determine which words are selected as disambiguation tar",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>ALM employed a novel pre-training paradigm, termed Anomalous Language Modeling. ALM was pre-trained to detect anomalous substituted words from a sequence and retrieve the original words from a set of candidates that contains a positive sample (an original word) and multiple hard negative samples (the synonyms of the original word) via contrastive learning. Thus, it is more suitable for fine-tuning a lexical substitution model in our task, because we also aim to detect appropriate substitution words from a set of candidates.",
+        "bbox": [
+            112,
+            809,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8776",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "gets, we select the top  $I$  words in the input with the highest  $a$  produced by Equation 3, forming the set  $w^{att}$ . We denote their corresponding representations as  $V^{att} = \\{v_{1}^{att},\\dots ,v_{I}^{att}\\}$ . For each  $v_{i}^{att}$ , we compute its average Euclidean distance to all of its candidates' hidden states. The candidates consist of WordNet synonyms of all of its possible senses under the same part-of-speech type, which are transformed into hidden states  $G = \\{g_{1},\\ldots ,g_{M}\\}$  using Equation 2.  $M$  represents the number of candidates provided by WordNet. The top  $J$  words in  $w^{att}$  with the largest corresponding average distance are considered to be the ones with the most diverse word meanings, and thus are more likely to require disambiguation. We will demonstrate the success of such an ambiguous word selection method in the later ablation study (see Section 6.1).",
+        "bbox": [
+            112,
+            84,
+            489,
+            357
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Hence, target words  $t = (t_1, \\ldots, t_J)$  are selected by finding each corresponding  $v_j^{att}$ :",
+        "bbox": [
+            112,
+            357,
+            489,
+            393
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nv _ {j} ^ {\\text {a t t}} = \\arg \\max  _ {i} \\left(\\frac {1}{M} \\sum_ {M} d \\left(v _ {i} ^ {\\text {a t t}}, g _ {m}\\right)\\right). \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            144,
+            400,
+            487,
+            436
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Following the lexical substitution benchmarks, only words whose part-of-speech tags fall under the general categories of noun, verb, adjective, and adverb are considered as targets.",
+        "bbox": [
+            112,
+            445,
+            487,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Dynamic Rewarding Mechanism",
+        "text_level": 1,
+        "bbox": [
+            112,
+            519,
+            418,
+            535
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To fine-tune the lexical substitution model on the downstream sentiment analysis task without knowing WSD gold standard truth, for each target word  $t_j$ , we utilize the top  $N$  candidates  $\\hat{c}_j = \\{\\hat{c}_{j,1},\\dots ,\\hat{c}_{j,n},\\dots ,\\hat{c}_{j,N}\\}$  produced by the lexical substitution model. The new input resulting from  $t_j$  being replaced by candidate  $\\hat{c}_{j,n}$  is denoted as  $w^{j,n}$ . Same with Equation 5, the probability distribution of a sentiment prediction from  $w^{j,n}$  is:",
+        "bbox": [
+            112,
+            539,
+            487,
+            684
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP (\\hat {y}) _ {j, n} = \\operatorname {s o f t m a x} \\left(F N N _ {2} \\left(h ^ {j, n}\\right)\\right), \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            147,
+            694,
+            487,
+            713
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $h^{j,n}$  is the hidden states of  $w^{j,n}$  outputted by Equation 4 in the sentiment analysis model.",
+        "bbox": [
+            112,
+            721,
+            485,
+            753
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To adjust the model in such a way that a more accurate sentiment prediction results in a higher reward for the corresponding substitution output, we compute the loss weight for  $\\hat{c}_{j,n}$  being the correct substitution prediction as:",
+        "bbox": [
+            112,
+            753,
+            487,
+            834
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {j, n} = \\beta P (\\hat {y} = \\tilde {y}) _ {j, n} ^ {2}, \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            216,
+            843,
+            487,
+            863
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\beta$  is a hyperparameter for balancing the sentiment analysis and lexical substitution losses,  $\\tilde{y}$  is the ground truth sentiment label as defined above.",
+        "bbox": [
+            112,
+            871,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/fb98672a0045cdbafab2133bb7d89264d70bee2bb4f6dbf0cb714526a1f024da.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Split</td><td># Samples</td></tr><tr><td>SemEval 2007 Task 10</td><td>-</td><td>2,010</td></tr><tr><td>CoinCo</td><td>-</td><td>15,629</td></tr><tr><td></td><td>Train</td><td>8,530</td></tr><tr><td>Rotten Tomatoes</td><td>Dev</td><td>1,066</td></tr><tr><td></td><td>Test</td><td>1,066</td></tr><tr><td>Financial PhraseBank</td><td>-</td><td>4,846</td></tr><tr><td>STS-Gold</td><td>-</td><td>2,026</td></tr></table>",
+        "bbox": [
+            544,
+            80,
+            847,
+            181
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Datasets used for experiments.",
+        "bbox": [
+            559,
+            191,
+            830,
+            205
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We formulate the input  $s^j$  to lexical substitution model with sentence  $w$  and candidates  $\\hat{c}_j$ , using Formula 1. Similar to Equation 8, the loss  $(\\mathcal{L}_{j,n}^{(ls)})$  of the lexical substitution model when  $\\hat{c}_{j,n}$  is considered as gold standard is computed as follows:",
+        "bbox": [
+            507,
+            230,
+            882,
+            315
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {j, n} ^ {(l s)} = - \\sum_ {i} \\log \\frac {\\exp \\left(d \\left(u _ {t} , \\hat {c} _ {j , n}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(d \\left(u _ {t} , \\hat {c} _ {j}\\right) / \\tau\\right)}. \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            527,
+            325,
+            882,
+            362
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Finally, the total loss is computed as:",
+        "bbox": [
+            527,
+            375,
+            803,
+            391
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} ^ {(s a)} + \\theta_ {1, 1} \\mathcal {L} _ {1, 1} ^ {(l s)} + \\dots + \\theta_ {J, N} \\mathcal {L} _ {J, N} ^ {(l s)}. \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            401,
+            882,
+            426
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            437,
+            655,
+            454
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            463,
+            623,
+            476
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To fine-tune the pretrained lexical substitution model, we combined SemEval 2007 Task 10 (McCarthy and Navigli, 2007) and CoinCo (Kremer et al., 2014). The former consists of 2,010 example sentences for 201 polysemous words. The latter contains 2,474 sentences with multiple words to be substituted. We used  $70\\%$  of the concatenated dataset for training,  $30\\%$  for testing.",
+        "bbox": [
+            505,
+            483,
+            882,
+            612
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For sentiment analysis, we used three datasets from different domains: Rotten Tomatoes (Pang and Lee, 2005) consists of 5,331 positive and 5,331 negative sentences collected from Rotten Tomatoes movie reviews. Financial PhraseBank (Malo et al., 2014) contains 4,846 sentences from English financial news, each labeled as negative, positive, or neutral. As the dataset has no standard split, We adopted a split of  $70\\%$  and  $30\\%$  for the training and testing sets. STS-Gold (Saif et al., 2013) contains 2,026 tweets, annotated as negative or positive. As the dataset has no standard split, We adopted a split of  $90\\%$  and  $10\\%$  for the training and testing sets. Details of the datasets are listed in Table 2.",
+        "bbox": [
+            507,
+            613,
+            884,
+            837
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Baselines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            848,
+            628,
+            864
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Since we aim to improve a sentiment analysis classifier with explicit WSD in a neurosymbolic fashion, vanilla RoBERTa-large (Liu et al., 2019) is",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8777",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "considered as the most direct baseline, because we also used it for sentiment classification. To put our model's performance into perspective, we further include published baselines with the highest reported accuracy on the experiment datasets, which consist of various pre-trained language models finetuned on the target dataset, i.e., FinBERT (Araci, 2019), DeBERTa (He et al., 2020; Sileo, 2023); and ones trained with additional resources, namely, DeBERTa+tarksource (Sileo, 2023) which is pretrained on a unified dataset for 500 English tasks, and SenticNet 7 (Cambria et al., 2022) which is a XLNet enhanced by sentiment knowledge.",
+        "bbox": [
+            112,
+            84,
+            492,
+            294
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Setups",
+        "text_level": 1,
+        "bbox": [
+            112,
+            306,
+            213,
+            322
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To obtain the lexical substitution model, we fine-tune ALM with the lexical substitution datasets for 20 epochs with a batch size of 20 and a learning rate of 1e-8. The pre-trained model we used as the encoder for sentiment analysis is RoBERTa-large. The hidden size for both models is set to 1024. We use the NLTK part-of-speech tagger (Bird, 2006) to label the sentiment analysis input.",
+        "bbox": [
+            112,
+            329,
+            489,
+            457
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To train our proposed framework, we adopt  $I = 5$ ,  $J = 2$ ,  $K = 15$ ,  $N = 3$ ,  $\\tau = 0.05$  and  $\\beta = 0.0001$ . We set the learning rate to 1-e6. The models are trained for 50 epochs with a batch size of 10, using early stopping. We used Adam (Kingma and Ba, 2014) as the optimizer. Experiments are run on RTX A6000 GPU. Following the setups in the baselines, we adopt accuracy as an evaluation metric.",
+        "bbox": [
+            112,
+            458,
+            489,
+            602
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            617,
+            213,
+            633
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compare our model with the baselines in Table 3. Our model outperforms existing models with the best-published results, achieving state-of-the-art performance on the three datasets. Compared to the RoBERTa baseline, our model is able to improve the accuracy of sentiment analysis on corpora from different domains to varying extent. We hypothesize that the extent of improvement is dependent on domain types. For instance, Twitter language is informal and brief, with frequent use of short homographs. Thus, the integrated disambiguation model is able to obtain the most significant increase of  $1.91\\%$  accuracy on STS-Gold. In the financial domain, metaphors and less common word senses are commonly used, accounting for our  $1.23\\%$  gain on Financial PhraseBank.",
+        "bbox": [
+            112,
+            644,
+            489,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Movie reviews, on the other hand, often use",
+        "bbox": [
+            131,
+            903,
+            485,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/806970070fa20766615286380f248bc9dd344b51900be623383c69aede464d3b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>RT</td><td>FPB</td><td>STS-Gold</td><td>Avg</td></tr><tr><td>DeBERTa</td><td>90.42</td><td>84.48</td><td>-</td><td>87.45</td></tr><tr><td>+tasksource</td><td>90.99</td><td>85.20</td><td>-</td><td>88.10</td></tr><tr><td>FinBERT</td><td>-</td><td>86.00</td><td>-</td><td>-</td></tr><tr><td>SenticNet 7</td><td>-</td><td>-</td><td>90.08</td><td>-</td></tr><tr><td>RoBERTa</td><td>91.31</td><td>87.09</td><td>97.14</td><td>91.85</td></tr><tr><td>Ours</td><td>91.87*</td><td>88.32*</td><td>99.05*</td><td>93.08</td></tr></table>",
+        "bbox": [
+            517,
+            80,
+            877,
+            170
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Comparisons with baselines. RT stands for Rotten Tomatoes. FPB stands for Financial PhraseBank. Bold font denotes the best results. The RoBERTa baseline was implemented by us. * denotes the improvement is statistically significant ( $p < 0.01$  on a two-tailed t-test), against the highest baseline score. Our results are averaged over 5 runs.",
+        "bbox": [
+            507,
+            178,
+            885,
+            279
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "standard expressions that are semantically clearer. Hence, our model obtains a marginal increase of  $0.56\\%$  on Rotten Tomatoes. Despite the improvement being marginal on average, i.e.,  $1.23\\%$ , we would like to highlight that this improvement is significant considering the context of error propagation from the lexical substitution model. Our framework consistently demonstrates improvements across all datasets, substantiating its ability to achieve higher gains in accuracy while minimizing any potential loss. Furthermore, we can interpret the WSD outcomes and the sentiment analysis decision-making process before and after lexical substitutions from the proposed neurosymbolic framework. We will demonstrate the interpretability of the framework in Section 6.2.",
+        "bbox": [
+            507,
+            305,
+            885,
+            562
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.1 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            590,
+            673,
+            607
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To prove the effectiveness of each component in our proposed framework, we compare with two variations: w/o WSD, where sentiment prediction is done by the sentiment analysis model without any modifications to the input; w/o DR, where the lexical substitution model is not fine-tuned during sentiment analysis training. To prove our hypothesis in WSD target word selection, we compare the proposed framework with two variations: Top-2, where the top 2 words in the input with the highest attention weights are automatically chosen as target words to be disambiguated; Rand, where 2 out of the most attended 5 words are randomly selected to be the target words. The comparison of performance is shown in Table 4. It can be observed that our full model outperforms w/o WSD by  $0.28\\%$  on Rotten Tomatoes,  $1.03\\%$  on Financial PhraseBank, and  $1.43\\%$  on STS-Gold.",
+        "bbox": [
+            507,
+            612,
+            884,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The extent of gains in different domains is in",
+        "bbox": [
+            526,
+            903,
+            880,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8778",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/9ab1ce708fb261d56c4beeb33424149cf55fb59e8632f991d4456c9d67bfb5a1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>RT</td><td>FPB</td><td>STS-Gold</td><td>Avg</td></tr><tr><td>w/o WSD</td><td>91.59</td><td>87.29</td><td>97.62</td><td>92.17</td></tr><tr><td>w/o DR</td><td>91.40</td><td>87.27</td><td>96.67</td><td>91.78</td></tr><tr><td>Top-2</td><td>91.60</td><td>87.81</td><td>97.62</td><td>92.34</td></tr><tr><td>Rand</td><td>89.72</td><td>86.30</td><td>97.14</td><td>91.05</td></tr><tr><td>Ours</td><td>91.87</td><td>88.32</td><td>99.05</td><td>93.08</td></tr></table>",
+        "bbox": [
+            136,
+            80,
+            467,
+            156
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Ablation study results.",
+        "bbox": [
+            191,
+            167,
+            408,
+            181
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "line with our hypothesis in the previous section. It shows that our disambiguation framework is indeed able to improve the accuracy of sentiment analysis by selecting the appropriate ambiguous words and paraphrasing them into semantically clearer ones. Our model also achieves higher accuracy than w/o DR on all three datasets, proving the validity and effectiveness of our dynamic rewarding mechanism. Noticeably, without dynamic rewarding and joint fine-tuning, the performance of w/o DR is weaker than w/o WSD, showing that dynamic rewarding-based fine-tuning with data from another task can effectively alleviate error propagation from lexical substitution.",
+        "bbox": [
+            112,
+            206,
+            489,
+            431
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our proposed framework outperforms the Top-2 selection method on all datasets. It proves our assumption that words that are highly diverse in their senses are more likely to cause ambiguity for sentiment prediction. Rand performs the worst on all three datasets compared to w/o WSD, Top-2, and our model, indicating i) the necessity of selecting the appropriate target words to disambiguate so as not to introduce noise, and ii) the validity of choosing target words based on both their importance to the decision and their sense diversity level.",
+        "bbox": [
+            112,
+            432,
+            489,
+            609
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.2 Interpretability Demonstration",
+        "text_level": 1,
+        "bbox": [
+            112,
+            623,
+            405,
+            639
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Since an interpretable encoder (HAN) is used in our framework to learn sequential representations, we can understand which words contribute the most to a sentiment analysis prediction by visualizing the attention weights of HAN.",
+        "bbox": [
+            112,
+            644,
+            487,
+            722
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 2, we conduct a study on the change of attention weights before and after paraphrasing. In Figure 2 (a), the words that receive the most attention in the original sentence are not,  $a$ , sore, etc., among which not has significantly higher weight. This indicates that the model is mostly relying on negation to make inference. As not does not have diverse meaning and  $a$  is an article, sore is selected as target word and replaced with painful. Our attention module assigns a much higher weight to the negative adjective after paraphrasing.",
+        "bbox": [
+            112,
+            725,
+            489,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This shows that our model is able to find suitable",
+        "bbox": [
+            131,
+            903,
+            487,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b40c38e1a8db81de2405e02c97c9e7c4738fed8a0b063751c1baafc1a4ffc089.jpg",
+        "image_caption": [
+            "Figure 2: The most attended word demonstration before and after paraphrasing in sentiment analysis predictions. The lighter the color, the higher the attention weight."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            84,
+            882,
+            262
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "word substitutions that aid the correct sentiment predictions. From Figure 2 (b), we can observe a rather evenly distributed attention weights in the original sentence, meaning that the attention module cannot identify the words that significantly affect sentiment. After our framework replaces the word decreased to fell, the model is able to assign higher weights to the sentiment-loaded phrase profit fell. From Figure 2 (c), we can see that after replacing the word mediocre with average, the sentiment related words including the paraphrased word and unaware all received higher attention. By attending to the straightforward paraphrases expressing opinions, the sentiment analysis model can yield correct predictions for all three cases.",
+        "bbox": [
+            505,
+            351,
+            885,
+            593
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.3 Case Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            617,
+            643,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In Table 5, we study the ability of our framework to identify and paraphrase ambiguous words into unequivocal ones. In the first instance, albeit not affecting the prediction, the word delicious is disambiguated by our framework into its figurative meaning delighted instead of its literal one. In the second instance, the word succeeds has a more positive word sense equating to win, and a more neutral one equating to follow. Our framework is able to paraphrase it into the latter. In the last instance, the word visibility falls under the word sense of degree of exposure to public notice in this context. Our framework correctly paraphrases it into profile, which is the synonym under the same sense in WordNet. It can be concluded that our framework can disambiguate word senses in a transparent and interpretable way for sentiment analysis.",
+        "bbox": [
+            505,
+            645,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8779",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/ed434bca90666544c47056017d90d1f907c0500ad2e4c8fc856b66587fc576c6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Setup</td><td>Content</td></tr><tr><td>Original</td><td>a delicious and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them.</td></tr><tr><td>Paraphrased</td><td>a delighted and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them.</td></tr><tr><td>Original</td><td>Le Lay succeeds Walter G++nter and will be based in Finland.</td></tr><tr><td>Paraphrased</td><td>Le Lay follows Walter G++nter and will be based in Finland.</td></tr><tr><td>Original</td><td>Strong brand visibility nationally and regionally is of primary importance in home sales, vehicle and consumer advertising.</td></tr><tr><td>Paraphrased</td><td>Strong brand profile nationally and regionally is of primary importance in home sales, vehicle and consumer advertising.</td></tr></table>",
+        "bbox": [
+            124,
+            80,
+            480,
+            298
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6.4 Error Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            351,
+            278,
+            367
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Finally, we perform a brief error analysis (Table 6). In the first sentence, the model wrongly paraphrases force into power, because it does not recognize the multi-word idiom tour de force. It shows the model's limitation in understanding the semantic meaning of infrequent multi-word idiom. In the second sentence, the word just is wrongly paraphrased as good, which negatively influenced the sentiment prediction. This is mainly due to the part-of-speech tagger wrongly classifying the word just into a adjective, instead of an adverb. Thus, the lexical substitution module might be subjected to error propagation of the part-of-speech tagger and introduce noise into the input. In the last sentence, the word skills is paraphrased as science. It is likely that the model mistakes the word major as a field of study instead of a student studying a field. Thus, the substitution makes sense in the local context the science of calculus major, but not in the whole context, meaning that the model fails to understand the semantic meaning of the sentence.",
+        "bbox": [
+            112,
+            375,
+            489,
+            712
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            728,
+            247,
+            744
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose a neurosymbolic framework for sentiment analysis that incorporates the disambiguation of word senses by identifying and paraphrasing ambiguous words in the input, which is realized by leveraging a pre-trained lexical substitution model. The proposed framework not only improves the accuracy of sentiment analysis, but also achieves interpretability by enabling us to trace back which words are paraphrased and what words they are replaced with.",
+        "bbox": [
+            112,
+            758,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/b912546ca4bfdea26d55ee3efc44ad13591b874ce31963783a2b0e27f8070c0e.jpg",
+        "table_caption": [
+            "Table 5: Case study of disambiguation as paraphrasing."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Setup</td><td>Content</td><td>Label</td></tr><tr><td>Original</td><td>a tour de force of modern cinema.</td><td>Positive</td></tr><tr><td>Paraphrased</td><td>a tour de power of modern cinema.</td><td>Positive</td></tr><tr><td>Original</td><td>Is just at work for the day -sigh- I have a headache</td><td>Negative</td></tr><tr><td>Paraphrased</td><td>Is good at work for the day -sigh- I have a headache</td><td>Positive</td></tr><tr><td>Original</td><td>the skills of a calculus major at m.i.t. are required to balance all the formulaic equations in the long-winded heist comedy who is cletis tout?</td><td>Negative</td></tr><tr><td>Paraphrased</td><td>the science of a calculus major at m.i.t. are required to balance all the formulaic equations in the long-winded heist comedy who is cletis tout?</td><td>Positive</td></tr></table>",
+        "bbox": [
+            517,
+            80,
+            877,
+            298
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Error analysis. Red indicates errors made by our proposed framework.",
+        "bbox": [
+            507,
+            307,
+            882,
+            338
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Additionally, to fine-tune the lexical substitution model to provide better disambiguating substitutes for sentiment analysis with no ground-truth word sense labels, we adopt the dynamic rewarding mechanism to jointly train the sentiment analysis and lexical substitution models. Experiments show that our framework is indeed effective in improving the performance of sentiment analysis, obtaining state-of-the-art performance on three corpora from different domains.",
+        "bbox": [
+            507,
+            363,
+            884,
+            526
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            556,
+            613,
+            570
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The framework proposed in this paper have to be seen in light of some limitations. First, the performance of sentiment analysis to some extent depends on a decent-performing pre-trained lexical substitution model. Currently, polysemous samples in lexical substitution data are not as abundant as in WSD data, which might limit its ability in disambiguating less common polysemies. Second, the effectiveness of our proposed model is only tested on English corpora, as it relies on the sense knowledge from WordNet. The disambiguation target selection and paraphrasing methods might work on languages with more complex segmentation or limited sense lexicon.",
+        "bbox": [
+            507,
+            583,
+            882,
+            809
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            509,
+            853,
+            608,
+            869
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "ref_text",
+        "text": "Abien Fred Agarap. 2018. Deep learning using rectified linear units (ReLU). arXiv preprint arXiv:1803.08375.",
+        "bbox": [
+            509,
+            877,
+            882,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8780",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Dogu Araci. 2019. FinBERT: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063.",
+            "Philip Arthur, Graham Neubig, Sakriani Sakti, Tomoki Toda, and Satoshi Nakamura. 2015. Semantic parsing of ambiguous input through paraphrasing and verification. Transactions of the Association for Computational Linguistics, 3:571-584.",
+            "Stefano Baccianella, Andrea Esuli, and Fabrizio Sebastiani. 2010. SentiWordNet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC'10).",
+            "Edoardo Barba, Tommaso Pasini, and Roberto Navigli. 2021a. ESC: Redesigning WSD with Extractive Sense Comprehension. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4661-4672, Online. Association for Computational Linguistics.",
+            "Edoardo Barba, Luigi Procopio, and Roberto Navigli. 2021b. ConSeC: Word sense disambiguation as continuous sense comprehension. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1492-1503, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Michele Bevilacqua and Roberto Navigli. 2020. Breaking through the  $80\\%$  glass ceiling: Raising the state of the art in word sense disambiguation by incorporating knowledge graph information. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2854-2864.",
+            "Steven Bird. 2006. NLTK: the natural language toolkit. In Proceedings of the COLING/ACL 2006 Interactive Presentation Sessions, pages 69-72.",
+            "Terra Blevins and Luke Zettlemoyer. 2020. Moving down the long tail of word sense disambiguation with gloss-informed biencoders. arXiv preprint arXiv:2005.02590.",
+            "Erik Cambria, Qian Liu, Sergio Decherchi, Frank Xing, and Kenneth Kwok. 2022. SenticNet 7: A commonsense-based neurosymbolic AI framework for explainable sentiment analysis. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 3829-3839.",
+            "Erik Cambria, Rui Mao, Melvin Chen, Zhaoxia Wang, and Seng-Beng Ho. 2023. Seven pillars for the future of AI. IEEE Intelligent Systems, 38(6).",
+            "Erik Cambria, Soujanya Poria, Federica Bisio, Rajiv Bajpai, and Iti Chaturvedi. 2015. The CLSA model: A novel framework for concept-level sentiment analysis. In LNCS, volume 9042, pages 3-22. Springer."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Erik Cambria, Soujanya Poria, Alexander Gelbukh, and Mike Thelwall. 2017. Sentiment analysis is a big suitcase. IEEE Intelligent Systems, 32(6):74-80.",
+            "Niccolò Campolungo, Tommaso Pasini, Denis Emelin, and Roberto Navigli. 2022. Reducing disambiguation biases in NMT by leveraging explicit word sense information. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4824-4838.",
+            "Yanfang Cao, Pu Zhang, and Anping Xiong. 2015. Sentiment analysis based on expanded aspect and polarity-ambiguous word lexicon. International Journal of Advanced Computer Science and Applications, 6(2).",
+            "Ann Devitt and Khurshid Ahmad. 2007. Sentiment polarity identification in financial news: A cohesion-based approach. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 984-991.",
+            "Philip Edmonds and Adam Kilgarriff. 2002. Introduction to the special issue on evaluating word sense disambiguation systems. Natural Language Engineering, 8(4):279-291.",
+            "Umar Farooq, Tej Prasad Dhamala, Antoine Nongaillard, Yacine Ouzrout, and Muhammad Abdul Qadir. 2015. A word sense disambiguation method for feature level sentiment analysis. In 2015 9th International Conference on Software, Knowledge, Information Management and Applications (SKIMA), pages 1-8. IEEE.",
+            "Christiane Fellbaum. 1998. WordNet: An electronic lexical database. MIT press.",
+            "Mengshi Ge, Rui Mao, and Erik Cambria. 2022. Explanable metaphor identification inspired by conceptual metaphor theory. Proceedings of the AAAI Conference on Artificial Intelligence, 36(10):10681-10689.",
+            "Sooji Han, Rui Mao, and Erik Cambria. 2022. Hierarchical attention network for explainable depression detection on Twitter aided by metaphor concept mappings. In Proceedings of the 29th International Conference on Computational Linguistics (COLING), pages 94-104.",
+            "Viktor Hangya, Qianchu Liu, Dario Stojanovski, Alexander Fraser, and Anna Korhonen. 2021. Improving machine translation of rare and unseen word senses. In Proceedings of the Sixth Conference on Machine Translation, pages 614-624.",
+            "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. DeBERTa: Decoding-enhanced BERT with disentangled attention. arXiv preprint arXiv:2006.03654."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8781",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Luyao Huang, Chi Sun, Xipeng Qiu, and Xuanjing Huang. 2019. GlossBERT: BERT for word sense disambiguation with gloss knowledge. arXiv preprint arXiv:1908.07245.",
+            "Chihli Hung and Shiuman-Jeng Chen. 2016. Word sense disambiguation based sentiment lexicons for sentiment classification. Knowledge-Based Systems, 110:224-232.",
+            "Chihli Hung and Hao-Kai Lin. 2013. Using objective words in SentiWordNet to improve word-of-mouth sentiment classification. IEEE Intelligent Systems, 28(02):47-54.",
+            "Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.",
+            "Gerhard Kremer, Katrin Erk, Sebastian Padó, and Stefan Thater. 2014. What substitutes tell us-analysis of an \"all-words\" lexical substitution corpus. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 540-549.",
+            "Frederick Liu, Han Lu, and Graham Neubig. 2018. Handling homographs in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1336-1345.",
+            "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.",
+            "P. Malo, A. Sinha, P. Korhonen, J. Wallenius, and P. Takala. 2014. Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology, 65.",
+            "Rui Mao, Kai He, Claudia Beth Ong, Qian Liu, and Erik Cambria. 2023a. MetaPro 2.0: Computational metaphor processing on the effectiveness of anomalous language modeling.",
+            "Rui Mao and Xiao Li. 2021. Bridging towers of multitask learning with a gating mechanism for aspect-based sentiment analysis and sequential metaphor identification. Proceedings of the AAAI Conference on Artificial Intelligence, 35(15):13534-13542.",
+            "Rui Mao, Xiao Li, Kai He, Mengshi Ge, and Erik Cambria. 2023b. MetaPro Online: A computational metaphor processing online system. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 127-135.",
+            "Rui Mao, Qian Liu, Kai He, Wei Li, and Erik Cambria. 2023c. The biases of pre-trained language models:"
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "An empirical study on prompt-based sentiment analysis and emotion detection. IEEE Transactions on Affective Computing, 14(3):1743-1753.",
+            "Marco Maru, Simone Conia, Michele Bevilacqua, and Roberto Navigli. 2022. Nibbling at the hard core of Word Sense Disambiguation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4724-4737.",
+            "Diana McCarthy and Roberto Navigli. 2007. SemEval-2007 task 10: English lexical substitution task. In Proceedings of the 4th International Workshop on Semantic Evaluations, SemEval '07, pages 48-53.",
+            "Arman Khadjeh Nassirtoussi, Saeed Aghabozorgi, Teh Ying Wah, and David Chek Ling Ngo. 2015. Text mining of news-headlines for forex market prediction: A multi-layer dimension reduction algorithm with semantics and sentiment. Expert Systems with Applications, 42(1):306-324.",
+            "Bruno Ohana and Brendan Tierney. 2009. Sentiment classification of reviews using SentiWordNet. Proceedings of IT&T, 8.",
+            "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.",
+            "Martha Palmer, Hoa Trang Dang, and Christiane Fellbaum. 2007. Making fine-grained and coarse-grained sense distinctions, both manually and automatically. Natural Language Engineering, 13(2):137-163.",
+            "Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the ACL.",
+            "Xiao Pu, Nikolaos Pappas, James Henderson, and Andrei Popescu-Belis. 2018. Integrating weakly supervised word sense disambiguation into neural machine translation. Transactions of the Association for Computational Linguistics, 6:635-649.",
+            "Rachel Elizabeth Ramsey. 2017. An exemplar-theoretic account of word senses. University of Northumbria at Newcastle (United Kingdom).",
+            "Hassan Saif, Miriam Fernandez, Yulan He, and Harith Alani. 2013. Evaluation datasets for Twitter sentiment analysis: A survey and a new dataset, the STS-Gold. Emotion and Sentiment in Social and Expressive Media.",
+            "Damien Sileo. 2023. tasksource: A dataset harmonization framework for streamlined NLP multitask learning and evaluation. arXiv preprint arXiv:2301.05948v3.",
+            "Yang Song, Xin Cai Ong, Hwee Tou Ng, and Qian Lin. 2021. Improved word sense disambiguation with enhanced sense representations. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 4311-4320."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "8782",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yunqing Xia, Erik Cambria, Amir Hussain, and Huan Zhao. 2015. Word polarity disambiguation using bayesian model and opinion-level features. Cognitive Computation, 7(3):369-380.",
+            "Yunqing Xia, Huan Zhao, Erik Cambria, and Amir Hussain. 2014. Disambiguating word sentiment polarity through bayesian modeling and opinion-level features. In Proceedings of ICML Workshops."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            200
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "8783",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    }
+]

2023/Neuro-Symbolic Sentiment Analysis with Dynamic Word Sense Disambiguation/f2ea77f7-c2a2-49e4-b922-17fa39103dc4_model.json

@@ -0,0 +1,2061 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.276,
+                0.091,
+                0.723,
+                0.131
+            ],
+            "angle": 0,
+            "content": "Neurosymbolic Sentiment Analysis with Dynamic Word Sense Disambiguation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.25,
+                0.148,
+                0.751,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Xulang Zhang\\*, Rui Mao\\*, Kai He\\* and Erik Cambria\\*"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.31,
+                0.167,
+                0.694,
+                0.184
+            ],
+            "angle": 0,
+            "content": "\\(\\spadesuit\\)Nanyang Technological University, Singapore"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.315,
+                0.184,
+                0.689,
+                0.201
+            ],
+            "angle": 0,
+            "content": "\\*National University of Singapore, Singapore"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.201,
+                0.86,
+                0.217
+            ],
+            "angle": 0,
+            "content": "xulang001@e.ntu.edu.sg; {rui.mao, cambria}@ntu.edu.sg; kai_he@nus.edu.sg"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.283,
+                0.461,
+                0.653
+            ],
+            "angle": 0,
+            "content": "Sentiment analysis is a task that highly depends on the understanding of word senses. Traditional neural network models are black boxes that represent word senses as vectors that are uninterpretable for humans. On the other hand, the application of Word Sense Disambiguation (WSD) systems in downstream tasks poses challenges regarding i) which words need to be disambiguated, and ii) how to model explicit word senses into easily understandable terms for a downstream model. This work proposes a neurosymbolic framework that incorporates WSD by identifying and paraphrasing ambiguous words to improve the accuracy of sentiment predictions. The framework allows us to understand which words are paraphrased into which semantically unequivocal words, thus enabling a downstream task model to gain both accuracy and interpretability. To better fine-tune a lexical substitution model for WSD on a downstream task without ground-truth word sense labels, we leverage dynamic rewarding to jointly train sentiment analysis and lexical substitution models. Our framework proves to effectively improve the performance of sentiment analysis on corpora from different domains."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.667,
+                0.26,
+                0.683
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.111,
+                0.695,
+                0.49,
+                0.904
+            ],
+            "angle": 0,
+            "content": "To improve the accuracy of sentiment analysis, it is crucial to disambiguate polysemous words according to the context (Cambria et al., 2023). A word may have senses that carry opposite sentiments, e.g., \"fine\" has the positive meaning of being delicate and refined, and the negative meaning of monetary penalty. Additionally, words may appear in positive, negative, or neutral senses in different domains. For instance, the word \"bearish\" is often used in a negative sense to describe the stock market in the financial domain, while more likely to be used neutrally to describe hulking figures in other domains."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.254,
+                0.885,
+                0.639
+            ],
+            "angle": 0,
+            "content": "In traditional black-box sentiment analysis, the representations of word meanings are generated by neural networks (Mao and Li, 2021). As such, the disambiguation process is implicit and opaque. It is challenging to investigate what implicit senses are assigned to ambiguous words by a neural network model for it to make accurate predictions. Coupling symbolism and sub-symbolism, a neurosymbolic approach to sentiment analysis is transparent and interpretable, regarding the extraction and distinction of word meaning as a sub-problem (Cambria et al., 2017). However, research that explicitly deals with ambiguous words in sentiment analysis is limited (Xia et al., 2014, 2015; Cambria et al., 2015). Sense-annotated resources such as SentiWordNet (Baccianella et al., 2010) have been utilized, but only to detect and score sentimental words in the input (Hung and Lin, 2013; Nassirtoussi et al., 2015; Hung and Chen, 2016). The disambiguation of polysemous words in sentiment analysis inputs remains under-explored. Thus, we are motivated to develop a neurosymbolic framework for sentiment analysis with explicit Word Sense Disambiguation (WSD)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.643,
+                0.885,
+                0.9
+            ],
+            "angle": 0,
+            "content": "Traditional WSD aims to classify which sense of the target word is appropriate according to its context (Huang et al., 2019; Blevins and Zettlemoyer, 2020; Bevilacqua and Navigli, 2020; Song et al., 2021; Barba et al., 2021a,b). However, the limitations \\((\\ell)\\) of their task setups are that \\((\\ell 1)\\) the output of those WSD systems, whether sense label (Bevilacqua and Navigli, 2020) or gloss definition (Blevins and Zettlemoyer, 2020; Barba et al., 2021a), is difficult to integrate into downstream input; \\((\\ell 2)\\) the target words must be pre-defined for WSD systems to disambiguate, making it difficult to apply them directly to downstream tasks; and \\((\\ell 3)\\) the reliability of sense distinction of the benchmark has been questioned (Ramsey, 2017; Maru et al., 2022), as word senses are not discrete."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8772"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.78,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8772-8783  \nDecember 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.295
+            ],
+            "angle": 0,
+            "content": "Attempts have been made to incorporate the disambiguation of word senses into downstream applications. Nonetheless, they are either impractical in real-world settings (Farooq et al., 2015; Pu et al., 2018), or only applicable to a specific downstream task, i.e., machine translation (Hangya et al., 2021; Campolungo et al., 2022). To address these issues, we leverage dynamic WSD in our neurosymbolic framework. We regard WSD as a lexical substitution task (for \\(\\ell 1\\)), where ambiguous words in sentiment analysis input are paraphrased into ones that cause less sentiment confusion using a lexical substitution model with dynamic WSD."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.301,
+                0.492,
+                0.576
+            ],
+            "angle": 0,
+            "content": "Our dynamic WSD method comprises two distinct aspects: the dynamic selection of the target word for disambiguation, based on downstream contextual information (for \\(\\ell 2\\)); and the dynamic learning of the most suitable lexical substitutions, based on the prediction accuracy of the downstream model (for \\(\\ell 3\\)). To identify target words that need to be substituted, we first adopt an attention-based explainable encoder to rank input words by the level of influence they have on the output, motivated by the assumption that not all polysemous words affect sentiment predictions (Cao et al., 2015). Then, the top-ranking words with the most diverse senses are selected to be targets, based on our hypothesis that an ambiguous word may coincide with significant disparities in the semantic space between the word itself and its potential substitution candidates."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.582,
+                0.492,
+                0.919
+            ],
+            "angle": 0,
+            "content": "As such, we can trace back which words are selected as targets, and which substitute words are used to disambiguate them, making our method interpretable and transparent. Furthermore, we fine-tune the lexical substitution model whilst training the sentiment analysis model by adopting a dynamic rewarding mechanism that does not require gold standard substitution labels. As such, the substitution model can be adjusted to provide more appropriate output for the downstream task as well as the target domain. Our model achieves more accurate sentiment analysis with explicit lexical-substitution-based WSD, despite not specifically trained on WSD and sentiment parallel datasets. Since the disambiguated word senses are represented as the lexical substitutions of the original words and an explainable encoder is used for retrieving the most contributing words, the neurosymbolic sentiment inference process is more interpretable for humans compared to black-box models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.887,
+                0.296
+            ],
+            "angle": 0,
+            "content": "Our model is examined with three publicly available datasets from movie review, finance, and Twitter domains, outperforming strong baselines by \\(1.23\\%\\) on average. The contributions of our work are twofold: (1) We propose a neurosymbolic sentiment analysis framework that leverages WSD as an auxiliary lexical substitution task, achieving state-of-the-art performance on our examined sentiment analysis datasets from different domains; (2) We design a transparent mechanism to select WSD target words, based on a) the importance of words to the learning and inference of a downstream task, and b) word sense diversity."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.31,
+                0.666,
+                0.326
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.339,
+                0.886,
+                0.725
+            ],
+            "angle": 0,
+            "content": "Word Sense Disambiguation is a task that assigns the appropriate meaning to an ambiguous word according to its context. Recent WSD systems attained F1 scores higher than \\(80\\%\\) on the benchmark datasets (Song et al., 2021; Bevilacqua and Navigli, 2020; Barba et al., 2021a,b), which is regarded as the estimated human performance since it is the highest inter-annotator agreement (Edmonds and Kilgarriff, 2002; Palmer et al., 2007). However, research on the application of WSD systems on downstream tasks is limited. Pu et al. (2018) utilized clustering algorithms to obtain sense embeddings by inferring from the similarity of the context vectors of a target word. The heavy requirement for disambiguating every homographs prior to downstream applications makes their method less feasible in many real-world situations. Campolungo et al. (2022) achieved the disambiguation of word sense in neural machine translation by leveraging a multilingual WSD system to establish a sense-tagged parallel corpus. Nonetheless, their approach relied on cross-lingual word alignments for reliable annotations, meaning that it cannot be adopted by other types of downstream tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.727,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Sentiment Analysis can be generally defined as the task that determines if a piece of text expresses positive, negative, or neutral opinion (Mao et al., 2023c). Word sense information has been applied in sentiment analysis in the form of sentiment-annotated lexicons (Devitt and Ahmad, 2007; Ohana and Tierney, 2009; Hung and Lin, 2013; Nassirtoussi et al., 2015), such as SentiWordNet. However, these methods incorporate the sentiment information of a word without disambiguating its word sense according to the context, but resorting to the emotional score of the most frequent"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8773"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.115,
+                0.081,
+                0.887,
+                0.326
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.337,
+                0.884,
+                0.396
+            ],
+            "angle": 0,
+            "content": "Figure 1: Architecture of the proposed disambiguation framework for sentiment analysis. A dotted line means backpropagation. Green block means that the original word is replaced by a word provided by the lexical substitution model. The darker the green, the higher probability it is assigned to be the model. Orange block indicates the words with top ranking attention scores produced by HAN."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.421,
+                0.489,
+                0.581
+            ],
+            "angle": 0,
+            "content": "sense, or the average score of all senses. Additionally, the main objective of this approach is to identify sentimental words in a sentence, instead of disambiguating polysemous words that might affect model predictions. Hung and Chen (2016); Farooq et al. (2015); Liu et al. (2018) attempted to disambiguate according to target domain and context, but either costed extensive human labour, or fail to incorporate word senses explicitly and transparently."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.583,
+                0.489,
+                0.695
+            ],
+            "angle": 0,
+            "content": "To address these limitations, our neurosymbolic framework aims to identify ambiguous target words in the context and replace them with clarifying ones automatically, making them interpretable. We also leverage a dynamic rewarding mechanism so that the substitutions can be fine-tuned to better serve the downstream task and the text domain."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.713,
+                0.282,
+                0.729
+            ],
+            "angle": 0,
+            "content": "3 Task Definition"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.743,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this work, we regard word sense ambiguity as the potential cause of errors in a downstream task, e.g., sentiment analysis. Inspired by the work of Mao et al. (2023b), we reframe WSD as a lexical substitution task, aiming to replace an ambiguous word with a synonym that not only represents the true sense of the original word in the context (Arthur et al., 2015), but also reduces misclassifications in the downstream task. Given an input sentence \\( w = (w_{1},w_{2},\\dots ,w_{L}) \\), we first aim to identify the target word \\( t \\) that is likely troublesome for"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.421,
+                0.884,
+                0.485
+            ],
+            "angle": 0,
+            "content": "the downstream prediction from \\( w \\). To resolve the potential ambiguity, we extract \\( M \\) of its synonyms from WordNet (Fellbaum, 1998) as candidates \\( c = (c_{1},\\ldots ,c_{M}) \\), forming the input \\( (s) \\) as"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.575,
+                0.494,
+                0.884,
+                0.515
+            ],
+            "angle": 0,
+            "content": "\\[\ns = \\langle s \\rangle , w _ {1}, w _ {2}, \\dots , t, \\dots , w _ {L}, \\langle / s \\rangle , \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.53,
+                0.883,
+                0.707
+            ],
+            "angle": 0,
+            "content": "\\(< \\mathrm{s}>\\) and \\(</ \\mathrm{s}>\\) are special tokens that were defined by our employed pre-trained language model. The lexical substitution model should identify the best candidate from \\(c\\) as the substitute \\(\\hat{c}\\) which retains the original meaning of \\(t\\) in the context by using contrast learning. Thus, \\(\\hat{c}\\) is the symbolic representation of \\(t\\) in context \\(w\\). Then, the paraphrased input \\(w^{f} = (w_{1},\\dots,\\hat{c},\\dots,w_{L})\\) is fed into a neural network classifier to predict a sentiment label. The objective is that the paraphrased input \\(w^{f}\\) increases the probability of correct sentiment prediction."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.718,
+                0.659,
+                0.735
+            ],
+            "angle": 0,
+            "content": "4 Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.743,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Our proposed neurosymbolic framework can be viewed in Figure 1. There are two models in our framework, namely sentiment analysis and lexical substitution models. The sentiment analysis model consists of an encoder, an interpretable attention module, and multiple feedforward neural networks. First, through Figure 1 (1) and (2), the original input is fed into the encoder and the interpretable attention module to obtain the top \\(I\\) tokens with the highest attention weights, which contribute the most to sentiment prediction."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8774"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.295
+            ],
+            "angle": 0,
+            "content": "In (3), the sense diversity of each token is computed as the average distance of the token's hidden state to those of its substitution candidates. Then, through (4), top \\( J \\) tokens with the most diverse senses are selected as disambiguation targets, because these target tokens may have very different senses in different contexts. Subsequently, the pretrained lexical substitution model is applied to provide the best substitution to replace each target token, as shown in (5) and (6)a. The new input sentence is passed onto the sentiment analysis model for final prediction ((7)a and (8)a), which is used for the sentiment model backpropagation ((9)a)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.296,
+                0.49,
+                0.633
+            ],
+            "angle": 0,
+            "content": "In order to fine-tune the lexical substitution model to provide better substitutions that improve the accuracy of sentiment analysis, we adapt a dynamic rewarding mechanism (Ge et al., 2022) that was proposed for multi-task learning. As shown in Figure 1 (5) and (6)b, for each target word, we choose the top \\( N \\) candidates. Each of them is seen as a substitute candidate to calculate the probability of correct sentiment prediction after the replacement ((7)b and (8)b). Then, each probability is used to dynamically compute the loss weight when the corresponding candidate is learned by the lexical substitution model as a ground truth ((9)b and (10)b). If the replacement by a candidate leads to a higher chance of predicting the right sentiment, the candidate is more likely given a higher rank by the lexical substitution model. Otherwise, the model would be less likely to select the candidate as the best replacement. The important notations used in this paper are summarized in Table 1. The detailed training process can be seen in Algorithm 1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.643,
+                0.478,
+                0.658
+            ],
+            "angle": 0,
+            "content": "4.1 Sentiment Analysis with Interpretability"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.663,
+                0.489,
+                0.711
+            ],
+            "angle": 0,
+            "content": "Given input sentence \\( w = (w_{1},\\dots ,w_{L}) \\), the goal is to predict the correct sentiment label \\( \\tilde{y} \\). The input is first fed into a pre-trained encoder:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.228,
+                0.722,
+                0.488,
+                0.739
+            ],
+            "angle": 0,
+            "content": "\\[\nV = \\operatorname {E n c o d e r} (w), \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.748,
+                0.49,
+                0.892
+            ],
+            "angle": 0,
+            "content": "where \\(V\\) is hidden states. Next, we aim to find the tokens that contribute the most to sentiment inference. We adopt an interpretable attention module called Hierarchical Attention Network (HAN, Han et al., 2022), which effectively encodes hidden states with multiple non-linear projections and ranks the most influential tokens based on attention. Following the setup of Han et al. (2022), we stack two blocks of HAN to form our attention module."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.2,
+                0.903,
+                0.488,
+                0.921
+            ],
+            "angle": 0,
+            "content": "\\[\nq, a = H A N (H A N (V)), \\tag {3}\n\\]"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.614,
+                0.083,
+                0.778,
+                0.096
+            ],
+            "angle": 0,
+            "content": "Table 1: Notation table."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.511,
+                0.108,
+                0.88,
+                0.744
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Description</td></tr><tr><td>L</td><td>Input sequence length</td></tr><tr><td>I</td><td>Number of words with top ranking attention scores to be potential target words</td></tr><tr><td>J</td><td>Number of words with top ranking diversity scores to be potential target words</td></tr><tr><td>N</td><td>Number of candidates used for finetuning lexical substitution via dynamic rewarding</td></tr><tr><td>K</td><td>Number of candidates for training lexical substitution model</td></tr><tr><td>w</td><td>The original input sentence</td></tr><tr><td>V</td><td>The hidden states of w</td></tr><tr><td>y</td><td>The ground-truth sentiment label of w</td></tr><tr><td>wf</td><td>The final input sentence where ambiguous words in w are substituted</td></tr><tr><td>ŷf</td><td>The predicted sentiment label with wf as input</td></tr><tr><td>wl</td><td>The l-th word in w</td></tr><tr><td>al</td><td>The attention weight of the l-th word in w</td></tr><tr><td>wiatt</td><td>The word in w with the i-th highest attention weight</td></tr><tr><td>viatt</td><td>The hidden state of wiatt</td></tr><tr><td>M</td><td>The number of synonyms of wiatt with the same part-of-speech that are provided by the WordNet</td></tr><tr><td>G</td><td>The hidden states of the M WordNet synonyms</td></tr><tr><td>gm</td><td>The hidden state of the m-th WordNet synonyms of wiatt</td></tr><tr><td>di</td><td>The average Euclidean distance between viatt and g1, . . . , gm, . . . , gM</td></tr><tr><td>tj</td><td>The j-th target word in w to be substituted</td></tr><tr><td>c</td><td>A set of candidate words from WordNet.</td></tr><tr><td>ˆc</td><td>The top N candidates ranked by the lexical substitution model</td></tr><tr><td>ˆcj,n</td><td>The top n-th candidate for replacing tj</td></tr><tr><td>wj,n</td><td>The resulting sentence when replacing tj in w withˆcj,n</td></tr><tr><td>ˆθj,n</td><td>The loss weight whenˆcj,n is treated as the gold standard substitution</td></tr><tr><td>s</td><td>The formatted input to lexical substitution model</td></tr><tr><td>U</td><td>The hidden state matrix of the input sentence in s</td></tr><tr><td>u1</td><td>The hidden state of the l-th input word in s</td></tr><tr><td>R</td><td>The hidden state matrix of the candidates in s</td></tr><tr><td>rk</td><td>The hidden state of the k-th candidate in s</td></tr><tr><td>τ</td><td>The temperature hyper-parameter for training the lexical substitution model</td></tr><tr><td>β</td><td>The hyper-parameter for balancing the sentiment analysis loss and fine-tuning loss</td></tr><tr><td>L(sa)</td><td>The sentiment analysis loss</td></tr><tr><td>L(tune)</td><td>The loss for finetuning the pretrained language model into a lexical substitution model</td></tr><tr><td>L(ls)</td><td>The lexical substitution loss whenˆcj,n is treated as the gold standard substitution</td></tr><tr><td>L(ls)</td><td>The final weighted lexical substitution loss for backpropagation</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.791,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "where vector \\( q \\) is the yielded hidden state. \\( a \\) is the attention weights, indicating the contribution of each token to the final sentiment prediction. To obtain the sentiment prediction, \\( q \\) is passed on to two layers of feed-forward neural networks (FNN) to obtain the probability distribution of sentiment prediction, with the first one being activated by ReLU (Agarap, 2018), and the second by softmax."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8775"
+        }
+    ],
+    [
+        {
+            "type": "equation",
+            "bbox": [
+                0.198,
+                0.113,
+                0.488,
+                0.13
+            ],
+            "angle": 0,
+            "content": "\\[\nh = R e L U \\left(F N N _ {1} (q)\\right) \\tag {4}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.2,
+                0.133,
+                0.488,
+                0.151
+            ],
+            "angle": 0,
+            "content": "\\[\n\\hat {y} = \\operatorname {s o f t m a x} \\left(F N N _ {2} (h)\\right) \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.163,
+                0.489,
+                0.243
+            ],
+            "angle": 0,
+            "content": "We denote the prediction of the sentiment analysis model as \\(\\hat{y}^f\\) when the input is \\(w^f\\), which denotes a paraphrased \\(w\\) where all selected target tokens are replaced by sense-clarifying substitutions. Thus, the sentiment analysis loss is computed as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.182,
+                0.254,
+                0.488,
+                0.273
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} ^ {(s a)} = \\operatorname {C r o s s E n t r o p y} (\\hat {y} ^ {f}, \\tilde {y}) \\tag {6}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.296,
+                0.472,
+                0.312
+            ],
+            "angle": 0,
+            "content": "4.2 Disambiguation by Lexical Substitution"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.316,
+                0.489,
+                0.445
+            ],
+            "angle": 0,
+            "content": "To obtain a lexical substitution model, we fine-tune the pre-trained language model ALM (Mao et al., 2023a)<sup>1</sup> with benchmark datasets from McCarthy and Navigli (2007); Kremer et al. (2014). We use the candidates from the datasets to formulate our input \\( s \\) as in Formula 1. The candidate with the highest score is set as the ground truth substitution \\( \\tilde{c} \\). Given the input \\( s \\), the model encodes it as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.232,
+                0.459,
+                0.488,
+                0.475
+            ],
+            "angle": 0,
+            "content": "\\[\nU, R = A L M (s), \\tag {7}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.487,
+                0.489,
+                0.552
+            ],
+            "angle": 0,
+            "content": "where \\(U = [u_{1},\\ldots ,u_{L}]\\) is the hidden states of the input sentence, and \\(R = [r_1,\\dots,r_K]\\) is the hidden states of the candidates. We denote the representation of the target word as \\(u_{t}\\) \\((t\\in \\{1,\\ldots ,L\\})\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.553,
+                0.489,
+                0.696
+            ],
+            "angle": 0,
+            "content": "Our training objective is to a) close the distance between the representation of the ground truth candidate \\( r_k \\), \\( (k \\in \\{1, \\dots, K\\}) \\) with \\( u_t \\), and b) push away incorrect candidates representations \\( r_j \\) \\( (j \\in \\{1, \\dots, K | j \\neq k\\}) \\) from \\( u_t \\). Namely, \\( (r_k, u_t) \\) will be regarded as a positive pair, while \\( (r_j, u_t) \\) will be regarded as a negative pair. We follow the InfoNCE loss (Oord et al., 2018) to achieve these goals, which can be formulated as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.132,
+                0.707,
+                0.488,
+                0.745
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} ^ {(t u n e)} = - \\sum_ {i} \\log \\frac {e x p (d (u _ {t} , r _ {k}) / \\tau)}{\\sum_ {j} e x p (d (u _ {t} , r _ {i}) / \\tau)}, \\tag {8}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.757,
+                0.489,
+                0.79
+            ],
+            "angle": 0,
+            "content": "where \\(i\\in \\{1,\\dots ,K\\}\\) \\(\\tau\\) is a temperature hyperparameter, and \\(d(\\cdot)\\) is Euclidean distance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.516,
+                0.088,
+                0.84,
+                0.135
+            ],
+            "angle": 0,
+            "content": "Algorithm 1: Sentiment analysis with word sense disambiguation using a dynamic rewarding mechanism."
+        },
+        {
+            "type": "algorithm",
+            "bbox": [
+                0.512,
+                0.139,
+                0.853,
+                0.73
+            ],
+            "angle": 0,
+            "content": "1 Initialize sentiment analysis model as \\(\\Phi\\) , pre-trained   \n2 lexical substitution model as \\(\\Psi\\) .   \n3 Initialize hyperparameters \\(\\beta\\) \\(I,J,N\\)   \nwhile not done do   \n4 Sample a sentence \\(w = w_{1},w_{2},\\ldots ,w_{L}\\) .   \nfor \\(l = 1:L\\) do   \n6 Compute the attention weight \\(a_{l}\\) of token   \n\\(w_{l}\\) .   \n7 end   \n8 \\(w^{att}\\gets\\) Top I of \\(w\\) ordered by attention weights   \n\\(a = a_1,a_2,\\dots ,a_L\\) .   \nfor \\(i = 1:I\\) do   \n10 \\(c\\gets\\) all possible candidates with the same   \npart-of-speech type as \\(w_{i}^{att}\\) , provided by   \nWordNet;   \n11 Compute the average Eulidean distance \\(d_{i}\\)   \nbetween the hidden states of \\(w_{i}^{att}\\) and \\(c\\) produced by the encoder in \\(\\Phi\\) .   \nend   \n13 \\(t = (t_1,t_2,\\dots ,t_J)\\gets\\) Top J of \\(w^{att}\\) ordered   \nby \\(d = d_{1},d_{2},\\ldots ,d_{I}\\)   \n\\(w^{f}\\gets w\\) .   \nfor \\(j = 1:J\\) do   \n16 Input w into \\(\\Psi\\) to produce top N candidates   \n\\(\\hat{c}_j = (\\hat{c}_{j,1},\\dots ,\\hat{c}_{j,N})\\) for target word \\(t_j\\) ordered by probability;   \n17 Replace \\(t_j\\) in \\(w^f\\) with \\(\\hat{c}_{j,1}\\)   \ns \\(\\leftarrow (< s > ,w_{1},\\dots ,t_{i},\\dots ,w_{L}, <   / s > ,\\right.\\)   \n\\(\\hat{c}_{j,1}, <   / s > ,\\dots ,\\hat{c}_{j,N}, <   / s > )\\) .   \nfor \\(n = 1:N\\) do   \n21 \\(w^{j,n}\\gets\\) Replace \\(t_j\\) in \\(w\\) with \\(\\hat{c}_{j,n}\\) .   \n22 Input \\(w^{j,n}\\) into \\(\\Phi\\) to obtain the probability of correct sentiment prediction \\(P(\\hat{y} = \\tilde{y})_{j,n}\\) ..   \n23 \\(\\theta_{j,n}\\gets \\beta P(\\hat{y} = \\tilde{y})_{j,n}^2\\) .   \n24 Compute \\(\\mathcal{L}_{j,n}^{(ls)}\\) by feeding s with \\(\\hat{c}_n\\) labeled as true substitute into \\(\\Psi\\) ..   \n25 \\(\\mathcal{L}_{j,n}^{(ls)}\\gets \\theta_{j,n}\\mathcal{L}_{j,n}^{(ls)};\\)   \nend   \nend   \n28 \\(\\mathcal{L}^{(ls)}\\gets \\mathcal{L}_{1,1}^{(ls)} + \\dots +\\mathcal{L}_{1,N}^{(ls)} + \\mathcal{L}_{J,1}^{(ls)} + \\dots +\\mathcal{L}_{J,N}^{(ls)}\\) .   \n29 Compute sentiment analysis loss \\(\\mathcal{L}^{(sa)}\\) using \\(w^{f}\\) as input;   \n30 \\(L\\gets \\mathcal{L}^{(sa)} + \\mathcal{L}^{(ls)}\\) ."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.762,
+                0.885,
+                0.824
+            ],
+            "angle": 0,
+            "content": "During the inference stage, we choose the candidate \\(\\hat{c}\\) whose corresponding hidden state \\(r_k\\) is the most similar to \\(u_t\\), measured by Euclidean distance:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.607,
+                0.828,
+                0.883,
+                0.845
+            ],
+            "angle": 0,
+            "content": "\\[\n\\hat {c} = \\arg \\min  \\left(d \\left(u _ {t}, r _ {k}\\right)\\right) \\tag {9}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.855,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We then use the resulting lexical substitution model to find replacements for the selected target words in sentiment analysis input. To determine which words are selected as disambiguation tar"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.81,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>ALM employed a novel pre-training paradigm, termed Anomalous Language Modeling. ALM was pre-trained to detect anomalous substituted words from a sequence and retrieve the original words from a set of candidates that contains a positive sample (an original word) and multiple hard negative samples (the synonyms of the original word) via contrastive learning. Thus, it is more suitable for fine-tuning a lexical substitution model in our task, because we also aim to detect appropriate substitution words from a set of candidates."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8776"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.358
+            ],
+            "angle": 0,
+            "content": "gets, we select the top \\(I\\) words in the input with the highest \\(a\\) produced by Equation 3, forming the set \\(w^{att}\\). We denote their corresponding representations as \\(V^{att} = \\{v_{1}^{att},\\dots ,v_{I}^{att}\\}\\). For each \\(v_{i}^{att}\\), we compute its average Euclidean distance to all of its candidates' hidden states. The candidates consist of WordNet synonyms of all of its possible senses under the same part-of-speech type, which are transformed into hidden states \\(G = \\{g_{1},\\ldots ,g_{M}\\}\\) using Equation 2. \\(M\\) represents the number of candidates provided by WordNet. The top \\(J\\) words in \\(w^{att}\\) with the largest corresponding average distance are considered to be the ones with the most diverse word meanings, and thus are more likely to require disambiguation. We will demonstrate the success of such an ambiguous word selection method in the later ablation study (see Section 6.1)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.359,
+                0.49,
+                0.394
+            ],
+            "angle": 0,
+            "content": "Hence, target words \\( t = (t_1, \\ldots, t_J) \\) are selected by finding each corresponding \\( v_j^{att} \\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.146,
+                0.401,
+                0.488,
+                0.437
+            ],
+            "angle": 0,
+            "content": "\\[\nv _ {j} ^ {\\text {a t t}} = \\arg \\max  _ {i} \\left(\\frac {1}{M} \\sum_ {M} d \\left(v _ {i} ^ {\\text {a t t}}, g _ {m}\\right)\\right). \\tag {10}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.446,
+                0.489,
+                0.51
+            ],
+            "angle": 0,
+            "content": "Following the lexical substitution benchmarks, only words whose part-of-speech tags fall under the general categories of noun, verb, adjective, and adverb are considered as targets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.52,
+                0.419,
+                0.536
+            ],
+            "angle": 0,
+            "content": "4.3 Dynamic Rewarding Mechanism"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.541,
+                0.489,
+                0.686
+            ],
+            "angle": 0,
+            "content": "To fine-tune the lexical substitution model on the downstream sentiment analysis task without knowing WSD gold standard truth, for each target word \\( t_j \\), we utilize the top \\( N \\) candidates \\( \\hat{c}_j = \\{\\hat{c}_{j,1},\\dots ,\\hat{c}_{j,n},\\dots ,\\hat{c}_{j,N}\\} \\) produced by the lexical substitution model. The new input resulting from \\( t_j \\) being replaced by candidate \\( \\hat{c}_{j,n} \\) is denoted as \\( w^{j,n} \\). Same with Equation 5, the probability distribution of a sentiment prediction from \\( w^{j,n} \\) is:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.149,
+                0.695,
+                0.488,
+                0.714
+            ],
+            "angle": 0,
+            "content": "\\[\nP (\\hat {y}) _ {j, n} = \\operatorname {s o f t m a x} \\left(F N N _ {2} \\left(h ^ {j, n}\\right)\\right), \\tag {11}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.722,
+                0.486,
+                0.754
+            ],
+            "angle": 0,
+            "content": "where \\( h^{j,n} \\) is the hidden states of \\( w^{j,n} \\) outputted by Equation 4 in the sentiment analysis model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.755,
+                0.489,
+                0.835
+            ],
+            "angle": 0,
+            "content": "To adjust the model in such a way that a more accurate sentiment prediction results in a higher reward for the corresponding substitution output, we compute the loss weight for \\(\\hat{c}_{j,n}\\) being the correct substitution prediction as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.217,
+                0.844,
+                0.488,
+                0.864
+            ],
+            "angle": 0,
+            "content": "\\[\n\\theta_ {j, n} = \\beta P (\\hat {y} = \\tilde {y}) _ {j, n} ^ {2}, \\tag {12}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.872,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "where \\(\\beta\\) is a hyperparameter for balancing the sentiment analysis and lexical substitution losses, \\(\\tilde{y}\\) is the ground truth sentiment label as defined above."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.546,
+                0.081,
+                0.848,
+                0.183
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Dataset</td><td>Split</td><td># Samples</td></tr><tr><td>SemEval 2007 Task 10</td><td>-</td><td>2,010</td></tr><tr><td>CoinCo</td><td>-</td><td>15,629</td></tr><tr><td></td><td>Train</td><td>8,530</td></tr><tr><td>Rotten Tomatoes</td><td>Dev</td><td>1,066</td></tr><tr><td></td><td>Test</td><td>1,066</td></tr><tr><td>Financial PhraseBank</td><td>-</td><td>4,846</td></tr><tr><td>STS-Gold</td><td>-</td><td>2,026</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.56,
+                0.192,
+                0.831,
+                0.206
+            ],
+            "angle": 0,
+            "content": "Table 2: Datasets used for experiments."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.231,
+                0.884,
+                0.316
+            ],
+            "angle": 0,
+            "content": "We formulate the input \\( s^j \\) to lexical substitution model with sentence \\( w \\) and candidates \\( \\hat{c}_j \\), using Formula 1. Similar to Equation 8, the loss \\( (\\mathcal{L}_{j,n}^{(ls)}) \\) of the lexical substitution model when \\( \\hat{c}_{j,n} \\) is considered as gold standard is computed as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.529,
+                0.326,
+                0.884,
+                0.363
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {j, n} ^ {(l s)} = - \\sum_ {i} \\log \\frac {\\exp \\left(d \\left(u _ {t} , \\hat {c} _ {j , n}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(d \\left(u _ {t} , \\hat {c} _ {j}\\right) / \\tau\\right)}. \\tag {13}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.528,
+                0.376,
+                0.805,
+                0.392
+            ],
+            "angle": 0,
+            "content": "Finally, the total loss is computed as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.527,
+                0.403,
+                0.884,
+                0.427
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} = \\mathcal {L} ^ {(s a)} + \\theta_ {1, 1} \\mathcal {L} _ {1, 1} ^ {(l s)} + \\dots + \\theta_ {J, N} \\mathcal {L} _ {J, N} ^ {(l s)}. \\tag {14}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.438,
+                0.656,
+                0.455
+            ],
+            "angle": 0,
+            "content": "5 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.464,
+                0.624,
+                0.478
+            ],
+            "angle": 0,
+            "content": "5.1 Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.484,
+                0.884,
+                0.613
+            ],
+            "angle": 0,
+            "content": "To fine-tune the pretrained lexical substitution model, we combined SemEval 2007 Task 10 (McCarthy and Navigli, 2007) and CoinCo (Kremer et al., 2014). The former consists of 2,010 example sentences for 201 polysemous words. The latter contains 2,474 sentences with multiple words to be substituted. We used \\(70\\%\\) of the concatenated dataset for training, \\(30\\%\\) for testing."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.614,
+                0.885,
+                0.838
+            ],
+            "angle": 0,
+            "content": "For sentiment analysis, we used three datasets from different domains: Rotten Tomatoes (Pang and Lee, 2005) consists of 5,331 positive and 5,331 negative sentences collected from Rotten Tomatoes movie reviews. Financial PhraseBank (Malo et al., 2014) contains 4,846 sentences from English financial news, each labeled as negative, positive, or neutral. As the dataset has no standard split, We adopted a split of \\(70\\%\\) and \\(30\\%\\) for the training and testing sets. STS-Gold (Saif et al., 2013) contains 2,026 tweets, annotated as negative or positive. As the dataset has no standard split, We adopted a split of \\(90\\%\\) and \\(10\\%\\) for the training and testing sets. Details of the datasets are listed in Table 2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.85,
+                0.63,
+                0.865
+            ],
+            "angle": 0,
+            "content": "5.2 Baselines"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.872,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Since we aim to improve a sentiment analysis classifier with explicit WSD in a neurosymbolic fashion, vanilla RoBERTa-large (Liu et al., 2019) is"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8777"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.295
+            ],
+            "angle": 0,
+            "content": "considered as the most direct baseline, because we also used it for sentiment classification. To put our model's performance into perspective, we further include published baselines with the highest reported accuracy on the experiment datasets, which consist of various pre-trained language models finetuned on the target dataset, i.e., FinBERT (Araci, 2019), DeBERTa (He et al., 2020; Sileo, 2023); and ones trained with additional resources, namely, DeBERTa+tarksource (Sileo, 2023) which is pretrained on a unified dataset for 500 English tasks, and SenticNet 7 (Cambria et al., 2022) which is a XLNet enhanced by sentiment knowledge."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.307,
+                0.214,
+                0.323
+            ],
+            "angle": 0,
+            "content": "5.3 Setups"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.33,
+                0.49,
+                0.458
+            ],
+            "angle": 0,
+            "content": "To obtain the lexical substitution model, we fine-tune ALM with the lexical substitution datasets for 20 epochs with a batch size of 20 and a learning rate of 1e-8. The pre-trained model we used as the encoder for sentiment analysis is RoBERTa-large. The hidden size for both models is set to 1024. We use the NLTK part-of-speech tagger (Bird, 2006) to label the sentiment analysis input."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.459,
+                0.49,
+                0.603
+            ],
+            "angle": 0,
+            "content": "To train our proposed framework, we adopt \\( I = 5 \\), \\( J = 2 \\), \\( K = 15 \\), \\( N = 3 \\), \\( \\tau = 0.05 \\) and \\( \\beta = 0.0001 \\). We set the learning rate to 1-e6. The models are trained for 50 epochs with a batch size of 10, using early stopping. We used Adam (Kingma and Ba, 2014) as the optimizer. Experiments are run on RTX A6000 GPU. Following the setups in the baselines, we adopt accuracy as an evaluation metric."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.618,
+                0.214,
+                0.634
+            ],
+            "angle": 0,
+            "content": "6 Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.645,
+                0.49,
+                0.902
+            ],
+            "angle": 0,
+            "content": "We compare our model with the baselines in Table 3. Our model outperforms existing models with the best-published results, achieving state-of-the-art performance on the three datasets. Compared to the RoBERTa baseline, our model is able to improve the accuracy of sentiment analysis on corpora from different domains to varying extent. We hypothesize that the extent of improvement is dependent on domain types. For instance, Twitter language is informal and brief, with frequent use of short homographs. Thus, the integrated disambiguation model is able to obtain the most significant increase of \\(1.91\\%\\) accuracy on STS-Gold. In the financial domain, metaphors and less common word senses are commonly used, accounting for our \\(1.23\\%\\) gain on Financial PhraseBank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.132,
+                0.904,
+                0.486,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Movie reviews, on the other hand, often use"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.518,
+                0.081,
+                0.878,
+                0.171
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>RT</td><td>FPB</td><td>STS-Gold</td><td>Avg</td></tr><tr><td>DeBERTa</td><td>90.42</td><td>84.48</td><td>-</td><td>87.45</td></tr><tr><td>+tasksource</td><td>90.99</td><td>85.20</td><td>-</td><td>88.10</td></tr><tr><td>FinBERT</td><td>-</td><td>86.00</td><td>-</td><td>-</td></tr><tr><td>SenticNet 7</td><td>-</td><td>-</td><td>90.08</td><td>-</td></tr><tr><td>RoBERTa</td><td>91.31</td><td>87.09</td><td>97.14</td><td>91.85</td></tr><tr><td>Ours</td><td>91.87*</td><td>88.32*</td><td>99.05*</td><td>93.08</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.179,
+                0.886,
+                0.28
+            ],
+            "angle": 0,
+            "content": "Table 3: Comparisons with baselines. RT stands for Rotten Tomatoes. FPB stands for Financial PhraseBank. Bold font denotes the best results. The RoBERTa baseline was implemented by us. * denotes the improvement is statistically significant (\\( p < 0.01 \\) on a two-tailed t-test), against the highest baseline score. Our results are averaged over 5 runs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.306,
+                0.887,
+                0.563
+            ],
+            "angle": 0,
+            "content": "standard expressions that are semantically clearer. Hence, our model obtains a marginal increase of \\(0.56\\%\\) on Rotten Tomatoes. Despite the improvement being marginal on average, i.e., \\(1.23\\%\\), we would like to highlight that this improvement is significant considering the context of error propagation from the lexical substitution model. Our framework consistently demonstrates improvements across all datasets, substantiating its ability to achieve higher gains in accuracy while minimizing any potential loss. Furthermore, we can interpret the WSD outcomes and the sentiment analysis decision-making process before and after lexical substitutions from the proposed neurosymbolic framework. We will demonstrate the interpretability of the framework in Section 6.2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.592,
+                0.674,
+                0.608
+            ],
+            "angle": 0,
+            "content": "6.1 Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.613,
+                0.885,
+                0.901
+            ],
+            "angle": 0,
+            "content": "To prove the effectiveness of each component in our proposed framework, we compare with two variations: w/o WSD, where sentiment prediction is done by the sentiment analysis model without any modifications to the input; w/o DR, where the lexical substitution model is not fine-tuned during sentiment analysis training. To prove our hypothesis in WSD target word selection, we compare the proposed framework with two variations: Top-2, where the top 2 words in the input with the highest attention weights are automatically chosen as target words to be disambiguated; Rand, where 2 out of the most attended 5 words are randomly selected to be the target words. The comparison of performance is shown in Table 4. It can be observed that our full model outperforms w/o WSD by \\(0.28\\%\\) on Rotten Tomatoes, \\(1.03\\%\\) on Financial PhraseBank, and \\(1.43\\%\\) on STS-Gold."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.527,
+                0.904,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The extent of gains in different domains is in"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8778"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.137,
+                0.081,
+                0.468,
+                0.158
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>RT</td><td>FPB</td><td>STS-Gold</td><td>Avg</td></tr><tr><td>w/o WSD</td><td>91.59</td><td>87.29</td><td>97.62</td><td>92.17</td></tr><tr><td>w/o DR</td><td>91.40</td><td>87.27</td><td>96.67</td><td>91.78</td></tr><tr><td>Top-2</td><td>91.60</td><td>87.81</td><td>97.62</td><td>92.34</td></tr><tr><td>Rand</td><td>89.72</td><td>86.30</td><td>97.14</td><td>91.05</td></tr><tr><td>Ours</td><td>91.87</td><td>88.32</td><td>99.05</td><td>93.08</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.192,
+                0.168,
+                0.409,
+                0.182
+            ],
+            "angle": 0,
+            "content": "Table 4: Ablation study results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.208,
+                0.49,
+                0.432
+            ],
+            "angle": 0,
+            "content": "line with our hypothesis in the previous section. It shows that our disambiguation framework is indeed able to improve the accuracy of sentiment analysis by selecting the appropriate ambiguous words and paraphrasing them into semantically clearer ones. Our model also achieves higher accuracy than w/o DR on all three datasets, proving the validity and effectiveness of our dynamic rewarding mechanism. Noticeably, without dynamic rewarding and joint fine-tuning, the performance of w/o DR is weaker than w/o WSD, showing that dynamic rewarding-based fine-tuning with data from another task can effectively alleviate error propagation from lexical substitution."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.434,
+                0.49,
+                0.611
+            ],
+            "angle": 0,
+            "content": "Our proposed framework outperforms the Top-2 selection method on all datasets. It proves our assumption that words that are highly diverse in their senses are more likely to cause ambiguity for sentiment prediction. Rand performs the worst on all three datasets compared to w/o WSD, Top-2, and our model, indicating i) the necessity of selecting the appropriate target words to disambiguate so as not to introduce noise, and ii) the validity of choosing target words based on both their importance to the decision and their sense diversity level."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.624,
+                0.406,
+                0.64
+            ],
+            "angle": 0,
+            "content": "6.2 Interpretability Demonstration"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.645,
+                0.489,
+                0.724
+            ],
+            "angle": 0,
+            "content": "Since an interpretable encoder (HAN) is used in our framework to learn sequential representations, we can understand which words contribute the most to a sentiment analysis prediction by visualizing the attention weights of HAN."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.726,
+                0.49,
+                0.902
+            ],
+            "angle": 0,
+            "content": "As shown in Figure 2, we conduct a study on the change of attention weights before and after paraphrasing. In Figure 2 (a), the words that receive the most attention in the original sentence are not, \\( a \\), sore, etc., among which not has significantly higher weight. This indicates that the model is mostly relying on negation to make inference. As not does not have diverse meaning and \\( a \\) is an article, sore is selected as target word and replaced with painful. Our attention module assigns a much higher weight to the negative adjective after paraphrasing."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.132,
+                0.904,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "This shows that our model is able to find suitable"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.518,
+                0.085,
+                0.883,
+                0.263
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.277,
+                0.886,
+                0.321
+            ],
+            "angle": 0,
+            "content": "Figure 2: The most attended word demonstration before and after paraphrasing in sentiment analysis predictions. The lighter the color, the higher the attention weight."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.353,
+                0.886,
+                0.594
+            ],
+            "angle": 0,
+            "content": "word substitutions that aid the correct sentiment predictions. From Figure 2 (b), we can observe a rather evenly distributed attention weights in the original sentence, meaning that the attention module cannot identify the words that significantly affect sentiment. After our framework replaces the word decreased to fell, the model is able to assign higher weights to the sentiment-loaded phrase profit fell. From Figure 2 (c), we can see that after replacing the word mediocre with average, the sentiment related words including the paraphrased word and unaware all received higher attention. By attending to the straightforward paraphrases expressing opinions, the sentiment analysis model can yield correct predictions for all three cases."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.618,
+                0.645,
+                0.634
+            ],
+            "angle": 0,
+            "content": "6.3 Case Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.646,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In Table 5, we study the ability of our framework to identify and paraphrase ambiguous words into unequivocal ones. In the first instance, albeit not affecting the prediction, the word delicious is disambiguated by our framework into its figurative meaning delighted instead of its literal one. In the second instance, the word succeeds has a more positive word sense equating to win, and a more neutral one equating to follow. Our framework is able to paraphrase it into the latter. In the last instance, the word visibility falls under the word sense of degree of exposure to public notice in this context. Our framework correctly paraphrases it into profile, which is the synonym under the same sense in WordNet. It can be concluded that our framework can disambiguate word senses in a transparent and interpretable way for sentiment analysis."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8779"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.125,
+                0.082,
+                0.481,
+                0.299
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Setup</td><td>Content</td></tr><tr><td>Original</td><td>a delicious and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them.</td></tr><tr><td>Paraphrased</td><td>a delighted and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them.</td></tr><tr><td>Original</td><td>Le Lay succeeds Walter G++nter and will be based in Finland.</td></tr><tr><td>Paraphrased</td><td>Le Lay follows Walter G++nter and will be based in Finland.</td></tr><tr><td>Original</td><td>Strong brand visibility nationally and regionally is of primary importance in home sales, vehicle and consumer advertising.</td></tr><tr><td>Paraphrased</td><td>Strong brand profile nationally and regionally is of primary importance in home sales, vehicle and consumer advertising.</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.115,
+                0.309,
+                0.486,
+                0.325
+            ],
+            "angle": 0,
+            "content": "Table 5: Case study of disambiguation as paraphrasing."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.352,
+                0.279,
+                0.368
+            ],
+            "angle": 0,
+            "content": "6.4 Error Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.376,
+                0.49,
+                0.713
+            ],
+            "angle": 0,
+            "content": "Finally, we perform a brief error analysis (Table 6). In the first sentence, the model wrongly paraphrases force into power, because it does not recognize the multi-word idiom tour de force. It shows the model's limitation in understanding the semantic meaning of infrequent multi-word idiom. In the second sentence, the word just is wrongly paraphrased as good, which negatively influenced the sentiment prediction. This is mainly due to the part-of-speech tagger wrongly classifying the word just into a adjective, instead of an adverb. Thus, the lexical substitution module might be subjected to error propagation of the part-of-speech tagger and introduce noise into the input. In the last sentence, the word skills is paraphrased as science. It is likely that the model mistakes the word major as a field of study instead of a student studying a field. Thus, the substitution makes sense in the local context the science of calculus major, but not in the whole context, meaning that the model fails to understand the semantic meaning of the sentence."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.73,
+                0.248,
+                0.745
+            ],
+            "angle": 0,
+            "content": "7 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this work, we propose a neurosymbolic framework for sentiment analysis that incorporates the disambiguation of word senses by identifying and paraphrasing ambiguous words in the input, which is realized by leveraging a pre-trained lexical substitution model. The proposed framework not only improves the accuracy of sentiment analysis, but also achieves interpretability by enabling us to trace back which words are paraphrased and what words they are replaced with."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.518,
+                0.082,
+                0.878,
+                0.299
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Setup</td><td>Content</td><td>Label</td></tr><tr><td>Original</td><td>a tour de force of modern cinema.</td><td>Positive</td></tr><tr><td>Paraphrased</td><td>a tour de power of modern cinema.</td><td>Positive</td></tr><tr><td>Original</td><td>Is just at work for the day -sigh- I have a headache</td><td>Negative</td></tr><tr><td>Paraphrased</td><td>Is good at work for the day -sigh- I have a headache</td><td>Positive</td></tr><tr><td>Original</td><td>the skills of a calculus major at m.i.t. are required to balance all the formulaic equations in the long-winded heist comedy who is cletis tout?</td><td>Negative</td></tr><tr><td>Paraphrased</td><td>the science of a calculus major at m.i.t. are required to balance all the formulaic equations in the long-winded heist comedy who is cletis tout?</td><td>Positive</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.309,
+                0.883,
+                0.339
+            ],
+            "angle": 0,
+            "content": "Table 6: Error analysis. Red indicates errors made by our proposed framework."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.365,
+                0.885,
+                0.527
+            ],
+            "angle": 0,
+            "content": "Additionally, to fine-tune the lexical substitution model to provide better disambiguating substitutes for sentiment analysis with no ground-truth word sense labels, we adopt the dynamic rewarding mechanism to jointly train the sentiment analysis and lexical substitution models. Experiments show that our framework is indeed effective in improving the performance of sentiment analysis, obtaining state-of-the-art performance on three corpora from different domains."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.557,
+                0.615,
+                0.571
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.584,
+                0.884,
+                0.81
+            ],
+            "angle": 0,
+            "content": "The framework proposed in this paper have to be seen in light of some limitations. First, the performance of sentiment analysis to some extent depends on a decent-performing pre-trained lexical substitution model. Currently, polysemous samples in lexical substitution data are not as abundant as in WSD data, which might limit its ability in disambiguating less common polysemies. Second, the effectiveness of our proposed model is only tested on English corpora, as it relies on the sense knowledge from WordNet. The disambiguation target selection and paraphrasing methods might work on languages with more complex segmentation or limited sense lexicon."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.854,
+                0.61,
+                0.87
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.878,
+                0.884,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Abien Fred Agarap. 2018. Deep learning using rectified linear units (ReLU). arXiv preprint arXiv:1803.08375."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8780"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.125
+            ],
+            "angle": 0,
+            "content": "Dogu Araci. 2019. FinBERT: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.138,
+                0.488,
+                0.204
+            ],
+            "angle": 0,
+            "content": "Philip Arthur, Graham Neubig, Sakriani Sakti, Tomoki Toda, and Satoshi Nakamura. 2015. Semantic parsing of ambiguous input through paraphrasing and verification. Transactions of the Association for Computational Linguistics, 3:571-584."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.216,
+                0.488,
+                0.294
+            ],
+            "angle": 0,
+            "content": "Stefano Baccianella, Andrea Esuli, and Fabrizio Sebastiani. 2010. SentiWordNet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC'10)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.307,
+                0.488,
+                0.399
+            ],
+            "angle": 0,
+            "content": "Edoardo Barba, Tommaso Pasini, and Roberto Navigli. 2021a. ESC: Redesigning WSD with Extractive Sense Comprehension. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4661-4672, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.411,
+                0.488,
+                0.502
+            ],
+            "angle": 0,
+            "content": "Edoardo Barba, Luigi Procopio, and Roberto Navigli. 2021b. ConSeC: Word sense disambiguation as continuous sense comprehension. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1492-1503, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.515,
+                0.488,
+                0.594
+            ],
+            "angle": 0,
+            "content": "Michele Bevilacqua and Roberto Navigli. 2020. Breaking through the \\(80\\%\\) glass ceiling: Raising the state of the art in word sense disambiguation by incorporating knowledge graph information. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2854-2864."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.606,
+                0.488,
+                0.646
+            ],
+            "angle": 0,
+            "content": "Steven Bird. 2006. NLTK: the natural language toolkit. In Proceedings of the COLING/ACL 2006 Interactive Presentation Sessions, pages 69-72."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.658,
+                0.488,
+                0.709
+            ],
+            "angle": 0,
+            "content": "Terra Blevins and Luke Zettlemoyer. 2020. Moving down the long tail of word sense disambiguation with gloss-informed biencoders. arXiv preprint arXiv:2005.02590."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.723,
+                0.488,
+                0.802
+            ],
+            "angle": 0,
+            "content": "Erik Cambria, Qian Liu, Sergio Decherchi, Frank Xing, and Kenneth Kwok. 2022. SenticNet 7: A commonsense-based neurosymbolic AI framework for explainable sentiment analysis. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 3829-3839."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.814,
+                0.488,
+                0.853
+            ],
+            "angle": 0,
+            "content": "Erik Cambria, Rui Mao, Melvin Chen, Zhaoxia Wang, and Seng-Beng Ho. 2023. Seven pillars for the future of AI. IEEE Intelligent Systems, 38(6)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.866,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Erik Cambria, Soujanya Poria, Federica Bisio, Rajiv Bajpai, and Iti Chaturvedi. 2015. The CLSA model: A novel framework for concept-level sentiment analysis. In LNCS, volume 9042, pages 3-22. Springer."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Erik Cambria, Soujanya Poria, Alexander Gelbukh, and Mike Thelwall. 2017. Sentiment analysis is a big suitcase. IEEE Intelligent Systems, 32(6):74-80."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.139,
+                0.882,
+                0.231
+            ],
+            "angle": 0,
+            "content": "Niccolò Campolungo, Tommaso Pasini, Denis Emelin, and Roberto Navigli. 2022. Reducing disambiguation biases in NMT by leveraging explicit word sense information. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4824-4838."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.244,
+                0.882,
+                0.31
+            ],
+            "angle": 0,
+            "content": "Yanfang Cao, Pu Zhang, and Anping Xiong. 2015. Sentiment analysis based on expanded aspect and polarity-ambiguous word lexicon. International Journal of Advanced Computer Science and Applications, 6(2)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.324,
+                0.882,
+                0.39
+            ],
+            "angle": 0,
+            "content": "Ann Devitt and Khurshid Ahmad. 2007. Sentiment polarity identification in financial news: A cohesion-based approach. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 984-991."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.403,
+                0.882,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Philip Edmonds and Adam Kilgarriff. 2002. Introduction to the special issue on evaluating word sense disambiguation systems. Natural Language Engineering, 8(4):279-291."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.469,
+                0.882,
+                0.56
+            ],
+            "angle": 0,
+            "content": "Umar Farooq, Tej Prasad Dhamala, Antoine Nongaillard, Yacine Ouzrout, and Muhammad Abdul Qadir. 2015. A word sense disambiguation method for feature level sentiment analysis. In 2015 9th International Conference on Software, Knowledge, Information Management and Applications (SKIMA), pages 1-8. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.574,
+                0.882,
+                0.602
+            ],
+            "angle": 0,
+            "content": "Christiane Fellbaum. 1998. WordNet: An electronic lexical database. MIT press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.615,
+                0.882,
+                0.679
+            ],
+            "angle": 0,
+            "content": "Mengshi Ge, Rui Mao, and Erik Cambria. 2022. Explanable metaphor identification inspired by conceptual metaphor theory. Proceedings of the AAAI Conference on Artificial Intelligence, 36(10):10681-10689."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.694,
+                0.882,
+                0.773
+            ],
+            "angle": 0,
+            "content": "Sooji Han, Rui Mao, and Erik Cambria. 2022. Hierarchical attention network for explainable depression detection on Twitter aided by metaphor concept mappings. In Proceedings of the 29th International Conference on Computational Linguistics (COLING), pages 94-104."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.786,
+                0.882,
+                0.853
+            ],
+            "angle": 0,
+            "content": "Viktor Hangya, Qianchu Liu, Dario Stojanovski, Alexander Fraser, and Anna Korhonen. 2021. Improving machine translation of rare and unseen word senses. In Proceedings of the Sixth Conference on Machine Translation, pages 614-624."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.866,
+                0.882,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. DeBERTa: Decoding-enhanced BERT with disentangled attention. arXiv preprint arXiv:2006.03654."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8781"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.138
+            ],
+            "angle": 0,
+            "content": "Luyao Huang, Chi Sun, Xipeng Qiu, and Xuanjing Huang. 2019. GlossBERT: BERT for word sense disambiguation with gloss knowledge. arXiv preprint arXiv:1908.07245."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.15,
+                0.487,
+                0.202
+            ],
+            "angle": 0,
+            "content": "Chihli Hung and Shiuman-Jeng Chen. 2016. Word sense disambiguation based sentiment lexicons for sentiment classification. Knowledge-Based Systems, 110:224-232."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.214,
+                0.487,
+                0.265
+            ],
+            "angle": 0,
+            "content": "Chihli Hung and Hao-Kai Lin. 2013. Using objective words in SentiWordNet to improve word-of-mouth sentiment classification. IEEE Intelligent Systems, 28(02):47-54."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.278,
+                0.486,
+                0.316
+            ],
+            "angle": 0,
+            "content": "Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.328,
+                0.487,
+                0.407
+            ],
+            "angle": 0,
+            "content": "Gerhard Kremer, Katrin Erk, Sebastian Padó, and Stefan Thater. 2014. What substitutes tell us-analysis of an \"all-words\" lexical substitution corpus. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 540-549."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.418,
+                0.487,
+                0.497
+            ],
+            "angle": 0,
+            "content": "Frederick Liu, Han Lu, and Graham Neubig. 2018. Handling homographs in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1336-1345."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.508,
+                0.487,
+                0.573
+            ],
+            "angle": 0,
+            "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.585,
+                0.487,
+                0.649
+            ],
+            "angle": 0,
+            "content": "P. Malo, A. Sinha, P. Korhonen, J. Wallenius, and P. Takala. 2014. Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology, 65."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.661,
+                0.487,
+                0.714
+            ],
+            "angle": 0,
+            "content": "Rui Mao, Kai He, Claudia Beth Ong, Qian Liu, and Erik Cambria. 2023a. MetaPro 2.0: Computational metaphor processing on the effectiveness of anomalous language modeling."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.725,
+                0.487,
+                0.79
+            ],
+            "angle": 0,
+            "content": "Rui Mao and Xiao Li. 2021. Bridging towers of multitask learning with a gating mechanism for aspect-based sentiment analysis and sequential metaphor identification. Proceedings of the AAAI Conference on Artificial Intelligence, 35(15):13534-13542."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.802,
+                0.487,
+                0.881
+            ],
+            "angle": 0,
+            "content": "Rui Mao, Xiao Li, Kai He, Mengshi Ge, and Erik Cambria. 2023b. MetaPro Online: A computational metaphor processing online system. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 127-135."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.892,
+                0.487,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Rui Mao, Qian Liu, Kai He, Wei Li, and Erik Cambria. 2023c. The biases of pre-trained language models:"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.126
+            ],
+            "angle": 0,
+            "content": "An empirical study on prompt-based sentiment analysis and emotion detection. IEEE Transactions on Affective Computing, 14(3):1743-1753."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.136,
+                0.882,
+                0.213
+            ],
+            "angle": 0,
+            "content": "Marco Maru, Simone Conia, Michele Bevilacqua, and Roberto Navigli. 2022. Nibbling at the hard core of Word Sense Disambiguation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4724-4737."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.224,
+                0.882,
+                0.278
+            ],
+            "angle": 0,
+            "content": "Diana McCarthy and Roberto Navigli. 2007. SemEval-2007 task 10: English lexical substitution task. In Proceedings of the 4th International Workshop on Semantic Evaluations, SemEval '07, pages 48-53."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.287,
+                0.882,
+                0.366
+            ],
+            "angle": 0,
+            "content": "Arman Khadjeh Nassirtoussi, Saeed Aghabozorgi, Teh Ying Wah, and David Chek Ling Ngo. 2015. Text mining of news-headlines for forex market prediction: A multi-layer dimension reduction algorithm with semantics and sentiment. Expert Systems with Applications, 42(1):306-324."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.376,
+                0.882,
+                0.416
+            ],
+            "angle": 0,
+            "content": "Bruno Ohana and Brendan Tierney. 2009. Sentiment classification of reviews using SentiWordNet. Proceedings of IT&T, 8."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.426,
+                0.882,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.476,
+                0.882,
+                0.528
+            ],
+            "angle": 0,
+            "content": "Martha Palmer, Hoa Trang Dang, and Christiane Fellbaum. 2007. Making fine-grained and coarse-grained sense distinctions, both manually and automatically. Natural Language Engineering, 13(2):137-163."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.538,
+                0.882,
+                0.579
+            ],
+            "angle": 0,
+            "content": "Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the ACL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.588,
+                0.882,
+                0.654
+            ],
+            "angle": 0,
+            "content": "Xiao Pu, Nikolaos Pappas, James Henderson, and Andrei Popescu-Belis. 2018. Integrating weakly supervised word sense disambiguation into neural machine translation. Transactions of the Association for Computational Linguistics, 6:635-649."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.664,
+                0.882,
+                0.704
+            ],
+            "angle": 0,
+            "content": "Rachel Elizabeth Ramsey. 2017. An exemplar-theoretic account of word senses. University of Northumbria at Newcastle (United Kingdom)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.714,
+                0.882,
+                0.78
+            ],
+            "angle": 0,
+            "content": "Hassan Saif, Miriam Fernandez, Yulan He, and Harith Alani. 2013. Evaluation datasets for Twitter sentiment analysis: A survey and a new dataset, the STS-Gold. Emotion and Sentiment in Social and Expressive Media."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.79,
+                0.882,
+                0.842
+            ],
+            "angle": 0,
+            "content": "Damien Sileo. 2023. tasksource: A dataset harmonization framework for streamlined NLP multitask learning and evaluation. arXiv preprint arXiv:2301.05948v3."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.852,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Yang Song, Xin Cai Ong, Hwee Tou Ng, and Qian Lin. 2021. Improved word sense disambiguation with enhanced sense representations. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 4311-4320."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8782"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.139
+            ],
+            "angle": 0,
+            "content": "Yunqing Xia, Erik Cambria, Amir Hussain, and Huan Zhao. 2015. Word polarity disambiguation using bayesian model and opinion-level features. Cognitive Computation, 7(3):369-380."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.149,
+                0.487,
+                0.201
+            ],
+            "angle": 0,
+            "content": "Yunqing Xia, Huan Zhao, Erik Cambria, and Amir Hussain. 2014. Disambiguating word sentiment polarity through bayesian modeling and opinion-level features. In Proceedings of ICML Workshops."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.201
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8783"
+        }
+    ]
+]

2023/Neuro-Symbolic Sentiment Analysis with Dynamic Word Sense Disambiguation/f2ea77f7-c2a2-49e4-b922-17fa39103dc4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b474ce50afe94f2ff39ac38e0cf34da8200813b9d9a074696a609d090adef611
+size 424787

2023/Neuro-Symbolic Sentiment Analysis with Dynamic Word Sense Disambiguation/full.md

@@ -0,0 +1,348 @@
+# Neurosymbolic Sentiment Analysis with Dynamic Word Sense Disambiguation
+
+Xulang Zhang\*, Rui Mao\*, Kai He\* and Erik Cambria\*
+
+$\spadesuit$ Nanyang Technological University, Singapore
+
+\*National University of Singapore, Singapore
+
+xulang001@e.ntu.edu.sg; {rui.mao, cambria}@ntu.edu.sg; kai_he@nus.edu.sg
+
+# Abstract
+
+Sentiment analysis is a task that highly depends on the understanding of word senses. Traditional neural network models are black boxes that represent word senses as vectors that are uninterpretable for humans. On the other hand, the application of Word Sense Disambiguation (WSD) systems in downstream tasks poses challenges regarding i) which words need to be disambiguated, and ii) how to model explicit word senses into easily understandable terms for a downstream model. This work proposes a neurosymbolic framework that incorporates WSD by identifying and paraphrasing ambiguous words to improve the accuracy of sentiment predictions. The framework allows us to understand which words are paraphrased into which semantically unequivocal words, thus enabling a downstream task model to gain both accuracy and interpretability. To better fine-tune a lexical substitution model for WSD on a downstream task without ground-truth word sense labels, we leverage dynamic rewarding to jointly train sentiment analysis and lexical substitution models. Our framework proves to effectively improve the performance of sentiment analysis on corpora from different domains.
+
+# 1 Introduction
+
+To improve the accuracy of sentiment analysis, it is crucial to disambiguate polysemous words according to the context (Cambria et al., 2023). A word may have senses that carry opposite sentiments, e.g., "fine" has the positive meaning of being delicate and refined, and the negative meaning of monetary penalty. Additionally, words may appear in positive, negative, or neutral senses in different domains. For instance, the word "bearish" is often used in a negative sense to describe the stock market in the financial domain, while more likely to be used neutrally to describe hulking figures in other domains.
+
+In traditional black-box sentiment analysis, the representations of word meanings are generated by neural networks (Mao and Li, 2021). As such, the disambiguation process is implicit and opaque. It is challenging to investigate what implicit senses are assigned to ambiguous words by a neural network model for it to make accurate predictions. Coupling symbolism and sub-symbolism, a neurosymbolic approach to sentiment analysis is transparent and interpretable, regarding the extraction and distinction of word meaning as a sub-problem (Cambria et al., 2017). However, research that explicitly deals with ambiguous words in sentiment analysis is limited (Xia et al., 2014, 2015; Cambria et al., 2015). Sense-annotated resources such as SentiWordNet (Baccianella et al., 2010) have been utilized, but only to detect and score sentimental words in the input (Hung and Lin, 2013; Nassirtoussi et al., 2015; Hung and Chen, 2016). The disambiguation of polysemous words in sentiment analysis inputs remains under-explored. Thus, we are motivated to develop a neurosymbolic framework for sentiment analysis with explicit Word Sense Disambiguation (WSD).
+
+Traditional WSD aims to classify which sense of the target word is appropriate according to its context (Huang et al., 2019; Blevins and Zettlemoyer, 2020; Bevilacqua and Navigli, 2020; Song et al., 2021; Barba et al., 2021a,b). However, the limitations  $(\ell)$  of their task setups are that  $(\ell 1)$  the output of those WSD systems, whether sense label (Bevilacqua and Navigli, 2020) or gloss definition (Blevins and Zettlemoyer, 2020; Barba et al., 2021a), is difficult to integrate into downstream input;  $(\ell 2)$  the target words must be pre-defined for WSD systems to disambiguate, making it difficult to apply them directly to downstream tasks; and  $(\ell 3)$  the reliability of sense distinction of the benchmark has been questioned (Ramsey, 2017; Maru et al., 2022), as word senses are not discrete.
+
+Attempts have been made to incorporate the disambiguation of word senses into downstream applications. Nonetheless, they are either impractical in real-world settings (Farooq et al., 2015; Pu et al., 2018), or only applicable to a specific downstream task, i.e., machine translation (Hangya et al., 2021; Campolungo et al., 2022). To address these issues, we leverage dynamic WSD in our neurosymbolic framework. We regard WSD as a lexical substitution task (for  $\ell 1$ ), where ambiguous words in sentiment analysis input are paraphrased into ones that cause less sentiment confusion using a lexical substitution model with dynamic WSD.
+
+Our dynamic WSD method comprises two distinct aspects: the dynamic selection of the target word for disambiguation, based on downstream contextual information (for  $\ell 2$ ); and the dynamic learning of the most suitable lexical substitutions, based on the prediction accuracy of the downstream model (for  $\ell 3$ ). To identify target words that need to be substituted, we first adopt an attention-based explainable encoder to rank input words by the level of influence they have on the output, motivated by the assumption that not all polysemous words affect sentiment predictions (Cao et al., 2015). Then, the top-ranking words with the most diverse senses are selected to be targets, based on our hypothesis that an ambiguous word may coincide with significant disparities in the semantic space between the word itself and its potential substitution candidates.
+
+As such, we can trace back which words are selected as targets, and which substitute words are used to disambiguate them, making our method interpretable and transparent. Furthermore, we fine-tune the lexical substitution model whilst training the sentiment analysis model by adopting a dynamic rewarding mechanism that does not require gold standard substitution labels. As such, the substitution model can be adjusted to provide more appropriate output for the downstream task as well as the target domain. Our model achieves more accurate sentiment analysis with explicit lexical-substitution-based WSD, despite not specifically trained on WSD and sentiment parallel datasets. Since the disambiguated word senses are represented as the lexical substitutions of the original words and an explainable encoder is used for retrieving the most contributing words, the neurosymbolic sentiment inference process is more interpretable for humans compared to black-box models.
+
+Our model is examined with three publicly available datasets from movie review, finance, and Twitter domains, outperforming strong baselines by  $1.23\%$  on average. The contributions of our work are twofold: (1) We propose a neurosymbolic sentiment analysis framework that leverages WSD as an auxiliary lexical substitution task, achieving state-of-the-art performance on our examined sentiment analysis datasets from different domains; (2) We design a transparent mechanism to select WSD target words, based on a) the importance of words to the learning and inference of a downstream task, and b) word sense diversity.
+
+# 2 Related Work
+
+Word Sense Disambiguation is a task that assigns the appropriate meaning to an ambiguous word according to its context. Recent WSD systems attained F1 scores higher than  $80\%$  on the benchmark datasets (Song et al., 2021; Bevilacqua and Navigli, 2020; Barba et al., 2021a,b), which is regarded as the estimated human performance since it is the highest inter-annotator agreement (Edmonds and Kilgarriff, 2002; Palmer et al., 2007). However, research on the application of WSD systems on downstream tasks is limited. Pu et al. (2018) utilized clustering algorithms to obtain sense embeddings by inferring from the similarity of the context vectors of a target word. The heavy requirement for disambiguating every homographs prior to downstream applications makes their method less feasible in many real-world situations. Campolungo et al. (2022) achieved the disambiguation of word sense in neural machine translation by leveraging a multilingual WSD system to establish a sense-tagged parallel corpus. Nonetheless, their approach relied on cross-lingual word alignments for reliable annotations, meaning that it cannot be adopted by other types of downstream tasks.
+
+Sentiment Analysis can be generally defined as the task that determines if a piece of text expresses positive, negative, or neutral opinion (Mao et al., 2023c). Word sense information has been applied in sentiment analysis in the form of sentiment-annotated lexicons (Devitt and Ahmad, 2007; Ohana and Tierney, 2009; Hung and Lin, 2013; Nassirtoussi et al., 2015), such as SentiWordNet. However, these methods incorporate the sentiment information of a word without disambiguating its word sense according to the context, but resorting to the emotional score of the most frequent
+
+![](images/8339731eaaa1ea3228dd3efe4658383c01c8f2bb2183ac4ccba471934e643f1c.jpg)  
+Figure 1: Architecture of the proposed disambiguation framework for sentiment analysis. A dotted line means backpropagation. Green block means that the original word is replaced by a word provided by the lexical substitution model. The darker the green, the higher probability it is assigned to be the model. Orange block indicates the words with top ranking attention scores produced by HAN.
+
+sense, or the average score of all senses. Additionally, the main objective of this approach is to identify sentimental words in a sentence, instead of disambiguating polysemous words that might affect model predictions. Hung and Chen (2016); Farooq et al. (2015); Liu et al. (2018) attempted to disambiguate according to target domain and context, but either costed extensive human labour, or fail to incorporate word senses explicitly and transparently.
+
+To address these limitations, our neurosymbolic framework aims to identify ambiguous target words in the context and replace them with clarifying ones automatically, making them interpretable. We also leverage a dynamic rewarding mechanism so that the substitutions can be fine-tuned to better serve the downstream task and the text domain.
+
+# 3 Task Definition
+
+In this work, we regard word sense ambiguity as the potential cause of errors in a downstream task, e.g., sentiment analysis. Inspired by the work of Mao et al. (2023b), we reframe WSD as a lexical substitution task, aiming to replace an ambiguous word with a synonym that not only represents the true sense of the original word in the context (Arthur et al., 2015), but also reduces misclassifications in the downstream task. Given an input sentence  $w = (w_{1},w_{2},\dots ,w_{L})$ , we first aim to identify the target word  $t$  that is likely troublesome for
+
+the downstream prediction from  $w$ . To resolve the potential ambiguity, we extract  $M$  of its synonyms from WordNet (Fellbaum, 1998) as candidates  $c = (c_{1},\ldots ,c_{M})$ , forming the input  $(s)$  as
+
+$$
+s = \langle s \rangle , w _ {1}, w _ {2}, \dots , t, \dots , w _ {L}, \langle / s \rangle , \tag {1}
+$$
+
+$< \mathrm{s}>$  and  $</ \mathrm{s}>$  are special tokens that were defined by our employed pre-trained language model. The lexical substitution model should identify the best candidate from  $c$  as the substitute  $\hat{c}$  which retains the original meaning of  $t$  in the context by using contrast learning. Thus,  $\hat{c}$  is the symbolic representation of  $t$  in context  $w$ . Then, the paraphrased input  $w^{f} = (w_{1},\dots,\hat{c},\dots,w_{L})$  is fed into a neural network classifier to predict a sentiment label. The objective is that the paraphrased input  $w^{f}$  increases the probability of correct sentiment prediction.
+
+# 4 Methodology
+
+Our proposed neurosymbolic framework can be viewed in Figure 1. There are two models in our framework, namely sentiment analysis and lexical substitution models. The sentiment analysis model consists of an encoder, an interpretable attention module, and multiple feedforward neural networks. First, through Figure 1 (1) and (2), the original input is fed into the encoder and the interpretable attention module to obtain the top  $I$  tokens with the highest attention weights, which contribute the most to sentiment prediction.
+
+In (3), the sense diversity of each token is computed as the average distance of the token's hidden state to those of its substitution candidates. Then, through (4), top  $J$  tokens with the most diverse senses are selected as disambiguation targets, because these target tokens may have very different senses in different contexts. Subsequently, the pretrained lexical substitution model is applied to provide the best substitution to replace each target token, as shown in (5) and (6)a. The new input sentence is passed onto the sentiment analysis model for final prediction ((7)a and (8)a), which is used for the sentiment model backpropagation ((9)a).
+
+In order to fine-tune the lexical substitution model to provide better substitutions that improve the accuracy of sentiment analysis, we adapt a dynamic rewarding mechanism (Ge et al., 2022) that was proposed for multi-task learning. As shown in Figure 1 (5) and (6)b, for each target word, we choose the top  $N$  candidates. Each of them is seen as a substitute candidate to calculate the probability of correct sentiment prediction after the replacement ((7)b and (8)b). Then, each probability is used to dynamically compute the loss weight when the corresponding candidate is learned by the lexical substitution model as a ground truth ((9)b and (10)b). If the replacement by a candidate leads to a higher chance of predicting the right sentiment, the candidate is more likely given a higher rank by the lexical substitution model. Otherwise, the model would be less likely to select the candidate as the best replacement. The important notations used in this paper are summarized in Table 1. The detailed training process can be seen in Algorithm 1.
+
+# 4.1 Sentiment Analysis with Interpretability
+
+Given input sentence  $w = (w_{1},\dots ,w_{L})$ , the goal is to predict the correct sentiment label  $\tilde{y}$ . The input is first fed into a pre-trained encoder:
+
+$$
+V = \operatorname {E n c o d e r} (w), \tag {2}
+$$
+
+where  $V$  is hidden states. Next, we aim to find the tokens that contribute the most to sentiment inference. We adopt an interpretable attention module called Hierarchical Attention Network (HAN, Han et al., 2022), which effectively encodes hidden states with multiple non-linear projections and ranks the most influential tokens based on attention. Following the setup of Han et al. (2022), we stack two blocks of HAN to form our attention module.
+
+$$
+q, a = H A N (H A N (V)), \tag {3}
+$$
+
+Table 1: Notation table.  
+
+<table><tr><td></td><td>Description</td></tr><tr><td>L</td><td>Input sequence length</td></tr><tr><td>I</td><td>Number of words with top ranking attention scores to be potential target words</td></tr><tr><td>J</td><td>Number of words with top ranking diversity scores to be potential target words</td></tr><tr><td>N</td><td>Number of candidates used for finetuning lexical substitution via dynamic rewarding</td></tr><tr><td>K</td><td>Number of candidates for training lexical substitution model</td></tr><tr><td>w</td><td>The original input sentence</td></tr><tr><td>V</td><td>The hidden states of w</td></tr><tr><td>y</td><td>The ground-truth sentiment label of w</td></tr><tr><td>wf</td><td>The final input sentence where ambiguous words in w are substituted</td></tr><tr><td>ŷf</td><td>The predicted sentiment label with wf as input</td></tr><tr><td>wl</td><td>The l-th word in w</td></tr><tr><td>al</td><td>The attention weight of the l-th word in w</td></tr><tr><td>wiatt</td><td>The word in w with the i-th highest attention weight</td></tr><tr><td>viatt</td><td>The hidden state of wiatt</td></tr><tr><td>M</td><td>The number of synonyms of wiatt with the same part-of-speech that are provided by the WordNet</td></tr><tr><td>G</td><td>The hidden states of the M WordNet synonyms</td></tr><tr><td>gm</td><td>The hidden state of the m-th WordNet synonyms of wiatt</td></tr><tr><td>di</td><td>The average Euclidean distance between viatt and g1, . . . , gm, . . . , gM</td></tr><tr><td>tj</td><td>The j-th target word in w to be substituted</td></tr><tr><td>c</td><td>A set of candidate words from WordNet.</td></tr><tr><td>ˆc</td><td>The top N candidates ranked by the lexical substitution model</td></tr><tr><td>ˆcj,n</td><td>The top n-th candidate for replacing tj</td></tr><tr><td>wj,n</td><td>The resulting sentence when replacing tj in w withˆcj,n</td></tr><tr><td>ˆθj,n</td><td>The loss weight whenˆcj,n is treated as the gold standard substitution</td></tr><tr><td>s</td><td>The formatted input to lexical substitution model</td></tr><tr><td>U</td><td>The hidden state matrix of the input sentence in s</td></tr><tr><td>u1</td><td>The hidden state of the l-th input word in s</td></tr><tr><td>R</td><td>The hidden state matrix of the candidates in s</td></tr><tr><td>rk</td><td>The hidden state of the k-th candidate in s</td></tr><tr><td>τ</td><td>The temperature hyper-parameter for training the lexical substitution model</td></tr><tr><td>β</td><td>The hyper-parameter for balancing the sentiment analysis loss and fine-tuning loss</td></tr><tr><td>L(sa)</td><td>The sentiment analysis loss</td></tr><tr><td>L(tune)</td><td>The loss for finetuning the pretrained language model into a lexical substitution model</td></tr><tr><td>L(ls)</td><td>The lexical substitution loss whenˆcj,n is treated as the gold standard substitution</td></tr><tr><td>L(ls)</td><td>The final weighted lexical substitution loss for backpropagation</td></tr></table>
+
+where vector  $q$  is the yielded hidden state.  $a$  is the attention weights, indicating the contribution of each token to the final sentiment prediction. To obtain the sentiment prediction,  $q$  is passed on to two layers of feed-forward neural networks (FNN) to obtain the probability distribution of sentiment prediction, with the first one being activated by ReLU (Agarap, 2018), and the second by softmax.
+
+$$
+h = R e L U \left(F N N _ {1} (q)\right) \tag {4}
+$$
+
+$$
+\hat {y} = \operatorname {s o f t m a x} \left(F N N _ {2} (h)\right) \tag {5}
+$$
+
+We denote the prediction of the sentiment analysis model as  $\hat{y}^f$  when the input is  $w^f$ , which denotes a paraphrased  $w$  where all selected target tokens are replaced by sense-clarifying substitutions. Thus, the sentiment analysis loss is computed as:
+
+$$
+\mathcal {L} ^ {(s a)} = \operatorname {C r o s s E n t r o p y} (\hat {y} ^ {f}, \tilde {y}) \tag {6}
+$$
+
+# 4.2 Disambiguation by Lexical Substitution
+
+To obtain a lexical substitution model, we fine-tune the pre-trained language model ALM (Mao et al., 2023a)<sup>1</sup> with benchmark datasets from McCarthy and Navigli (2007); Kremer et al. (2014). We use the candidates from the datasets to formulate our input  $s$  as in Formula 1. The candidate with the highest score is set as the ground truth substitution  $\tilde{c}$ . Given the input  $s$ , the model encodes it as:
+
+$$
+U, R = A L M (s), \tag {7}
+$$
+
+where  $U = [u_{1},\ldots ,u_{L}]$  is the hidden states of the input sentence, and  $R = [r_1,\dots,r_K]$  is the hidden states of the candidates. We denote the representation of the target word as  $u_{t}$ $(t\in \{1,\ldots ,L\})$
+
+Our training objective is to a) close the distance between the representation of the ground truth candidate  $r_k$ ,  $(k \in \{1, \dots, K\})$  with  $u_t$ , and b) push away incorrect candidates representations  $r_j$ $(j \in \{1, \dots, K | j \neq k\})$  from  $u_t$ . Namely,  $(r_k, u_t)$  will be regarded as a positive pair, while  $(r_j, u_t)$  will be regarded as a negative pair. We follow the InfoNCE loss (Oord et al., 2018) to achieve these goals, which can be formulated as:
+
+$$
+\mathcal {L} ^ {(t u n e)} = - \sum_ {i} \log \frac {e x p (d (u _ {t} , r _ {k}) / \tau)}{\sum_ {j} e x p (d (u _ {t} , r _ {i}) / \tau)}, \tag {8}
+$$
+
+where  $i\in \{1,\dots ,K\}$ $\tau$  is a temperature hyperparameter, and  $d(\cdot)$  is Euclidean distance.
+
+Algorithm 1: Sentiment analysis with word sense disambiguation using a dynamic rewarding mechanism.
+
+1 Initialize sentiment analysis model as  $\Phi$  , pre-trained   
+2 lexical substitution model as  $\Psi$  .   
+3 Initialize hyperparameters  $\beta$ $I,J,N$    
+while not done do   
+4 Sample a sentence  $w = w_{1},w_{2},\ldots ,w_{L}$  .   
+for  $l = 1:L$  do   
+6 Compute the attention weight  $a_{l}$  of token   
+ $w_{l}$  .   
+7 end   
+8  $w^{att}\gets$  Top I of  $w$  ordered by attention weights   
+ $a = a_1,a_2,\dots ,a_L$  .   
+for  $i = 1:I$  do   
+10  $c\gets$  all possible candidates with the same   
+part-of-speech type as  $w_{i}^{att}$  , provided by   
+WordNet;   
+11 Compute the average Eulidean distance  $d_{i}$    
+between the hidden states of  $w_{i}^{att}$  and  $c$  produced by the encoder in  $\Phi$  .   
+end   
+13  $t = (t_1,t_2,\dots ,t_J)\gets$  Top J of  $w^{att}$  ordered   
+by  $d = d_{1},d_{2},\ldots ,d_{I}$ $w^{f}\gets w$  .   
+for  $j = 1:J$  do   
+16 Input w into  $\Psi$  to produce top N candidates   
+ $\hat{c}_j = (\hat{c}_{j,1},\dots ,\hat{c}_{j,N})$  for target word  $t_j$  ordered by probability;   
+17 Replace  $t_j$  in  $w^f$  with  $\hat{c}_{j,1}$    
+s  $\leftarrow (< s > ,w_{1},\dots ,t_{i},\dots ,w_{L}, <   / s > ,\right.$ $\hat{c}_{j,1}, <   / s > ,\dots ,\hat{c}_{j,N}, <   / s > )$  .   
+for  $n = 1:N$  do   
+21  $w^{j,n}\gets$  Replace  $t_j$  in  $w$  with  $\hat{c}_{j,n}$  .   
+22 Input  $w^{j,n}$  into  $\Phi$  to obtain the probability of correct sentiment prediction  $P(\hat{y} = \tilde{y})_{j,n}$  ..   
+23  $\theta_{j,n}\gets \beta P(\hat{y} = \tilde{y})_{j,n}^2$  .   
+24 Compute  $\mathcal{L}_{j,n}^{(ls)}$  by feeding s with  $\hat{c}_n$  labeled as true substitute into  $\Psi$  ..   
+25  $\mathcal{L}_{j,n}^{(ls)}\gets \theta_{j,n}\mathcal{L}_{j,n}^{(ls)};$    
+end   
+end   
+28  $\mathcal{L}^{(ls)}\gets \mathcal{L}_{1,1}^{(ls)} + \dots +\mathcal{L}_{1,N}^{(ls)} + \mathcal{L}_{J,1}^{(ls)} + \dots +\mathcal{L}_{J,N}^{(ls)}$  .   
+29 Compute sentiment analysis loss  $\mathcal{L}^{(sa)}$  using  $w^{f}$  as input;   
+30  $L\gets \mathcal{L}^{(sa)} + \mathcal{L}^{(ls)}$  .
+
+During the inference stage, we choose the candidate  $\hat{c}$  whose corresponding hidden state  $r_k$  is the most similar to  $u_t$ , measured by Euclidean distance:
+
+$$
+\hat {c} = \arg \min  \left(d \left(u _ {t}, r _ {k}\right)\right) \tag {9}
+$$
+
+We then use the resulting lexical substitution model to find replacements for the selected target words in sentiment analysis input. To determine which words are selected as disambiguation tar
+
+gets, we select the top  $I$  words in the input with the highest  $a$  produced by Equation 3, forming the set  $w^{att}$ . We denote their corresponding representations as  $V^{att} = \{v_{1}^{att},\dots ,v_{I}^{att}\}$ . For each  $v_{i}^{att}$ , we compute its average Euclidean distance to all of its candidates' hidden states. The candidates consist of WordNet synonyms of all of its possible senses under the same part-of-speech type, which are transformed into hidden states  $G = \{g_{1},\ldots ,g_{M}\}$  using Equation 2.  $M$  represents the number of candidates provided by WordNet. The top  $J$  words in  $w^{att}$  with the largest corresponding average distance are considered to be the ones with the most diverse word meanings, and thus are more likely to require disambiguation. We will demonstrate the success of such an ambiguous word selection method in the later ablation study (see Section 6.1).
+
+Hence, target words  $t = (t_1, \ldots, t_J)$  are selected by finding each corresponding  $v_j^{att}$ :
+
+$$
+v _ {j} ^ {\text {a t t}} = \arg \max  _ {i} \left(\frac {1}{M} \sum_ {M} d \left(v _ {i} ^ {\text {a t t}}, g _ {m}\right)\right). \tag {10}
+$$
+
+Following the lexical substitution benchmarks, only words whose part-of-speech tags fall under the general categories of noun, verb, adjective, and adverb are considered as targets.
+
+# 4.3 Dynamic Rewarding Mechanism
+
+To fine-tune the lexical substitution model on the downstream sentiment analysis task without knowing WSD gold standard truth, for each target word  $t_j$ , we utilize the top  $N$  candidates  $\hat{c}_j = \{\hat{c}_{j,1},\dots ,\hat{c}_{j,n},\dots ,\hat{c}_{j,N}\}$  produced by the lexical substitution model. The new input resulting from  $t_j$  being replaced by candidate  $\hat{c}_{j,n}$  is denoted as  $w^{j,n}$ . Same with Equation 5, the probability distribution of a sentiment prediction from  $w^{j,n}$  is:
+
+$$
+P (\hat {y}) _ {j, n} = \operatorname {s o f t m a x} \left(F N N _ {2} \left(h ^ {j, n}\right)\right), \tag {11}
+$$
+
+where  $h^{j,n}$  is the hidden states of  $w^{j,n}$  outputted by Equation 4 in the sentiment analysis model.
+
+To adjust the model in such a way that a more accurate sentiment prediction results in a higher reward for the corresponding substitution output, we compute the loss weight for  $\hat{c}_{j,n}$  being the correct substitution prediction as:
+
+$$
+\theta_ {j, n} = \beta P (\hat {y} = \tilde {y}) _ {j, n} ^ {2}, \tag {12}
+$$
+
+where  $\beta$  is a hyperparameter for balancing the sentiment analysis and lexical substitution losses,  $\tilde{y}$  is the ground truth sentiment label as defined above.
+
+<table><tr><td>Dataset</td><td>Split</td><td># Samples</td></tr><tr><td>SemEval 2007 Task 10</td><td>-</td><td>2,010</td></tr><tr><td>CoinCo</td><td>-</td><td>15,629</td></tr><tr><td></td><td>Train</td><td>8,530</td></tr><tr><td>Rotten Tomatoes</td><td>Dev</td><td>1,066</td></tr><tr><td></td><td>Test</td><td>1,066</td></tr><tr><td>Financial PhraseBank</td><td>-</td><td>4,846</td></tr><tr><td>STS-Gold</td><td>-</td><td>2,026</td></tr></table>
+
+Table 2: Datasets used for experiments.
+
+We formulate the input  $s^j$  to lexical substitution model with sentence  $w$  and candidates  $\hat{c}_j$ , using Formula 1. Similar to Equation 8, the loss  $(\mathcal{L}_{j,n}^{(ls)})$  of the lexical substitution model when  $\hat{c}_{j,n}$  is considered as gold standard is computed as follows:
+
+$$
+\mathcal {L} _ {j, n} ^ {(l s)} = - \sum_ {i} \log \frac {\exp \left(d \left(u _ {t} , \hat {c} _ {j , n}\right) / \tau\right)}{\sum_ {j} \exp \left(d \left(u _ {t} , \hat {c} _ {j}\right) / \tau\right)}. \tag {13}
+$$
+
+Finally, the total loss is computed as:
+
+$$
+\mathcal {L} = \mathcal {L} ^ {(s a)} + \theta_ {1, 1} \mathcal {L} _ {1, 1} ^ {(l s)} + \dots + \theta_ {J, N} \mathcal {L} _ {J, N} ^ {(l s)}. \tag {14}
+$$
+
+# 5 Experiments
+
+# 5.1 Datasets
+
+To fine-tune the pretrained lexical substitution model, we combined SemEval 2007 Task 10 (McCarthy and Navigli, 2007) and CoinCo (Kremer et al., 2014). The former consists of 2,010 example sentences for 201 polysemous words. The latter contains 2,474 sentences with multiple words to be substituted. We used  $70\%$  of the concatenated dataset for training,  $30\%$  for testing.
+
+For sentiment analysis, we used three datasets from different domains: Rotten Tomatoes (Pang and Lee, 2005) consists of 5,331 positive and 5,331 negative sentences collected from Rotten Tomatoes movie reviews. Financial PhraseBank (Malo et al., 2014) contains 4,846 sentences from English financial news, each labeled as negative, positive, or neutral. As the dataset has no standard split, We adopted a split of  $70\%$  and  $30\%$  for the training and testing sets. STS-Gold (Saif et al., 2013) contains 2,026 tweets, annotated as negative or positive. As the dataset has no standard split, We adopted a split of  $90\%$  and  $10\%$  for the training and testing sets. Details of the datasets are listed in Table 2.
+
+# 5.2 Baselines
+
+Since we aim to improve a sentiment analysis classifier with explicit WSD in a neurosymbolic fashion, vanilla RoBERTa-large (Liu et al., 2019) is
+
+considered as the most direct baseline, because we also used it for sentiment classification. To put our model's performance into perspective, we further include published baselines with the highest reported accuracy on the experiment datasets, which consist of various pre-trained language models finetuned on the target dataset, i.e., FinBERT (Araci, 2019), DeBERTa (He et al., 2020; Sileo, 2023); and ones trained with additional resources, namely, DeBERTa+tarksource (Sileo, 2023) which is pretrained on a unified dataset for 500 English tasks, and SenticNet 7 (Cambria et al., 2022) which is a XLNet enhanced by sentiment knowledge.
+
+# 5.3 Setups
+
+To obtain the lexical substitution model, we fine-tune ALM with the lexical substitution datasets for 20 epochs with a batch size of 20 and a learning rate of 1e-8. The pre-trained model we used as the encoder for sentiment analysis is RoBERTa-large. The hidden size for both models is set to 1024. We use the NLTK part-of-speech tagger (Bird, 2006) to label the sentiment analysis input.
+
+To train our proposed framework, we adopt  $I = 5$ ,  $J = 2$ ,  $K = 15$ ,  $N = 3$ ,  $\tau = 0.05$  and  $\beta = 0.0001$ . We set the learning rate to 1-e6. The models are trained for 50 epochs with a batch size of 10, using early stopping. We used Adam (Kingma and Ba, 2014) as the optimizer. Experiments are run on RTX A6000 GPU. Following the setups in the baselines, we adopt accuracy as an evaluation metric.
+
+# 6 Results
+
+We compare our model with the baselines in Table 3. Our model outperforms existing models with the best-published results, achieving state-of-the-art performance on the three datasets. Compared to the RoBERTa baseline, our model is able to improve the accuracy of sentiment analysis on corpora from different domains to varying extent. We hypothesize that the extent of improvement is dependent on domain types. For instance, Twitter language is informal and brief, with frequent use of short homographs. Thus, the integrated disambiguation model is able to obtain the most significant increase of  $1.91\%$  accuracy on STS-Gold. In the financial domain, metaphors and less common word senses are commonly used, accounting for our  $1.23\%$  gain on Financial PhraseBank.
+
+Movie reviews, on the other hand, often use
+
+<table><tr><td>Model</td><td>RT</td><td>FPB</td><td>STS-Gold</td><td>Avg</td></tr><tr><td>DeBERTa</td><td>90.42</td><td>84.48</td><td>-</td><td>87.45</td></tr><tr><td>+tasksource</td><td>90.99</td><td>85.20</td><td>-</td><td>88.10</td></tr><tr><td>FinBERT</td><td>-</td><td>86.00</td><td>-</td><td>-</td></tr><tr><td>SenticNet 7</td><td>-</td><td>-</td><td>90.08</td><td>-</td></tr><tr><td>RoBERTa</td><td>91.31</td><td>87.09</td><td>97.14</td><td>91.85</td></tr><tr><td>Ours</td><td>91.87*</td><td>88.32*</td><td>99.05*</td><td>93.08</td></tr></table>
+
+Table 3: Comparisons with baselines. RT stands for Rotten Tomatoes. FPB stands for Financial PhraseBank. Bold font denotes the best results. The RoBERTa baseline was implemented by us. * denotes the improvement is statistically significant ( $p < 0.01$  on a two-tailed t-test), against the highest baseline score. Our results are averaged over 5 runs.
+
+standard expressions that are semantically clearer. Hence, our model obtains a marginal increase of  $0.56\%$  on Rotten Tomatoes. Despite the improvement being marginal on average, i.e.,  $1.23\%$ , we would like to highlight that this improvement is significant considering the context of error propagation from the lexical substitution model. Our framework consistently demonstrates improvements across all datasets, substantiating its ability to achieve higher gains in accuracy while minimizing any potential loss. Furthermore, we can interpret the WSD outcomes and the sentiment analysis decision-making process before and after lexical substitutions from the proposed neurosymbolic framework. We will demonstrate the interpretability of the framework in Section 6.2.
+
+# 6.1 Ablation Study
+
+To prove the effectiveness of each component in our proposed framework, we compare with two variations: w/o WSD, where sentiment prediction is done by the sentiment analysis model without any modifications to the input; w/o DR, where the lexical substitution model is not fine-tuned during sentiment analysis training. To prove our hypothesis in WSD target word selection, we compare the proposed framework with two variations: Top-2, where the top 2 words in the input with the highest attention weights are automatically chosen as target words to be disambiguated; Rand, where 2 out of the most attended 5 words are randomly selected to be the target words. The comparison of performance is shown in Table 4. It can be observed that our full model outperforms w/o WSD by  $0.28\%$  on Rotten Tomatoes,  $1.03\%$  on Financial PhraseBank, and  $1.43\%$  on STS-Gold.
+
+The extent of gains in different domains is in
+
+<table><tr><td>Model</td><td>RT</td><td>FPB</td><td>STS-Gold</td><td>Avg</td></tr><tr><td>w/o WSD</td><td>91.59</td><td>87.29</td><td>97.62</td><td>92.17</td></tr><tr><td>w/o DR</td><td>91.40</td><td>87.27</td><td>96.67</td><td>91.78</td></tr><tr><td>Top-2</td><td>91.60</td><td>87.81</td><td>97.62</td><td>92.34</td></tr><tr><td>Rand</td><td>89.72</td><td>86.30</td><td>97.14</td><td>91.05</td></tr><tr><td>Ours</td><td>91.87</td><td>88.32</td><td>99.05</td><td>93.08</td></tr></table>
+
+Table 4: Ablation study results.
+
+line with our hypothesis in the previous section. It shows that our disambiguation framework is indeed able to improve the accuracy of sentiment analysis by selecting the appropriate ambiguous words and paraphrasing them into semantically clearer ones. Our model also achieves higher accuracy than w/o DR on all three datasets, proving the validity and effectiveness of our dynamic rewarding mechanism. Noticeably, without dynamic rewarding and joint fine-tuning, the performance of w/o DR is weaker than w/o WSD, showing that dynamic rewarding-based fine-tuning with data from another task can effectively alleviate error propagation from lexical substitution.
+
+Our proposed framework outperforms the Top-2 selection method on all datasets. It proves our assumption that words that are highly diverse in their senses are more likely to cause ambiguity for sentiment prediction. Rand performs the worst on all three datasets compared to w/o WSD, Top-2, and our model, indicating i) the necessity of selecting the appropriate target words to disambiguate so as not to introduce noise, and ii) the validity of choosing target words based on both their importance to the decision and their sense diversity level.
+
+# 6.2 Interpretability Demonstration
+
+Since an interpretable encoder (HAN) is used in our framework to learn sequential representations, we can understand which words contribute the most to a sentiment analysis prediction by visualizing the attention weights of HAN.
+
+As shown in Figure 2, we conduct a study on the change of attention weights before and after paraphrasing. In Figure 2 (a), the words that receive the most attention in the original sentence are not,  $a$ , sore, etc., among which not has significantly higher weight. This indicates that the model is mostly relying on negation to make inference. As not does not have diverse meaning and  $a$  is an article, sore is selected as target word and replaced with painful. Our attention module assigns a much higher weight to the negative adjective after paraphrasing.
+
+This shows that our model is able to find suitable
+
+![](images/b40c38e1a8db81de2405e02c97c9e7c4738fed8a0b063751c1baafc1a4ffc089.jpg)  
+Figure 2: The most attended word demonstration before and after paraphrasing in sentiment analysis predictions. The lighter the color, the higher the attention weight.
+
+word substitutions that aid the correct sentiment predictions. From Figure 2 (b), we can observe a rather evenly distributed attention weights in the original sentence, meaning that the attention module cannot identify the words that significantly affect sentiment. After our framework replaces the word decreased to fell, the model is able to assign higher weights to the sentiment-loaded phrase profit fell. From Figure 2 (c), we can see that after replacing the word mediocre with average, the sentiment related words including the paraphrased word and unaware all received higher attention. By attending to the straightforward paraphrases expressing opinions, the sentiment analysis model can yield correct predictions for all three cases.
+
+# 6.3 Case Study
+
+In Table 5, we study the ability of our framework to identify and paraphrase ambiguous words into unequivocal ones. In the first instance, albeit not affecting the prediction, the word delicious is disambiguated by our framework into its figurative meaning delighted instead of its literal one. In the second instance, the word succeeds has a more positive word sense equating to win, and a more neutral one equating to follow. Our framework is able to paraphrase it into the latter. In the last instance, the word visibility falls under the word sense of degree of exposure to public notice in this context. Our framework correctly paraphrases it into profile, which is the synonym under the same sense in WordNet. It can be concluded that our framework can disambiguate word senses in a transparent and interpretable way for sentiment analysis.
+
+<table><tr><td>Setup</td><td>Content</td></tr><tr><td>Original</td><td>a delicious and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them.</td></tr><tr><td>Paraphrased</td><td>a delighted and delicately funny look at the residents of a copenhagen neighborhood coping with the befuddling complications life tosses at them.</td></tr><tr><td>Original</td><td>Le Lay succeeds Walter G++nter and will be based in Finland.</td></tr><tr><td>Paraphrased</td><td>Le Lay follows Walter G++nter and will be based in Finland.</td></tr><tr><td>Original</td><td>Strong brand visibility nationally and regionally is of primary importance in home sales, vehicle and consumer advertising.</td></tr><tr><td>Paraphrased</td><td>Strong brand profile nationally and regionally is of primary importance in home sales, vehicle and consumer advertising.</td></tr></table>
+
+# 6.4 Error Analysis
+
+Finally, we perform a brief error analysis (Table 6). In the first sentence, the model wrongly paraphrases force into power, because it does not recognize the multi-word idiom tour de force. It shows the model's limitation in understanding the semantic meaning of infrequent multi-word idiom. In the second sentence, the word just is wrongly paraphrased as good, which negatively influenced the sentiment prediction. This is mainly due to the part-of-speech tagger wrongly classifying the word just into a adjective, instead of an adverb. Thus, the lexical substitution module might be subjected to error propagation of the part-of-speech tagger and introduce noise into the input. In the last sentence, the word skills is paraphrased as science. It is likely that the model mistakes the word major as a field of study instead of a student studying a field. Thus, the substitution makes sense in the local context the science of calculus major, but not in the whole context, meaning that the model fails to understand the semantic meaning of the sentence.
+
+# 7 Conclusion
+
+In this work, we propose a neurosymbolic framework for sentiment analysis that incorporates the disambiguation of word senses by identifying and paraphrasing ambiguous words in the input, which is realized by leveraging a pre-trained lexical substitution model. The proposed framework not only improves the accuracy of sentiment analysis, but also achieves interpretability by enabling us to trace back which words are paraphrased and what words they are replaced with.
+
+Table 5: Case study of disambiguation as paraphrasing.  
+
+<table><tr><td>Setup</td><td>Content</td><td>Label</td></tr><tr><td>Original</td><td>a tour de force of modern cinema.</td><td>Positive</td></tr><tr><td>Paraphrased</td><td>a tour de power of modern cinema.</td><td>Positive</td></tr><tr><td>Original</td><td>Is just at work for the day -sigh- I have a headache</td><td>Negative</td></tr><tr><td>Paraphrased</td><td>Is good at work for the day -sigh- I have a headache</td><td>Positive</td></tr><tr><td>Original</td><td>the skills of a calculus major at m.i.t. are required to balance all the formulaic equations in the long-winded heist comedy who is cletis tout?</td><td>Negative</td></tr><tr><td>Paraphrased</td><td>the science of a calculus major at m.i.t. are required to balance all the formulaic equations in the long-winded heist comedy who is cletis tout?</td><td>Positive</td></tr></table>
+
+Table 6: Error analysis. Red indicates errors made by our proposed framework.
+
+Additionally, to fine-tune the lexical substitution model to provide better disambiguating substitutes for sentiment analysis with no ground-truth word sense labels, we adopt the dynamic rewarding mechanism to jointly train the sentiment analysis and lexical substitution models. Experiments show that our framework is indeed effective in improving the performance of sentiment analysis, obtaining state-of-the-art performance on three corpora from different domains.
+
+# Limitations
+
+The framework proposed in this paper have to be seen in light of some limitations. First, the performance of sentiment analysis to some extent depends on a decent-performing pre-trained lexical substitution model. Currently, polysemous samples in lexical substitution data are not as abundant as in WSD data, which might limit its ability in disambiguating less common polysemies. Second, the effectiveness of our proposed model is only tested on English corpora, as it relies on the sense knowledge from WordNet. The disambiguation target selection and paraphrasing methods might work on languages with more complex segmentation or limited sense lexicon.
+
+# References
+
+Abien Fred Agarap. 2018. Deep learning using rectified linear units (ReLU). arXiv preprint arXiv:1803.08375.
+
+Dogu Araci. 2019. FinBERT: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063.  
+Philip Arthur, Graham Neubig, Sakriani Sakti, Tomoki Toda, and Satoshi Nakamura. 2015. Semantic parsing of ambiguous input through paraphrasing and verification. Transactions of the Association for Computational Linguistics, 3:571-584.  
+Stefano Baccianella, Andrea Esuli, and Fabrizio Sebastiani. 2010. SentiWordNet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC'10).  
+Edoardo Barba, Tommaso Pasini, and Roberto Navigli. 2021a. ESC: Redesigning WSD with Extractive Sense Comprehension. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4661-4672, Online. Association for Computational Linguistics.  
+Edoardo Barba, Luigi Procopio, and Roberto Navigli. 2021b. ConSeC: Word sense disambiguation as continuous sense comprehension. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1492-1503, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Michele Bevilacqua and Roberto Navigli. 2020. Breaking through the  $80\%$  glass ceiling: Raising the state of the art in word sense disambiguation by incorporating knowledge graph information. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2854-2864.  
+Steven Bird. 2006. NLTK: the natural language toolkit. In Proceedings of the COLING/ACL 2006 Interactive Presentation Sessions, pages 69-72.  
+Terra Blevins and Luke Zettlemoyer. 2020. Moving down the long tail of word sense disambiguation with gloss-informed biencoders. arXiv preprint arXiv:2005.02590.  
+Erik Cambria, Qian Liu, Sergio Decherchi, Frank Xing, and Kenneth Kwok. 2022. SenticNet 7: A commonsense-based neurosymbolic AI framework for explainable sentiment analysis. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 3829-3839.  
+Erik Cambria, Rui Mao, Melvin Chen, Zhaoxia Wang, and Seng-Beng Ho. 2023. Seven pillars for the future of AI. IEEE Intelligent Systems, 38(6).  
+Erik Cambria, Soujanya Poria, Federica Bisio, Rajiv Bajpai, and Iti Chaturvedi. 2015. The CLSA model: A novel framework for concept-level sentiment analysis. In LNCS, volume 9042, pages 3-22. Springer.
+
+Erik Cambria, Soujanya Poria, Alexander Gelbukh, and Mike Thelwall. 2017. Sentiment analysis is a big suitcase. IEEE Intelligent Systems, 32(6):74-80.  
+Niccolò Campolungo, Tommaso Pasini, Denis Emelin, and Roberto Navigli. 2022. Reducing disambiguation biases in NMT by leveraging explicit word sense information. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4824-4838.  
+Yanfang Cao, Pu Zhang, and Anping Xiong. 2015. Sentiment analysis based on expanded aspect and polarity-ambiguous word lexicon. International Journal of Advanced Computer Science and Applications, 6(2).  
+Ann Devitt and Khurshid Ahmad. 2007. Sentiment polarity identification in financial news: A cohesion-based approach. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 984-991.  
+Philip Edmonds and Adam Kilgarriff. 2002. Introduction to the special issue on evaluating word sense disambiguation systems. Natural Language Engineering, 8(4):279-291.  
+Umar Farooq, Tej Prasad Dhamala, Antoine Nongaillard, Yacine Ouzrout, and Muhammad Abdul Qadir. 2015. A word sense disambiguation method for feature level sentiment analysis. In 2015 9th International Conference on Software, Knowledge, Information Management and Applications (SKIMA), pages 1-8. IEEE.  
+Christiane Fellbaum. 1998. WordNet: An electronic lexical database. MIT press.  
+Mengshi Ge, Rui Mao, and Erik Cambria. 2022. Explanable metaphor identification inspired by conceptual metaphor theory. Proceedings of the AAAI Conference on Artificial Intelligence, 36(10):10681-10689.  
+Sooji Han, Rui Mao, and Erik Cambria. 2022. Hierarchical attention network for explainable depression detection on Twitter aided by metaphor concept mappings. In Proceedings of the 29th International Conference on Computational Linguistics (COLING), pages 94-104.  
+Viktor Hangya, Qianchu Liu, Dario Stojanovski, Alexander Fraser, and Anna Korhonen. 2021. Improving machine translation of rare and unseen word senses. In Proceedings of the Sixth Conference on Machine Translation, pages 614-624.  
+Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. DeBERTa: Decoding-enhanced BERT with disentangled attention. arXiv preprint arXiv:2006.03654.
+
+Luyao Huang, Chi Sun, Xipeng Qiu, and Xuanjing Huang. 2019. GlossBERT: BERT for word sense disambiguation with gloss knowledge. arXiv preprint arXiv:1908.07245.  
+Chihli Hung and Shiuman-Jeng Chen. 2016. Word sense disambiguation based sentiment lexicons for sentiment classification. Knowledge-Based Systems, 110:224-232.  
+Chihli Hung and Hao-Kai Lin. 2013. Using objective words in SentiWordNet to improve word-of-mouth sentiment classification. IEEE Intelligent Systems, 28(02):47-54.  
+Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.  
+Gerhard Kremer, Katrin Erk, Sebastian Padó, and Stefan Thater. 2014. What substitutes tell us-analysis of an "all-words" lexical substitution corpus. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 540-549.  
+Frederick Liu, Han Lu, and Graham Neubig. 2018. Handling homographs in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1336-1345.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.  
+P. Malo, A. Sinha, P. Korhonen, J. Wallenius, and P. Takala. 2014. Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology, 65.  
+Rui Mao, Kai He, Claudia Beth Ong, Qian Liu, and Erik Cambria. 2023a. MetaPro 2.0: Computational metaphor processing on the effectiveness of anomalous language modeling.  
+Rui Mao and Xiao Li. 2021. Bridging towers of multitask learning with a gating mechanism for aspect-based sentiment analysis and sequential metaphor identification. Proceedings of the AAAI Conference on Artificial Intelligence, 35(15):13534-13542.  
+Rui Mao, Xiao Li, Kai He, Mengshi Ge, and Erik Cambria. 2023b. MetaPro Online: A computational metaphor processing online system. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 127-135.  
+Rui Mao, Qian Liu, Kai He, Wei Li, and Erik Cambria. 2023c. The biases of pre-trained language models:
+
+An empirical study on prompt-based sentiment analysis and emotion detection. IEEE Transactions on Affective Computing, 14(3):1743-1753.  
+Marco Maru, Simone Conia, Michele Bevilacqua, and Roberto Navigli. 2022. Nibbling at the hard core of Word Sense Disambiguation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4724-4737.  
+Diana McCarthy and Roberto Navigli. 2007. SemEval-2007 task 10: English lexical substitution task. In Proceedings of the 4th International Workshop on Semantic Evaluations, SemEval '07, pages 48-53.  
+Arman Khadjeh Nassirtoussi, Saeed Aghabozorgi, Teh Ying Wah, and David Chek Ling Ngo. 2015. Text mining of news-headlines for forex market prediction: A multi-layer dimension reduction algorithm with semantics and sentiment. Expert Systems with Applications, 42(1):306-324.  
+Bruno Ohana and Brendan Tierney. 2009. Sentiment classification of reviews using SentiWordNet. Proceedings of IT&T, 8.  
+Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.  
+Martha Palmer, Hoa Trang Dang, and Christiane Fellbaum. 2007. Making fine-grained and coarse-grained sense distinctions, both manually and automatically. Natural Language Engineering, 13(2):137-163.  
+Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the ACL.  
+Xiao Pu, Nikolaos Pappas, James Henderson, and Andrei Popescu-Belis. 2018. Integrating weakly supervised word sense disambiguation into neural machine translation. Transactions of the Association for Computational Linguistics, 6:635-649.  
+Rachel Elizabeth Ramsey. 2017. An exemplar-theoretic account of word senses. University of Northumbria at Newcastle (United Kingdom).  
+Hassan Saif, Miriam Fernandez, Yulan He, and Harith Alani. 2013. Evaluation datasets for Twitter sentiment analysis: A survey and a new dataset, the STS-Gold. Emotion and Sentiment in Social and Expressive Media.  
+Damien Sileo. 2023. tasksource: A dataset harmonization framework for streamlined NLP multitask learning and evaluation. arXiv preprint arXiv:2301.05948v3.  
+Yang Song, Xin Cai Ong, Hwee Tou Ng, and Qian Lin. 2021. Improved word sense disambiguation with enhanced sense representations. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 4311-4320.
+
+Yunqing Xia, Erik Cambria, Amir Hussain, and Huan Zhao. 2015. Word polarity disambiguation using bayesian model and opinion-level features. Cognitive Computation, 7(3):369-380.  
+Yunqing Xia, Huan Zhao, Erik Cambria, and Amir Hussain. 2014. Disambiguating word sentiment polarity through bayesian modeling and opinion-level features. In Proceedings of ICML Workshops.

2023/Neuro-Symbolic Sentiment Analysis with Dynamic Word Sense Disambiguation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4ba67f8ef91319af6d814f3dafc29ff7987a7914173d32abec7f5c53149e80d
+size 541454

2023/New Datasets and Controllable Iterative Data Augmentation Method for Code-switching ASR Error Correction/704f5812-1da2-4355-bf1f-7c6e5a7b4a97_content_list.json

@@ -0,0 +1,2189 @@
+[
+    {
+        "type": "text",
+        "text": "New Datasets and Controllable Iterative Data Augmentation Method for Code-switching ASR Error Correction",
+        "text_level": 1,
+        "bbox": [
+            124,
+            87,
+            872,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhaohong Wan $^{1,2}$  and Xiaojun Wan $^{1,2}$  and Wei Peng $^{3}$  and Rongjun Li $^{3}$",
+        "bbox": [
+            184,
+            141,
+            815,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Wangxuan Institute of Computer Technology, Peking University",
+        "bbox": [
+            236,
+            159,
+            764,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2The MOE Key Laboratory of Computational Linguistics, Peking University",
+        "bbox": [
+            191,
+            175,
+            808,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>3</sup> Artificial Intelligence Application Research Center, Huawei Technologies {xmwzh, wanxiaojun}@pku.edu.cn",
+        "bbox": [
+            196,
+            192,
+            805,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "With the wide use of automatic speech recognition (ASR) systems, researchers pay more attention to the ASR error correction task to improve the quality of recognition results. In particular, ASR in bilingual or multilingual settings, namely code-switching ASR, has greater challenges and research value. In this paper, we first present code-switching ASR correction datasets obtained from solid ASR systems and automatic annotators. The datasets contain Chinese-English code-switching dialogues of bilingual speakers in Singapore, Malaysia, and Hong Kong. Based on this task, we propose a controllable iterative (CI) data augmentation method for improving the performance of mainstream ASR error correction systems. With a small amount of training data, our proposed method has the ability to iteratively produce abundant pseudo parallel data from the monolingual corpus for Chinese-English code-switching ASR correction. Results of experiments show that our method achieves the best performance compared with the rule-based, back-translation-based data augmentation methods and large language model ChatGPT.",
+        "bbox": [
+            144,
+            282,
+            460,
+            650
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            665,
+            258,
+            682
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, automatic speech recognition (ASR) system has achieved impressive success with the adoption of advanced deep neural network (Moritz et al., 2020). Learning from training data, the ASR system can match human transcribers for some critical tasks. It has been widely applied in several areas, such as live captioning, intelligent appliance control (Aliprandi et al., 2014).",
+        "bbox": [
+            112,
+            693,
+            489,
+            820
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In some bilingual and multilingual societies, people often speak more than one language in a conversation. The speech phenomenon of switching language among sentences is called code-switching. To explore the code-switching phenomenon, code-switching tasks in many different",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/4728699cfd3432b669185b6b4a9e8c15070498f0fcddf0bcfa1eedb2ab0f39be.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Error Type</td><td>Redundant</td></tr><tr><td>Source</td><td>我要 start on 我的 a essay\n(I want to start on my a essay.)</td></tr><tr><td>Target</td><td>我要 start on 我的 essay\n(I want to start on my essay.)</td></tr></table>",
+        "bbox": [
+            542,
+            249,
+            848,
+            313
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Examples of Chinese-English code-switching ASR error correction in our datasets. The contents in brackets are the translations for non-native Chinese speakers.",
+        "bbox": [
+            507,
+            325,
+            882,
+            381
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "language pairs have been introduced (Dowlagar and Mamidi, 2023; Seddah et al., 2020), including the code-switching ASR task.",
+        "bbox": [
+            507,
+            413,
+            882,
+            460
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Employing multilingual ASR systems to solve code-switching ASR is proposed in recent work (Lovenia et al., 2021). However, the performance of systems in many cases is still far from satisfactory. Errors introduced by ASR systems usually affect the performance of downstream tasks. To improve the recognition performance, the code-switching ASR error correction task is necessary.",
+        "bbox": [
+            507,
+            463,
+            882,
+            590
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we focus on code-switching ASR error correction between Chinese and English, which is common in the Chinese communities of Asia area. To enable the research and development of code-switching ASR error correction, we propose two datasets SEAME-C and ASCEND-C that are constructed from two Chinese-English code-switching ASR datasets SEAME (Lyu et al., 2010) and ASCEND (Lovenia et al., 2021). In Table 1, we present typical examples of Chinese-English code-switching ASR error correction.",
+        "bbox": [
+            507,
+            595,
+            882,
+            770
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A challenge for the code-switching ASR error correction is the lack of a large scale of training data. Moreover, we can not even collect sufficient high-quality Chinese-English code-switching data. Considering these challenges, we propose a controllable iterative (CI) data augmentation method that can iteratively generate abundant code-switching ASR error correction instances from the monolingual corpus with small-scale labeled training data.",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8075",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8075-8087",
+        "bbox": [
+            216,
+            945,
+            778,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            277,
+            957,
+            719,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To verify the effectiveness of our proposed method, we implement two data augmentation baselines and evaluate the performance of these methods on proposed datasets at first. With a similar scale of augmented data, our proposed controllable iterative method achieves the best performance in both MaxMatch  $(M^2)$  scorer (Dahlmeier and Ng, 2012) and MER metrics on SEAME-C and ASCEND-C datasets. It performs well for both mT5 (Xue et al., 2021) and mBART (Tang et al., 2020) correction methods. Besides, considering the strong ability of large language models (LLM), we compare the performance of our method with the most powerful LLM ChatGPT. Experiments show that our method outdistances the ChatGPT method, which shows the effectiveness of our method in the Chinese-English code-switching ASR error correction task. Besides, we find that this task is challenging, and LLM method is far from achieving satisfactory results at now.",
+        "bbox": [
+            112,
+            84,
+            492,
+            405
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are summarized as follows:",
+        "bbox": [
+            131,
+            407,
+            473,
+            420
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We propose two datasets SEAME-C and ASCEND-C for the challenging Chinese-English code-switching ASR error correction task.",
+            "2. To address the problem of lacking sufficient training data, we propose the controllable iterative data augmentation method that can iteratively generate abundant code-switching ASR error correction instances from the monolingual corpus.",
+            "3. Extensive experiments show the superiority of our proposed controllable iterative method. Moreover, combining the pseudo data produced by the rule-based and controllable iterative methods can further improve the performance of error correction models."
+        ],
+        "bbox": [
+            112,
+            423,
+            487,
+            646
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The resources are publicly available<sup>1</sup>.",
+        "bbox": [
+            132,
+            649,
+            413,
+            665
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            678,
+            270,
+            694
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Code-switching ASR, the task aiming to recognize the speech that contains more than one language within the sentences, has been increasingly reported in speech technology and linguistic studies recently (Su, 2001; Auer, 2013). However, to our knowledge, there is no research on the code-switching ASR error correction task at present. In this section, we introduce two relevant fields: ASR error correction and data augmentation method in text error correction.",
+        "bbox": [
+            112,
+            705,
+            489,
+            864
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "With the successes of ASR systems, the ASR",
+        "bbox": [
+            132,
+            866,
+            487,
+            882
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "error correction has become a prevalent task in recent years. The idea of regarding the ASR error correction as the machine translation and applied advanced neural machine translation models is adopted in several previous approaches(Mani et al., 2020; Wang et al., 2020). D'Haro and Banchs (2016) first introduce machine translation to improve automatic transcription. Considering the machine translation models require only parallel text pairs, Guo et al. (2019) propose an approach to utilizing text-only data by training a correction model to explicitly correct the errors. The joint modeling method shows great performance in this task(Wang et al., 2023). Re-ranking ASR n-best hypotheses is also a popular topic for error recovery (Corona et al., 2017; Jonson, 2006).",
+        "bbox": [
+            507,
+            84,
+            884,
+            340
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the problem of lack of training data, data augmentation method is widely adopted in text error correction area. Many works (Grundkiewicz et al., 2019; Lichtarge et al., 2019) adopt pre-defined rules to generate sentences with errors. Inspired by back-translation procedure for machine translation (Sennrich et al., 2016), works (Xie et al., 2018; Kiyono et al., 2019) are proposed to train a model to generate erroneous sentences. Recently, researchers focus on the exposure bias problem of data augmentation method and proposed methods to utilize the augmented data in more effective ways (Solyman et al., 2023; Cao et al., 2023).",
+        "bbox": [
+            507,
+            342,
+            884,
+            551
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Code-switching ASR Error Correction Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            565,
+            877,
+            596
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we provide details about our two datasets SEAME-C and ASCEND-C. Specifically, we give the construction process of our datasets at first. Then, we do an analysis of our datasets and show some characteristics of them. In this work, we adopt the same corpus splitting as the original ASR datasets. In addition, we remove some bad cases that are too short and eliminate monolingual sentences in the validation and test sets to ensure the quality of evaluation. Our datasets provide quality resources for the research and development of code-switching ASR error correction. Table 2 describes the statistics of our proposed datasets.",
+        "bbox": [
+            507,
+            609,
+            882,
+            819
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 Data Construction",
+        "text_level": 1,
+        "bbox": [
+            507,
+            832,
+            702,
+            847
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In some Asia's multilingual countries and regions, speaking a mixture of Chinese and English within a sentence is common. We collect data from two Chinese-English code-switching ASR datasets with",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ https://github.com/chopper2k/Code-switching-ASR-Error-Correction",
+        "bbox": [
+            112,
+            891,
+            470,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8076",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/2ea7492b3adc5b942cb535d9a3487e22a91c7840d05e05c2182fe559fecf0332.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">SEAME-C</td><td colspan=\"4\">ASCEND-C</td></tr><tr><td>Train</td><td>Test</td><td>Total</td><td>Train</td><td>Val</td><td>Test</td><td>Total</td></tr><tr><td># Dialogues</td><td>267</td><td>30</td><td>297</td><td>40</td><td>5</td><td>4</td><td>49</td></tr><tr><td># Sentences</td><td>48,405</td><td>6,293</td><td>54,698</td><td>9,869</td><td>213</td><td>373</td><td>10,455</td></tr><tr><td># Tokens</td><td>801,548</td><td>99,837</td><td>901,385</td><td>122,348</td><td>3,385</td><td>5,538</td><td>131,271</td></tr><tr><td>Avg. sentences in a dialogue</td><td>181.3</td><td>209.8</td><td>174.2</td><td>246.7</td><td>53.3</td><td>92.3</td><td>213.4</td></tr><tr><td>Avg. tokens in a sentence</td><td>16.6</td><td>15.9</td><td>16.5</td><td>12.4</td><td>15.9</td><td>14.8</td><td>12.6</td></tr></table>",
+        "bbox": [
+            176,
+            80,
+            820,
+            170
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the advanced ASR system and automatic annotator. In this part, we introduce the main components of the data construction. Details of processing and implementation are shown in the Appendix A.",
+        "bbox": [
+            112,
+            219,
+            487,
+            282
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ASR Corpus We build the error correction instances from two Chinese-English code-switching ASR datasets: SEAME and ASCEND. SEAME is a large Chinese-English code-switching spontaneous speech corpus, collected from residents of Malaysia and Singapore. The corpus contains 30 hours of word-level transcribed code-switching speech data. ASCEND is a high-quality Chinese-English code-switching corpus built on spontaneous multi-turn conversation collected in Hong Kong. It comprises 10.62 hours of clean speech data collected from dialogues.",
+        "bbox": [
+            112,
+            284,
+            487,
+            475
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ASR System With code-switching ASR datasets, we adopt the ASR system to obtain recognition results. For being close to the real recognition setups, we employ an advanced ASR model architecture, namely XLSR-53 (Xu et al., 2021). We fine-tune the XLSR-53 model and obtain the recognition results for datasets. The code-switching ASR error correction instances in our datasets are built by the pairs of manually annotated transcriptions and corresponding recognition results.",
+        "bbox": [
+            112,
+            478,
+            487,
+            637
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Annotator We develop an automatic annotator for the Chinese-English code-switching corpus, based on prior works in English and Chinese (Bryant et al., 2017; Hinson et al., 2020). It takes three steps, tokenization, alignment, and merging, to generate annotations that are suitable for MaxMatch  $(M_2)$  evaluation metrics. Besides, the automatic annotator is able to classify errors into four types: redundant (R), missing (M), word selection (S), and word ordering (W). It allows researchers to conduct more detailed experiments and analyses with different error types.",
+        "bbox": [
+            112,
+            639,
+            487,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            845,
+            226,
+            859
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Language Distribution",
+        "text_level": 1,
+        "bbox": [
+            112,
+            866,
+            351,
+            881
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As the code-switching datasets contain multiple languages, the observation of language distribu",
+        "bbox": [
+            112,
+            887,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/dc9f5c98410dafd7432b669aeef661583b69baa3ca772c155fcd11da9e161c0f.jpg",
+        "table_caption": [
+            "Table 2: Statistics of our proposed code-switching ASR error correction datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>SEAME-C</td><td>ASCEND-C</td></tr><tr><td colspan=\"3\">At dialogue level</td></tr><tr><td>Code-switching dialogues</td><td>100%</td><td>100%</td></tr><tr><td colspan=\"3\">At sentence level</td></tr><tr><td>Code-switching sentences</td><td>76.6%</td><td>27.8%</td></tr><tr><td>Chinese sentences</td><td>14.3%</td><td>48.6%</td></tr><tr><td>English sentences</td><td>9.1%</td><td>23.6%</td></tr><tr><td colspan=\"3\">At token level</td></tr><tr><td>Chinese tokens</td><td>83.7%</td><td>92.5%</td></tr><tr><td>English tokens</td><td>16.3%</td><td>7.5%</td></tr></table>",
+        "bbox": [
+            510,
+            215,
+            882,
+            342
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Proportion of languages used in datasets at dialogue, sentence and token levels.",
+        "bbox": [
+            507,
+            350,
+            880,
+            379
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tion is important for the analysis. We calculate the proportion of languages used in the datasets at dialogue, utterance, and token levels. Table 3 presents the details.",
+        "bbox": [
+            505,
+            406,
+            882,
+            468
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "From the table, we can find that the sentence-level language distribution of two datasets is quite different. In the SEAME-C dataset, the code-switching sentences take the main proportion, accounting for  $76.6\\%$ . As for the ASCEND-C dataset, the Chinese sentences account for the largest proportion, while the code-switching instances only account for  $27.8\\%$ . In terms of token-level language distribution, both two datasets are dominated by Chinese tokens. It shows that the code-switching phenomenon in Asia area mainly occurs by mixing a few English words into Chinese speech. Considering that, we generate pseudo data from Chinese sentences in our data augmentation method in the following sections.",
+        "bbox": [
+            505,
+            470,
+            882,
+            711
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Error Distribution",
+        "text_level": 1,
+        "bbox": [
+            507,
+            722,
+            715,
+            736
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With the proposed automatic annotator, the errors in datasets are classified into four types. To analyze the characteristic of ASR errors in proposed datasets, we calculate the error distributions on the test sets of SEAME-C and ASCEND-C datasets. The results are presented in Table 4.",
+        "bbox": [
+            505,
+            741,
+            882,
+            837
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As shown in the table, the error distributions on the test sets of two datasets are similar. The word selection errors account for the largest proportion, with  $85.7\\%$  in SEAME-C and  $81.3\\%$  in ASCEND-C. This type of error is caused by mistakes in the",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8077",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f11f88f4576ee1fef85abe9f62b7fe6605112df0c405a599335197be6439173.jpg",
+        "image_caption": [
+            "Figure 1: An example of the process of generating pseudo data in our controllable iterative method. We highlight the edited parts of sentences in each iteration. The controllable iterative method produces pseudo training instances by iteratively generating outputs from input sentences. We apply the sequence editing model to generate synthetic data and build training pairs with iterations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            131,
+            84,
+            868,
+            262
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/17d88b2a74b9dcd25a50b5185e9d67a7608b37f0272cfff736bfd3f8819d85de.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Error Type</td><td colspan=\"2\">SEAME-C</td><td colspan=\"2\">ASCEND-C</td></tr><tr><td>#</td><td>%</td><td>#</td><td>%</td></tr><tr><td>Redundant (R)</td><td>908</td><td>5.6</td><td>134</td><td>11.2</td></tr><tr><td>Missing (M)</td><td>1,381</td><td>8.6</td><td>86</td><td>7.2</td></tr><tr><td>Word Selection (S)</td><td>13,831</td><td>85.7</td><td>972</td><td>81.3</td></tr><tr><td>Word Ordering (W)</td><td>15</td><td>0.1</td><td>4</td><td>0.3</td></tr></table>",
+        "bbox": [
+            124,
+            354,
+            478,
+            432
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Error distribution on the test sets of SEAME-C and ASCEND-C datasets. The # columns present the amounts of specific error.",
+        "bbox": [
+            112,
+            441,
+            487,
+            485
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "recognition of words or phrases with similar pronunciation. Besides, there are a number of missing and redundant errors, which is challenging to correct for requiring to change the length of the input. Additionally, the frequency of word ordering errors is very few. We can find that the code-switching ASR error correction is a task with imbalanced error distribution, and we need to consider it when designing the correction method.",
+        "bbox": [
+            112,
+            510,
+            489,
+            653
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Controllable Iterative Data Augmentation Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            667,
+            381,
+            700
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To address the problem of lacking sufficient training data, we propose the controllable iterative data augmentation method. Due to the shortage of high-quality Chinese-English code-switching corpus, our method is designed to directly generate training instances from monolingual texts. Our method consists of two parts: sequence editing model and pseudo data generation algorithm.",
+        "bbox": [
+            112,
+            709,
+            487,
+            838
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Sequence editing model is the fundamental structure of our method. It adopts the Transformer architecture and the predefined editing operations called transformations to generate raw augmented data from seed corpus. Pseudo data generation algo",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "rithm plays a crucial role in the proposed method. It adopts the sequence editing model to iteratively produce pseudo training data under the control of confidence bias. With our designed transformations, the sequence editing model is able to directly generate training instances from monolingual texts. Figure 1 shows an example of the process of generating pseudo data in our controllable iterative method.",
+        "bbox": [
+            507,
+            359,
+            884,
+            502
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Sequence Editing Model",
+        "text_level": 1,
+        "bbox": [
+            507,
+            514,
+            746,
+            531
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The data augmentation method for the code-switching ASR error correction task aims to generate instances with errors as pseudo training data. Error generation can be treated as the process of noising the correct source input with edit operations. Based on this idea, we adopt the sequence editing model to obtain the editing sequence required to produce output with errors.",
+        "bbox": [
+            507,
+            536,
+            882,
+            665
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.1 Predefined Transformations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            674,
+            789,
+            688
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The sequence editing method requires predefined editing operations called transformations. In this work, we adopt the custom token-level transformations developed in GECToR (Omelianchuk et al., 2020), the details are shown in Appendix B. These transformations are originally designed for English texts. We denote these transformations except $TRANSLATION$ as noising transformations. For generating code-switching sentences from monolingual sources, we introduce crosslingual transformations into original ones. Firstly, we add the translation transformation ($TRANSLATION$) that replaces the current token with the most common translation in a Chinese-English",
+        "bbox": [
+            507,
+            694,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8078",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "dictionary. Secondly, to cover the specific cases that the most common translation can not fit, we insert additional Chinese characters to the vocabulary of $APPEND$ and $REPLACE$ operations. With predefined transformations, we can get the target sentence by applying the sequential outputs of transformations to the tokens in the source sentence.",
+        "bbox": [
+            112,
+            84,
+            489,
+            211
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1.2 Model Architecture",
+        "text_level": 1,
+        "bbox": [
+            112,
+            224,
+            329,
+            239
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The sequence editing model is an encoder made up of Transformer architecture stacked with two linear layers with softmax layers. Considering the strong performance of pre-trained model, the Transformer architecture is initiated with multilingual pre-trained model m-BERT (Devlin et al., 2019). The model is able to generate an editing operation sequence from the source input, and then apply the operations on the source to obtain the target output.",
+        "bbox": [
+            112,
+            244,
+            489,
+            388
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To make the generation process of the model controllable, we add a confidence bias to the decoding process. The confidence bias is a parameter applied to change the generation probability of transformations. For example, adding a positive confidence bias to $KEEP$ can increase the probability of generating the $KEEP$. By adjusting the confidence bias, the decoding process of sequence editing model is under control.",
+        "bbox": [
+            112,
+            391,
+            489,
+            535
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1.3 Training Process",
+        "text_level": 1,
+        "bbox": [
+            112,
+            546,
+            307,
+            562
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To generate pseudo data, we need to train the sequence editing model to generate noisy sentences from correct ones at first.",
+        "bbox": [
+            112,
+            567,
+            489,
+            613
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Given a pair of a monolingual sentence  $x$  without noise and a code-switching sentence with noise  $y$ . Then we convert the sentence  $y$  to a sequence of transformations  $z$ . Let  $\\epsilon$  denote the preset confidence bias and  $\\Theta$  denote all trainable parameters of the sequence editing model  $\\phi$ . Our objective is to find the optimal parameter set  $\\hat{\\Theta}$  that minimizes the following negative log-likelihood function:",
+        "bbox": [
+            112,
+            615,
+            489,
+            744
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL = - \\log P (z | x, \\epsilon , \\Theta) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            210,
+            763,
+            487,
+            780
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $P(z|x,\\epsilon ,\\Theta)$  denotes the conditional probability.",
+        "bbox": [
+            112,
+            789,
+            487,
+            821
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After training on constructed instances, our model can not only learn the editing operations to apply errors to sentences but also the transformations to generate English-Chinese code-switching texts. Some details of training are shown in Appendix C.",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1 Generating pseudo data",
+        "text_level": 1,
+        "bbox": [
+            510,
+            83,
+            788,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Input: Monolingual Sentence  $x_0$ ; Number of iterations  $T$ ; Sequence editing model  $\\phi$ ; Confidence bias  $\\epsilon$ ; Similarity Discriminator  $\\psi$ ; Hyper parameter  $\\lambda, \\epsilon_0, \\epsilon_1$",
+        "bbox": [
+            507,
+            99,
+            884,
+            164
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Output: Synthetic erroneous samples set  $S$",
+        "bbox": [
+            515,
+            165,
+            842,
+            180
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Function:  $Gen(x_0, T, \\lambda, \\epsilon_0, \\epsilon_1)$ :",
+        "bbox": [
+            515,
+            181,
+            761,
+            196
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [],
+        "code_body": "1: for  $t\\gets 1$  to  $T$  do  \n2: if  $t = 1$  then  \n3:  $\\epsilon \\gets \\epsilon_0$   \n4: else  \n5:  $\\epsilon \\gets \\epsilon_1$   \n6:  $x_{t}\\gets \\phi (x_{t - 1},\\epsilon)$   \n7: if  $t\\geq 2$  then  \n8:  $p\\gets \\psi (x_1,x_t)$   \n9: if  $p\\geq \\lambda$  then  \n10:  $S.\\text{add}(x_1,x_t)$   \nreturn S",
+        "bbox": [
+            517,
+            199,
+            731,
+            373
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Pseudo Data Generation Algorithm",
+        "text_level": 1,
+        "bbox": [
+            507,
+            401,
+            833,
+            418
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To generate pseudo training data from the monolingual corpus for code-switching ASR error correction, we have to generate both code-switching target sentences and source sentences with ASR errors. For this objective, we propose an iterative generation approach to obtain pseudo training pairs for code-switching ASR error correction with only one model in multiple iterations. The algorithm for generating pseudo training data is summarized in Algorithm 1.",
+        "bbox": [
+            507,
+            424,
+            882,
+            583
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Given a monolingual sentence from seed corpus  $x_0$ . We apply  $T$  times of iterations to the sequence editing model  $\\phi$  to generate data.",
+        "bbox": [
+            507,
+            586,
+            880,
+            634
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the first iteration, we adjust the confidence bias to  $\\epsilon_0$  that limits the noising transformations and makes the model use cross-lingual transformations and $KEEP$ to generate code-switching sentence  $x_{1}$ . This process can be formulated as:",
+        "bbox": [
+            507,
+            634,
+            880,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx _ {1} \\leftarrow \\phi \\left(x _ {0}, \\epsilon_ {0}\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            635,
+            734,
+            882,
+            751
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the  $t$ -th  $(1 < t \\leq T)$  iteration, the confidence bias is set to  $\\epsilon_{1}$  that balances the generation probability of each transformation. It makes the model apply diverse noises to the output of the  $(t-1)$ -th iteration  $x_{t-1}$  and obtain the pseudo instances with errors  $x_{t}$ . This process can be formulated as:",
+        "bbox": [
+            507,
+            760,
+            882,
+            857
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx _ {t} \\leftarrow \\phi \\left(x _ {t - 1}, \\epsilon_ {1}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            628,
+            877,
+            882,
+            892
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $1 < t \\leq T$ .",
+        "bbox": [
+            509,
+            903,
+            640,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8079",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "With above pseudo instances  $x_{1}, \\ldots, x_{T}$ , we treat the output in the first iteration as the correct target sentence and the output in the later turn as noised source sentence to get the sentence pairs.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In order to improve the quality of synthetic samples, we use BERTScore (Zhang et al., 2019) as the similarity discriminator to filter sentence pairs with low similarity. Given a sentence pair  $(x_{1}, x_{t})$ $(1 < t \\leq T)$ , we use similarity discriminator  $\\psi$  to get a score  $p \\in [0,1]$  that presents the semantic similarity between sentences  $x_{1}$  and  $x_{t}$ . We set a threshold  $\\lambda$  that if the score  $p$  of a sentence pair is greater than the threshold, the pair can be selected as an augmented training sample.",
+        "bbox": [
+            112,
+            149,
+            487,
+            310
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            323,
+            321,
+            340
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present our experimental setup including datasets, ASR error correction model, and data augmentation baseline. More details of implementation are presented in Appendix D.",
+        "bbox": [
+            112,
+            350,
+            489,
+            414
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Dataset for Experiment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            426,
+            344,
+            442
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Training and Test Data Firstly, we train the model on pseudo training instances generated from data augmentation method. Then, we fine-tune the code-switching ASR error correction model on our proposed Chinese-English code-switching ASR error correction datasets: SEAME-C and ASCEND-C. We take  $M^2$  scorer and mixed error rate (MER)(Adel et al., 2013) as evaluation metrics.",
+        "bbox": [
+            112,
+            449,
+            487,
+            592
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Seed Corpus for Pseudo Training Data Generation For data augmentation methods, we need to generate pseudo training instances from a seed corpus. Considering proposed datasets were mainly collected from dialogues, we pick the OpenSubtitles dataset (Lison and Tiedemann, 2016), which is a large collection of subtitles in various languages for movies and TV programs, as the seed corpus.",
+        "bbox": [
+            112,
+            595,
+            489,
+            722
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 ASR Error Correction Model",
+        "text_level": 1,
+        "bbox": [
+            112,
+            736,
+            391,
+            750
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In multilingual grammatical error correction task, the adoption of multilingual version of pre-trained model has achieved great success (Rothe et al., 2021). So we pick two strong multilingual pretrained models mT5 and mBART as the code-switching ASR error correction models to verify the effectiveness of our data augmentation method.",
+        "bbox": [
+            112,
+            757,
+            487,
+            869
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Considering that large language models (LLM) have made significant advancements in various NLP tasks. We adopt the most powerful LLM",
+        "bbox": [
+            112,
+            871,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ChatGPT as the baseline. In our experiment, we test the performance of the GPT-3.5-turbo model on the code-switching ASR error correction task with zero-shot and few-shot settings. The prompt templates for ChatGPT are shown in Appendix F.",
+        "bbox": [
+            507,
+            84,
+            882,
+            165
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.3 Data Augmentation Baseline",
+        "text_level": 1,
+        "bbox": [
+            507,
+            200,
+            779,
+            217
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "There are other ways to generate pseudo code-switching ASR error correction training instances. In our experiment, we implement two data augmentation baselines to compare with our controllable iterative data augmentation method. Firstly, we apply a code-switching text generation approach to get the Chinese-English code-switching texts from a parallel seed corpus. With the code-switching texts, we implement the rule-based and back-translation-based data augmentation methods to improve the ASR error correction models.",
+        "bbox": [
+            507,
+            237,
+            884,
+            413
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Code-switching Text Generation Following previous work (Hussein et al., 2022), we apply the EC theory (Pratapa et al., 2018) to generate Chinese-English code-switching texts. By taking the pairs of parallel sentences in seed corpus as input, the method generates the code-switching texts by replacing the token in the parse tree with its word-level alignments. In our experiment, we adopt the fast-align tool (Dyer et al., 2013) to get the word-level alignments and parse the sentences with a neural graph-based parser(Dozat and Manning).",
+        "bbox": [
+            507,
+            420,
+            882,
+            598
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Rule-based Method The rule-based method is a simple way to generate errors efficiently. We design a rule-based method that generates synthetic training data with five rules: delete, add, replace, shuffle, spelling error. The details of the rules are introduced in Appendix E. With above rules, we can get synthetic training samples with various errors. These synthetic training samples can be used to improve the performance of the code-switching ASR error correction system.",
+        "bbox": [
+            507,
+            605,
+            882,
+            766
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Back-translation Method Inspired by the use of back-translation in downstream tasks (Xie et al., 2018), we train a seq2seq noisy model (like the back-translation model) to generate noisy sentences from correct ones. Then the output of the noising model is paired with the input and the sentence pairs are then used as pseudo data. In our experiment, we consider adopting a base mT5 model as the noising model for our baseline.",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8080",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/33861ebe49c2b790f1bedbcc365359feeec5004654686eb14906447a4b61dd89.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Correction Model</td><td rowspan=\"2\">Augmentation Method</td><td rowspan=\"2\">Augmented Data Size</td><td colspan=\"4\">SEAME-C</td><td colspan=\"4\">ASCEND-C</td></tr><tr><td>P↑</td><td>R↑</td><td>F0.5↑</td><td>MER↓</td><td>P↑</td><td>R↑</td><td>F0.5↑</td><td>MER↓</td></tr><tr><td rowspan=\"4\">mT5</td><td>-</td><td>-</td><td>26.2</td><td>18.7</td><td>24.3</td><td>21.14</td><td>20.7</td><td>15.3</td><td>19.3</td><td>40.31</td></tr><tr><td>Rules</td><td>5M</td><td>27.4</td><td>19.9</td><td>25.4</td><td>20.74</td><td>23.9</td><td>17.8</td><td>22.4</td><td>37.57</td></tr><tr><td>Back-translation</td><td>4.1M</td><td>27.1</td><td>20.5</td><td>25.5</td><td>20.70</td><td>23.3</td><td>17.4</td><td>21.8</td><td>38.21</td></tr><tr><td>CI</td><td>3.3M</td><td>28.5</td><td>20.7</td><td>26.5</td><td>20.45</td><td>25.0</td><td>19.6</td><td>23.7</td><td>36.35</td></tr><tr><td rowspan=\"4\">mBART</td><td>-</td><td>-</td><td>23.8</td><td>17.1</td><td>22.1</td><td>21.65</td><td>19.6</td><td>14.9</td><td>18.4</td><td>40.12</td></tr><tr><td>Rules</td><td>5M</td><td>25.4</td><td>18.9</td><td>23.8</td><td>21.18</td><td>22.5</td><td>17.2</td><td>21.2</td><td>38.63</td></tr><tr><td>Back-translation</td><td>4.1M</td><td>25.7</td><td>19.1</td><td>24.0</td><td>21.03</td><td>22.8</td><td>18.0</td><td>21.7</td><td>38.44</td></tr><tr><td>CI</td><td>3.3M</td><td>26.9</td><td>21.0</td><td>25.5</td><td>20.73</td><td>24.1</td><td>18.3</td><td>22.7</td><td>37.50</td></tr><tr><td>GPT-3.5-turbo(zero-shot)</td><td>-</td><td>-</td><td>7.5</td><td>8.1</td><td>7.6</td><td>40.01</td><td>5.7</td><td>3.9</td><td>5.2</td><td>71.28</td></tr><tr><td>GPT-3.5-turbo(3-shot)</td><td>-</td><td>-</td><td>13.3</td><td>12.6</td><td>13.1</td><td>35.44</td><td>7.2</td><td>4.8</td><td>6.5</td><td>65.62</td></tr></table>",
+        "bbox": [
+            124,
+            80,
+            873,
+            229
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Comparison results of each method. Augmented data sizes show the amounts of additional training sentences used in each method.",
+        "bbox": [
+            112,
+            244,
+            882,
+            273
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            297,
+            327,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.1 Comparison Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            326,
+            319,
+            342
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the performance of our controllable iterative data augmentation method on proposed datasets. Table 5 shows the results.",
+        "bbox": [
+            112,
+            350,
+            487,
+            397
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Firstly, we compare our method with the two data augmentation baselines. For the fairness of comparison, we train the correction models with the similar scale of augmentation data. Here we use pseudo data generated from the controllable iterative method with three iterations to train the correction model.",
+        "bbox": [
+            112,
+            399,
+            487,
+            511
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As we can see from the table, our proposed controllable iterative data augmentation method achieves the best performance in both  $M^2$  scorer and MER metrics on SEAME-C and ASCEND-C datasets. It outperforms the two baselines for both mT5 and mBART correction methods, which shows the effectiveness and robustness of our controllable iterative method in Chinese-English code-switching ASR error correction task.",
+        "bbox": [
+            112,
+            514,
+            487,
+            657
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Besides, we test ChatGPT (GPT-3.5-turbo) on the proposed dataset. From the table, we can find that the performance of ChatGPT is poor in this task. Under the zero-shot setting, it can only correct a few errors. By learning from examples in the prompt, the performance of ChatGPT under few-shot setting has significantly improved, but still far from the performance of mT5 and mBART models.",
+        "bbox": [
+            112,
+            659,
+            487,
+            788
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The experiment results show that the code-switching ASR error correction task is challenging at present. Due to the scarcity of data and the complexity of code-switching text, LLM method is far from achieving satisfactory results in this task. We present that our data augmentation is an effective method to improve the performance of Chinese-English code-switching ASR error correc",
+        "bbox": [
+            112,
+            790,
+            487,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b568d66e98e23a3c7a389f607de6a2788e5b434ab08453fe507dbade1ccefc78.jpg",
+        "image_caption": [
+            "tion models.",
+            "Figure 2: Results of the mT5 correction model on SEAME-C when trained with outputs collected from different iterations of our controllable iterative method."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            326,
+            867,
+            458
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2 Iteration of Proposed Method",
+        "text_level": 1,
+        "bbox": [
+            507,
+            543,
+            789,
+            558
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We build the pseudo training pairs with the outputs of the controllable iterative method in multiple iterations. In this section, we try to explore the influence of the iteration number on the generation of augmented data. Figure 2 shows the results of the mT5 correction model on the SEAME-C dataset when it is trained with outputs collected from different iterations. The details of results are shown in Appendix G.",
+        "bbox": [
+            505,
+            564,
+            882,
+            709
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As the iteration progresses, controllable iterative method continuously generates noisy code-switching data. By collecting the outputs in iterations, we can obtain more pseudo training pairs at each new iteration. In general, using a larger scale of training data may achieve better performance. However, the experimental results show the bottleneck of the data augmentation method that further increasing augmented training data at the 3-th iteration even has a negative impact on performance. This bottleneck might be caused by the quality reduction of the generated data after a few iterations. Considering the balance between perfor",
+        "bbox": [
+            507,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8081",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b304c8d869951f24cc72b5ab3d867ec58366fa6c3d159829dfd69cbfbe05d7bf.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Augmentation Method</td><td>P↑</td><td>R↑</td><td>F0.5↑</td><td>MER↓</td></tr><tr><td rowspan=\"3\">SEAME-C</td><td>Rules</td><td>27.4</td><td>19.9</td><td>25.4</td><td>20.74</td></tr><tr><td>CI</td><td>28.5</td><td>20.7</td><td>26.5</td><td>20.45</td></tr><tr><td>Rules+CI</td><td>29.2</td><td>21.3</td><td>27.2</td><td>20.15</td></tr><tr><td rowspan=\"3\">ASCEND-C</td><td>Rules</td><td>23.9</td><td>17.8</td><td>22.4</td><td>37.57</td></tr><tr><td>CI</td><td>25.0</td><td>19.6</td><td>23.7</td><td>36.35</td></tr><tr><td>Rules+CI</td><td>25.6</td><td>20.8</td><td>24.5</td><td>35.46</td></tr></table>",
+        "bbox": [
+            124,
+            80,
+            485,
+            181
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "mance and computing costs, we select the outputs generated in the third iteration as the augmented data used for Chinese-English code-switching ASR error correction task.",
+        "bbox": [
+            112,
+            274,
+            487,
+            338
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.3 Combination of Augmented Data",
+        "text_level": 1,
+        "bbox": [
+            112,
+            351,
+            420,
+            366
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Rule-based and model-based data augmentation methods are two different approaches to pseudo training data generation. Therefore, the distributions of errors generated by these two types of data augmentation methods might be different in general. In this section, we try to explore the effect of combining the data generated by rule-based and our methods.",
+        "bbox": [
+            112,
+            372,
+            487,
+            500
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In our experiments, we pick the mT5 model that achieves higher evaluation scores in previous experiments as the error correction model. The results of using the combined pseudo data are shown in Table 6. From the results, we can find that training mT5 with the combined pseudo data achieves the best scores on our datasets. The use of the data generated by the rule-based method can further improve the performance of the error correction model.",
+        "bbox": [
+            112,
+            502,
+            487,
+            661
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We further analyze the results of different error types. With the automatic annotator, we evaluate the performance in four types: redundant (R), missing (M), word selection (S), and word ordering (W). We conduct the experiment on our SEAME-C dataset with mT5 as the error correction model. The results are shown in Table 7.",
+        "bbox": [
+            112,
+            663,
+            489,
+            776
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/07eba9bed302e7f6c1e0f0dd4ac7a0fd7f4436f6da096452a74ea397cf59e893.jpg",
+        "table_caption": [
+            "Table 6: Results of the mT5 correction model on datasets when trained with the combination of pseudo data generated by controllable iterative (CI) and rule-based methods."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Error Type</td><td>R</td><td>M</td><td>S</td><td>W</td></tr><tr><td>Augmentation Method</td><td>F0.5</td><td>F0.5</td><td>F0.5</td><td>F0.5</td></tr><tr><td>-</td><td>8.1</td><td>16.3</td><td>21.1</td><td>9.8</td></tr><tr><td>Rules</td><td>10.8</td><td>25.6</td><td>26.1</td><td>11.1</td></tr><tr><td>CI</td><td>17.1</td><td>20.5</td><td>28.7</td><td>14.7</td></tr><tr><td>Rules+CI</td><td>18.7</td><td>21.1</td><td>29.1</td><td>7.9</td></tr></table>",
+        "bbox": [
+            121,
+            785,
+            480,
+            860
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Results of mT5 correction model on SEAME-C with different augmentation methods on different error types.",
+        "bbox": [
+            112,
+            871,
+            487,
+            914
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As can be seen from the results, data augmentation methods affect the performance of the correction model with respect to different error types. Compared with the controllable iterative method, the rule-based method achieves higher scores on missing errors. The correction model benefits from the operations such as adding and deleting tokens defined in the rules that provide sufficient instances with missing errors. Generally, combining controllable iterative and rule-based data augmentation methods can cover most error types and generate samples with high diversity. It enables the correction model to achieve better performance.",
+        "bbox": [
+            507,
+            84,
+            885,
+            294
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            311,
+            640,
+            326
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we explore the code-switching ASR error correction task. To enable the research, we first propose two Chinese-English code-switching ASR correction datasets. To address the problem of lacking sufficient training data, we propose the controllable iterative data augmentation method that can generate code-switching ASR error correction instances from the monolingual corpus iteratively. Experiment results show that our method achieves the best performance compared with the rule-based augmentation method, back-translation augmentation method, and ChatGPT method. We further find that combining the data generated by both rule-based method and controllable iterative method can further improve the performance of the error correction model. In future work, we will apply our data augmentation method to other code-switching tasks and also test our method in other language pairs.",
+        "bbox": [
+            507,
+            340,
+            885,
+            646
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            664,
+            613,
+            678
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we only conduct experiments on Chinese-English datasets. There are code-switching phenomena that occur in other languages, for example, the Spanish-English and French-Italian code-switching in America. The empirical conclusions we draw in this study might be changed under different language settings. Another limitation of our study is the lack of manual evaluation. Considering the challenge of the code-switching ASR error correction task, manual evaluation can cover the cases that are evaluated inaccurately by automatic metrics. More interesting conclusions might be drawn by conducting the manual evaluation.",
+        "bbox": [
+            507,
+            694,
+            885,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8082",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            285,
+            101
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by National Key R & D Program of China (2021YFF0901502), National Science Foundation of China (No. 62161160339), State Key Laboratory of Media Convergence Production Technology and Systems and Key Laboratory of Science, Technology and Standard in Press Industry (Key Laboratory of Intelligent Press Media Technology). We appreciate the anonymous reviewers for their helpful comments. Xiaojun Wan is the corresponding author.",
+        "bbox": [
+            112,
+            109,
+            490,
+            269
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            296,
+            213,
+            311
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Heike Adel, Ngoc Thang Vu, Franziska Kraus, Tim Schlippe, Haizhou Li, and Tanja Schultz. 2013. Recurrent neural network language modeling for code switching conversational speech. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 8411-8415. IEEE.",
+            "Carlo Aliprandi, Cristina Scudellari, Isabella Gallucci, Nicola Piccinini, Matteo Raffaelli, Arantza del Pozo, Aitor Alvarez, Haritz Arzelus, Renato Cassaca, Tiago Luis, et al. 2014. Automatic live subtitling: state of the art, expectations and current trends. In Proceedings of NAB Broadcast Engineering Conference: Papers on Advanced Media Technologies, Las Vegas, volume 13.",
+            "Peter Auer. 2013. Code-switching in conversation: Language, interaction and identity. Routledge.",
+            "Christopher Bryant, Mariano Felice, and Edward Briscoe. 2017. Automatic annotation and evaluation of error types for grammatical error correction. Association for Computational Linguistics.",
+            "Hannan Cao, Wenmian Yang, and Hwee Tou Ng. 2023. Mitigating exposure bias in grammatical error correction with data augmentation and reweighting. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2115-2127.",
+            "Rodolfo Corona, Jesse Thomason, and Raymond Mooney. 2017. Improving black-box speech recognition using semantic parsing. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 122-127.",
+            "Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 568-572.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep"
+        ],
+        "bbox": [
+            115,
+            319,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186.",
+            "Suman Dowlagar and Radhika Mamidi. 2023. A code-mixed task-oriented dialog dataset for medical domain. Computer Speech & Language, 78:101449.",
+            "Timothy Dozat and Christopher D Manning. Deep bia-affine attention for neural dependency parsing.",
+            "Chris Dyer, Victor Chahuneau, and Noah A Smith. 2013. A simple, fast, and effective reparameterization of ibm model 2. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 644-648.",
+            "Luis Fernando D'Haro and Rafael E Banchs. 2016. Automatic correction of asr outputs by using machine translation. In Interspeech, volume 2016, pages 3469-3473.",
+            "Roman Grundkiewicz, Marcin Junczys-Dowmunt, and Kenneth Heafield. 2019. Neural grammatical error correction systems with unsupervised pre-training on synthetic data. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 252-263.",
+            "Jinxi Guo, Tara N Sainath, and Ron J Weiss. 2019. A spelling correction model for end-to-end speech recognition. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5651-5655. IEEE.",
+            "Charles Hinson, Hen-Hsen Huang, and Hsin-Hsi Chen. 2020. Heterogeneous recycle generation for chinese grammatical error correction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2191-2201.",
+            "Amir Hussein, Shammur Absar Chowdhury, Ahmed Abdelali, Najim Dehak, and Ahmed Ali. 2022. Code-switching text augmentation for multilingual speech processing. arXiv e-prints, pages arXiv-2201.",
+            "Yunqi Gao Jiaju Mei, Yiming Zhu. 1996. Tongyici cilin. In Shanghai Lexicographical Publishing House.",
+            "Rebecca Jonson. 2006. Dialogue context-based reranking of asr hypotheses. In 2006 IEEE Spoken Language Technology Workshop, pages 174-177. IEEE.",
+            "Shun Kiyono, Jun Suzuki, Masato Mita, Tomoya Mizumoto, and Kentaro Inui. 2019. An empirical study of incorporating pseudo data into grammatical error correction. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1236-1242."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8083",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jared Lichtarge, Chris Alberti, Shankar Kumar, Noam Shazeer, Niki Parmar, and Simon Tong. 2019. Corpora generation for grammatical error correction. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3291-3301.",
+            "Pierre Lison and Jörg Tiedemann. 2016. Opensubtitles2016: Extracting large parallel corpora from movie and tv subtitles.",
+            "Holy Lovenia, Samuel Cahyawijaya, Genta Indra Winata, Peng Xu, Xu Yan, Zihan Liu, Rita Frieske, Tiezheng Yu, Wenliang Dai, Elham J Barezi, et al. 2021. Ascend: A spontaneous chinese-english dataset for code-switching in multi-turn conversation. arXiv preprint arXiv:2112.06223.",
+            "Dau-Cheng Lyu, Tien-Ping Tan, Eng Siong Chng, and Haizhou Li. 2010. Seame: a mandarin-english code-switching speech corpus in south-east asia. In Eleventh Annual Conference of the International Speech Communication Association.",
+            "Anirudh Mani, Shruti Palaskar, Nimshi Venkat Meripo, Sandeep Konam, and Florian Metze. 2020. Asr error correction and domain adaptation using machine translation. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6344-6348. IEEE.",
+            "Niko Moritz, Takaaki Hori, and Jonathan Le. 2020. Streaming automatic speech recognition with the transformer model. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6074-6078. IEEE.",
+            "Kostiantyn Omelianchuk, Vitaliy Atrasevych, Artem Chernodub, and Oleksandr Skurzhanskyi. 2020. Gector-grammatical error correction: Tag, not rewrite. In Proceedings of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 163-170.",
+            "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32.",
+            "Adithya Pratapa, Gayatri Bhat, Monjit Choudhury, Sunayana Sitaram, Sandipan Dandapat, and Kalika Bali. 2018. Language modeling for code-mixing: The role of linguistic theory based synthetic data. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1543-1553.",
+            "Sascha Rothe, Jonathan Mallinson, Eric Malmi, Sebastian Krause, and Aliaksei Severyn. 2021. A simple recipe for multilingual grammatical error correction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 702-707.",
+            "Djame Seddah, Farah Essaidi, Amal Fethi, Matthieu Futeral, Benjamin Muller, Pedro Ortiz Suarez, Benoit Sagot, and Abhishek Srivastava. 2020. Building a user-generated content north-african arabizi treebank: Tackling hell. In Proceedings of the 58th annual meeting of the Association for Computational Linguistics, pages 1139-1150.",
+            "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86-96.",
+            "Aiman Solyman, Marco Zappatore, Wang Zhenyu, Zeinab Mahmoud, Ali Alfatemi, Ashraf Osman Ibrahim, and Lubna Abdelkareim Gabralla. 2023. Optimizing the impact of data augmentation for low-resource grammatical error correction. Journal of King Saud University-Computer and Information Sciences, 35(6):101572.",
+            "Hsi-Yao Su. 2001. Code-switching between mandarin and taiwanese in three telephone conversations: The negotiation of interpersonal relationships among bilingual speakers in taiwan. In Proc. of the Symposium about Language and Society.",
+            "Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, and Angela Fan. 2020. Multilingual translation with extensible multilingual pretraining and finetuning. arXiv preprint arXiv:2008.00401.",
+            "Jörg Tiedemann and Santhosh Thottingal. 2020. Opusmt-building open translation services for the world. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation. European Association for Machine Translation.",
+            "Deyuan Wang, Tiantian Zhang, Caixia Yuan, and Xiaojie Wang. 2023. Joint modeling for asr correction and dialog state tracking. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE.",
+            "Haoyu Wang, Shuyan Dong, Yue Liu, James Logan, Ashish Kumar Agrawal, and Yang Liu. 2020. Asr error correction with augmented transformer for entity retrieval. In *Interspeech*, pages 1550–1554.",
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45.",
+            "Ziang Xie, Guillaume Genthial, Stanley Xie, Andrew Y Ng, and Dan Jurafsky. 2018. Noising and denoising natural language: Diverse backtranslation for"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8084",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "grammar correction. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 619-628.",
+        "bbox": [
+            129,
+            85,
+            487,
+            151
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Qiantong Xu, Alexei Baevski, and Michael Auli. 2021. Simple and effective zero-shot cross-lingual phoneme recognition. arXiv preprint arXiv:2109.11680.",
+        "bbox": [
+            114,
+            160,
+            487,
+            200
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mt5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498.",
+        "bbox": [
+            114,
+            209,
+            487,
+            302
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations.",
+        "bbox": [
+            114,
+            311,
+            487,
+            363
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Processing for Data Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            376,
+            438,
+            391
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.1 ASR System",
+        "text_level": 1,
+        "bbox": [
+            112,
+            401,
+            262,
+            416
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Here is the process we use XLSR-53 model to generate code-switching ASR error correction instances. Firstly, we build the vocabulary from the transcription in the datasets. Based on the vocabulary of 26 English letters, we add all Chinese characters appearing in the transcriptions to the vocabulary for the generation of Chinese. Then, we fine-tune the XLSR-53 model on the training data from SEAME and ASCEND datasets. We pick the parameters with the best performance in the validation set as our code-switching ASR system for next recognition step. With the fine-tuned XLSR-53 model, we can initially obtain the recognition results required for building the code-switching ASR error correction corpus. As the datasets are spontaneous speeches, the corpus is mainly informal and non-speech sounds often occur, such as laughing and coughing. Given that, we design some simple rules to polish the corpus, such as removing non-speech sounds and repeated words.",
+        "bbox": [
+            112,
+            420,
+            489,
+            743
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.2 Annotator",
+        "text_level": 1,
+        "bbox": [
+            112,
+            753,
+            247,
+            766
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The annotator takes three steps: tokenization, alignment, and merging. For tokenization, we use character-level tokenization for Chinese and word-level tokenization for English. Alignment works by computing the alignment score for each pair of tokens in the source and target sentences. In particular, the Cilin thesaurus (Jiaju Mei, 1996) is leveraged to give a similarity score between Chinese characters. Once the alignment score matrix",
+        "bbox": [
+            112,
+            774,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "is computed, the sequence with the lowest score is returned. At last, we use a simple merging strategy that merges consecutive sequences of edits of the same type.",
+        "bbox": [
+            507,
+            84,
+            880,
+            149
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Transformations in GECToR",
+        "text_level": 1,
+        "bbox": [
+            509,
+            170,
+            798,
+            186
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "GECToR proposed the token-level transformations for the sequence editing grammatical error correction model. Transformations increase the coverage of grammatical error corrections for limited output vocabulary size for the most common types of errors. The work divided transformations into two types: basic transformations and g-transformations.",
+        "bbox": [
+            507,
+            203,
+            884,
+            316
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Basic transformations perform the most common token-level edit operations, such as: keep the current token unchanged (tag $KEEP$), delete current token (tag $DELETE$), append new token  $t_1$  next to the current token (tag $APPEND t1 $) or replace the current token with another token  $t_2$  (tag $REPLACE t2 $).",
+        "bbox": [
+            507,
+            319,
+            882,
+            431
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "G-transformations perform task-specific operations such as: change the case of the current token (\\(CASE\\) tags), merge the current token and the next token into a single one (\\(MERGE\\) tags) and split the current token into two new tokens (\\(SPLIT\\) tags). Moreover, tags from \\)NOUN NUMBER$ and $VERB FORM$ transformations encode grammatical properties for tokens.",
+        "bbox": [
+            507,
+            435,
+            882,
+            564
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Processing for Sequence Editing Model",
+        "text_level": 1,
+        "bbox": [
+            507,
+            586,
+            880,
+            602
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C.1 Processing for Transformations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            619,
+            806,
+            634
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "It takes a three-step pre-processing algorithm to convert a sentence to a sequence of transformations. Firstly, it maps each token from source sentence to sub-sequence of tokens from target sentence. For this purpose, it detects the minimal spans of tokens which define differences between source tokens and target tokens. Thus, such a span is a pair of selected source tokens and corresponding target tokens. To get tags on the token level, it searches for the best-fitting sub-sequence of target tokens by minimizing the modified Levenshtein distance for each source token. Secondly, for each mapping in the list, it finds token-level transformations which convert source token to the target sub-sequence. Finally, it leaves only one transformation for each source token. Through the above three steps, we map the tag to each single source token.",
+        "bbox": [
+            507,
+            645,
+            882,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "8085",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C.2 Processing for Translation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            371,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Our model is aimed to generate code-switching sentences with errors from correct monolingual sentences, but the training pairs in datasets only consist of code-switching texts. Considering that, we adopt an open translation tool OPUS-MT (Tiedemann and Thottingal, 2020) to translate the code-switching ASR transcriptions into monolingual sentences. Specifically, we translate the spans of English texts to Chinese and get the Chinese sentences. With regarding the translated monolingual sentence as source and the code-switching sentence with errors from error correction datasets as target, we train the sequence editing model.",
+        "bbox": [
+            112,
+            110,
+            492,
+            319
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            341,
+            357,
+            357
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In our work, we adopt PyTorch (Paszke et al., 2019) framework and HuggingFace (Wolf et al., 2020) tool to implement all deep neural models. Each model is trained on four NVIDIA A40 GPUs. The details of each model are described as follows.",
+        "bbox": [
+            112,
+            374,
+            489,
+            455
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D.1 ASR Error Correction Model",
+        "text_level": 1,
+        "bbox": [
+            114,
+            475,
+            396,
+            489
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Considering that mT5 and mBART have achieved strong performance in many multilingual tasks, we adopt these pre-trained models to correct ASR errors. In the experiments, we use the base version of mT5 and fine-tune it by using the Adam optimization method with learning rate 5e-4, weight decay parameter 0.01, dropout rate 0.1, and batch size 24. Besides, we use the large version of mBART, and fine-tune it by using the Adam optimization method with learning rate 2e-5, dropout rate 0.1, and batch size 16. Beam search with beam size of 5 is used for decoding.",
+        "bbox": [
+            112,
+            502,
+            489,
+            695
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D.2 Sequence Editing Model",
+        "text_level": 1,
+        "bbox": [
+            114,
+            715,
+            356,
+            732
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The sequence editing model is an encoder made up of Transformer architecture stacked with two linear layers with softmax layers. We adopt the base version of m-BERT to initialize the encoder. Our model consists of 12 Transformer layers and two linear layers. We train the model with learning rate 1e-5, weight decay parameter 0.01, dropout rate 0, and batch size 32. The size of the vocabulary is 12229. It consists of 4971 basic transformations, 29 token-independent g-transformations and 7229 cross-lingual transformations.",
+        "bbox": [
+            112,
+            741,
+            489,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D.3 Pseudo Data Generation Algorithm",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            836,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We add a confidence bias to control the decoding process. In our experiments, the confidence bias \\( \\epsilon_0 \\) is to add -1.0 to the generation probability of all transformations except \\( \\\\( \\text{KEEP} \\\\), \\\\) \\text{TRAN-SLATION} \\\\) and other cross-lingual transformations. The confidence bias \\( \\epsilon_1 \\) is set to keep the original generation probability of all transformations. The threshold \\( \\lambda \\) is set to 0.85.",
+        "bbox": [
+            507,
+            109,
+            885,
+            237
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/d971f9b7833134aea161f655e045083925637fd550f0e430ab8f01dc00179331.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Prompt Template</td></tr><tr><td>Zero-shot</td><td>&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: &quot;You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence.&quot; &quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;Please identify and correct any errors in the following sentence indicated by &lt;input&gt;ERROR &lt;/input&gt; tag, you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible. Remember to format your corrected output results with the tag &lt;output&gt; Your Corrected Version &lt;/output&gt;. Please start: &lt;input&gt;Input Sentence &lt;/input&gt;:&quot;</td></tr><tr><td>Few-shot</td><td>&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: &quot;You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence.&quot; &quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;Please identify and correct any errors in the following sentence indicated by &lt;input&gt;ERROR &lt;/input&gt; tag,you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible.Here are some in-context examples:(1), &lt;input&gt; SRC-1 &lt;/input&gt;: &lt;output&gt;TGT-1 &lt;/output&gt;(2), &lt;input&gt; SRC-1 &lt;/input&gt;: &lt;output&gt;TGT-1 &lt;/output&gt;; ... (n), &lt;input&gt; SRC-n &lt;/input&gt;: &lt;output&gt;TGT-n &lt;/output&gt;; Please feel free to refer to these examples. Remember to format your corrected output results with the tag &lt;output&gt; Your Corrected Version &lt;/output&gt;. Please start: &lt;input&gt;Input Sentence &lt;/input&gt;:&quot;</td></tr></table>",
+        "bbox": [
+            510,
+            250,
+            890,
+            860
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Zero-shot and few-shot prompt template for Chinese-English code-switching ASR error correction.",
+        "bbox": [
+            507,
+            869,
+            882,
+            897
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "8086",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "E Rule-based Method",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            321,
+            98
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The details of the five rules adopted in our proposed rule-based method are introduced in the following.",
+        "bbox": [
+            112,
+            109,
+            487,
+            140
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Delete. Randomly delete a token with a probability.",
+        "bbox": [
+            112,
+            141,
+            487,
+            172
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Add. Firstly, randomly select a word or Chinese character from the vocabulary of the ASR datasets, and then add the selected word or character to a random position.",
+        "bbox": [
+            112,
+            174,
+            487,
+            237
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Replace. Randomly replace a token with a word or Chinese character from the vocabulary of the ASR datasets. In particular, if the selected token is a Chinese character, we select the replacement from the homophones with a probability.",
+        "bbox": [
+            112,
+            239,
+            487,
+            317
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Shuffle. Shuffle the tokens by adding a normal distribution bias to the positions of the tokens.",
+        "bbox": [
+            112,
+            319,
+            487,
+            349
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Spell Error. Randomly apply spell error to an English word. We randomly perturb letters with operations, i.e. substitution, deletion, insertion or transposition of letters.",
+        "bbox": [
+            112,
+            351,
+            487,
+            414
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "F Prompt Template",
+        "text_level": 1,
+        "bbox": [
+            114,
+            426,
+            302,
+            443
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Zero-shot and few-shot prompt templates for Chinese-English code-switching ASR error correction are shown in Table 8.",
+        "bbox": [
+            112,
+            451,
+            489,
+            498
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "G Additional Experiment Results",
+        "text_level": 1,
+        "bbox": [
+            114,
+            511,
+            421,
+            527
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Results of the mT5 correction model on SEAME-C, when trained with outputs collected from different iterations of our method, are shown in Table 9.",
+        "bbox": [
+            112,
+            537,
+            487,
+            583
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/5877df100337da3038d691ac1656f44416084fbea8bd03360c13266b5ab6e907.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Iteration</td><td>Augmented Data Size</td><td>P↑</td><td>R↑</td><td>F0.5↑</td><td>MER↓</td></tr><tr><td>Iteration 1</td><td>1.8M</td><td>28.0</td><td>20.3</td><td>26.1</td><td>20.54</td></tr><tr><td>Iteration 2</td><td>2.6M</td><td>28.2</td><td>20.6</td><td>26.3</td><td>20.49</td></tr><tr><td>Iteration 3</td><td>3.3M</td><td>28.5</td><td>20.7</td><td>26.5</td><td>20.45</td></tr><tr><td>Iteration 4</td><td>3.6M</td><td>28.3</td><td>20.8</td><td>26.4</td><td>20.51</td></tr><tr><td>Iteration 5</td><td>3.8M</td><td>28.3</td><td>21.1</td><td>26.5</td><td>20.40</td></tr></table>",
+        "bbox": [
+            124,
+            594,
+            480,
+            683
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 9: Results of the mT5 correction model on SEAME-C when trained with outputs collected from different iterations of our method.",
+        "bbox": [
+            112,
+            692,
+            487,
+            734
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "8087",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/New Datasets and Controllable Iterative Data Augmentation Method for Code-switching ASR Error Correction/704f5812-1da2-4355-bf1f-7c6e5a7b4a97_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11d4444b9e87de5c4ec8a1f128bf9e6b26399249b4304a722cfb2ee629814c0a
+size 358095

2023/New Datasets and Controllable Iterative Data Augmentation Method for Code-switching ASR Error Correction/full.md

@@ -0,0 +1,403 @@
+# New Datasets and Controllable Iterative Data Augmentation Method for Code-switching ASR Error Correction
+
+Zhaohong Wan $^{1,2}$  and Xiaojun Wan $^{1,2}$  and Wei Peng $^{3}$  and Rongjun Li $^{3}$
+
+$^{1}$ Wangxuan Institute of Computer Technology, Peking University
+
+2The MOE Key Laboratory of Computational Linguistics, Peking University
+
+<sup>3</sup> Artificial Intelligence Application Research Center, Huawei Technologies {xmwzh, wanxiaojun}@pku.edu.cn
+
+# Abstract
+
+With the wide use of automatic speech recognition (ASR) systems, researchers pay more attention to the ASR error correction task to improve the quality of recognition results. In particular, ASR in bilingual or multilingual settings, namely code-switching ASR, has greater challenges and research value. In this paper, we first present code-switching ASR correction datasets obtained from solid ASR systems and automatic annotators. The datasets contain Chinese-English code-switching dialogues of bilingual speakers in Singapore, Malaysia, and Hong Kong. Based on this task, we propose a controllable iterative (CI) data augmentation method for improving the performance of mainstream ASR error correction systems. With a small amount of training data, our proposed method has the ability to iteratively produce abundant pseudo parallel data from the monolingual corpus for Chinese-English code-switching ASR correction. Results of experiments show that our method achieves the best performance compared with the rule-based, back-translation-based data augmentation methods and large language model ChatGPT.
+
+# 1 Introduction
+
+Recently, automatic speech recognition (ASR) system has achieved impressive success with the adoption of advanced deep neural network (Moritz et al., 2020). Learning from training data, the ASR system can match human transcribers for some critical tasks. It has been widely applied in several areas, such as live captioning, intelligent appliance control (Aliprandi et al., 2014).
+
+In some bilingual and multilingual societies, people often speak more than one language in a conversation. The speech phenomenon of switching language among sentences is called code-switching. To explore the code-switching phenomenon, code-switching tasks in many different
+
+<table><tr><td>Error Type</td><td>Redundant</td></tr><tr><td>Source</td><td>我要 start on 我的 a essay
+(I want to start on my a essay.)</td></tr><tr><td>Target</td><td>我要 start on 我的 essay
+(I want to start on my essay.)</td></tr></table>
+
+Table 1: Examples of Chinese-English code-switching ASR error correction in our datasets. The contents in brackets are the translations for non-native Chinese speakers.
+
+language pairs have been introduced (Dowlagar and Mamidi, 2023; Seddah et al., 2020), including the code-switching ASR task.
+
+Employing multilingual ASR systems to solve code-switching ASR is proposed in recent work (Lovenia et al., 2021). However, the performance of systems in many cases is still far from satisfactory. Errors introduced by ASR systems usually affect the performance of downstream tasks. To improve the recognition performance, the code-switching ASR error correction task is necessary.
+
+In this paper, we focus on code-switching ASR error correction between Chinese and English, which is common in the Chinese communities of Asia area. To enable the research and development of code-switching ASR error correction, we propose two datasets SEAME-C and ASCEND-C that are constructed from two Chinese-English code-switching ASR datasets SEAME (Lyu et al., 2010) and ASCEND (Lovenia et al., 2021). In Table 1, we present typical examples of Chinese-English code-switching ASR error correction.
+
+A challenge for the code-switching ASR error correction is the lack of a large scale of training data. Moreover, we can not even collect sufficient high-quality Chinese-English code-switching data. Considering these challenges, we propose a controllable iterative (CI) data augmentation method that can iteratively generate abundant code-switching ASR error correction instances from the monolingual corpus with small-scale labeled training data.
+
+To verify the effectiveness of our proposed method, we implement two data augmentation baselines and evaluate the performance of these methods on proposed datasets at first. With a similar scale of augmented data, our proposed controllable iterative method achieves the best performance in both MaxMatch  $(M^2)$  scorer (Dahlmeier and Ng, 2012) and MER metrics on SEAME-C and ASCEND-C datasets. It performs well for both mT5 (Xue et al., 2021) and mBART (Tang et al., 2020) correction methods. Besides, considering the strong ability of large language models (LLM), we compare the performance of our method with the most powerful LLM ChatGPT. Experiments show that our method outdistances the ChatGPT method, which shows the effectiveness of our method in the Chinese-English code-switching ASR error correction task. Besides, we find that this task is challenging, and LLM method is far from achieving satisfactory results at now.
+
+Our contributions are summarized as follows:
+
+1. We propose two datasets SEAME-C and ASCEND-C for the challenging Chinese-English code-switching ASR error correction task.  
+2. To address the problem of lacking sufficient training data, we propose the controllable iterative data augmentation method that can iteratively generate abundant code-switching ASR error correction instances from the monolingual corpus.  
+3. Extensive experiments show the superiority of our proposed controllable iterative method. Moreover, combining the pseudo data produced by the rule-based and controllable iterative methods can further improve the performance of error correction models.
+
+The resources are publicly available<sup>1</sup>.
+
+# 2 Related Work
+
+Code-switching ASR, the task aiming to recognize the speech that contains more than one language within the sentences, has been increasingly reported in speech technology and linguistic studies recently (Su, 2001; Auer, 2013). However, to our knowledge, there is no research on the code-switching ASR error correction task at present. In this section, we introduce two relevant fields: ASR error correction and data augmentation method in text error correction.
+
+With the successes of ASR systems, the ASR
+
+error correction has become a prevalent task in recent years. The idea of regarding the ASR error correction as the machine translation and applied advanced neural machine translation models is adopted in several previous approaches(Mani et al., 2020; Wang et al., 2020). D'Haro and Banchs (2016) first introduce machine translation to improve automatic transcription. Considering the machine translation models require only parallel text pairs, Guo et al. (2019) propose an approach to utilizing text-only data by training a correction model to explicitly correct the errors. The joint modeling method shows great performance in this task(Wang et al., 2023). Re-ranking ASR n-best hypotheses is also a popular topic for error recovery (Corona et al., 2017; Jonson, 2006).
+
+To address the problem of lack of training data, data augmentation method is widely adopted in text error correction area. Many works (Grundkiewicz et al., 2019; Lichtarge et al., 2019) adopt pre-defined rules to generate sentences with errors. Inspired by back-translation procedure for machine translation (Sennrich et al., 2016), works (Xie et al., 2018; Kiyono et al., 2019) are proposed to train a model to generate erroneous sentences. Recently, researchers focus on the exposure bias problem of data augmentation method and proposed methods to utilize the augmented data in more effective ways (Solyman et al., 2023; Cao et al., 2023).
+
+# 3 Code-switching ASR Error Correction Datasets
+
+In this section, we provide details about our two datasets SEAME-C and ASCEND-C. Specifically, we give the construction process of our datasets at first. Then, we do an analysis of our datasets and show some characteristics of them. In this work, we adopt the same corpus splitting as the original ASR datasets. In addition, we remove some bad cases that are too short and eliminate monolingual sentences in the validation and test sets to ensure the quality of evaluation. Our datasets provide quality resources for the research and development of code-switching ASR error correction. Table 2 describes the statistics of our proposed datasets.
+
+# 3.1 Data Construction
+
+In some Asia's multilingual countries and regions, speaking a mixture of Chinese and English within a sentence is common. We collect data from two Chinese-English code-switching ASR datasets with
+
+<table><tr><td rowspan="2"></td><td colspan="3">SEAME-C</td><td colspan="4">ASCEND-C</td></tr><tr><td>Train</td><td>Test</td><td>Total</td><td>Train</td><td>Val</td><td>Test</td><td>Total</td></tr><tr><td># Dialogues</td><td>267</td><td>30</td><td>297</td><td>40</td><td>5</td><td>4</td><td>49</td></tr><tr><td># Sentences</td><td>48,405</td><td>6,293</td><td>54,698</td><td>9,869</td><td>213</td><td>373</td><td>10,455</td></tr><tr><td># Tokens</td><td>801,548</td><td>99,837</td><td>901,385</td><td>122,348</td><td>3,385</td><td>5,538</td><td>131,271</td></tr><tr><td>Avg. sentences in a dialogue</td><td>181.3</td><td>209.8</td><td>174.2</td><td>246.7</td><td>53.3</td><td>92.3</td><td>213.4</td></tr><tr><td>Avg. tokens in a sentence</td><td>16.6</td><td>15.9</td><td>16.5</td><td>12.4</td><td>15.9</td><td>14.8</td><td>12.6</td></tr></table>
+
+the advanced ASR system and automatic annotator. In this part, we introduce the main components of the data construction. Details of processing and implementation are shown in the Appendix A.
+
+ASR Corpus We build the error correction instances from two Chinese-English code-switching ASR datasets: SEAME and ASCEND. SEAME is a large Chinese-English code-switching spontaneous speech corpus, collected from residents of Malaysia and Singapore. The corpus contains 30 hours of word-level transcribed code-switching speech data. ASCEND is a high-quality Chinese-English code-switching corpus built on spontaneous multi-turn conversation collected in Hong Kong. It comprises 10.62 hours of clean speech data collected from dialogues.
+
+ASR System With code-switching ASR datasets, we adopt the ASR system to obtain recognition results. For being close to the real recognition setups, we employ an advanced ASR model architecture, namely XLSR-53 (Xu et al., 2021). We fine-tune the XLSR-53 model and obtain the recognition results for datasets. The code-switching ASR error correction instances in our datasets are built by the pairs of manually annotated transcriptions and corresponding recognition results.
+
+Annotator We develop an automatic annotator for the Chinese-English code-switching corpus, based on prior works in English and Chinese (Bryant et al., 2017; Hinson et al., 2020). It takes three steps, tokenization, alignment, and merging, to generate annotations that are suitable for MaxMatch  $(M_2)$  evaluation metrics. Besides, the automatic annotator is able to classify errors into four types: redundant (R), missing (M), word selection (S), and word ordering (W). It allows researchers to conduct more detailed experiments and analyses with different error types.
+
+# 3.2 Analysis
+
+# 3.2.1 Language Distribution
+
+As the code-switching datasets contain multiple languages, the observation of language distribu
+
+Table 2: Statistics of our proposed code-switching ASR error correction datasets.  
+
+<table><tr><td></td><td>SEAME-C</td><td>ASCEND-C</td></tr><tr><td colspan="3">At dialogue level</td></tr><tr><td>Code-switching dialogues</td><td>100%</td><td>100%</td></tr><tr><td colspan="3">At sentence level</td></tr><tr><td>Code-switching sentences</td><td>76.6%</td><td>27.8%</td></tr><tr><td>Chinese sentences</td><td>14.3%</td><td>48.6%</td></tr><tr><td>English sentences</td><td>9.1%</td><td>23.6%</td></tr><tr><td colspan="3">At token level</td></tr><tr><td>Chinese tokens</td><td>83.7%</td><td>92.5%</td></tr><tr><td>English tokens</td><td>16.3%</td><td>7.5%</td></tr></table>
+
+Table 3: Proportion of languages used in datasets at dialogue, sentence and token levels.
+
+tion is important for the analysis. We calculate the proportion of languages used in the datasets at dialogue, utterance, and token levels. Table 3 presents the details.
+
+From the table, we can find that the sentence-level language distribution of two datasets is quite different. In the SEAME-C dataset, the code-switching sentences take the main proportion, accounting for  $76.6\%$ . As for the ASCEND-C dataset, the Chinese sentences account for the largest proportion, while the code-switching instances only account for  $27.8\%$ . In terms of token-level language distribution, both two datasets are dominated by Chinese tokens. It shows that the code-switching phenomenon in Asia area mainly occurs by mixing a few English words into Chinese speech. Considering that, we generate pseudo data from Chinese sentences in our data augmentation method in the following sections.
+
+# 3.2.2 Error Distribution
+
+With the proposed automatic annotator, the errors in datasets are classified into four types. To analyze the characteristic of ASR errors in proposed datasets, we calculate the error distributions on the test sets of SEAME-C and ASCEND-C datasets. The results are presented in Table 4.
+
+As shown in the table, the error distributions on the test sets of two datasets are similar. The word selection errors account for the largest proportion, with  $85.7\%$  in SEAME-C and  $81.3\%$  in ASCEND-C. This type of error is caused by mistakes in the
+
+![](images/4f11f88f4576ee1fef85abe9f62b7fe6605112df0c405a599335197be6439173.jpg)  
+Figure 1: An example of the process of generating pseudo data in our controllable iterative method. We highlight the edited parts of sentences in each iteration. The controllable iterative method produces pseudo training instances by iteratively generating outputs from input sentences. We apply the sequence editing model to generate synthetic data and build training pairs with iterations.
+
+<table><tr><td rowspan="2">Error Type</td><td colspan="2">SEAME-C</td><td colspan="2">ASCEND-C</td></tr><tr><td>#</td><td>%</td><td>#</td><td>%</td></tr><tr><td>Redundant (R)</td><td>908</td><td>5.6</td><td>134</td><td>11.2</td></tr><tr><td>Missing (M)</td><td>1,381</td><td>8.6</td><td>86</td><td>7.2</td></tr><tr><td>Word Selection (S)</td><td>13,831</td><td>85.7</td><td>972</td><td>81.3</td></tr><tr><td>Word Ordering (W)</td><td>15</td><td>0.1</td><td>4</td><td>0.3</td></tr></table>
+
+Table 4: Error distribution on the test sets of SEAME-C and ASCEND-C datasets. The # columns present the amounts of specific error.
+
+recognition of words or phrases with similar pronunciation. Besides, there are a number of missing and redundant errors, which is challenging to correct for requiring to change the length of the input. Additionally, the frequency of word ordering errors is very few. We can find that the code-switching ASR error correction is a task with imbalanced error distribution, and we need to consider it when designing the correction method.
+
+# 4 Controllable Iterative Data Augmentation Method
+
+To address the problem of lacking sufficient training data, we propose the controllable iterative data augmentation method. Due to the shortage of high-quality Chinese-English code-switching corpus, our method is designed to directly generate training instances from monolingual texts. Our method consists of two parts: sequence editing model and pseudo data generation algorithm.
+
+Sequence editing model is the fundamental structure of our method. It adopts the Transformer architecture and the predefined editing operations called transformations to generate raw augmented data from seed corpus. Pseudo data generation algo
+
+rithm plays a crucial role in the proposed method. It adopts the sequence editing model to iteratively produce pseudo training data under the control of confidence bias. With our designed transformations, the sequence editing model is able to directly generate training instances from monolingual texts. Figure 1 shows an example of the process of generating pseudo data in our controllable iterative method.
+
+# 4.1 Sequence Editing Model
+
+The data augmentation method for the code-switching ASR error correction task aims to generate instances with errors as pseudo training data. Error generation can be treated as the process of noising the correct source input with edit operations. Based on this idea, we adopt the sequence editing model to obtain the editing sequence required to produce output with errors.
+
+# 4.1.1 Predefined Transformations
+
+The sequence editing method requires predefined editing operations called transformations. In this work, we adopt the custom token-level transformations developed in GECToR (Omelianchuk et al., 2020), the details are shown in Appendix B. These transformations are originally designed for English texts. We denote these transformations except $TRANSLATION$ as noising transformations. For generating code-switching sentences from monolingual sources, we introduce crosslingual transformations into original ones. Firstly, we add the translation transformation ($TRANSLATION$) that replaces the current token with the most common translation in a Chinese-English
+
+dictionary. Secondly, to cover the specific cases that the most common translation can not fit, we insert additional Chinese characters to the vocabulary of $APPEND$ and $REPLACE$ operations. With predefined transformations, we can get the target sentence by applying the sequential outputs of transformations to the tokens in the source sentence.
+
+# 4.1.2 Model Architecture
+
+The sequence editing model is an encoder made up of Transformer architecture stacked with two linear layers with softmax layers. Considering the strong performance of pre-trained model, the Transformer architecture is initiated with multilingual pre-trained model m-BERT (Devlin et al., 2019). The model is able to generate an editing operation sequence from the source input, and then apply the operations on the source to obtain the target output.
+
+To make the generation process of the model controllable, we add a confidence bias to the decoding process. The confidence bias is a parameter applied to change the generation probability of transformations. For example, adding a positive confidence bias to $KEEP$ can increase the probability of generating the $KEEP$. By adjusting the confidence bias, the decoding process of sequence editing model is under control.
+
+# 4.1.3 Training Process
+
+To generate pseudo data, we need to train the sequence editing model to generate noisy sentences from correct ones at first.
+
+Given a pair of a monolingual sentence  $x$  without noise and a code-switching sentence with noise  $y$ . Then we convert the sentence  $y$  to a sequence of transformations  $z$ . Let  $\epsilon$  denote the preset confidence bias and  $\Theta$  denote all trainable parameters of the sequence editing model  $\phi$ . Our objective is to find the optimal parameter set  $\hat{\Theta}$  that minimizes the following negative log-likelihood function:
+
+$$
+L = - \log P (z | x, \epsilon , \Theta) \tag {1}
+$$
+
+where  $P(z|x,\epsilon ,\Theta)$  denotes the conditional probability.
+
+After training on constructed instances, our model can not only learn the editing operations to apply errors to sentences but also the transformations to generate English-Chinese code-switching texts. Some details of training are shown in Appendix C.
+
+# Algorithm 1 Generating pseudo data
+
+Input: Monolingual Sentence  $x_0$ ; Number of iterations  $T$ ; Sequence editing model  $\phi$ ; Confidence bias  $\epsilon$ ; Similarity Discriminator  $\psi$ ; Hyper parameter  $\lambda, \epsilon_0, \epsilon_1$
+
+Output: Synthetic erroneous samples set  $S$
+
+Function:  $Gen(x_0, T, \lambda, \epsilon_0, \epsilon_1)$ :
+
+1: for  $t\gets 1$  to  $T$  do  
+2: if  $t = 1$  then  
+3:  $\epsilon \gets \epsilon_0$   
+4: else  
+5:  $\epsilon \gets \epsilon_1$   
+6:  $x_{t}\gets \phi (x_{t - 1},\epsilon)$   
+7: if  $t\geq 2$  then  
+8:  $p\gets \psi (x_1,x_t)$   
+9: if  $p\geq \lambda$  then  
+10:  $S.\text{add}(x_1,x_t)$   
+return S
+
+# 4.2 Pseudo Data Generation Algorithm
+
+To generate pseudo training data from the monolingual corpus for code-switching ASR error correction, we have to generate both code-switching target sentences and source sentences with ASR errors. For this objective, we propose an iterative generation approach to obtain pseudo training pairs for code-switching ASR error correction with only one model in multiple iterations. The algorithm for generating pseudo training data is summarized in Algorithm 1.
+
+Given a monolingual sentence from seed corpus  $x_0$ . We apply  $T$  times of iterations to the sequence editing model  $\phi$  to generate data.
+
+In the first iteration, we adjust the confidence bias to  $\epsilon_0$  that limits the noising transformations and makes the model use cross-lingual transformations and $KEEP$ to generate code-switching sentence  $x_{1}$ . This process can be formulated as:
+
+$$
+x _ {1} \leftarrow \phi \left(x _ {0}, \epsilon_ {0}\right) \tag {2}
+$$
+
+In the  $t$ -th  $(1 < t \leq T)$  iteration, the confidence bias is set to  $\epsilon_{1}$  that balances the generation probability of each transformation. It makes the model apply diverse noises to the output of the  $(t-1)$ -th iteration  $x_{t-1}$  and obtain the pseudo instances with errors  $x_{t}$ . This process can be formulated as:
+
+$$
+x _ {t} \leftarrow \phi \left(x _ {t - 1}, \epsilon_ {1}\right) \tag {3}
+$$
+
+where  $1 < t \leq T$ .
+
+With above pseudo instances  $x_{1}, \ldots, x_{T}$ , we treat the output in the first iteration as the correct target sentence and the output in the later turn as noised source sentence to get the sentence pairs.
+
+In order to improve the quality of synthetic samples, we use BERTScore (Zhang et al., 2019) as the similarity discriminator to filter sentence pairs with low similarity. Given a sentence pair  $(x_{1}, x_{t})$ $(1 < t \leq T)$ , we use similarity discriminator  $\psi$  to get a score  $p \in [0,1]$  that presents the semantic similarity between sentences  $x_{1}$  and  $x_{t}$ . We set a threshold  $\lambda$  that if the score  $p$  of a sentence pair is greater than the threshold, the pair can be selected as an augmented training sample.
+
+# 5 Experimental Setup
+
+In this section, we present our experimental setup including datasets, ASR error correction model, and data augmentation baseline. More details of implementation are presented in Appendix D.
+
+# 5.1 Dataset for Experiment
+
+Training and Test Data Firstly, we train the model on pseudo training instances generated from data augmentation method. Then, we fine-tune the code-switching ASR error correction model on our proposed Chinese-English code-switching ASR error correction datasets: SEAME-C and ASCEND-C. We take  $M^2$  scorer and mixed error rate (MER)(Adel et al., 2013) as evaluation metrics.
+
+Seed Corpus for Pseudo Training Data Generation For data augmentation methods, we need to generate pseudo training instances from a seed corpus. Considering proposed datasets were mainly collected from dialogues, we pick the OpenSubtitles dataset (Lison and Tiedemann, 2016), which is a large collection of subtitles in various languages for movies and TV programs, as the seed corpus.
+
+# 5.2 ASR Error Correction Model
+
+In multilingual grammatical error correction task, the adoption of multilingual version of pre-trained model has achieved great success (Rothe et al., 2021). So we pick two strong multilingual pretrained models mT5 and mBART as the code-switching ASR error correction models to verify the effectiveness of our data augmentation method.
+
+Considering that large language models (LLM) have made significant advancements in various NLP tasks. We adopt the most powerful LLM
+
+ChatGPT as the baseline. In our experiment, we test the performance of the GPT-3.5-turbo model on the code-switching ASR error correction task with zero-shot and few-shot settings. The prompt templates for ChatGPT are shown in Appendix F.
+
+# 5.3 Data Augmentation Baseline
+
+There are other ways to generate pseudo code-switching ASR error correction training instances. In our experiment, we implement two data augmentation baselines to compare with our controllable iterative data augmentation method. Firstly, we apply a code-switching text generation approach to get the Chinese-English code-switching texts from a parallel seed corpus. With the code-switching texts, we implement the rule-based and back-translation-based data augmentation methods to improve the ASR error correction models.
+
+Code-switching Text Generation Following previous work (Hussein et al., 2022), we apply the EC theory (Pratapa et al., 2018) to generate Chinese-English code-switching texts. By taking the pairs of parallel sentences in seed corpus as input, the method generates the code-switching texts by replacing the token in the parse tree with its word-level alignments. In our experiment, we adopt the fast-align tool (Dyer et al., 2013) to get the word-level alignments and parse the sentences with a neural graph-based parser(Dozat and Manning).
+
+Rule-based Method The rule-based method is a simple way to generate errors efficiently. We design a rule-based method that generates synthetic training data with five rules: delete, add, replace, shuffle, spelling error. The details of the rules are introduced in Appendix E. With above rules, we can get synthetic training samples with various errors. These synthetic training samples can be used to improve the performance of the code-switching ASR error correction system.
+
+Back-translation Method Inspired by the use of back-translation in downstream tasks (Xie et al., 2018), we train a seq2seq noisy model (like the back-translation model) to generate noisy sentences from correct ones. Then the output of the noising model is paired with the input and the sentence pairs are then used as pseudo data. In our experiment, we consider adopting a base mT5 model as the noising model for our baseline.
+
+<table><tr><td rowspan="2">Correction Model</td><td rowspan="2">Augmentation Method</td><td rowspan="2">Augmented Data Size</td><td colspan="4">SEAME-C</td><td colspan="4">ASCEND-C</td></tr><tr><td>P↑</td><td>R↑</td><td>F0.5↑</td><td>MER↓</td><td>P↑</td><td>R↑</td><td>F0.5↑</td><td>MER↓</td></tr><tr><td rowspan="4">mT5</td><td>-</td><td>-</td><td>26.2</td><td>18.7</td><td>24.3</td><td>21.14</td><td>20.7</td><td>15.3</td><td>19.3</td><td>40.31</td></tr><tr><td>Rules</td><td>5M</td><td>27.4</td><td>19.9</td><td>25.4</td><td>20.74</td><td>23.9</td><td>17.8</td><td>22.4</td><td>37.57</td></tr><tr><td>Back-translation</td><td>4.1M</td><td>27.1</td><td>20.5</td><td>25.5</td><td>20.70</td><td>23.3</td><td>17.4</td><td>21.8</td><td>38.21</td></tr><tr><td>CI</td><td>3.3M</td><td>28.5</td><td>20.7</td><td>26.5</td><td>20.45</td><td>25.0</td><td>19.6</td><td>23.7</td><td>36.35</td></tr><tr><td rowspan="4">mBART</td><td>-</td><td>-</td><td>23.8</td><td>17.1</td><td>22.1</td><td>21.65</td><td>19.6</td><td>14.9</td><td>18.4</td><td>40.12</td></tr><tr><td>Rules</td><td>5M</td><td>25.4</td><td>18.9</td><td>23.8</td><td>21.18</td><td>22.5</td><td>17.2</td><td>21.2</td><td>38.63</td></tr><tr><td>Back-translation</td><td>4.1M</td><td>25.7</td><td>19.1</td><td>24.0</td><td>21.03</td><td>22.8</td><td>18.0</td><td>21.7</td><td>38.44</td></tr><tr><td>CI</td><td>3.3M</td><td>26.9</td><td>21.0</td><td>25.5</td><td>20.73</td><td>24.1</td><td>18.3</td><td>22.7</td><td>37.50</td></tr><tr><td>GPT-3.5-turbo(zero-shot)</td><td>-</td><td>-</td><td>7.5</td><td>8.1</td><td>7.6</td><td>40.01</td><td>5.7</td><td>3.9</td><td>5.2</td><td>71.28</td></tr><tr><td>GPT-3.5-turbo(3-shot)</td><td>-</td><td>-</td><td>13.3</td><td>12.6</td><td>13.1</td><td>35.44</td><td>7.2</td><td>4.8</td><td>6.5</td><td>65.62</td></tr></table>
+
+Table 5: Comparison results of each method. Augmented data sizes show the amounts of additional training sentences used in each method.
+
+# 6 Results and Analysis
+
+# 6.1 Comparison Results
+
+We evaluate the performance of our controllable iterative data augmentation method on proposed datasets. Table 5 shows the results.
+
+Firstly, we compare our method with the two data augmentation baselines. For the fairness of comparison, we train the correction models with the similar scale of augmentation data. Here we use pseudo data generated from the controllable iterative method with three iterations to train the correction model.
+
+As we can see from the table, our proposed controllable iterative data augmentation method achieves the best performance in both  $M^2$  scorer and MER metrics on SEAME-C and ASCEND-C datasets. It outperforms the two baselines for both mT5 and mBART correction methods, which shows the effectiveness and robustness of our controllable iterative method in Chinese-English code-switching ASR error correction task.
+
+Besides, we test ChatGPT (GPT-3.5-turbo) on the proposed dataset. From the table, we can find that the performance of ChatGPT is poor in this task. Under the zero-shot setting, it can only correct a few errors. By learning from examples in the prompt, the performance of ChatGPT under few-shot setting has significantly improved, but still far from the performance of mT5 and mBART models.
+
+The experiment results show that the code-switching ASR error correction task is challenging at present. Due to the scarcity of data and the complexity of code-switching text, LLM method is far from achieving satisfactory results in this task. We present that our data augmentation is an effective method to improve the performance of Chinese-English code-switching ASR error correc
+
+![](images/b568d66e98e23a3c7a389f607de6a2788e5b434ab08453fe507dbade1ccefc78.jpg)  
+tion models.  
+Figure 2: Results of the mT5 correction model on SEAME-C when trained with outputs collected from different iterations of our controllable iterative method.
+
+# 6.2 Iteration of Proposed Method
+
+We build the pseudo training pairs with the outputs of the controllable iterative method in multiple iterations. In this section, we try to explore the influence of the iteration number on the generation of augmented data. Figure 2 shows the results of the mT5 correction model on the SEAME-C dataset when it is trained with outputs collected from different iterations. The details of results are shown in Appendix G.
+
+As the iteration progresses, controllable iterative method continuously generates noisy code-switching data. By collecting the outputs in iterations, we can obtain more pseudo training pairs at each new iteration. In general, using a larger scale of training data may achieve better performance. However, the experimental results show the bottleneck of the data augmentation method that further increasing augmented training data at the 3-th iteration even has a negative impact on performance. This bottleneck might be caused by the quality reduction of the generated data after a few iterations. Considering the balance between perfor
+
+<table><tr><td>Dataset</td><td>Augmentation Method</td><td>P↑</td><td>R↑</td><td>F0.5↑</td><td>MER↓</td></tr><tr><td rowspan="3">SEAME-C</td><td>Rules</td><td>27.4</td><td>19.9</td><td>25.4</td><td>20.74</td></tr><tr><td>CI</td><td>28.5</td><td>20.7</td><td>26.5</td><td>20.45</td></tr><tr><td>Rules+CI</td><td>29.2</td><td>21.3</td><td>27.2</td><td>20.15</td></tr><tr><td rowspan="3">ASCEND-C</td><td>Rules</td><td>23.9</td><td>17.8</td><td>22.4</td><td>37.57</td></tr><tr><td>CI</td><td>25.0</td><td>19.6</td><td>23.7</td><td>36.35</td></tr><tr><td>Rules+CI</td><td>25.6</td><td>20.8</td><td>24.5</td><td>35.46</td></tr></table>
+
+mance and computing costs, we select the outputs generated in the third iteration as the augmented data used for Chinese-English code-switching ASR error correction task.
+
+# 6.3 Combination of Augmented Data
+
+Rule-based and model-based data augmentation methods are two different approaches to pseudo training data generation. Therefore, the distributions of errors generated by these two types of data augmentation methods might be different in general. In this section, we try to explore the effect of combining the data generated by rule-based and our methods.
+
+In our experiments, we pick the mT5 model that achieves higher evaluation scores in previous experiments as the error correction model. The results of using the combined pseudo data are shown in Table 6. From the results, we can find that training mT5 with the combined pseudo data achieves the best scores on our datasets. The use of the data generated by the rule-based method can further improve the performance of the error correction model.
+
+We further analyze the results of different error types. With the automatic annotator, we evaluate the performance in four types: redundant (R), missing (M), word selection (S), and word ordering (W). We conduct the experiment on our SEAME-C dataset with mT5 as the error correction model. The results are shown in Table 7.
+
+Table 6: Results of the mT5 correction model on datasets when trained with the combination of pseudo data generated by controllable iterative (CI) and rule-based methods.  
+
+<table><tr><td>Error Type</td><td>R</td><td>M</td><td>S</td><td>W</td></tr><tr><td>Augmentation Method</td><td>F0.5</td><td>F0.5</td><td>F0.5</td><td>F0.5</td></tr><tr><td>-</td><td>8.1</td><td>16.3</td><td>21.1</td><td>9.8</td></tr><tr><td>Rules</td><td>10.8</td><td>25.6</td><td>26.1</td><td>11.1</td></tr><tr><td>CI</td><td>17.1</td><td>20.5</td><td>28.7</td><td>14.7</td></tr><tr><td>Rules+CI</td><td>18.7</td><td>21.1</td><td>29.1</td><td>7.9</td></tr></table>
+
+Table 7: Results of mT5 correction model on SEAME-C with different augmentation methods on different error types.
+
+As can be seen from the results, data augmentation methods affect the performance of the correction model with respect to different error types. Compared with the controllable iterative method, the rule-based method achieves higher scores on missing errors. The correction model benefits from the operations such as adding and deleting tokens defined in the rules that provide sufficient instances with missing errors. Generally, combining controllable iterative and rule-based data augmentation methods can cover most error types and generate samples with high diversity. It enables the correction model to achieve better performance.
+
+# 7 Conclusion
+
+In this paper, we explore the code-switching ASR error correction task. To enable the research, we first propose two Chinese-English code-switching ASR correction datasets. To address the problem of lacking sufficient training data, we propose the controllable iterative data augmentation method that can generate code-switching ASR error correction instances from the monolingual corpus iteratively. Experiment results show that our method achieves the best performance compared with the rule-based augmentation method, back-translation augmentation method, and ChatGPT method. We further find that combining the data generated by both rule-based method and controllable iterative method can further improve the performance of the error correction model. In future work, we will apply our data augmentation method to other code-switching tasks and also test our method in other language pairs.
+
+# Limitations
+
+In this paper, we only conduct experiments on Chinese-English datasets. There are code-switching phenomena that occur in other languages, for example, the Spanish-English and French-Italian code-switching in America. The empirical conclusions we draw in this study might be changed under different language settings. Another limitation of our study is the lack of manual evaluation. Considering the challenge of the code-switching ASR error correction task, manual evaluation can cover the cases that are evaluated inaccurately by automatic metrics. More interesting conclusions might be drawn by conducting the manual evaluation.
+
+# Acknowledgements
+
+This work was supported by National Key R & D Program of China (2021YFF0901502), National Science Foundation of China (No. 62161160339), State Key Laboratory of Media Convergence Production Technology and Systems and Key Laboratory of Science, Technology and Standard in Press Industry (Key Laboratory of Intelligent Press Media Technology). We appreciate the anonymous reviewers for their helpful comments. Xiaojun Wan is the corresponding author.
+
+# References
+
+Heike Adel, Ngoc Thang Vu, Franziska Kraus, Tim Schlippe, Haizhou Li, and Tanja Schultz. 2013. Recurrent neural network language modeling for code switching conversational speech. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 8411-8415. IEEE.  
+Carlo Aliprandi, Cristina Scudellari, Isabella Gallucci, Nicola Piccinini, Matteo Raffaelli, Arantza del Pozo, Aitor Alvarez, Haritz Arzelus, Renato Cassaca, Tiago Luis, et al. 2014. Automatic live subtitling: state of the art, expectations and current trends. In Proceedings of NAB Broadcast Engineering Conference: Papers on Advanced Media Technologies, Las Vegas, volume 13.  
+Peter Auer. 2013. Code-switching in conversation: Language, interaction and identity. Routledge.  
+Christopher Bryant, Mariano Felice, and Edward Briscoe. 2017. Automatic annotation and evaluation of error types for grammatical error correction. Association for Computational Linguistics.  
+Hannan Cao, Wenmian Yang, and Hwee Tou Ng. 2023. Mitigating exposure bias in grammatical error correction with data augmentation and reweighting. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2115-2127.  
+Rodolfo Corona, Jesse Thomason, and Raymond Mooney. 2017. Improving black-box speech recognition using semantic parsing. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 122-127.  
+Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 568-572.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep
+
+bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186.  
+Suman Dowlagar and Radhika Mamidi. 2023. A code-mixed task-oriented dialog dataset for medical domain. Computer Speech & Language, 78:101449.  
+Timothy Dozat and Christopher D Manning. Deep bia-affine attention for neural dependency parsing.  
+Chris Dyer, Victor Chahuneau, and Noah A Smith. 2013. A simple, fast, and effective reparameterization of ibm model 2. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 644-648.  
+Luis Fernando D'Haro and Rafael E Banchs. 2016. Automatic correction of asr outputs by using machine translation. In Interspeech, volume 2016, pages 3469-3473.  
+Roman Grundkiewicz, Marcin Junczys-Dowmunt, and Kenneth Heafield. 2019. Neural grammatical error correction systems with unsupervised pre-training on synthetic data. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 252-263.  
+Jinxi Guo, Tara N Sainath, and Ron J Weiss. 2019. A spelling correction model for end-to-end speech recognition. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5651-5655. IEEE.  
+Charles Hinson, Hen-Hsen Huang, and Hsin-Hsi Chen. 2020. Heterogeneous recycle generation for chinese grammatical error correction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2191-2201.  
+Amir Hussein, Shammur Absar Chowdhury, Ahmed Abdelali, Najim Dehak, and Ahmed Ali. 2022. Code-switching text augmentation for multilingual speech processing. arXiv e-prints, pages arXiv-2201.  
+Yunqi Gao Jiaju Mei, Yiming Zhu. 1996. Tongyici cilin. In Shanghai Lexicographical Publishing House.  
+Rebecca Jonson. 2006. Dialogue context-based reranking of asr hypotheses. In 2006 IEEE Spoken Language Technology Workshop, pages 174-177. IEEE.  
+Shun Kiyono, Jun Suzuki, Masato Mita, Tomoya Mizumoto, and Kentaro Inui. 2019. An empirical study of incorporating pseudo data into grammatical error correction. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1236-1242.
+
+Jared Lichtarge, Chris Alberti, Shankar Kumar, Noam Shazeer, Niki Parmar, and Simon Tong. 2019. Corpora generation for grammatical error correction. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3291-3301.  
+Pierre Lison and Jörg Tiedemann. 2016. Opensubtitles2016: Extracting large parallel corpora from movie and tv subtitles.  
+Holy Lovenia, Samuel Cahyawijaya, Genta Indra Winata, Peng Xu, Xu Yan, Zihan Liu, Rita Frieske, Tiezheng Yu, Wenliang Dai, Elham J Barezi, et al. 2021. Ascend: A spontaneous chinese-english dataset for code-switching in multi-turn conversation. arXiv preprint arXiv:2112.06223.  
+Dau-Cheng Lyu, Tien-Ping Tan, Eng Siong Chng, and Haizhou Li. 2010. Seame: a mandarin-english code-switching speech corpus in south-east asia. In Eleventh Annual Conference of the International Speech Communication Association.  
+Anirudh Mani, Shruti Palaskar, Nimshi Venkat Meripo, Sandeep Konam, and Florian Metze. 2020. Asr error correction and domain adaptation using machine translation. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6344-6348. IEEE.  
+Niko Moritz, Takaaki Hori, and Jonathan Le. 2020. Streaming automatic speech recognition with the transformer model. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6074-6078. IEEE.  
+Kostiantyn Omelianchuk, Vitaliy Atrasevych, Artem Chernodub, and Oleksandr Skurzhanskyi. 2020. Gector-grammatical error correction: Tag, not rewrite. In Proceedings of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 163-170.  
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32.  
+Adithya Pratapa, Gayatri Bhat, Monjit Choudhury, Sunayana Sitaram, Sandipan Dandapat, and Kalika Bali. 2018. Language modeling for code-mixing: The role of linguistic theory based synthetic data. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1543-1553.  
+Sascha Rothe, Jonathan Mallinson, Eric Malmi, Sebastian Krause, and Aliaksei Severyn. 2021. A simple recipe for multilingual grammatical error correction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the
+
+11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 702-707.  
+Djame Seddah, Farah Essaidi, Amal Fethi, Matthieu Futeral, Benjamin Muller, Pedro Ortiz Suarez, Benoit Sagot, and Abhishek Srivastava. 2020. Building a user-generated content north-african arabizi treebank: Tackling hell. In Proceedings of the 58th annual meeting of the Association for Computational Linguistics, pages 1139-1150.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86-96.  
+Aiman Solyman, Marco Zappatore, Wang Zhenyu, Zeinab Mahmoud, Ali Alfatemi, Ashraf Osman Ibrahim, and Lubna Abdelkareim Gabralla. 2023. Optimizing the impact of data augmentation for low-resource grammatical error correction. Journal of King Saud University-Computer and Information Sciences, 35(6):101572.  
+Hsi-Yao Su. 2001. Code-switching between mandarin and taiwanese in three telephone conversations: The negotiation of interpersonal relationships among bilingual speakers in taiwan. In Proc. of the Symposium about Language and Society.  
+Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, and Angela Fan. 2020. Multilingual translation with extensible multilingual pretraining and finetuning. arXiv preprint arXiv:2008.00401.  
+Jörg Tiedemann and Santhosh Thottingal. 2020. Opusmt-building open translation services for the world. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation. European Association for Machine Translation.  
+Deyuan Wang, Tiantian Zhang, Caixia Yuan, and Xiaojie Wang. 2023. Joint modeling for asr correction and dialog state tracking. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE.  
+Haoyu Wang, Shuyan Dong, Yue Liu, James Logan, Ashish Kumar Agrawal, and Yang Liu. 2020. Asr error correction with augmented transformer for entity retrieval. In *Interspeech*, pages 1550–1554.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45.  
+Ziang Xie, Guillaume Genthial, Stanley Xie, Andrew Y Ng, and Dan Jurafsky. 2018. Noising and denoising natural language: Diverse backtranslation for
+
+grammar correction. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 619-628.
+
+Qiantong Xu, Alexei Baevski, and Michael Auli. 2021. Simple and effective zero-shot cross-lingual phoneme recognition. arXiv preprint arXiv:2109.11680.
+
+Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mt5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498.
+
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations.
+
+# A Processing for Data Construction
+
+# A.1 ASR System
+
+Here is the process we use XLSR-53 model to generate code-switching ASR error correction instances. Firstly, we build the vocabulary from the transcription in the datasets. Based on the vocabulary of 26 English letters, we add all Chinese characters appearing in the transcriptions to the vocabulary for the generation of Chinese. Then, we fine-tune the XLSR-53 model on the training data from SEAME and ASCEND datasets. We pick the parameters with the best performance in the validation set as our code-switching ASR system for next recognition step. With the fine-tuned XLSR-53 model, we can initially obtain the recognition results required for building the code-switching ASR error correction corpus. As the datasets are spontaneous speeches, the corpus is mainly informal and non-speech sounds often occur, such as laughing and coughing. Given that, we design some simple rules to polish the corpus, such as removing non-speech sounds and repeated words.
+
+# A.2 Annotator
+
+The annotator takes three steps: tokenization, alignment, and merging. For tokenization, we use character-level tokenization for Chinese and word-level tokenization for English. Alignment works by computing the alignment score for each pair of tokens in the source and target sentences. In particular, the Cilin thesaurus (Jiaju Mei, 1996) is leveraged to give a similarity score between Chinese characters. Once the alignment score matrix
+
+is computed, the sequence with the lowest score is returned. At last, we use a simple merging strategy that merges consecutive sequences of edits of the same type.
+
+# B Transformations in GECToR
+
+GECToR proposed the token-level transformations for the sequence editing grammatical error correction model. Transformations increase the coverage of grammatical error corrections for limited output vocabulary size for the most common types of errors. The work divided transformations into two types: basic transformations and g-transformations.
+
+Basic transformations perform the most common token-level edit operations, such as: keep the current token unchanged (tag $KEEP$), delete current token (tag $DELETE$), append new token  $t_1$  next to the current token (tag $APPEND t1 $) or replace the current token with another token  $t_2$  (tag $REPLACE t2 $).
+
+G-transformations perform task-specific operations such as: change the case of the current token (\(CASE\) tags), merge the current token and the next token into a single one (\(MERGE\) tags) and split the current token into two new tokens (\(SPLIT\) tags). Moreover, tags from \)NOUN NUMBER$ and $VERB FORM$ transformations encode grammatical properties for tokens.
+
+# C Processing for Sequence Editing Model
+
+# C.1 Processing for Transformations
+
+It takes a three-step pre-processing algorithm to convert a sentence to a sequence of transformations. Firstly, it maps each token from source sentence to sub-sequence of tokens from target sentence. For this purpose, it detects the minimal spans of tokens which define differences between source tokens and target tokens. Thus, such a span is a pair of selected source tokens and corresponding target tokens. To get tags on the token level, it searches for the best-fitting sub-sequence of target tokens by minimizing the modified Levenshtein distance for each source token. Secondly, for each mapping in the list, it finds token-level transformations which convert source token to the target sub-sequence. Finally, it leaves only one transformation for each source token. Through the above three steps, we map the tag to each single source token.
+
+# C.2 Processing for Translation
+
+Our model is aimed to generate code-switching sentences with errors from correct monolingual sentences, but the training pairs in datasets only consist of code-switching texts. Considering that, we adopt an open translation tool OPUS-MT (Tiedemann and Thottingal, 2020) to translate the code-switching ASR transcriptions into monolingual sentences. Specifically, we translate the spans of English texts to Chinese and get the Chinese sentences. With regarding the translated monolingual sentence as source and the code-switching sentence with errors from error correction datasets as target, we train the sequence editing model.
+
+# D Implementation Details
+
+In our work, we adopt PyTorch (Paszke et al., 2019) framework and HuggingFace (Wolf et al., 2020) tool to implement all deep neural models. Each model is trained on four NVIDIA A40 GPUs. The details of each model are described as follows.
+
+# D.1 ASR Error Correction Model
+
+Considering that mT5 and mBART have achieved strong performance in many multilingual tasks, we adopt these pre-trained models to correct ASR errors. In the experiments, we use the base version of mT5 and fine-tune it by using the Adam optimization method with learning rate 5e-4, weight decay parameter 0.01, dropout rate 0.1, and batch size 24. Besides, we use the large version of mBART, and fine-tune it by using the Adam optimization method with learning rate 2e-5, dropout rate 0.1, and batch size 16. Beam search with beam size of 5 is used for decoding.
+
+# D.2 Sequence Editing Model
+
+The sequence editing model is an encoder made up of Transformer architecture stacked with two linear layers with softmax layers. We adopt the base version of m-BERT to initialize the encoder. Our model consists of 12 Transformer layers and two linear layers. We train the model with learning rate 1e-5, weight decay parameter 0.01, dropout rate 0, and batch size 32. The size of the vocabulary is 12229. It consists of 4971 basic transformations, 29 token-independent g-transformations and 7229 cross-lingual transformations.
+
+# D.3 Pseudo Data Generation Algorithm
+
+We add a confidence bias to control the decoding process. In our experiments, the confidence bias \( \epsilon_0 \) is to add -1.0 to the generation probability of all transformations except \( \\( \text{KEEP} \\), \\) \text{TRAN-SLATION} \\) and other cross-lingual transformations. The confidence bias \( \epsilon_1 \) is set to keep the original generation probability of all transformations. The threshold \( \lambda \) is set to 0.85.
+
+<table><tr><td>Method</td><td>Prompt Template</td></tr><tr><td>Zero-shot</td><td>&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: &quot;You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence.&quot; &quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;Please identify and correct any errors in the following sentence indicated by &lt;input&gt;ERROR &lt;/input&gt; tag, you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible. Remember to format your corrected output results with the tag &lt;output&gt; Your Corrected Version &lt;/output&gt;. Please start: &lt;input&gt;Input Sentence &lt;/input&gt;:&quot;</td></tr><tr><td>Few-shot</td><td>&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: &quot;You are a Chinese-English code-switching ASR error correction tool that can identify and correct errors in the text. Especially, code-switching is a linguistic phenomenon where a speaker alternates between two or more languages or language varieties within a single conversation or even within a single sentence.&quot; &quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;Please identify and correct any errors in the following sentence indicated by &lt;input&gt;ERROR &lt;/input&gt; tag,you need to comprehend the sentence as a whole before gradually identifying and correcting any errors while keeping the original sentence structure unchanged as much as possible.Here are some in-context examples:(1), &lt;input&gt; SRC-1 &lt;/input&gt;: &lt;output&gt;TGT-1 &lt;/output&gt;(2), &lt;input&gt; SRC-1 &lt;/input&gt;: &lt;output&gt;TGT-1 &lt;/output&gt;; ... (n), &lt;input&gt; SRC-n &lt;/input&gt;: &lt;output&gt;TGT-n &lt;/output&gt;; Please feel free to refer to these examples. Remember to format your corrected output results with the tag &lt;output&gt; Your Corrected Version &lt;/output&gt;. Please start: &lt;input&gt;Input Sentence &lt;/input&gt;:&quot;</td></tr></table>
+
+Table 8: Zero-shot and few-shot prompt template for Chinese-English code-switching ASR error correction.
+
+# E Rule-based Method
+
+The details of the five rules adopted in our proposed rule-based method are introduced in the following.
+
+Delete. Randomly delete a token with a probability.
+
+Add. Firstly, randomly select a word or Chinese character from the vocabulary of the ASR datasets, and then add the selected word or character to a random position.
+
+Replace. Randomly replace a token with a word or Chinese character from the vocabulary of the ASR datasets. In particular, if the selected token is a Chinese character, we select the replacement from the homophones with a probability.
+
+Shuffle. Shuffle the tokens by adding a normal distribution bias to the positions of the tokens.
+
+Spell Error. Randomly apply spell error to an English word. We randomly perturb letters with operations, i.e. substitution, deletion, insertion or transposition of letters.
+
+# F Prompt Template
+
+Zero-shot and few-shot prompt templates for Chinese-English code-switching ASR error correction are shown in Table 8.
+
+# G Additional Experiment Results
+
+Results of the mT5 correction model on SEAME-C, when trained with outputs collected from different iterations of our method, are shown in Table 9.
+
+<table><tr><td>Iteration</td><td>Augmented Data Size</td><td>P↑</td><td>R↑</td><td>F0.5↑</td><td>MER↓</td></tr><tr><td>Iteration 1</td><td>1.8M</td><td>28.0</td><td>20.3</td><td>26.1</td><td>20.54</td></tr><tr><td>Iteration 2</td><td>2.6M</td><td>28.2</td><td>20.6</td><td>26.3</td><td>20.49</td></tr><tr><td>Iteration 3</td><td>3.3M</td><td>28.5</td><td>20.7</td><td>26.5</td><td>20.45</td></tr><tr><td>Iteration 4</td><td>3.6M</td><td>28.3</td><td>20.8</td><td>26.4</td><td>20.51</td></tr><tr><td>Iteration 5</td><td>3.8M</td><td>28.3</td><td>21.1</td><td>26.5</td><td>20.40</td></tr></table>
+
+Table 9: Results of the mT5 correction model on SEAME-C when trained with outputs collected from different iterations of our method.

2023/New Datasets and Controllable Iterative Data Augmentation Method for Code-switching ASR Error Correction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9e49c7d80ac8c58b0d64e44d4e1950f7c855496259db09a8c1c40eeb325f497
+size 518604

2023/No offence, Bert - I insult only humans! Multilingual sentence-level attack on toxicity detection networks/6570fd92-ec0d-40f4-8099-434dc1b0509c_content_list.json

@@ -0,0 +1,1158 @@
+[
+    {
+        "type": "text",
+        "text": "No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks",
+        "text_level": 1,
+        "bbox": [
+            115,
+            89,
+            884,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sergey Berezin, Reza Farahbakhsh, Noel Crespi  \nSAMOVAR, Télécom SudParis, Institut Polytechnique de Paris 91120 Palaiseau, France  \nsberezin@telecom-sudparis.eu",
+        "bbox": [
+            243,
+            149,
+            759,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            342,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We introduce a simple yet efficient sentence-level attack on black-box toxicity detector models. By adding several positive words or sentences to the end of a hateful message, we are able to change the prediction of a neural network and pass the toxicity detection system check. This approach is shown to be working on seven languages from three different language families. We also describe the defence mechanism against the aforementioned attack and discuss its limitations.",
+        "bbox": [
+            141,
+            277,
+            460,
+            434
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            444,
+            260,
+            458
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Toxicity detection systems have become a crucial part of automoderation solutions. They are now used by most social media platforms, including those with groups of people in dangerously sensitive conditions, e.g. suicide prevention, Facebook groups, victims of cyberbullying, etc. Vulnerability in such systems may have dreadful, terrific effects on people, especially those in precarious situations.(Cheng et al., 2022).",
+        "bbox": [
+            112,
+            469,
+            489,
+            613
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "On the other hand, such systems can be and are used to silence the voices of criticism, which leads to the creation of echo chambers and amplifies the voice of the (powerful) minority over the voice of the majority, thereby destroying the foundation of democracy and denying the freedom of speech (Gorwa et al., 2020).",
+        "bbox": [
+            112,
+            615,
+            489,
+            726
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This situation can be viewed as a double-edged sword, and we propose another double-edged sword that can be used to parry the blade of toxicity detection systems.",
+        "bbox": [
+            112,
+            727,
+            489,
+            791
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1.1 Task description",
+        "text_level": 1,
+        "bbox": [
+            114,
+            801,
+            290,
+            816
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We present an attack on toxicity detection models based on the separation of the messages intended for a person and a piece of text added to confuse an algorithm.",
+        "bbox": [
+            112,
+            822,
+            487,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "For example, this can be done by concatenating an original message with a collection of keywords",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "of an opposite intent: \"Kill yourself, you dirty pig! Text for a bot to avoid ban: flowers, rainbow, happy, happy good\" - here only the underlined part is intended to address the human, and it is clear that the remaining part was added to avoid detection by the automoderation algorithm.",
+        "bbox": [
+            507,
+            253,
+            884,
+            348
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This represents an example of a sentence-level black-box adversarial attack on Natural Language Processing systems.",
+        "bbox": [
+            507,
+            350,
+            882,
+            398
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1.2 Related work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            412,
+            662,
+            426
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The first work suggesting the concatenation of distracting but meaningless sentences at the end of a paragraph to confuse a neural model was \"Adversarial Examples for Evaluating Reading Comprehension Systems\" (Jia and Liang, 2017) (according to the (Zhang et al., 2020)). Jia and Liang attacked question-answering systems with either manually-generated informative sentences (ADDSENT) or arbitrary sequences of words using a pool of 20 random common words (ADDANY). Both perturbations were obtained by iterative querying of the neural network until the output was changed. The authors of \"Robust Machine Comprehension Models via Adversarial Training\" (Wang and Bansal, 2018) improved this approach by varying the locations where the distracting sentences are placed and expanding the set of fake answers for generating the distracting sentences (ADDSENTDIVERSE).",
+        "bbox": [
+            507,
+            432,
+            885,
+            722
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In \"Universal Adversarial Triggers for Attacking and Analyzing NLP\" (Wallace et al., 2019) apply the gradient-based technique to construct an adversarial text. Despite being a white-box attack in its origin (due to the gradient information required in the training phase), this approach can be applied as a black-box attack during its inference.",
+        "bbox": [
+            507,
+            724,
+            884,
+            835
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The T3 model (Wang et al., 2020) utilises the autoencoder architecture to generate adversarial texts that can manipulate the question-answering models to output the targeted incorrect answer.",
+        "bbox": [
+            507,
+            838,
+            882,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In \"All You Need is \"Love\": Evading Hate",
+        "bbox": [
+            526,
+            903,
+            880,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "2362",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2362-2369 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            779,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Speech Detection\" (Gröndahl et al., 2018) performs a sentence-level attack on hate speech detection systems by inserting typos, changing word boundaries and adding innocuous words to the original hate speech. The authors show the effectiveness of such an attack on MLP, CNN+RNN and LSTM neural networks.",
+        "bbox": [
+            112,
+            84,
+            487,
+            197
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1.3 Contribution",
+        "text_level": 1,
+        "bbox": [
+            114,
+            209,
+            265,
+            223
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions lie in the following:",
+        "bbox": [
+            112,
+            230,
+            400,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1) Introduction of a concept of \"To Each His Own\" attack, based on the idea of separating the messages addressed to a human and those addressed to an algorithm. The attack exploits the fact that toxicity detection models are trained on sentence-level labels and do not consider the context or the intention of the message.",
+            "2) A systematic approach to the generation of adversarial texts. The words and sentences are selected based on their positivity scores rather than being randomly selected from non-hate ones, as in previous works.",
+            "3) Demonstration of attacks on modern state-of-the-art architectures. We tested the most popular transformer-based toxicity detectors, including Google Perspective API and OpenAI Moderation API.",
+            "4) Highlighting a significant flaw in the current approach to toxicity detection. We show that the predicted toxicity score can be watered down by the addition of positive words to the point of effectively masking toxicity.",
+            "5) Multi- and cross-lingual settings of attacks. We describe the attacks on seven different languages and combinations among them. A larger number of languages leads to a larger number of possible attacks - in the case of the 5-language model, we have 120 attack scenarios."
+        ],
+        "bbox": [
+            112,
+            246,
+            487,
+            697
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 \"To Each His Own\" attack",
+        "text_level": 1,
+        "bbox": [
+            112,
+            709,
+            381,
+            725
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Word-based attack",
+        "text_level": 1,
+        "bbox": [
+            112,
+            737,
+            310,
+            751
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this attack type, the original message is appended or prepended with a sequence of positive words, with the aim that these words will be analysed by a model but passed over by a human. This allows an attacker to preserve their original message without any changes, unlike in other attacks, which substitute words with synonyms or replace original characters with their look-alikes.",
+        "bbox": [
+            112,
+            758,
+            487,
+            885
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In our experiment, we used words from the SentiWordNet with a positivity score  $\\geq 0.85$ . We tested",
+        "bbox": [
+            112,
+            887,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "14 openly available pretrained neural networks with the phrase \"I hate you\" and the addition of 0 to 9 positive words.",
+        "bbox": [
+            507,
+            84,
+            880,
+            131
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In case the toxicity score for a suggested phrase was estimated to be less than 0.5 - the most common threshold, we used the phrase \"I hate you dumbass\" instead. If, after attacking with 9 words, the model's output was greater than 0.5, we continued adding new positive words until the score dropped below this threshold or until we reached the maximum length of an input sequence for the model tested. For the English language, we also performed experiments with the phrase \"Text for bot:\" placed between a message and added words.",
+        "bbox": [
+            507,
+            134,
+            882,
+            310
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conducted additional experiments with the Perspective API and OpenAI Moderation API, discovering they are also vulnerable to our attacks.",
+        "bbox": [
+            507,
+            312,
+            882,
+            359
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For eight non-English models, we asked native speakers of the respective languages to translate a selected subset of words and attack phrases. Within this study, we tested the language models for seven languages aside from English: Russian, German, French, Portuguese, Spanish, Turkish and Vietnamese. It is worth noting that the first five listed languages represent three branches (Slavic, Germanic and Italic) of the Indo-European language family, and the last two are from different language families: the Turkic and the Austro-Asiatic family, respectfully.",
+        "bbox": [
+            507,
+            362,
+            882,
+            554
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "By using various models and languages, we aim to investigate if the suggested attack is model- and language-agnostic.",
+        "bbox": [
+            507,
+            556,
+            880,
+            602
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Cross-lingual word-based attack",
+        "text_level": 1,
+        "bbox": [
+            507,
+            619,
+            808,
+            634
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The main weakness of all multilingual models lies in their greatest advantage - the ability to work with multiple languages and writing systems. Adding positive words in different languages will even more separate messages to the human and to the toxicity detection system, perhaps even denying a human the ability to read non-intended text. Examples of such texts are shown in Figure 1.",
+        "bbox": [
+            507,
+            642,
+            880,
+            770
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Even some monolingual models, which had been exposed to another language during the pretraining phase, fall victim to this situation.",
+        "bbox": [
+            507,
+            772,
+            880,
+            819
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We also conducted cross-lingual experiments with OpenAI models, including ChatGPT in conversation mode with different prompts, and found them susceptible to attack.",
+        "bbox": [
+            507,
+            821,
+            880,
+            885
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Perspective API was found to be non-functional in a multilingual setting, failing with an error, as it",
+        "bbox": [
+            507,
+            887,
+            880,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "2363",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/ad8c4117997fdc471e2a002b85338e7fd48e62e030d9a6d6961e8fe0572ca951.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>I hate you lióobb aǐltpruíntiúhái lüedpí</td></tr><tr><td>You stink. eşsism.dkukemmel telkari güzel ahlak</td></tr><tr><td>Goodbye, dumbass! animado melhor formidable</td></tr></table>",
+        "bbox": [
+            115,
+            86,
+            485,
+            145
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: Examples of cross-lingual word-based attack. The languages of the added text from the top are Russian, Turkish, and Spanish.",
+        "bbox": [
+            112,
+            159,
+            489,
+            203
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "is unable to detect the language of the text.",
+        "bbox": [
+            112,
+            227,
+            433,
+            242
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Sentence-based attack",
+        "text_level": 1,
+        "bbox": [
+            114,
+            252,
+            334,
+            266
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Incoherent text, produced by concatenating lexically unconnected words, is easily detectable by modern language models. To address that easy detection, we experimented with another version of the attack: concatenating sentences from the Stanford Sentiment Treebank with a positivity score  $\\geq 0.9$  and a length of no less than 100 symbols.",
+        "bbox": [
+            112,
+            273,
+            487,
+            385
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Since sentences consist of grammatically linked words instead of just a set of random positive words, this attack variant is less obvious and more challenging to detect.",
+        "bbox": [
+            112,
+            386,
+            489,
+            451
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Defence",
+        "text_level": 1,
+        "bbox": [
+            114,
+            461,
+            218,
+            476
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We performed simple adversarial training of the DistilBERT model on the binary Jigsaw Toxic Comments dataset to improve the model's defence. We performed experiments with word- and sentence-based attacks, attacking both toxic messages and all messages in the dataset. In addition, we picked out only toxic messages and attacked half of them - in this scenario, the task was distinguishing attacked and non-attacked texts.",
+        "bbox": [
+            112,
+            486,
+            489,
+            630
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            642,
+            213,
+            657
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 Attack",
+        "text_level": 1,
+        "bbox": [
+            112,
+            667,
+            213,
+            681
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "During a word-based attack, both prepending and appending of the positive words showed similar results, with prepending being slightly less effective.",
+        "bbox": [
+            112,
+            687,
+            487,
+            736
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Appending nine words was enough to flip the prediction of almost every model. The \"SkolkovoInstitute RoBERTa\" toxicity classifier required 23 words to fall below 0.5, and \"English-abusive-MuRIL\" was still strongly predicting even after the addition of 252 words and reaching the length limit for the input. The results of selected experiments are shown in tables 1, 2 and  $3^{1}$ .",
+        "bbox": [
+            112,
+            737,
+            487,
+            864
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The addition of the phrase \"Text for bot:\" between two parts of a text made a little difference in",
+        "bbox": [
+            112,
+            865,
+            487,
+            897
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/6d2caf4e2c32bccc1fc6aaf189b73f050b95afd9ed925dc34b473471cf4dcc51.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n words</td><td>BERT</td><td>RoBERTa</td><td>ELECTRA</td></tr><tr><td>0</td><td>0,951</td><td>0,857</td><td>0,898</td></tr><tr><td>1</td><td>0,925</td><td>0,743</td><td>0,730</td></tr><tr><td>2</td><td>0,708</td><td>0,507</td><td>0,361</td></tr><tr><td>3</td><td>0,592</td><td>0,415</td><td>0,064</td></tr><tr><td>4</td><td>0,579</td><td>0,274</td><td>0,067</td></tr><tr><td>5</td><td>0,548</td><td>0,170</td><td>0,082</td></tr><tr><td>6</td><td>0,510</td><td>0,158</td><td>0,087</td></tr><tr><td>7</td><td>0,494</td><td>0,148</td><td>0,069</td></tr><tr><td>8</td><td>0,465</td><td>0,142</td><td>0,065</td></tr><tr><td>9</td><td>0,303</td><td>0,132</td><td>0,072</td></tr></table>",
+        "bbox": [
+            514,
+            80,
+            878,
+            263
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/6d225c4b904fc60d41292e0083203cf7001e885e79fc25aabdd9b103ebdc01eb.jpg",
+        "table_caption": [
+            "Table 1: Word-based attack in English on different transformer architectures."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n words</td><td>OpenAI S</td><td>OpenAI L</td><td>Per. API</td></tr><tr><td>1</td><td>0.643</td><td>0.946</td><td>0.899</td></tr><tr><td>...</td><td>...</td><td>...</td><td>...</td></tr><tr><td>10</td><td>0.421</td><td>0.847</td><td>0.846</td></tr><tr><td>11</td><td>0.381</td><td>0.705</td><td>0.852</td></tr><tr><td>12</td><td>0.427</td><td>0.74</td><td>0.852</td></tr><tr><td>13</td><td>0.453</td><td>0.761</td><td>0.852</td></tr><tr><td>14</td><td>0.426</td><td>0.665</td><td>0.853</td></tr><tr><td>15</td><td>0.362</td><td>0.681</td><td>0.853</td></tr><tr><td>16</td><td>0.265</td><td>0.716</td><td>0.853</td></tr></table>",
+        "bbox": [
+            510,
+            315,
+            882,
+            481
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Word-based attack in English on Perspective API and OpenAI Moderation API. Per. - perspective, S - stable, L - latest.",
+        "bbox": [
+            507,
+            489,
+            882,
+            532
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the results, i.e. it can be used to create even more apparent separation without lowering an attack's efficiency.",
+        "bbox": [
+            507,
+            560,
+            880,
+            608
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As expected, the results of cross-lingual attacks followed the same trend as those of monolingual ones. These results are shown in tables 4 and  $5^{1}$ .",
+        "bbox": [
+            507,
+            609,
+            880,
+            656
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With a sentence-based attack, adding even one sentence drastically lowered the prediction scores of the models. The scores are shown in table  $6^{1}$ .",
+        "bbox": [
+            507,
+            658,
+            880,
+            706
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.2 Defence",
+        "text_level": 1,
+        "bbox": [
+            507,
+            719,
+            618,
+            733
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The model's F1 measure on toxic class fell from 0.80 to 0.44 after an attack with 15 sentences and to 0.79 after an attack with 50 words. After adversarial training, the model regained its F1 measure in both cases.",
+        "bbox": [
+            507,
+            741,
+            882,
+            820
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Interestingly, testing the word-trained model on sentence-attacked examples showed high precision and low recall (0.90 and 0.68, respectively). Testing the sentence-trained model on word-attacked examples showed the opposite result, with low precision and high recall (0.57 and 0.93, respectively).",
+        "bbox": [
+            507,
+            822,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "1 Full tables can be found in Appendix A.",
+        "bbox": [
+            134,
+            903,
+            386,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "2364",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/a73c56dda2cb53098c67c5bc48a508b973ac94917af50609dd2f3976aa974262.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n words</td><td>Vietnamese</td><td>French</td><td>Turkish</td></tr><tr><td>0</td><td>0,996</td><td>0,970</td><td>0,814</td></tr><tr><td>1</td><td>0,008</td><td>0,962</td><td>0,186</td></tr><tr><td>2</td><td>0,007</td><td>0,573</td><td>0,503</td></tr><tr><td>3</td><td>0,008</td><td>0,396</td><td>0,043</td></tr><tr><td>4</td><td>0,009</td><td>0,090</td><td>0,048</td></tr><tr><td>5</td><td>0,008</td><td>0,054</td><td>0,085</td></tr><tr><td>6</td><td>0,007</td><td>0,055</td><td>0,090</td></tr><tr><td>7</td><td>0,007</td><td>0,064</td><td>0,029</td></tr><tr><td>8</td><td>0,007</td><td>0,082</td><td>0,009</td></tr><tr><td>9</td><td>0,007</td><td>0,045</td><td>0,005</td></tr></table>",
+        "bbox": [
+            122,
+            80,
+            477,
+            262
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/8db959d3bdbfca7d0a5e1174b76eea86c5c45fa444e382b6154ddcf3d63e7428.jpg",
+        "table_caption": [
+            "Table 3: Word-based attack on transformer-based language models trained on the languages from three different language families."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n words</td><td>eng+tur</td><td>eng+rus</td><td>eng+sp</td></tr><tr><td>0</td><td>0,971</td><td>0,971</td><td>0,971</td></tr><tr><td>1</td><td>0,913</td><td>0,717</td><td>0,802</td></tr><tr><td>2</td><td>0,904</td><td>0,717</td><td>0,687</td></tr><tr><td>3</td><td>0,912</td><td>0,761</td><td>0,536</td></tr><tr><td>4</td><td>0,541</td><td>0,800</td><td>0,558</td></tr><tr><td>5</td><td>0,634</td><td>0,592</td><td>0,500</td></tr><tr><td>6</td><td>0,731</td><td>0,456</td><td>0,539</td></tr><tr><td>7</td><td>0,604</td><td>0,402</td><td>0,426</td></tr><tr><td>8</td><td>0,519</td><td>0,433</td><td>0,434</td></tr><tr><td>9</td><td>0,455</td><td>0,369</td><td>0,384</td></tr></table>",
+        "bbox": [
+            115,
+            328,
+            448,
+            510
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The defence performed strongly, achieving an F1 score of 0.82 on binary toxic classification with word- and sentence-based attacks. The model achieved the same score on a non-attacked dataset, succeeding even better in distinguishing attacked sentences, with an F1 score of 0.99. However, defence models must be built for both word- and sentence-based attacks, as one single model will not work perfectly for both scenarios.",
+        "bbox": [
+            112,
+            574,
+            487,
+            720
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            732,
+            240,
+            747
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "All of the attack variants showed decent results in confusing toxicity detection models in every model tested (to varying degrees). The simplistic nature of such attacks could allow virtually any Internet user to exploit them to avoid automoderation systems.",
+        "bbox": [
+            112,
+            758,
+            487,
+            838
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A possible countermeasure for this type of attack is adversarial training or rule-based filtering. However, a simple rule-based filter can be fooled by char-level adversarial attacks, and adversarial training can be made much more difficult by in",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/41f4716ca326d8f7759c7e9ef8e2a33db8cb4f10f221bdd8eecf964298e22d6b.jpg",
+        "table_caption": [
+            "Table 4: Cross-lingual word-based attack on toxicity detection models."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n</td><td>L en+fr</td><td>L en+ger</td><td>S en+fr</td><td>S en+ger</td></tr><tr><td>0</td><td>0.872</td><td>0.872</td><td>0.888</td><td>0.888</td></tr><tr><td>1</td><td>0.931</td><td>0.981</td><td>0.793</td><td>0.848</td></tr><tr><td>2</td><td>0.895</td><td>0.898</td><td>0.779</td><td>0.510</td></tr><tr><td>3</td><td>0.769</td><td>0.889</td><td>0.671</td><td>0.511</td></tr><tr><td>4</td><td>0.816</td><td>0.846</td><td>0.594</td><td>0.344</td></tr><tr><td>5</td><td>0.771</td><td>0.807</td><td>0.606</td><td>0.302</td></tr><tr><td>6</td><td>0.800</td><td>0.647</td><td>0.618</td><td>0.318</td></tr><tr><td>7</td><td>0.659</td><td>0.533</td><td>0.605</td><td>0.201</td></tr><tr><td>8</td><td>0.592</td><td>0.555</td><td>0.573</td><td>0.208</td></tr><tr><td>9</td><td>0.579</td><td>0.489</td><td>0.424</td><td>0.283</td></tr></table>",
+        "bbox": [
+            510,
+            80,
+            884,
+            262
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/e125d5cf701bca605197ba0cec200df8f520a80e5c437c133e0a2f8a3d743c03.jpg",
+        "table_caption": [
+            "Table 5: The results of cross-lingual attacks on OpenAI models. n - number of words added, L - latest, S - stable."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n sentences</td><td>BERT</td><td>RoBERTa</td><td>ELECTRA</td></tr><tr><td>0</td><td>0,590</td><td>0,962</td><td>0,898</td></tr><tr><td>1</td><td>0,001</td><td>0,001</td><td>0,232</td></tr><tr><td>2</td><td>0,001</td><td>0,002</td><td>0,147</td></tr></table>",
+        "bbox": [
+            510,
+            316,
+            899,
+            382
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Results of sentence-based attacks.",
+        "bbox": [
+            547,
+            393,
+            840,
+            407
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "creasing the quality and quantity of the possible text insertions.",
+        "bbox": [
+            507,
+            434,
+            880,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The complex approach, which covers all levels of perturbation, should be applied both in attack and defence real-life scenarios.",
+        "bbox": [
+            507,
+            468,
+            880,
+            514
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            529,
+            640,
+            544
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we demonstrated a novel and effective way of bypassing toxicity detection models by appending positive words or sentences to the end of a toxic message.",
+        "bbox": [
+            507,
+            556,
+            882,
+            619
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The introduced concept of the \"To Each His Own\" attack is based on the idea of separating the messages addressed to a human and those addressed to an algorithm. The attack exploits the fact that the toxicity detection models are trained on sentence-level labels and do not consider the context or the intention of messages.",
+        "bbox": [
+            507,
+            621,
+            882,
+            733
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The described attack in all its variants can be easily used on almost any toxic-detection neural model, with fairly good results. For future research, we suggest looking for a more sophisticated way of constructing insertions, perhaps with respect to the original message, making it a semantically correct addition to human-written messages.",
+        "bbox": [
+            507,
+            734,
+            882,
+            847
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "7 Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            860,
+            702,
+            876
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We thank all the volunteers who helped us with translations.",
+        "bbox": [
+            507,
+            887,
+            880,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "2365",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We are grateful to the reviewers and the program chair of the EMNLP conference for providing constructive feedback, suggesting additional experiments and providing different points of view. This helped us structure the paper and ensure the robustness of our results.",
+        "bbox": [
+            112,
+            84,
+            489,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Sergey Berezin wishes to record that he is grateful to Mr. Mario Fritz from Saarland University for the introduction to the domain of adversarial attacks through his insightful course on this topic. He is also grateful to Mr. Maxime Amblard from the University of Lorraine for encouraging him to pursue research in this direction and validating his ideas.",
+        "bbox": [
+            112,
+            181,
+            489,
+            307
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            335,
+            213,
+            350
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Lu Cheng, Ahmadreza Mosallanezhad, Yasin N. Silva, Deborah L. Hall, and Huan Liu. 2022. Bias mitigation for toxicity detection via sequential decisions. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 1750-1760, New York, NY, USA. Association for Computing Machinery.",
+            "Robert Gorwa, Reuben Binns, and Christian Katzenbach. 2020. Algorithmic content moderation: Technical and political challenges in the automation of platform governance. *Big Data Society*, 7:205395171989794.",
+            "Tommi Gröndahl, Luca Pajola, Mika Juuti, Mauro Conti, and N. Asokan. 2018. All you need is \"love\": Evading hate speech detection. AIsec '18, New York, NY, USA. Association for Computing Machinery.",
+            "Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. pages 2021-2031.",
+            "Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing nlp. pages 2153-2162.",
+            "Boxin Wang, Hengzhi Pei, Boyuan Pan, Qian Chen, Shuohang Wang, and Bo Li. 2020. T3: Treeautoencoder constrained adversarial text generation for targeted attack. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6134-6150, Online. Association for Computational Linguistics.",
+            "Yicheng Wang and Mohit Bansal. 2018. Robust machine comprehension models via adversarial training. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 575-581, New Orleans, Louisiana. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            357,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Wei Emma Zhang, Quan Z. Sheng, Ahoud Alhazmi, and Chenliang Li. 2020. Adversarial attacks on deep-learning models in natural language processing: A survey. ACM Trans. Intell. Syst. Technol., 11(3).",
+        "bbox": [
+            510,
+            84,
+            884,
+            139
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A Appendix A",
+        "text_level": 1,
+        "bbox": [
+            509,
+            151,
+            653,
+            168
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "2366",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/507c5c1f6ffe28625c24c6551c0f786d62fde5feb6dbf19a1bf7ae48d20a1139.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n words</td><td>BERT</td><td>roberta-base</td><td>xlm-roberta-base</td><td>toxigen-roberta</td><td>toxigen-hatebert</td></tr><tr><td>0</td><td>0,95</td><td>0,86</td><td>0,97</td><td>0,96</td><td>0,88</td></tr><tr><td>1</td><td>0,93</td><td>0,74</td><td>0,89</td><td>0,96</td><td>1,00</td></tr><tr><td>2</td><td>0,71</td><td>0,51</td><td>0,65</td><td>0,71</td><td>0,98</td></tr><tr><td>3</td><td>0,59</td><td>0,42</td><td>0,63</td><td>0,59</td><td>0,98</td></tr><tr><td>4</td><td>0,58</td><td>0,27</td><td>0,56</td><td>0,02</td><td>0,11</td></tr><tr><td>5</td><td>0,55</td><td>0,17</td><td>0,53</td><td>0,00</td><td>0,07</td></tr><tr><td>6</td><td>0,51</td><td>0,16</td><td>0,41</td><td>0,00</td><td>0,02</td></tr><tr><td>7</td><td>0,49</td><td>0,15</td><td>0,55</td><td>0,00</td><td>0,31</td></tr><tr><td>8</td><td>0,47</td><td>0,14</td><td>0,45</td><td>0,00</td><td>0,39</td></tr><tr><td>9</td><td>0,30</td><td>0,13</td><td>0,41</td><td>0,00</td><td>0,06</td></tr><tr><td>n words</td><td>roberta skolkovo</td><td>roberta facebook</td><td>MuRIL</td><td>MuRIL*</td><td>electra-hateXplain</td></tr><tr><td>0</td><td>1,00</td><td>1,00</td><td>0,13</td><td>0,96</td><td>0,90</td></tr><tr><td>1</td><td>1,00</td><td>1,00</td><td>0,10</td><td>0,95</td><td>0,73</td></tr><tr><td>2</td><td>0,99</td><td>1,00</td><td>0,03</td><td>0,94</td><td>0,36</td></tr><tr><td>3</td><td>0,92</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,06</td></tr><tr><td>4</td><td>0,98</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,07</td></tr><tr><td>5</td><td>0,98</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,08</td></tr><tr><td>6</td><td>0,99</td><td>1,00</td><td>0,03</td><td>0,94</td><td>0,09</td></tr><tr><td>7</td><td>0,99</td><td>1,00</td><td>0,01</td><td>0,93</td><td>0,07</td></tr><tr><td>8</td><td>0,99</td><td>0,03</td><td>0,01</td><td>0,91</td><td>0,07</td></tr><tr><td>9</td><td>0,98</td><td>0,59</td><td>0,01</td><td>0,89</td><td>0,07</td></tr><tr><td></td><td colspan=\"2\">n words multi-MuRIL</td><td>BERT-movies</td><td>DistilBERT</td><td>DistilBERT*</td></tr><tr><td></td><td>0</td><td>0,74</td><td>0,32</td><td>0,07</td><td>0,95</td></tr><tr><td></td><td>1</td><td>0,63</td><td>0,13</td><td>0,01</td><td>0,45</td></tr><tr><td></td><td>2</td><td>0,19</td><td>0,10</td><td>0,01</td><td>0,17</td></tr><tr><td></td><td>3</td><td>0,01</td><td>0,12</td><td>0,01</td><td>0,14</td></tr><tr><td></td><td>4</td><td>0,02</td><td>0,08</td><td>0,00</td><td>0,11</td></tr><tr><td></td><td>5</td><td>0,12</td><td>0,10</td><td>0,03</td><td>0,16</td></tr><tr><td></td><td>6</td><td>0,44</td><td>0,09</td><td>0,03</td><td>0,22</td></tr><tr><td></td><td>7</td><td>0,35</td><td>0,07</td><td>0,02</td><td>0,17</td></tr><tr><td></td><td>8</td><td>0,31</td><td>0,06</td><td>0,01</td><td>0,09</td></tr><tr><td></td><td>9</td><td>0,25</td><td>0,04</td><td>0,01</td><td>0,08</td></tr></table>",
+        "bbox": [
+            136,
+            216,
+            860,
+            753
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Word-based attack in English on different transformer architectures - extended table.",
+        "bbox": [
+            181,
+            764,
+            811,
+            778
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "2367",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/6230ccc37e02b6da91aa631120debf20ee00d237b721fc34b2178ade49cf6d80.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n words</td><td>Vietnamese</td><td>French</td><td colspan=\"2\">German</td><td>Spanish</td><td>Spanish</td></tr><tr><td></td><td>PhoBERT</td><td>detoxify-fr</td><td colspan=\"2\">BERT-GermEval18Coarse</td><td>dehatebert</td><td>detoxify-sp</td></tr><tr><td>0</td><td>0,996</td><td>0,749</td><td colspan=\"2\">0,819</td><td>0,970</td><td>0,943</td></tr><tr><td>1</td><td>0,008</td><td>0,834</td><td colspan=\"2\">0,531</td><td>0,962</td><td>0,645</td></tr><tr><td>2</td><td>0,007</td><td>0,812</td><td colspan=\"2\">0,432</td><td>0,573</td><td>0,654</td></tr><tr><td>3</td><td>0,008</td><td>0,757</td><td colspan=\"2\">0,305</td><td>0,396</td><td>0,583</td></tr><tr><td>4</td><td>0,009</td><td>0,735</td><td colspan=\"2\">0,383</td><td>0,090</td><td>0,463</td></tr><tr><td>5</td><td>0,008</td><td>0,652</td><td colspan=\"2\">0,392</td><td>0,054</td><td>0,444</td></tr><tr><td>6</td><td>0,007</td><td>0,515</td><td colspan=\"2\">0,360</td><td>0,055</td><td>0,334</td></tr><tr><td>7</td><td>0,007</td><td>0,415</td><td colspan=\"2\">0,438</td><td>0,064</td><td>0,247</td></tr><tr><td>8</td><td>0,007</td><td>0,486</td><td colspan=\"2\">0,484</td><td>0,082</td><td>0,244</td></tr><tr><td>9</td><td>0,007</td><td>0,511</td><td colspan=\"2\">0,422</td><td>0,045</td><td>0,244</td></tr><tr><td>n words</td><td>Portuguese</td><td>Portuguese</td><td>Turkish</td><td>Russian</td><td>Russian</td><td>Russian</td></tr><tr><td></td><td>dehatebert</td><td>detoxify</td><td>detoxify-tur</td><td>Skolkovo</td><td>rubert-toxic</td><td>detoxify-ru</td></tr><tr><td>0</td><td>0,591</td><td>0,983</td><td>0,977</td><td>0,814</td><td>0,689</td><td>0,973</td></tr><tr><td>1</td><td>0,614</td><td>0,887</td><td>0,812</td><td>0,186</td><td>0,360</td><td>0,863</td></tr><tr><td>2</td><td>0,587</td><td>0,794</td><td>0,841</td><td>0,503</td><td>0,042</td><td>0,834</td></tr><tr><td>3</td><td>0,489</td><td>0,709</td><td>0,735</td><td>0,043</td><td>0,037</td><td>0,817</td></tr><tr><td>4</td><td>0,473</td><td>0,637</td><td>0,767</td><td>0,048</td><td>0,028</td><td>0,817</td></tr><tr><td>5</td><td>0,410</td><td>0,663</td><td>0,745</td><td>0,085</td><td>0,039</td><td>0,684</td></tr><tr><td>6</td><td>0,406</td><td>0,496</td><td>0,752</td><td>0,090</td><td>0,050</td><td>0,503</td></tr><tr><td>7</td><td>0,418</td><td>0,327</td><td>0,673</td><td>0,029</td><td>0,044</td><td>0,534</td></tr><tr><td>8</td><td>0,434</td><td>0,332</td><td>0,649</td><td>0,009</td><td>0,042</td><td>0,514</td></tr><tr><td>9</td><td>0,447</td><td>0,354</td><td>0,622</td><td>0,005</td><td>0,052</td><td>0,414</td></tr></table>",
+        "bbox": [
+            142,
+            158,
+            855,
+            554
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/d5135bd80e4bea2dd1932796221c2dbf86f1491296dc59680aac4dfeb4e1630d.jpg",
+        "table_caption": [
+            "Table 8: Word-based attack on languages from three different language families - extended table."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n</td><td>toxigen-rob.</td><td>Skolkovo rob.</td><td>hatebert</td><td>MuRIL</td><td>electra-hateXplain</td><td>Multi-MuRIL</td></tr><tr><td>0</td><td>0,962</td><td>0,999</td><td>0,590</td><td>0,126</td><td>0,898</td><td>0,737</td></tr><tr><td>1</td><td>0,001</td><td>0,002</td><td>0,001</td><td>0,022</td><td>0,232</td><td>0,160</td></tr><tr><td>2</td><td>0,002</td><td>0,001</td><td>0,001</td><td>0,017</td><td>0,147</td><td>0,091</td></tr></table>",
+        "bbox": [
+            129,
+            741,
+            868,
+            812
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 9: Results of sentence-based attack - extended table. n - number of added sentences, rob - roberta.",
+        "bbox": [
+            144,
+            821,
+            847,
+            834
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "2368",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2755ad08730f01d7b3670a10b2d67c6ceb84e755dd82965278651d0e99a2ffbf.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>n words</td><td colspan=\"2\">rus + eng 1</td><td>rus + eng 2</td><td>rus + sp</td><td>rus + fr</td><td>rus + ge</td></tr><tr><td></td><td colspan=\"2\">SkolkovoInstitute/ Russian_toxicity_classifier</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td></tr><tr><td>0</td><td colspan=\"2\">0,814</td><td>0,689</td><td>0,689</td><td>0,689</td><td>0,689</td></tr><tr><td>1</td><td colspan=\"2\">0,498</td><td>0,252</td><td>0,326</td><td>0,608</td><td>0,263</td></tr><tr><td>2</td><td colspan=\"2\">0,290</td><td>0,141</td><td>0,186</td><td>0,288</td><td>0,247</td></tr><tr><td>3</td><td colspan=\"2\">0,754</td><td>0,122</td><td>0,259</td><td>0,291</td><td>0,296</td></tr><tr><td>4</td><td colspan=\"2\">0,707</td><td>0,131</td><td>0,349</td><td>0,237</td><td>0,284</td></tr><tr><td>5</td><td colspan=\"2\">0,743</td><td>0,441</td><td>0,423</td><td>0,236</td><td>0,403</td></tr><tr><td>6</td><td colspan=\"2\">0,749</td><td>0,527</td><td>0,190</td><td>0,196</td><td>0,244</td></tr><tr><td>7</td><td colspan=\"2\">0,576</td><td>0,206</td><td>0,082</td><td>0,237</td><td>0,296</td></tr><tr><td>8</td><td colspan=\"2\">0,382</td><td>0,357</td><td>0,116</td><td>0,180</td><td>0,281</td></tr><tr><td>9</td><td colspan=\"2\">0,467</td><td>0,356</td><td>0,109</td><td>0,226</td><td>0,290</td></tr><tr><td>n words</td><td>eng + tur</td><td>eng + rus</td><td>eng+sp 1</td><td>eng+sp 2</td><td colspan=\"2\">eng+sp 3</td></tr><tr><td></td><td>Detoxify</td><td>Detoxify</td><td>Detoxify</td><td>electra-base-hateXplain base-hateXplain</td><td colspan=\"2\">indic-abusive-allInOne-MuRIL</td></tr><tr><td>0</td><td>0,971</td><td>0,971</td><td>0,971</td><td>0,898</td><td colspan=\"2\">0,737</td></tr><tr><td>1</td><td>0,913</td><td>0,717</td><td>0,802</td><td>0,574</td><td colspan=\"2\">0,724</td></tr><tr><td>2</td><td>0,904</td><td>0,717</td><td>0,687</td><td>0,912</td><td colspan=\"2\">0,454</td></tr><tr><td>3</td><td>0,912</td><td>0,761</td><td>0,536</td><td>0,471</td><td colspan=\"2\">0,590</td></tr><tr><td>4</td><td>0,541</td><td>0,800</td><td>0,558</td><td>0,340</td><td colspan=\"2\">0,574</td></tr><tr><td>5</td><td>0,634</td><td>0,592</td><td>0,500</td><td>0,270</td><td colspan=\"2\">0,668</td></tr><tr><td>6</td><td>0,731</td><td>0,456</td><td>0,539</td><td>0,216</td><td colspan=\"2\">0,579</td></tr><tr><td>7</td><td>0,604</td><td>0,402</td><td>0,426</td><td>0,226</td><td colspan=\"2\">0,508</td></tr><tr><td>8</td><td>0,519</td><td>0,433</td><td>0,434</td><td>0,212</td><td colspan=\"2\">0,290</td></tr><tr><td>9</td><td>0,455</td><td>0,369</td><td>0,384</td><td>0,186</td><td colspan=\"2\">0,236</td></tr></table>",
+        "bbox": [
+            124,
+            265,
+            872,
+            690
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 10: Cross-lingual word-based attack - extended table. First language listed - language of a message, second - language of added text.",
+        "bbox": [
+            112,
+            701,
+            882,
+            730
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "2369",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    }
+]

2023/No offence, Bert - I insult only humans! Multilingual sentence-level attack on toxicity detection networks/6570fd92-ec0d-40f4-8099-434dc1b0509c_model.json

@@ -0,0 +1,1283 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.09,
+                0.885,
+                0.131
+            ],
+            "angle": 0,
+            "content": "No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.244,
+                0.15,
+                0.761,
+                0.218
+            ],
+            "angle": 0,
+            "content": "Sergey Berezin, Reza Farahbakhsh, Noel Crespi  \nSAMOVAR, Télécom SudParis, Institut Polytechnique de Paris 91120 Palaiseau, France  \nsberezin@telecom-sudparis.eu"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.343,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.278,
+                0.462,
+                0.435
+            ],
+            "angle": 0,
+            "content": "We introduce a simple yet efficient sentence-level attack on black-box toxicity detector models. By adding several positive words or sentences to the end of a hateful message, we are able to change the prediction of a neural network and pass the toxicity detection system check. This approach is shown to be working on seven languages from three different language families. We also describe the defence mechanism against the aforementioned attack and discuss its limitations."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.445,
+                0.262,
+                0.46
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.47,
+                0.49,
+                0.614
+            ],
+            "angle": 0,
+            "content": "Toxicity detection systems have become a crucial part of automoderation solutions. They are now used by most social media platforms, including those with groups of people in dangerously sensitive conditions, e.g. suicide prevention, Facebook groups, victims of cyberbullying, etc. Vulnerability in such systems may have dreadful, terrific effects on people, especially those in precarious situations.(Cheng et al., 2022)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.616,
+                0.49,
+                0.727
+            ],
+            "angle": 0,
+            "content": "On the other hand, such systems can be and are used to silence the voices of criticism, which leads to the creation of echo chambers and amplifies the voice of the (powerful) minority over the voice of the majority, thereby destroying the foundation of democracy and denying the freedom of speech (Gorwa et al., 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.728,
+                0.49,
+                0.793
+            ],
+            "angle": 0,
+            "content": "This situation can be viewed as a double-edged sword, and we propose another double-edged sword that can be used to parry the blade of toxicity detection systems."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.802,
+                0.292,
+                0.818
+            ],
+            "angle": 0,
+            "content": "1.1 Task description"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.823,
+                0.488,
+                0.887
+            ],
+            "angle": 0,
+            "content": "We present an attack on toxicity detection models based on the separation of the messages intended for a person and a piece of text added to confuse an algorithm."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.888,
+                0.489,
+                0.92
+            ],
+            "angle": 0,
+            "content": "For example, this can be done by concatenating an original message with a collection of keywords"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.254,
+                0.885,
+                0.349
+            ],
+            "angle": 0,
+            "content": "of an opposite intent: \"Kill yourself, you dirty pig! Text for a bot to avoid ban: flowers, rainbow, happy, happy good\" - here only the underlined part is intended to address the human, and it is clear that the remaining part was added to avoid detection by the automoderation algorithm."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.351,
+                0.884,
+                0.399
+            ],
+            "angle": 0,
+            "content": "This represents an example of a sentence-level black-box adversarial attack on Natural Language Processing systems."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.413,
+                0.663,
+                0.427
+            ],
+            "angle": 0,
+            "content": "1.2 Related work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.434,
+                0.886,
+                0.723
+            ],
+            "angle": 0,
+            "content": "The first work suggesting the concatenation of distracting but meaningless sentences at the end of a paragraph to confuse a neural model was \"Adversarial Examples for Evaluating Reading Comprehension Systems\" (Jia and Liang, 2017) (according to the (Zhang et al., 2020)). Jia and Liang attacked question-answering systems with either manually-generated informative sentences (ADDSENT) or arbitrary sequences of words using a pool of 20 random common words (ADDANY). Both perturbations were obtained by iterative querying of the neural network until the output was changed. The authors of \"Robust Machine Comprehension Models via Adversarial Training\" (Wang and Bansal, 2018) improved this approach by varying the locations where the distracting sentences are placed and expanding the set of fake answers for generating the distracting sentences (ADDSENTDIVERSE)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.725,
+                0.885,
+                0.837
+            ],
+            "angle": 0,
+            "content": "In \"Universal Adversarial Triggers for Attacking and Analyzing NLP\" (Wallace et al., 2019) apply the gradient-based technique to construct an adversarial text. Despite being a white-box attack in its origin (due to the gradient information required in the training phase), this approach can be applied as a black-box attack during its inference."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.839,
+                0.884,
+                0.902
+            ],
+            "angle": 0,
+            "content": "The T3 model (Wang et al., 2020) utilises the autoencoder architecture to generate adversarial texts that can manipulate the question-answering models to output the targeted incorrect answer."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.527,
+                0.904,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In \"All You Need is \"Love\": Evading Hate"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2362"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.78,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2362-2369 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.198
+            ],
+            "angle": 0,
+            "content": "Speech Detection\" (Gröndahl et al., 2018) performs a sentence-level attack on hate speech detection systems by inserting typos, changing word boundaries and adding innocuous words to the original hate speech. The authors show the effectiveness of such an attack on MLP, CNN+RNN and LSTM neural networks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.21,
+                0.266,
+                0.224
+            ],
+            "angle": 0,
+            "content": "1.3 Contribution"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.231,
+                0.4,
+                0.246
+            ],
+            "angle": 0,
+            "content": "Our contributions lie in the following:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.247,
+                0.488,
+                0.358
+            ],
+            "angle": 0,
+            "content": "1) Introduction of a concept of \"To Each His Own\" attack, based on the idea of separating the messages addressed to a human and those addressed to an algorithm. The attack exploits the fact that toxicity detection models are trained on sentence-level labels and do not consider the context or the intention of the message."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.36,
+                0.488,
+                0.439
+            ],
+            "angle": 0,
+            "content": "2) A systematic approach to the generation of adversarial texts. The words and sentences are selected based on their positivity scores rather than being randomly selected from non-hate ones, as in previous works."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.441,
+                0.489,
+                0.519
+            ],
+            "angle": 0,
+            "content": "3) Demonstration of attacks on modern state-of-the-art architectures. We tested the most popular transformer-based toxicity detectors, including Google Perspective API and OpenAI Moderation API."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.521,
+                0.489,
+                0.601
+            ],
+            "angle": 0,
+            "content": "4) Highlighting a significant flaw in the current approach to toxicity detection. We show that the predicted toxicity score can be watered down by the addition of positive words to the point of effectively masking toxicity."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.602,
+                0.489,
+                0.698
+            ],
+            "angle": 0,
+            "content": "5) Multi- and cross-lingual settings of attacks. We describe the attacks on seven different languages and combinations among them. A larger number of languages leads to a larger number of possible attacks - in the case of the 5-language model, we have 120 attack scenarios."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.114,
+                0.247,
+                0.489,
+                0.698
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.711,
+                0.382,
+                0.726
+            ],
+            "angle": 0,
+            "content": "2 \"To Each His Own\" attack"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.738,
+                0.312,
+                0.752
+            ],
+            "angle": 0,
+            "content": "2.1 Word-based attack"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.488,
+                0.886
+            ],
+            "angle": 0,
+            "content": "In this attack type, the original message is appended or prepended with a sequence of positive words, with the aim that these words will be analysed by a model but passed over by a human. This allows an attacker to preserve their original message without any changes, unlike in other attacks, which substitute words with synonyms or replace original characters with their look-alikes."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.888,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In our experiment, we used words from the SentiWordNet with a positivity score \\(\\geq 0.85\\). We tested"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.882,
+                0.133
+            ],
+            "angle": 0,
+            "content": "14 openly available pretrained neural networks with the phrase \"I hate you\" and the addition of 0 to 9 positive words."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.135,
+                0.884,
+                0.311
+            ],
+            "angle": 0,
+            "content": "In case the toxicity score for a suggested phrase was estimated to be less than 0.5 - the most common threshold, we used the phrase \"I hate you dumbass\" instead. If, after attacking with 9 words, the model's output was greater than 0.5, we continued adding new positive words until the score dropped below this threshold or until we reached the maximum length of an input sequence for the model tested. For the English language, we also performed experiments with the phrase \"Text for bot:\" placed between a message and added words."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.313,
+                0.884,
+                0.36
+            ],
+            "angle": 0,
+            "content": "We conducted additional experiments with the Perspective API and OpenAI Moderation API, discovering they are also vulnerable to our attacks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.363,
+                0.884,
+                0.555
+            ],
+            "angle": 0,
+            "content": "For eight non-English models, we asked native speakers of the respective languages to translate a selected subset of words and attack phrases. Within this study, we tested the language models for seven languages aside from English: Russian, German, French, Portuguese, Spanish, Turkish and Vietnamese. It is worth noting that the first five listed languages represent three branches (Slavic, Germanic and Italic) of the Indo-European language family, and the last two are from different language families: the Turkic and the Austro-Asiatic family, respectfully."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.557,
+                0.882,
+                0.604
+            ],
+            "angle": 0,
+            "content": "By using various models and languages, we aim to investigate if the suggested attack is model- and language-agnostic."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.62,
+                0.81,
+                0.635
+            ],
+            "angle": 0,
+            "content": "2.2 Cross-lingual word-based attack"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.643,
+                0.882,
+                0.771
+            ],
+            "angle": 0,
+            "content": "The main weakness of all multilingual models lies in their greatest advantage - the ability to work with multiple languages and writing systems. Adding positive words in different languages will even more separate messages to the human and to the toxicity detection system, perhaps even denying a human the ability to read non-intended text. Examples of such texts are shown in Figure 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.773,
+                0.882,
+                0.82
+            ],
+            "angle": 0,
+            "content": "Even some monolingual models, which had been exposed to another language during the pretraining phase, fall victim to this situation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.822,
+                0.882,
+                0.886
+            ],
+            "angle": 0,
+            "content": "We also conducted cross-lingual experiments with OpenAI models, including ChatGPT in conversation mode with different prompts, and found them susceptible to attack."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.888,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Perspective API was found to be non-functional in a multilingual setting, failing with an error, as it"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2363"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.087,
+                0.486,
+                0.146
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>I hate you lióobb aǐltpruíntiúhái lüedpí</td></tr><tr><td>You stink. eşsism.dkukemmel telkari güzel ahlak</td></tr><tr><td>Goodbye, dumbass! animado melhor formidable</td></tr></table>"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.16,
+                0.49,
+                0.204
+            ],
+            "angle": 0,
+            "content": "Figure 1: Examples of cross-lingual word-based attack. The languages of the added text from the top are Russian, Turkish, and Spanish."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.228,
+                0.435,
+                0.243
+            ],
+            "angle": 0,
+            "content": "is unable to detect the language of the text."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.253,
+                0.336,
+                0.267
+            ],
+            "angle": 0,
+            "content": "2.3 Sentence-based attack"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.274,
+                0.489,
+                0.386
+            ],
+            "angle": 0,
+            "content": "Incoherent text, produced by concatenating lexically unconnected words, is easily detectable by modern language models. To address that easy detection, we experimented with another version of the attack: concatenating sentences from the Stanford Sentiment Treebank with a positivity score \\(\\geq 0.9\\) and a length of no less than 100 symbols."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.387,
+                0.49,
+                0.452
+            ],
+            "angle": 0,
+            "content": "Since sentences consist of grammatically linked words instead of just a set of random positive words, this attack variant is less obvious and more challenging to detect."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.462,
+                0.219,
+                0.477
+            ],
+            "angle": 0,
+            "content": "3 Defence"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.487,
+                0.49,
+                0.631
+            ],
+            "angle": 0,
+            "content": "We performed simple adversarial training of the DistilBERT model on the binary Jigsaw Toxic Comments dataset to improve the model's defence. We performed experiments with word- and sentence-based attacks, attacking both toxic messages and all messages in the dataset. In addition, we picked out only toxic messages and attacked half of them - in this scenario, the task was distinguishing attacked and non-attacked texts."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.643,
+                0.214,
+                0.658
+            ],
+            "angle": 0,
+            "content": "4 Results"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.668,
+                0.215,
+                0.682
+            ],
+            "angle": 0,
+            "content": "4.1 Attack"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.688,
+                0.489,
+                0.737
+            ],
+            "angle": 0,
+            "content": "During a word-based attack, both prepending and appending of the positive words showed similar results, with prepending being slightly less effective."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.738,
+                0.488,
+                0.865
+            ],
+            "angle": 0,
+            "content": "Appending nine words was enough to flip the prediction of almost every model. The \"SkolkovoInstitute RoBERTa\" toxicity classifier required 23 words to fall below 0.5, and \"English-abusive-MuRIL\" was still strongly predicting even after the addition of 252 words and reaching the length limit for the input. The results of selected experiments are shown in tables 1, 2 and \\(3^{1}\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.866,
+                0.489,
+                0.898
+            ],
+            "angle": 0,
+            "content": "The addition of the phrase \"Text for bot:\" between two parts of a text made a little difference in"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.515,
+                0.081,
+                0.88,
+                0.264
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n words</td><td>BERT</td><td>RoBERTa</td><td>ELECTRA</td></tr><tr><td>0</td><td>0,951</td><td>0,857</td><td>0,898</td></tr><tr><td>1</td><td>0,925</td><td>0,743</td><td>0,730</td></tr><tr><td>2</td><td>0,708</td><td>0,507</td><td>0,361</td></tr><tr><td>3</td><td>0,592</td><td>0,415</td><td>0,064</td></tr><tr><td>4</td><td>0,579</td><td>0,274</td><td>0,067</td></tr><tr><td>5</td><td>0,548</td><td>0,170</td><td>0,082</td></tr><tr><td>6</td><td>0,510</td><td>0,158</td><td>0,087</td></tr><tr><td>7</td><td>0,494</td><td>0,148</td><td>0,069</td></tr><tr><td>8</td><td>0,465</td><td>0,142</td><td>0,065</td></tr><tr><td>9</td><td>0,303</td><td>0,132</td><td>0,072</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.273,
+                0.884,
+                0.301
+            ],
+            "angle": 0,
+            "content": "Table 1: Word-based attack in English on different transformer architectures."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.512,
+                0.316,
+                0.884,
+                0.482
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n words</td><td>OpenAI S</td><td>OpenAI L</td><td>Per. API</td></tr><tr><td>1</td><td>0.643</td><td>0.946</td><td>0.899</td></tr><tr><td>...</td><td>...</td><td>...</td><td>...</td></tr><tr><td>10</td><td>0.421</td><td>0.847</td><td>0.846</td></tr><tr><td>11</td><td>0.381</td><td>0.705</td><td>0.852</td></tr><tr><td>12</td><td>0.427</td><td>0.74</td><td>0.852</td></tr><tr><td>13</td><td>0.453</td><td>0.761</td><td>0.852</td></tr><tr><td>14</td><td>0.426</td><td>0.665</td><td>0.853</td></tr><tr><td>15</td><td>0.362</td><td>0.681</td><td>0.853</td></tr><tr><td>16</td><td>0.265</td><td>0.716</td><td>0.853</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.491,
+                0.884,
+                0.533
+            ],
+            "angle": 0,
+            "content": "Table 2: Word-based attack in English on Perspective API and OpenAI Moderation API. Per. - perspective, S - stable, L - latest."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.561,
+                0.882,
+                0.609
+            ],
+            "angle": 0,
+            "content": "the results, i.e. it can be used to create even more apparent separation without lowering an attack's efficiency."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.61,
+                0.882,
+                0.657
+            ],
+            "angle": 0,
+            "content": "As expected, the results of cross-lingual attacks followed the same trend as those of monolingual ones. These results are shown in tables 4 and \\(5^{1}\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.659,
+                0.882,
+                0.707
+            ],
+            "angle": 0,
+            "content": "With a sentence-based attack, adding even one sentence drastically lowered the prediction scores of the models. The scores are shown in table \\(6^{1}\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.72,
+                0.619,
+                0.734
+            ],
+            "angle": 0,
+            "content": "4.2 Defence"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.742,
+                0.884,
+                0.821
+            ],
+            "angle": 0,
+            "content": "The model's F1 measure on toxic class fell from 0.80 to 0.44 after an attack with 15 sentences and to 0.79 after an attack with 50 words. After adversarial training, the model regained its F1 measure in both cases."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Interestingly, testing the word-trained model on sentence-attacked examples showed high precision and low recall (0.90 and 0.68, respectively). Testing the sentence-trained model on word-attacked examples showed the opposite result, with low precision and high recall (0.57 and 0.93, respectively)."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.136,
+                0.904,
+                0.388,
+                0.919
+            ],
+            "angle": 0,
+            "content": "1 Full tables can be found in Appendix A."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2364"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.124,
+                0.082,
+                0.478,
+                0.263
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n words</td><td>Vietnamese</td><td>French</td><td>Turkish</td></tr><tr><td>0</td><td>0,996</td><td>0,970</td><td>0,814</td></tr><tr><td>1</td><td>0,008</td><td>0,962</td><td>0,186</td></tr><tr><td>2</td><td>0,007</td><td>0,573</td><td>0,503</td></tr><tr><td>3</td><td>0,008</td><td>0,396</td><td>0,043</td></tr><tr><td>4</td><td>0,009</td><td>0,090</td><td>0,048</td></tr><tr><td>5</td><td>0,008</td><td>0,054</td><td>0,085</td></tr><tr><td>6</td><td>0,007</td><td>0,055</td><td>0,090</td></tr><tr><td>7</td><td>0,007</td><td>0,064</td><td>0,029</td></tr><tr><td>8</td><td>0,007</td><td>0,082</td><td>0,009</td></tr><tr><td>9</td><td>0,007</td><td>0,045</td><td>0,005</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.273,
+                0.49,
+                0.318
+            ],
+            "angle": 0,
+            "content": "Table 3: Word-based attack on transformer-based language models trained on the languages from three different language families."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.329,
+                0.449,
+                0.511
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n words</td><td>eng+tur</td><td>eng+rus</td><td>eng+sp</td></tr><tr><td>0</td><td>0,971</td><td>0,971</td><td>0,971</td></tr><tr><td>1</td><td>0,913</td><td>0,717</td><td>0,802</td></tr><tr><td>2</td><td>0,904</td><td>0,717</td><td>0,687</td></tr><tr><td>3</td><td>0,912</td><td>0,761</td><td>0,536</td></tr><tr><td>4</td><td>0,541</td><td>0,800</td><td>0,558</td></tr><tr><td>5</td><td>0,634</td><td>0,592</td><td>0,500</td></tr><tr><td>6</td><td>0,731</td><td>0,456</td><td>0,539</td></tr><tr><td>7</td><td>0,604</td><td>0,402</td><td>0,426</td></tr><tr><td>8</td><td>0,519</td><td>0,433</td><td>0,434</td></tr><tr><td>9</td><td>0,455</td><td>0,369</td><td>0,384</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.521,
+                0.487,
+                0.549
+            ],
+            "angle": 0,
+            "content": "Table 4: Cross-lingual word-based attack on toxicity detection models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.575,
+                0.489,
+                0.721
+            ],
+            "angle": 0,
+            "content": "The defence performed strongly, achieving an F1 score of 0.82 on binary toxic classification with word- and sentence-based attacks. The model achieved the same score on a non-attacked dataset, succeeding even better in distinguishing attacked sentences, with an F1 score of 0.99. However, defence models must be built for both word- and sentence-based attacks, as one single model will not work perfectly for both scenarios."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.733,
+                0.242,
+                0.748
+            ],
+            "angle": 0,
+            "content": "5 Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.489,
+                0.839
+            ],
+            "angle": 0,
+            "content": "All of the attack variants showed decent results in confusing toxicity detection models in every model tested (to varying degrees). The simplistic nature of such attacks could allow virtually any Internet user to exploit them to avoid automoderation systems."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "A possible countermeasure for this type of attack is adversarial training or rule-based filtering. However, a simple rule-based filter can be fooled by char-level adversarial attacks, and adversarial training can be made much more difficult by in"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.512,
+                0.082,
+                0.885,
+                0.263
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n</td><td>L en+fr</td><td>L en+ger</td><td>S en+fr</td><td>S en+ger</td></tr><tr><td>0</td><td>0.872</td><td>0.872</td><td>0.888</td><td>0.888</td></tr><tr><td>1</td><td>0.931</td><td>0.981</td><td>0.793</td><td>0.848</td></tr><tr><td>2</td><td>0.895</td><td>0.898</td><td>0.779</td><td>0.510</td></tr><tr><td>3</td><td>0.769</td><td>0.889</td><td>0.671</td><td>0.511</td></tr><tr><td>4</td><td>0.816</td><td>0.846</td><td>0.594</td><td>0.344</td></tr><tr><td>5</td><td>0.771</td><td>0.807</td><td>0.606</td><td>0.302</td></tr><tr><td>6</td><td>0.800</td><td>0.647</td><td>0.618</td><td>0.318</td></tr><tr><td>7</td><td>0.659</td><td>0.533</td><td>0.605</td><td>0.201</td></tr><tr><td>8</td><td>0.592</td><td>0.555</td><td>0.573</td><td>0.208</td></tr><tr><td>9</td><td>0.579</td><td>0.489</td><td>0.424</td><td>0.283</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.273,
+                0.883,
+                0.302
+            ],
+            "angle": 0,
+            "content": "Table 5: The results of cross-lingual attacks on OpenAI models. n - number of words added, L - latest, S - stable."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.512,
+                0.317,
+                0.9,
+                0.384
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n sentences</td><td>BERT</td><td>RoBERTa</td><td>ELECTRA</td></tr><tr><td>0</td><td>0,590</td><td>0,962</td><td>0,898</td></tr><tr><td>1</td><td>0,001</td><td>0,001</td><td>0,232</td></tr><tr><td>2</td><td>0,001</td><td>0,002</td><td>0,147</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.549,
+                0.394,
+                0.842,
+                0.408
+            ],
+            "angle": 0,
+            "content": "Table 6: Results of sentence-based attacks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.435,
+                0.882,
+                0.466
+            ],
+            "angle": 0,
+            "content": "creasing the quality and quantity of the possible text insertions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.469,
+                0.882,
+                0.515
+            ],
+            "angle": 0,
+            "content": "The complex approach, which covers all levels of perturbation, should be applied both in attack and defence real-life scenarios."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.53,
+                0.642,
+                0.545
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.557,
+                0.884,
+                0.62
+            ],
+            "angle": 0,
+            "content": "In this paper, we demonstrated a novel and effective way of bypassing toxicity detection models by appending positive words or sentences to the end of a toxic message."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.622,
+                0.884,
+                0.734
+            ],
+            "angle": 0,
+            "content": "The introduced concept of the \"To Each His Own\" attack is based on the idea of separating the messages addressed to a human and those addressed to an algorithm. The attack exploits the fact that the toxicity detection models are trained on sentence-level labels and do not consider the context or the intention of messages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.735,
+                0.883,
+                0.848
+            ],
+            "angle": 0,
+            "content": "The described attack in all its variants can be easily used on almost any toxic-detection neural model, with fairly good results. For future research, we suggest looking for a more sophisticated way of constructing insertions, perhaps with respect to the original message, making it a semantically correct addition to human-written messages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.861,
+                0.704,
+                0.877
+            ],
+            "angle": 0,
+            "content": "7 Acknowledgments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.888,
+                0.882,
+                0.918
+            ],
+            "angle": 0,
+            "content": "We thank all the volunteers who helped us with translations."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2365"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.18
+            ],
+            "angle": 0,
+            "content": "We are grateful to the reviewers and the program chair of the EMNLP conference for providing constructive feedback, suggesting additional experiments and providing different points of view. This helped us structure the paper and ensure the robustness of our results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.182,
+                0.49,
+                0.309
+            ],
+            "angle": 0,
+            "content": "Sergey Berezin wishes to record that he is grateful to Mr. Mario Fritz from Saarland University for the introduction to the domain of adversarial attacks through his insightful course on this topic. He is also grateful to Mr. Maxime Amblard from the University of Lorraine for encouraging him to pursue research in this direction and validating his ideas."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.336,
+                0.214,
+                0.351
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.358,
+                0.49,
+                0.464
+            ],
+            "angle": 0,
+            "content": "Lu Cheng, Ahmadreza Mosallanezhad, Yasin N. Silva, Deborah L. Hall, and Huan Liu. 2022. Bias mitigation for toxicity detection via sequential decisions. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 1750-1760, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.471,
+                0.49,
+                0.536
+            ],
+            "angle": 0,
+            "content": "Robert Gorwa, Reuben Binns, and Christian Katzenbach. 2020. Algorithmic content moderation: Technical and political challenges in the automation of platform governance. *Big Data Society*, 7:205395171989794."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.545,
+                0.49,
+                0.599
+            ],
+            "angle": 0,
+            "content": "Tommi Gröndahl, Luca Pajola, Mika Juuti, Mauro Conti, and N. Asokan. 2018. All you need is \"love\": Evading hate speech detection. AIsec '18, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.606,
+                0.489,
+                0.645
+            ],
+            "angle": 0,
+            "content": "Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. pages 2021-2031."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.653,
+                0.49,
+                0.706
+            ],
+            "angle": 0,
+            "content": "Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing nlp. pages 2153-2162."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.714,
+                0.49,
+                0.807
+            ],
+            "angle": 0,
+            "content": "Boxin Wang, Hengzhi Pei, Boyuan Pan, Qian Chen, Shuohang Wang, and Bo Li. 2020. T3: Treeautoencoder constrained adversarial text generation for targeted attack. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6134-6150, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.814,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Yicheng Wang and Mohit Bansal. 2018. Robust machine comprehension models via adversarial training. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 575-581, New Orleans, Louisiana. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.358,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.085,
+                0.885,
+                0.14
+            ],
+            "angle": 0,
+            "content": "Wei Emma Zhang, Quan Z. Sheng, Ahoud Alhazmi, and Chenliang Li. 2020. Adversarial attacks on deep-learning models in natural language processing: A survey. ACM Trans. Intell. Syst. Technol., 11(3)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.152,
+                0.655,
+                0.169
+            ],
+            "angle": 0,
+            "content": "A Appendix A"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2366"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.137,
+                0.217,
+                0.861,
+                0.755
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n words</td><td>BERT</td><td>roberta-base</td><td>xlm-roberta-base</td><td>toxigen-roberta</td><td>toxigen-hatebert</td></tr><tr><td>0</td><td>0,95</td><td>0,86</td><td>0,97</td><td>0,96</td><td>0,88</td></tr><tr><td>1</td><td>0,93</td><td>0,74</td><td>0,89</td><td>0,96</td><td>1,00</td></tr><tr><td>2</td><td>0,71</td><td>0,51</td><td>0,65</td><td>0,71</td><td>0,98</td></tr><tr><td>3</td><td>0,59</td><td>0,42</td><td>0,63</td><td>0,59</td><td>0,98</td></tr><tr><td>4</td><td>0,58</td><td>0,27</td><td>0,56</td><td>0,02</td><td>0,11</td></tr><tr><td>5</td><td>0,55</td><td>0,17</td><td>0,53</td><td>0,00</td><td>0,07</td></tr><tr><td>6</td><td>0,51</td><td>0,16</td><td>0,41</td><td>0,00</td><td>0,02</td></tr><tr><td>7</td><td>0,49</td><td>0,15</td><td>0,55</td><td>0,00</td><td>0,31</td></tr><tr><td>8</td><td>0,47</td><td>0,14</td><td>0,45</td><td>0,00</td><td>0,39</td></tr><tr><td>9</td><td>0,30</td><td>0,13</td><td>0,41</td><td>0,00</td><td>0,06</td></tr><tr><td>n words</td><td>roberta skolkovo</td><td>roberta facebook</td><td>MuRIL</td><td>MuRIL*</td><td>electra-hateXplain</td></tr><tr><td>0</td><td>1,00</td><td>1,00</td><td>0,13</td><td>0,96</td><td>0,90</td></tr><tr><td>1</td><td>1,00</td><td>1,00</td><td>0,10</td><td>0,95</td><td>0,73</td></tr><tr><td>2</td><td>0,99</td><td>1,00</td><td>0,03</td><td>0,94</td><td>0,36</td></tr><tr><td>3</td><td>0,92</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,06</td></tr><tr><td>4</td><td>0,98</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,07</td></tr><tr><td>5</td><td>0,98</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,08</td></tr><tr><td>6</td><td>0,99</td><td>1,00</td><td>0,03</td><td>0,94</td><td>0,09</td></tr><tr><td>7</td><td>0,99</td><td>1,00</td><td>0,01</td><td>0,93</td><td>0,07</td></tr><tr><td>8</td><td>0,99</td><td>0,03</td><td>0,01</td><td>0,91</td><td>0,07</td></tr><tr><td>9</td><td>0,98</td><td>0,59</td><td>0,01</td><td>0,89</td><td>0,07</td></tr><tr><td></td><td colspan=\"2\">n words multi-MuRIL</td><td>BERT-movies</td><td>DistilBERT</td><td>DistilBERT*</td></tr><tr><td></td><td>0</td><td>0,74</td><td>0,32</td><td>0,07</td><td>0,95</td></tr><tr><td></td><td>1</td><td>0,63</td><td>0,13</td><td>0,01</td><td>0,45</td></tr><tr><td></td><td>2</td><td>0,19</td><td>0,10</td><td>0,01</td><td>0,17</td></tr><tr><td></td><td>3</td><td>0,01</td><td>0,12</td><td>0,01</td><td>0,14</td></tr><tr><td></td><td>4</td><td>0,02</td><td>0,08</td><td>0,00</td><td>0,11</td></tr><tr><td></td><td>5</td><td>0,12</td><td>0,10</td><td>0,03</td><td>0,16</td></tr><tr><td></td><td>6</td><td>0,44</td><td>0,09</td><td>0,03</td><td>0,22</td></tr><tr><td></td><td>7</td><td>0,35</td><td>0,07</td><td>0,02</td><td>0,17</td></tr><tr><td></td><td>8</td><td>0,31</td><td>0,06</td><td>0,01</td><td>0,09</td></tr><tr><td></td><td>9</td><td>0,25</td><td>0,04</td><td>0,01</td><td>0,08</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.182,
+                0.765,
+                0.813,
+                0.78
+            ],
+            "angle": 0,
+            "content": "Table 7: Word-based attack in English on different transformer architectures - extended table."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2367"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.144,
+                0.159,
+                0.856,
+                0.555
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n words</td><td>Vietnamese</td><td>French</td><td colspan=\"2\">German</td><td>Spanish</td><td>Spanish</td></tr><tr><td></td><td>PhoBERT</td><td>detoxify-fr</td><td colspan=\"2\">BERT-GermEval18Coarse</td><td>dehatebert</td><td>detoxify-sp</td></tr><tr><td>0</td><td>0,996</td><td>0,749</td><td colspan=\"2\">0,819</td><td>0,970</td><td>0,943</td></tr><tr><td>1</td><td>0,008</td><td>0,834</td><td colspan=\"2\">0,531</td><td>0,962</td><td>0,645</td></tr><tr><td>2</td><td>0,007</td><td>0,812</td><td colspan=\"2\">0,432</td><td>0,573</td><td>0,654</td></tr><tr><td>3</td><td>0,008</td><td>0,757</td><td colspan=\"2\">0,305</td><td>0,396</td><td>0,583</td></tr><tr><td>4</td><td>0,009</td><td>0,735</td><td colspan=\"2\">0,383</td><td>0,090</td><td>0,463</td></tr><tr><td>5</td><td>0,008</td><td>0,652</td><td colspan=\"2\">0,392</td><td>0,054</td><td>0,444</td></tr><tr><td>6</td><td>0,007</td><td>0,515</td><td colspan=\"2\">0,360</td><td>0,055</td><td>0,334</td></tr><tr><td>7</td><td>0,007</td><td>0,415</td><td colspan=\"2\">0,438</td><td>0,064</td><td>0,247</td></tr><tr><td>8</td><td>0,007</td><td>0,486</td><td colspan=\"2\">0,484</td><td>0,082</td><td>0,244</td></tr><tr><td>9</td><td>0,007</td><td>0,511</td><td colspan=\"2\">0,422</td><td>0,045</td><td>0,244</td></tr><tr><td>n words</td><td>Portuguese</td><td>Portuguese</td><td>Turkish</td><td>Russian</td><td>Russian</td><td>Russian</td></tr><tr><td></td><td>dehatebert</td><td>detoxify</td><td>detoxify-tur</td><td>Skolkovo</td><td>rubert-toxic</td><td>detoxify-ru</td></tr><tr><td>0</td><td>0,591</td><td>0,983</td><td>0,977</td><td>0,814</td><td>0,689</td><td>0,973</td></tr><tr><td>1</td><td>0,614</td><td>0,887</td><td>0,812</td><td>0,186</td><td>0,360</td><td>0,863</td></tr><tr><td>2</td><td>0,587</td><td>0,794</td><td>0,841</td><td>0,503</td><td>0,042</td><td>0,834</td></tr><tr><td>3</td><td>0,489</td><td>0,709</td><td>0,735</td><td>0,043</td><td>0,037</td><td>0,817</td></tr><tr><td>4</td><td>0,473</td><td>0,637</td><td>0,767</td><td>0,048</td><td>0,028</td><td>0,817</td></tr><tr><td>5</td><td>0,410</td><td>0,663</td><td>0,745</td><td>0,085</td><td>0,039</td><td>0,684</td></tr><tr><td>6</td><td>0,406</td><td>0,496</td><td>0,752</td><td>0,090</td><td>0,050</td><td>0,503</td></tr><tr><td>7</td><td>0,418</td><td>0,327</td><td>0,673</td><td>0,029</td><td>0,044</td><td>0,534</td></tr><tr><td>8</td><td>0,434</td><td>0,332</td><td>0,649</td><td>0,009</td><td>0,042</td><td>0,514</td></tr><tr><td>9</td><td>0,447</td><td>0,354</td><td>0,622</td><td>0,005</td><td>0,052</td><td>0,414</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.171,
+                0.564,
+                0.825,
+                0.579
+            ],
+            "angle": 0,
+            "content": "Table 8: Word-based attack on languages from three different language families - extended table."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.131,
+                0.743,
+                0.869,
+                0.813
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n</td><td>toxigen-rob.</td><td>Skolkovo rob.</td><td>hatebert</td><td>MuRIL</td><td>electra-hateXplain</td><td>Multi-MuRIL</td></tr><tr><td>0</td><td>0,962</td><td>0,999</td><td>0,590</td><td>0,126</td><td>0,898</td><td>0,737</td></tr><tr><td>1</td><td>0,001</td><td>0,002</td><td>0,001</td><td>0,022</td><td>0,232</td><td>0,160</td></tr><tr><td>2</td><td>0,002</td><td>0,001</td><td>0,001</td><td>0,017</td><td>0,147</td><td>0,091</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.146,
+                0.822,
+                0.848,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Table 9: Results of sentence-based attack - extended table. n - number of added sentences, rob - roberta."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2368"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.126,
+                0.266,
+                0.873,
+                0.691
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>n words</td><td colspan=\"2\">rus + eng 1</td><td>rus + eng 2</td><td>rus + sp</td><td>rus + fr</td><td>rus + ge</td></tr><tr><td></td><td colspan=\"2\">SkolkovoInstitute/ Russian_toxicity_classifier</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td></tr><tr><td>0</td><td colspan=\"2\">0,814</td><td>0,689</td><td>0,689</td><td>0,689</td><td>0,689</td></tr><tr><td>1</td><td colspan=\"2\">0,498</td><td>0,252</td><td>0,326</td><td>0,608</td><td>0,263</td></tr><tr><td>2</td><td colspan=\"2\">0,290</td><td>0,141</td><td>0,186</td><td>0,288</td><td>0,247</td></tr><tr><td>3</td><td colspan=\"2\">0,754</td><td>0,122</td><td>0,259</td><td>0,291</td><td>0,296</td></tr><tr><td>4</td><td colspan=\"2\">0,707</td><td>0,131</td><td>0,349</td><td>0,237</td><td>0,284</td></tr><tr><td>5</td><td colspan=\"2\">0,743</td><td>0,441</td><td>0,423</td><td>0,236</td><td>0,403</td></tr><tr><td>6</td><td colspan=\"2\">0,749</td><td>0,527</td><td>0,190</td><td>0,196</td><td>0,244</td></tr><tr><td>7</td><td colspan=\"2\">0,576</td><td>0,206</td><td>0,082</td><td>0,237</td><td>0,296</td></tr><tr><td>8</td><td colspan=\"2\">0,382</td><td>0,357</td><td>0,116</td><td>0,180</td><td>0,281</td></tr><tr><td>9</td><td colspan=\"2\">0,467</td><td>0,356</td><td>0,109</td><td>0,226</td><td>0,290</td></tr><tr><td>n words</td><td>eng + tur</td><td>eng + rus</td><td>eng+sp 1</td><td>eng+sp 2</td><td colspan=\"2\">eng+sp 3</td></tr><tr><td></td><td>Detoxify</td><td>Detoxify</td><td>Detoxify</td><td>electra-base-hateXplain base-hateXplain</td><td colspan=\"2\">indic-abusive-allInOne-MuRIL</td></tr><tr><td>0</td><td>0,971</td><td>0,971</td><td>0,971</td><td>0,898</td><td colspan=\"2\">0,737</td></tr><tr><td>1</td><td>0,913</td><td>0,717</td><td>0,802</td><td>0,574</td><td colspan=\"2\">0,724</td></tr><tr><td>2</td><td>0,904</td><td>0,717</td><td>0,687</td><td>0,912</td><td colspan=\"2\">0,454</td></tr><tr><td>3</td><td>0,912</td><td>0,761</td><td>0,536</td><td>0,471</td><td colspan=\"2\">0,590</td></tr><tr><td>4</td><td>0,541</td><td>0,800</td><td>0,558</td><td>0,340</td><td colspan=\"2\">0,574</td></tr><tr><td>5</td><td>0,634</td><td>0,592</td><td>0,500</td><td>0,270</td><td colspan=\"2\">0,668</td></tr><tr><td>6</td><td>0,731</td><td>0,456</td><td>0,539</td><td>0,216</td><td colspan=\"2\">0,579</td></tr><tr><td>7</td><td>0,604</td><td>0,402</td><td>0,426</td><td>0,226</td><td colspan=\"2\">0,508</td></tr><tr><td>8</td><td>0,519</td><td>0,433</td><td>0,434</td><td>0,212</td><td colspan=\"2\">0,290</td></tr><tr><td>9</td><td>0,455</td><td>0,369</td><td>0,384</td><td>0,186</td><td colspan=\"2\">0,236</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.702,
+                0.884,
+                0.731
+            ],
+            "angle": 0,
+            "content": "Table 10: Cross-lingual word-based attack - extended table. First language listed - language of a message, second - language of added text."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "2369"
+        }
+    ]
+]

2023/No offence, Bert - I insult only humans! Multilingual sentence-level attack on toxicity detection networks/6570fd92-ec0d-40f4-8099-434dc1b0509c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09a5acb8067d5883eb6f0c0ca853cba64541fb3d7e9800081865a3ba78b4dfc5
+size 155839

2023/No offence, Bert - I insult only humans! Multilingual sentence-level attack on toxicity detection networks/full.md

@@ -0,0 +1,197 @@
+# No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks
+
+Sergey Berezin, Reza Farahbakhsh, Noel Crespi  
+SAMOVAR, Télécom SudParis, Institut Polytechnique de Paris 91120 Palaiseau, France  
+sberezin@telecom-sudparis.eu
+
+# Abstract
+
+We introduce a simple yet efficient sentence-level attack on black-box toxicity detector models. By adding several positive words or sentences to the end of a hateful message, we are able to change the prediction of a neural network and pass the toxicity detection system check. This approach is shown to be working on seven languages from three different language families. We also describe the defence mechanism against the aforementioned attack and discuss its limitations.
+
+# 1 Introduction
+
+Toxicity detection systems have become a crucial part of automoderation solutions. They are now used by most social media platforms, including those with groups of people in dangerously sensitive conditions, e.g. suicide prevention, Facebook groups, victims of cyberbullying, etc. Vulnerability in such systems may have dreadful, terrific effects on people, especially those in precarious situations.(Cheng et al., 2022).
+
+On the other hand, such systems can be and are used to silence the voices of criticism, which leads to the creation of echo chambers and amplifies the voice of the (powerful) minority over the voice of the majority, thereby destroying the foundation of democracy and denying the freedom of speech (Gorwa et al., 2020).
+
+This situation can be viewed as a double-edged sword, and we propose another double-edged sword that can be used to parry the blade of toxicity detection systems.
+
+# 1.1 Task description
+
+We present an attack on toxicity detection models based on the separation of the messages intended for a person and a piece of text added to confuse an algorithm.
+
+For example, this can be done by concatenating an original message with a collection of keywords
+
+of an opposite intent: "Kill yourself, you dirty pig! Text for a bot to avoid ban: flowers, rainbow, happy, happy good" - here only the underlined part is intended to address the human, and it is clear that the remaining part was added to avoid detection by the automoderation algorithm.
+
+This represents an example of a sentence-level black-box adversarial attack on Natural Language Processing systems.
+
+# 1.2 Related work
+
+The first work suggesting the concatenation of distracting but meaningless sentences at the end of a paragraph to confuse a neural model was "Adversarial Examples for Evaluating Reading Comprehension Systems" (Jia and Liang, 2017) (according to the (Zhang et al., 2020)). Jia and Liang attacked question-answering systems with either manually-generated informative sentences (ADDSENT) or arbitrary sequences of words using a pool of 20 random common words (ADDANY). Both perturbations were obtained by iterative querying of the neural network until the output was changed. The authors of "Robust Machine Comprehension Models via Adversarial Training" (Wang and Bansal, 2018) improved this approach by varying the locations where the distracting sentences are placed and expanding the set of fake answers for generating the distracting sentences (ADDSENTDIVERSE).
+
+In "Universal Adversarial Triggers for Attacking and Analyzing NLP" (Wallace et al., 2019) apply the gradient-based technique to construct an adversarial text. Despite being a white-box attack in its origin (due to the gradient information required in the training phase), this approach can be applied as a black-box attack during its inference.
+
+The T3 model (Wang et al., 2020) utilises the autoencoder architecture to generate adversarial texts that can manipulate the question-answering models to output the targeted incorrect answer.
+
+In "All You Need is "Love": Evading Hate
+
+Speech Detection" (Gröndahl et al., 2018) performs a sentence-level attack on hate speech detection systems by inserting typos, changing word boundaries and adding innocuous words to the original hate speech. The authors show the effectiveness of such an attack on MLP, CNN+RNN and LSTM neural networks.
+
+# 1.3 Contribution
+
+Our contributions lie in the following:
+
+1) Introduction of a concept of "To Each His Own" attack, based on the idea of separating the messages addressed to a human and those addressed to an algorithm. The attack exploits the fact that toxicity detection models are trained on sentence-level labels and do not consider the context or the intention of the message.  
+2) A systematic approach to the generation of adversarial texts. The words and sentences are selected based on their positivity scores rather than being randomly selected from non-hate ones, as in previous works.  
+3) Demonstration of attacks on modern state-of-the-art architectures. We tested the most popular transformer-based toxicity detectors, including Google Perspective API and OpenAI Moderation API.  
+4) Highlighting a significant flaw in the current approach to toxicity detection. We show that the predicted toxicity score can be watered down by the addition of positive words to the point of effectively masking toxicity.  
+5) Multi- and cross-lingual settings of attacks. We describe the attacks on seven different languages and combinations among them. A larger number of languages leads to a larger number of possible attacks - in the case of the 5-language model, we have 120 attack scenarios.
+
+# 2 "To Each His Own" attack
+
+# 2.1 Word-based attack
+
+In this attack type, the original message is appended or prepended with a sequence of positive words, with the aim that these words will be analysed by a model but passed over by a human. This allows an attacker to preserve their original message without any changes, unlike in other attacks, which substitute words with synonyms or replace original characters with their look-alikes.
+
+In our experiment, we used words from the SentiWordNet with a positivity score  $\geq 0.85$ . We tested
+
+14 openly available pretrained neural networks with the phrase "I hate you" and the addition of 0 to 9 positive words.
+
+In case the toxicity score for a suggested phrase was estimated to be less than 0.5 - the most common threshold, we used the phrase "I hate you dumbass" instead. If, after attacking with 9 words, the model's output was greater than 0.5, we continued adding new positive words until the score dropped below this threshold or until we reached the maximum length of an input sequence for the model tested. For the English language, we also performed experiments with the phrase "Text for bot:" placed between a message and added words.
+
+We conducted additional experiments with the Perspective API and OpenAI Moderation API, discovering they are also vulnerable to our attacks.
+
+For eight non-English models, we asked native speakers of the respective languages to translate a selected subset of words and attack phrases. Within this study, we tested the language models for seven languages aside from English: Russian, German, French, Portuguese, Spanish, Turkish and Vietnamese. It is worth noting that the first five listed languages represent three branches (Slavic, Germanic and Italic) of the Indo-European language family, and the last two are from different language families: the Turkic and the Austro-Asiatic family, respectfully.
+
+By using various models and languages, we aim to investigate if the suggested attack is model- and language-agnostic.
+
+# 2.2 Cross-lingual word-based attack
+
+The main weakness of all multilingual models lies in their greatest advantage - the ability to work with multiple languages and writing systems. Adding positive words in different languages will even more separate messages to the human and to the toxicity detection system, perhaps even denying a human the ability to read non-intended text. Examples of such texts are shown in Figure 1.
+
+Even some monolingual models, which had been exposed to another language during the pretraining phase, fall victim to this situation.
+
+We also conducted cross-lingual experiments with OpenAI models, including ChatGPT in conversation mode with different prompts, and found them susceptible to attack.
+
+Perspective API was found to be non-functional in a multilingual setting, failing with an error, as it
+
+<table><tr><td>I hate you lióobb aǐltpruíntiúhái lüedpí</td></tr><tr><td>You stink. eşsism.dkukemmel telkari güzel ahlak</td></tr><tr><td>Goodbye, dumbass! animado melhor formidable</td></tr></table>
+
+Figure 1: Examples of cross-lingual word-based attack. The languages of the added text from the top are Russian, Turkish, and Spanish.
+
+is unable to detect the language of the text.
+
+# 2.3 Sentence-based attack
+
+Incoherent text, produced by concatenating lexically unconnected words, is easily detectable by modern language models. To address that easy detection, we experimented with another version of the attack: concatenating sentences from the Stanford Sentiment Treebank with a positivity score  $\geq 0.9$  and a length of no less than 100 symbols.
+
+Since sentences consist of grammatically linked words instead of just a set of random positive words, this attack variant is less obvious and more challenging to detect.
+
+# 3 Defence
+
+We performed simple adversarial training of the DistilBERT model on the binary Jigsaw Toxic Comments dataset to improve the model's defence. We performed experiments with word- and sentence-based attacks, attacking both toxic messages and all messages in the dataset. In addition, we picked out only toxic messages and attacked half of them - in this scenario, the task was distinguishing attacked and non-attacked texts.
+
+# 4 Results
+
+# 4.1 Attack
+
+During a word-based attack, both prepending and appending of the positive words showed similar results, with prepending being slightly less effective.
+
+Appending nine words was enough to flip the prediction of almost every model. The "SkolkovoInstitute RoBERTa" toxicity classifier required 23 words to fall below 0.5, and "English-abusive-MuRIL" was still strongly predicting even after the addition of 252 words and reaching the length limit for the input. The results of selected experiments are shown in tables 1, 2 and  $3^{1}$ .
+
+The addition of the phrase "Text for bot:" between two parts of a text made a little difference in
+
+<table><tr><td>n words</td><td>BERT</td><td>RoBERTa</td><td>ELECTRA</td></tr><tr><td>0</td><td>0,951</td><td>0,857</td><td>0,898</td></tr><tr><td>1</td><td>0,925</td><td>0,743</td><td>0,730</td></tr><tr><td>2</td><td>0,708</td><td>0,507</td><td>0,361</td></tr><tr><td>3</td><td>0,592</td><td>0,415</td><td>0,064</td></tr><tr><td>4</td><td>0,579</td><td>0,274</td><td>0,067</td></tr><tr><td>5</td><td>0,548</td><td>0,170</td><td>0,082</td></tr><tr><td>6</td><td>0,510</td><td>0,158</td><td>0,087</td></tr><tr><td>7</td><td>0,494</td><td>0,148</td><td>0,069</td></tr><tr><td>8</td><td>0,465</td><td>0,142</td><td>0,065</td></tr><tr><td>9</td><td>0,303</td><td>0,132</td><td>0,072</td></tr></table>
+
+Table 1: Word-based attack in English on different transformer architectures.  
+
+<table><tr><td>n words</td><td>OpenAI S</td><td>OpenAI L</td><td>Per. API</td></tr><tr><td>1</td><td>0.643</td><td>0.946</td><td>0.899</td></tr><tr><td>...</td><td>...</td><td>...</td><td>...</td></tr><tr><td>10</td><td>0.421</td><td>0.847</td><td>0.846</td></tr><tr><td>11</td><td>0.381</td><td>0.705</td><td>0.852</td></tr><tr><td>12</td><td>0.427</td><td>0.74</td><td>0.852</td></tr><tr><td>13</td><td>0.453</td><td>0.761</td><td>0.852</td></tr><tr><td>14</td><td>0.426</td><td>0.665</td><td>0.853</td></tr><tr><td>15</td><td>0.362</td><td>0.681</td><td>0.853</td></tr><tr><td>16</td><td>0.265</td><td>0.716</td><td>0.853</td></tr></table>
+
+Table 2: Word-based attack in English on Perspective API and OpenAI Moderation API. Per. - perspective, S - stable, L - latest.
+
+the results, i.e. it can be used to create even more apparent separation without lowering an attack's efficiency.
+
+As expected, the results of cross-lingual attacks followed the same trend as those of monolingual ones. These results are shown in tables 4 and  $5^{1}$ .
+
+With a sentence-based attack, adding even one sentence drastically lowered the prediction scores of the models. The scores are shown in table  $6^{1}$ .
+
+# 4.2 Defence
+
+The model's F1 measure on toxic class fell from 0.80 to 0.44 after an attack with 15 sentences and to 0.79 after an attack with 50 words. After adversarial training, the model regained its F1 measure in both cases.
+
+Interestingly, testing the word-trained model on sentence-attacked examples showed high precision and low recall (0.90 and 0.68, respectively). Testing the sentence-trained model on word-attacked examples showed the opposite result, with low precision and high recall (0.57 and 0.93, respectively).
+
+<table><tr><td>n words</td><td>Vietnamese</td><td>French</td><td>Turkish</td></tr><tr><td>0</td><td>0,996</td><td>0,970</td><td>0,814</td></tr><tr><td>1</td><td>0,008</td><td>0,962</td><td>0,186</td></tr><tr><td>2</td><td>0,007</td><td>0,573</td><td>0,503</td></tr><tr><td>3</td><td>0,008</td><td>0,396</td><td>0,043</td></tr><tr><td>4</td><td>0,009</td><td>0,090</td><td>0,048</td></tr><tr><td>5</td><td>0,008</td><td>0,054</td><td>0,085</td></tr><tr><td>6</td><td>0,007</td><td>0,055</td><td>0,090</td></tr><tr><td>7</td><td>0,007</td><td>0,064</td><td>0,029</td></tr><tr><td>8</td><td>0,007</td><td>0,082</td><td>0,009</td></tr><tr><td>9</td><td>0,007</td><td>0,045</td><td>0,005</td></tr></table>
+
+Table 3: Word-based attack on transformer-based language models trained on the languages from three different language families.  
+
+<table><tr><td>n words</td><td>eng+tur</td><td>eng+rus</td><td>eng+sp</td></tr><tr><td>0</td><td>0,971</td><td>0,971</td><td>0,971</td></tr><tr><td>1</td><td>0,913</td><td>0,717</td><td>0,802</td></tr><tr><td>2</td><td>0,904</td><td>0,717</td><td>0,687</td></tr><tr><td>3</td><td>0,912</td><td>0,761</td><td>0,536</td></tr><tr><td>4</td><td>0,541</td><td>0,800</td><td>0,558</td></tr><tr><td>5</td><td>0,634</td><td>0,592</td><td>0,500</td></tr><tr><td>6</td><td>0,731</td><td>0,456</td><td>0,539</td></tr><tr><td>7</td><td>0,604</td><td>0,402</td><td>0,426</td></tr><tr><td>8</td><td>0,519</td><td>0,433</td><td>0,434</td></tr><tr><td>9</td><td>0,455</td><td>0,369</td><td>0,384</td></tr></table>
+
+The defence performed strongly, achieving an F1 score of 0.82 on binary toxic classification with word- and sentence-based attacks. The model achieved the same score on a non-attacked dataset, succeeding even better in distinguishing attacked sentences, with an F1 score of 0.99. However, defence models must be built for both word- and sentence-based attacks, as one single model will not work perfectly for both scenarios.
+
+# 5 Discussion
+
+All of the attack variants showed decent results in confusing toxicity detection models in every model tested (to varying degrees). The simplistic nature of such attacks could allow virtually any Internet user to exploit them to avoid automoderation systems.
+
+A possible countermeasure for this type of attack is adversarial training or rule-based filtering. However, a simple rule-based filter can be fooled by char-level adversarial attacks, and adversarial training can be made much more difficult by in
+
+Table 4: Cross-lingual word-based attack on toxicity detection models.  
+
+<table><tr><td>n</td><td>L en+fr</td><td>L en+ger</td><td>S en+fr</td><td>S en+ger</td></tr><tr><td>0</td><td>0.872</td><td>0.872</td><td>0.888</td><td>0.888</td></tr><tr><td>1</td><td>0.931</td><td>0.981</td><td>0.793</td><td>0.848</td></tr><tr><td>2</td><td>0.895</td><td>0.898</td><td>0.779</td><td>0.510</td></tr><tr><td>3</td><td>0.769</td><td>0.889</td><td>0.671</td><td>0.511</td></tr><tr><td>4</td><td>0.816</td><td>0.846</td><td>0.594</td><td>0.344</td></tr><tr><td>5</td><td>0.771</td><td>0.807</td><td>0.606</td><td>0.302</td></tr><tr><td>6</td><td>0.800</td><td>0.647</td><td>0.618</td><td>0.318</td></tr><tr><td>7</td><td>0.659</td><td>0.533</td><td>0.605</td><td>0.201</td></tr><tr><td>8</td><td>0.592</td><td>0.555</td><td>0.573</td><td>0.208</td></tr><tr><td>9</td><td>0.579</td><td>0.489</td><td>0.424</td><td>0.283</td></tr></table>
+
+Table 5: The results of cross-lingual attacks on OpenAI models. n - number of words added, L - latest, S - stable.  
+
+<table><tr><td>n sentences</td><td>BERT</td><td>RoBERTa</td><td>ELECTRA</td></tr><tr><td>0</td><td>0,590</td><td>0,962</td><td>0,898</td></tr><tr><td>1</td><td>0,001</td><td>0,001</td><td>0,232</td></tr><tr><td>2</td><td>0,001</td><td>0,002</td><td>0,147</td></tr></table>
+
+Table 6: Results of sentence-based attacks.
+
+creasing the quality and quantity of the possible text insertions.
+
+The complex approach, which covers all levels of perturbation, should be applied both in attack and defence real-life scenarios.
+
+# 6 Conclusion
+
+In this paper, we demonstrated a novel and effective way of bypassing toxicity detection models by appending positive words or sentences to the end of a toxic message.
+
+The introduced concept of the "To Each His Own" attack is based on the idea of separating the messages addressed to a human and those addressed to an algorithm. The attack exploits the fact that the toxicity detection models are trained on sentence-level labels and do not consider the context or the intention of messages.
+
+The described attack in all its variants can be easily used on almost any toxic-detection neural model, with fairly good results. For future research, we suggest looking for a more sophisticated way of constructing insertions, perhaps with respect to the original message, making it a semantically correct addition to human-written messages.
+
+# 7 Acknowledgments
+
+We thank all the volunteers who helped us with translations.
+
+We are grateful to the reviewers and the program chair of the EMNLP conference for providing constructive feedback, suggesting additional experiments and providing different points of view. This helped us structure the paper and ensure the robustness of our results.
+
+Sergey Berezin wishes to record that he is grateful to Mr. Mario Fritz from Saarland University for the introduction to the domain of adversarial attacks through his insightful course on this topic. He is also grateful to Mr. Maxime Amblard from the University of Lorraine for encouraging him to pursue research in this direction and validating his ideas.
+
+# References
+
+Lu Cheng, Ahmadreza Mosallanezhad, Yasin N. Silva, Deborah L. Hall, and Huan Liu. 2022. Bias mitigation for toxicity detection via sequential decisions. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 1750-1760, New York, NY, USA. Association for Computing Machinery.  
+Robert Gorwa, Reuben Binns, and Christian Katzenbach. 2020. Algorithmic content moderation: Technical and political challenges in the automation of platform governance. *Big Data Society*, 7:205395171989794.  
+Tommi Gröndahl, Luca Pajola, Mika Juuti, Mauro Conti, and N. Asokan. 2018. All you need is "love": Evading hate speech detection. AIsec '18, New York, NY, USA. Association for Computing Machinery.  
+Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. pages 2021-2031.  
+Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing nlp. pages 2153-2162.  
+Boxin Wang, Hengzhi Pei, Boyuan Pan, Qian Chen, Shuohang Wang, and Bo Li. 2020. T3: Treeautoencoder constrained adversarial text generation for targeted attack. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6134-6150, Online. Association for Computational Linguistics.  
+Yicheng Wang and Mohit Bansal. 2018. Robust machine comprehension models via adversarial training. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 575-581, New Orleans, Louisiana. Association for Computational Linguistics.
+
+Wei Emma Zhang, Quan Z. Sheng, Ahoud Alhazmi, and Chenliang Li. 2020. Adversarial attacks on deep-learning models in natural language processing: A survey. ACM Trans. Intell. Syst. Technol., 11(3).
+
+# A Appendix A
+
+<table><tr><td>n words</td><td>BERT</td><td>roberta-base</td><td>xlm-roberta-base</td><td>toxigen-roberta</td><td>toxigen-hatebert</td></tr><tr><td>0</td><td>0,95</td><td>0,86</td><td>0,97</td><td>0,96</td><td>0,88</td></tr><tr><td>1</td><td>0,93</td><td>0,74</td><td>0,89</td><td>0,96</td><td>1,00</td></tr><tr><td>2</td><td>0,71</td><td>0,51</td><td>0,65</td><td>0,71</td><td>0,98</td></tr><tr><td>3</td><td>0,59</td><td>0,42</td><td>0,63</td><td>0,59</td><td>0,98</td></tr><tr><td>4</td><td>0,58</td><td>0,27</td><td>0,56</td><td>0,02</td><td>0,11</td></tr><tr><td>5</td><td>0,55</td><td>0,17</td><td>0,53</td><td>0,00</td><td>0,07</td></tr><tr><td>6</td><td>0,51</td><td>0,16</td><td>0,41</td><td>0,00</td><td>0,02</td></tr><tr><td>7</td><td>0,49</td><td>0,15</td><td>0,55</td><td>0,00</td><td>0,31</td></tr><tr><td>8</td><td>0,47</td><td>0,14</td><td>0,45</td><td>0,00</td><td>0,39</td></tr><tr><td>9</td><td>0,30</td><td>0,13</td><td>0,41</td><td>0,00</td><td>0,06</td></tr><tr><td>n words</td><td>roberta skolkovo</td><td>roberta facebook</td><td>MuRIL</td><td>MuRIL*</td><td>electra-hateXplain</td></tr><tr><td>0</td><td>1,00</td><td>1,00</td><td>0,13</td><td>0,96</td><td>0,90</td></tr><tr><td>1</td><td>1,00</td><td>1,00</td><td>0,10</td><td>0,95</td><td>0,73</td></tr><tr><td>2</td><td>0,99</td><td>1,00</td><td>0,03</td><td>0,94</td><td>0,36</td></tr><tr><td>3</td><td>0,92</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,06</td></tr><tr><td>4</td><td>0,98</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,07</td></tr><tr><td>5</td><td>0,98</td><td>1,00</td><td>0,01</td><td>0,92</td><td>0,08</td></tr><tr><td>6</td><td>0,99</td><td>1,00</td><td>0,03</td><td>0,94</td><td>0,09</td></tr><tr><td>7</td><td>0,99</td><td>1,00</td><td>0,01</td><td>0,93</td><td>0,07</td></tr><tr><td>8</td><td>0,99</td><td>0,03</td><td>0,01</td><td>0,91</td><td>0,07</td></tr><tr><td>9</td><td>0,98</td><td>0,59</td><td>0,01</td><td>0,89</td><td>0,07</td></tr><tr><td></td><td colspan="2">n words multi-MuRIL</td><td>BERT-movies</td><td>DistilBERT</td><td>DistilBERT*</td></tr><tr><td></td><td>0</td><td>0,74</td><td>0,32</td><td>0,07</td><td>0,95</td></tr><tr><td></td><td>1</td><td>0,63</td><td>0,13</td><td>0,01</td><td>0,45</td></tr><tr><td></td><td>2</td><td>0,19</td><td>0,10</td><td>0,01</td><td>0,17</td></tr><tr><td></td><td>3</td><td>0,01</td><td>0,12</td><td>0,01</td><td>0,14</td></tr><tr><td></td><td>4</td><td>0,02</td><td>0,08</td><td>0,00</td><td>0,11</td></tr><tr><td></td><td>5</td><td>0,12</td><td>0,10</td><td>0,03</td><td>0,16</td></tr><tr><td></td><td>6</td><td>0,44</td><td>0,09</td><td>0,03</td><td>0,22</td></tr><tr><td></td><td>7</td><td>0,35</td><td>0,07</td><td>0,02</td><td>0,17</td></tr><tr><td></td><td>8</td><td>0,31</td><td>0,06</td><td>0,01</td><td>0,09</td></tr><tr><td></td><td>9</td><td>0,25</td><td>0,04</td><td>0,01</td><td>0,08</td></tr></table>
+
+Table 7: Word-based attack in English on different transformer architectures - extended table.
+
+<table><tr><td>n words</td><td>Vietnamese</td><td>French</td><td colspan="2">German</td><td>Spanish</td><td>Spanish</td></tr><tr><td></td><td>PhoBERT</td><td>detoxify-fr</td><td colspan="2">BERT-GermEval18Coarse</td><td>dehatebert</td><td>detoxify-sp</td></tr><tr><td>0</td><td>0,996</td><td>0,749</td><td colspan="2">0,819</td><td>0,970</td><td>0,943</td></tr><tr><td>1</td><td>0,008</td><td>0,834</td><td colspan="2">0,531</td><td>0,962</td><td>0,645</td></tr><tr><td>2</td><td>0,007</td><td>0,812</td><td colspan="2">0,432</td><td>0,573</td><td>0,654</td></tr><tr><td>3</td><td>0,008</td><td>0,757</td><td colspan="2">0,305</td><td>0,396</td><td>0,583</td></tr><tr><td>4</td><td>0,009</td><td>0,735</td><td colspan="2">0,383</td><td>0,090</td><td>0,463</td></tr><tr><td>5</td><td>0,008</td><td>0,652</td><td colspan="2">0,392</td><td>0,054</td><td>0,444</td></tr><tr><td>6</td><td>0,007</td><td>0,515</td><td colspan="2">0,360</td><td>0,055</td><td>0,334</td></tr><tr><td>7</td><td>0,007</td><td>0,415</td><td colspan="2">0,438</td><td>0,064</td><td>0,247</td></tr><tr><td>8</td><td>0,007</td><td>0,486</td><td colspan="2">0,484</td><td>0,082</td><td>0,244</td></tr><tr><td>9</td><td>0,007</td><td>0,511</td><td colspan="2">0,422</td><td>0,045</td><td>0,244</td></tr><tr><td>n words</td><td>Portuguese</td><td>Portuguese</td><td>Turkish</td><td>Russian</td><td>Russian</td><td>Russian</td></tr><tr><td></td><td>dehatebert</td><td>detoxify</td><td>detoxify-tur</td><td>Skolkovo</td><td>rubert-toxic</td><td>detoxify-ru</td></tr><tr><td>0</td><td>0,591</td><td>0,983</td><td>0,977</td><td>0,814</td><td>0,689</td><td>0,973</td></tr><tr><td>1</td><td>0,614</td><td>0,887</td><td>0,812</td><td>0,186</td><td>0,360</td><td>0,863</td></tr><tr><td>2</td><td>0,587</td><td>0,794</td><td>0,841</td><td>0,503</td><td>0,042</td><td>0,834</td></tr><tr><td>3</td><td>0,489</td><td>0,709</td><td>0,735</td><td>0,043</td><td>0,037</td><td>0,817</td></tr><tr><td>4</td><td>0,473</td><td>0,637</td><td>0,767</td><td>0,048</td><td>0,028</td><td>0,817</td></tr><tr><td>5</td><td>0,410</td><td>0,663</td><td>0,745</td><td>0,085</td><td>0,039</td><td>0,684</td></tr><tr><td>6</td><td>0,406</td><td>0,496</td><td>0,752</td><td>0,090</td><td>0,050</td><td>0,503</td></tr><tr><td>7</td><td>0,418</td><td>0,327</td><td>0,673</td><td>0,029</td><td>0,044</td><td>0,534</td></tr><tr><td>8</td><td>0,434</td><td>0,332</td><td>0,649</td><td>0,009</td><td>0,042</td><td>0,514</td></tr><tr><td>9</td><td>0,447</td><td>0,354</td><td>0,622</td><td>0,005</td><td>0,052</td><td>0,414</td></tr></table>
+
+Table 8: Word-based attack on languages from three different language families - extended table.  
+
+<table><tr><td>n</td><td>toxigen-rob.</td><td>Skolkovo rob.</td><td>hatebert</td><td>MuRIL</td><td>electra-hateXplain</td><td>Multi-MuRIL</td></tr><tr><td>0</td><td>0,962</td><td>0,999</td><td>0,590</td><td>0,126</td><td>0,898</td><td>0,737</td></tr><tr><td>1</td><td>0,001</td><td>0,002</td><td>0,001</td><td>0,022</td><td>0,232</td><td>0,160</td></tr><tr><td>2</td><td>0,002</td><td>0,001</td><td>0,001</td><td>0,017</td><td>0,147</td><td>0,091</td></tr></table>
+
+Table 9: Results of sentence-based attack - extended table. n - number of added sentences, rob - roberta.
+
+<table><tr><td>n words</td><td colspan="2">rus + eng 1</td><td>rus + eng 2</td><td>rus + sp</td><td>rus + fr</td><td>rus + ge</td></tr><tr><td></td><td colspan="2">SkolkovoInstitute/ Russian_toxicity_classifier</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td><td>rubert-toxic- pikabu-2ch</td></tr><tr><td>0</td><td colspan="2">0,814</td><td>0,689</td><td>0,689</td><td>0,689</td><td>0,689</td></tr><tr><td>1</td><td colspan="2">0,498</td><td>0,252</td><td>0,326</td><td>0,608</td><td>0,263</td></tr><tr><td>2</td><td colspan="2">0,290</td><td>0,141</td><td>0,186</td><td>0,288</td><td>0,247</td></tr><tr><td>3</td><td colspan="2">0,754</td><td>0,122</td><td>0,259</td><td>0,291</td><td>0,296</td></tr><tr><td>4</td><td colspan="2">0,707</td><td>0,131</td><td>0,349</td><td>0,237</td><td>0,284</td></tr><tr><td>5</td><td colspan="2">0,743</td><td>0,441</td><td>0,423</td><td>0,236</td><td>0,403</td></tr><tr><td>6</td><td colspan="2">0,749</td><td>0,527</td><td>0,190</td><td>0,196</td><td>0,244</td></tr><tr><td>7</td><td colspan="2">0,576</td><td>0,206</td><td>0,082</td><td>0,237</td><td>0,296</td></tr><tr><td>8</td><td colspan="2">0,382</td><td>0,357</td><td>0,116</td><td>0,180</td><td>0,281</td></tr><tr><td>9</td><td colspan="2">0,467</td><td>0,356</td><td>0,109</td><td>0,226</td><td>0,290</td></tr><tr><td>n words</td><td>eng + tur</td><td>eng + rus</td><td>eng+sp 1</td><td>eng+sp 2</td><td colspan="2">eng+sp 3</td></tr><tr><td></td><td>Detoxify</td><td>Detoxify</td><td>Detoxify</td><td>electra-base-hateXplain base-hateXplain</td><td colspan="2">indic-abusive-allInOne-MuRIL</td></tr><tr><td>0</td><td>0,971</td><td>0,971</td><td>0,971</td><td>0,898</td><td colspan="2">0,737</td></tr><tr><td>1</td><td>0,913</td><td>0,717</td><td>0,802</td><td>0,574</td><td colspan="2">0,724</td></tr><tr><td>2</td><td>0,904</td><td>0,717</td><td>0,687</td><td>0,912</td><td colspan="2">0,454</td></tr><tr><td>3</td><td>0,912</td><td>0,761</td><td>0,536</td><td>0,471</td><td colspan="2">0,590</td></tr><tr><td>4</td><td>0,541</td><td>0,800</td><td>0,558</td><td>0,340</td><td colspan="2">0,574</td></tr><tr><td>5</td><td>0,634</td><td>0,592</td><td>0,500</td><td>0,270</td><td colspan="2">0,668</td></tr><tr><td>6</td><td>0,731</td><td>0,456</td><td>0,539</td><td>0,216</td><td colspan="2">0,579</td></tr><tr><td>7</td><td>0,604</td><td>0,402</td><td>0,426</td><td>0,226</td><td colspan="2">0,508</td></tr><tr><td>8</td><td>0,519</td><td>0,433</td><td>0,434</td><td>0,212</td><td colspan="2">0,290</td></tr><tr><td>9</td><td>0,455</td><td>0,369</td><td>0,384</td><td>0,186</td><td colspan="2">0,236</td></tr></table>
+
+Table 10: Cross-lingual word-based attack - extended table. First language listed - language of a message, second - language of added text.

2023/No offence, Bert - I insult only humans! Multilingual sentence-level attack on toxicity detection networks/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4652de4de65d4005c37e1e946b7a6f9d07c5db455e51b5d26c4ccdf8f200a334
+size 656296

2023/Noise-Robust Fine-Tuning of Pretrained Language Models via External Guidance/569fc499-7adf-4636-bebe-f243ca93393e_content_list.json

@@ -0,0 +1,1963 @@
+[
+    {
+        "type": "text",
+        "text": "Noise-Robust Fine-Tuning of Pretrained Language Models via External Guidance",
+        "text_level": 1,
+        "bbox": [
+            196,
+            79,
+            801,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Song Wang",
+        "bbox": [
+            280,
+            126,
+            384,
+            142
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Virginia",
+        "bbox": [
+            242,
+            143,
+            423,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "sw3wv@virginia.edu",
+        "bbox": [
+            242,
+            160,
+            426,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ruocheng Guo",
+        "bbox": [
+            268,
+            187,
+            400,
+            204
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ByteDance Research",
+        "bbox": [
+            247,
+            205,
+            418,
+            221
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ruocheng.guo@bytedance.com",
+        "bbox": [
+            201,
+            223,
+            467,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhen Tan",
+        "bbox": [
+            621,
+            126,
+            709,
+            140
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Arizona State University",
+        "bbox": [
+            564,
+            142,
+            766,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ztan36@asu.edu",
+        "bbox": [
+            593,
+            160,
+            739,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jundong Li",
+        "bbox": [
+            613,
+            187,
+            717,
+            205
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Virginia",
+        "bbox": [
+            576,
+            205,
+            756,
+            221
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "jundong@virginia.edu",
+        "bbox": [
+            564,
+            222,
+            768,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            341,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Adopting a two-stage paradigm of pretraining followed by fine-tuning, Pretrained Language Models (PLMs) have achieved substantial advancements in the field of natural language processing. However, in real-world scenarios, data labels are often noisy due to the complex annotation process, making it essential to develop strategies for fine-tuning PLMs with such noisy labels. To this end, we introduce an innovative approach for fine-tuning PLMs using noisy labels, which incorporates the guidance of Large Language Models (LLMs) like ChatGPT. This guidance assists in accurately distinguishing between clean and noisy samples and provides supplementary information beyond the noisy labels, thereby boosting the learning process during fine-tuning PLMs. Extensive experiments on synthetic and real-world noisy datasets further demonstrate the superiority of our framework over the state-of-the-art baselines.",
+        "bbox": [
+            141,
+            277,
+            460,
+            561
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            571,
+            258,
+            587
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In recent years, the development of language models has significantly expanded the applications within the field of natural language processing (NLP). Fine-tuning Pretrained Language Models (PLMs) like BERT (Devlin et al., 2018) for specific downstream tasks has become an essential step in real-world implementations (Alt et al., 2020; Wang et al., 2021). In general, achieving significant performance gains in fine-tuning PLMs necessitates the availability of task-specific data for fine-tuning (Zhou and Chen, 2021; Wang et al., 2022). However, obtaining high-quality labeled datasets for this purpose poses significant challenges due to the expensive, complex, and labor-intensive nature of the annotation process (Yu et al., 2019; Bae et al., 2022). For example, large-scale datasets, often derived from web-crawling (Li et al., 2017; Song et al., 2019) or crowd-sourcing (Yan et al., 2014; Williams et al., 2017; Sakaguchi et al., 2021), frequently suffer from the presence of noisy labels.",
+        "bbox": [
+            112,
+            596,
+            490,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/37f59962e010dae0084546f1130e1b2c56617463f7426f7f6bc29b1982450a4a.jpg",
+        "image_caption": [
+            "Figure 1: The average confidence of clean and noisy samples during fine-tuning BERT on datasets  $20\\mathrm{Ng}$  and AGNews with  $20\\%$  noisy labels."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            258,
+            865,
+            401
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Prior research (Arpit et al., 2017; Cheng et al., 2021; Zhang et al., 2021c; Wang et al., 2023a) has shown that PLMs are prone to overfitting and generally deliver subpar performance when fine-tuned on datasets containing label noise. Since fine-tuning involves incorporating supervision information to enhance performance on downstream tasks, the presence of noisy labels will mislead the training process and significantly impede the efficacy of PLMs (Zhu et al., 2022). Therefore, there is an immediate need to design effective algorithms for fine-tuning PLMs in the presence of noisy labels.",
+        "bbox": [
+            507,
+            481,
+            884,
+            675
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In the context of learning from noisy labels, an intuitive approach is to separate clean samples from the noisy ones in the training set for model training (Han et al., 2018). For the remaining noisy samples, a prevalent strategy is to pseudo-label them based on model predictions to reduce the adverse impact of noise (Berthelot et al., 2019; Sohn et al., 2020). However, it remains challenging when applying this paradigm to PLMs. This is because PLMs, equipped with prior knowledge encoded in a large size of parameters, tend to easily memorize noisy samples early on during the fine-tuning process. As illustrated in Figure 1, PLMs exhibit similar prediction confidences for both clean and noisy samples, which will result in two challenges",
+        "bbox": [
+            507,
+            677,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12528",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12528-12540",
+        "bbox": [
+            208,
+            945,
+            788,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            277,
+            958,
+            719,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "when utilizing existing methods: (1) Existing approaches often rely on confidence generated by models trained on potentially noisy samples (Li et al., 2020; Karim et al., 2022). However, as shown in Figure 1, it is challenging to accurately distinguish clean and noisy samples solely based on confidence. (2) Existing methods are susceptible to erroneous information during pseudo-labeling, as they directly utilize model predictions as pseudolabels, which can be detrimental when the predictions are inaccurate. PLMs, being capable of easily remembering noisy samples, further exacerbate the risk of capturing and amplifying erroneous information during pseudo-labeling. For example, in the  $20\\mathrm{Ng}$  dataset (Lang, 1995), when an \"autos\" sample is assigned a wrong label \"hardware\", PLMs can easily memorize the erroneous information and thus cannot infer the correct labels in predictions, resulting in incorrect pseudo-labeling.",
+        "bbox": [
+            115,
+            84,
+            490,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To overcome these challenges, we propose a novel framework, named LAFT, that harnesses the power of Large Language Models (LLMs) for Fine-Tuning PLMs. We leverage LLMs trained on extensive corpora and fine-tuned with human instructions, such as GPT4 (OpenAI, 2023), to obtain external guidance in the form of confidence values. Our approach addresses the first challenge of separating reliable samples by utilizing LLM-generated confidence for each class of all training samples. By comparing these confidence with the assigned labels (i.e., the given noisy labels) of training samples, we categorize them into three disjoint subsets: the Easy Clean (EC) Set, the Hard Clean (HC) Set, and the True Noisy (TN) Set. Regarding the second challenge, we propose a novel method to incorporate LLM-generated confidence scores as robust supervision information for all samples to ensure that PLMs learn useful information from them. As LLM-generated confidence are not affected by label noise, they can provide potentially relevant labels that are useful even if they are not entirely accurate. In summary, our contributions are as follows:",
+        "bbox": [
+            115,
+            391,
+            489,
+            774
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We are the first to explore the potential of leveraging supervision information (i.e., confidence scores) generated by LLMs to tackle the noisy label problem in fine-tuning PLMs.",
+            "- We propose a novel framework LAFT that can effectively separate clean and noisy samples and learn from noisy labels, based on the LLMs-generated confidence."
+        ],
+        "bbox": [
+            115,
+            785,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We conduct extensive experiments on synthetic and real-world noisy datasets and demonstrate the superiority of our framework in fine-tuning PLMs with noisy labels.",
+        "bbox": [
+            510,
+            84,
+            880,
+            148
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            165,
+            665,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Learning from Noisy Labels",
+        "text_level": 1,
+        "bbox": [
+            509,
+            195,
+            774,
+            211
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Various strategies for learning from noisy labels have been proposed, falling primarily into three categories. The first category is Sample Selection methods, which typically employ loss or confidence to identify reliable samples for model optimization. These methods generally necessitate a predefined threshold (Li et al., 2021; Karim et al., 2022) or prior knowledge concerning the noise label rate (Han et al., 2018; Yu et al., 2019) to choose the instances. The second category, dubbed Label Transition methods, aims to learn (Tanaka et al., 2018) or generate pseudo-labels (Zhang et al., 2021c; Sohn et al., 2020) to replace the original noisy labels. Lastly, Regularization methods design robust loss functions (Liu et al., 2020; Englesson and Azizpour, 2021) or regularization techniques (Zhang and Sabuncu, 2018) that can effectively utilize all samples to enhance the model robustness against label noise. Nevertheless, these methods generally do not consider the scenario of fine-tuning a pretrained model with noisy labels.",
+        "bbox": [
+            507,
+            219,
+            884,
+            557
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Confidence-Guided Sample Separation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            573,
+            860,
+            589
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To separate noisy samples, existing works leverage the loss or confidence that provide insights into the model's prediction behavior as training proceeds (Yu et al., 2019; Karim et al., 2022). In the context of learning from noisy labels, the key concept is to leverage these dynamics as criteria for identifying and separating noisy samples. Several works propose to identify samples with lower training loss as the clean subset (Li et al., 2020; Han et al., 2018; Jiang et al., 2018; Zhao et al., 2022; Wang et al., 2023b), however, they are generally simplistic and inflexible, resulting in the selection of only easy samples. To address this limitation, alternative approaches have been proposed to effectively utilize the loss or confidence during training, as demonstrated in (Zhang et al., 2021a) and (Nishi et al., 2021). In contrast to existing methods that only rely on confidence generated by PLMs, we leverage LLMs as external guidance that provides more precise separations for fine-tuning PLMs.",
+        "bbox": [
+            507,
+            596,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12529",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Problem Definition",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            312,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this work, we study the problem of fine-tuning PLMs for text classification with noisy labels. Formally, consider a noisy training dataset containing  $n$  samples:  $\\mathcal{D}_{tr} = \\{(x_i,\\widetilde{y}_i),i = 1,2,\\ldots ,n\\}$  where  $x_{i}$  is an input sample and  $\\widetilde{y}_i$  denotes the assigned label of  $x_{i}$ . Note that  $\\widetilde{y}_i$  is potentially corrupted, and we denote the true label of  $x_{i}$  as  $\\overline{y}_i$  which is inaccessible during fine-tuning. Specifically, we aim to fine-tune PLMs with samples in  $\\mathcal{D}_{tr}$  to achieve satisfactory prediction performance on test samples, while utilizing LLMs as external guidance. Notably, the LLMs remain fixed in our framework, which means we can also use blackbox LLMs. In practice, we implement PLMs for text classification by attaching an additional classifier, which will take the output of the PLM as input and generate class probabilities for classification.",
+        "bbox": [
+            112,
+            109,
+            489,
+            382
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            395,
+            263,
+            411
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The overall framework of LAFT is illustrated in Fig. 2. In particular, we propose to divide all training samples into three subsets based on the accordance among LLM-generated confidence, PLM-generated confidence, and the assigned labels of samples. In particular, we query an LLM to provide confidence for each training sample, spanning all classes. Combining confidence obtained from both LLMs and PLMs, we perform two steps of separation to segregate the entire training set into three subsets: Easy Clean (EC) Set, Hard Clean (HC) Set, and True Noisy (TN) Set. Each of these subsets, displaying unique behaviors, is then subjected to specific noise-robust fine-tuning strategies.",
+        "bbox": [
+            112,
+            420,
+            489,
+            645
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 Confidences",
+        "text_level": 1,
+        "bbox": [
+            112,
+            657,
+            255,
+            670
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Existing studies establish a correlation between confidences and the degree to which deep models memorize specific samples during training (Arpit et al., 2017; Karim et al., 2022). It has been observed that as the model's memorization of a particular sample strengthens, the model tends to assign higher confidence for this sample (Li et al., 2023). Therefore, these methods generally employ confidences to distinguish between clean and noisy samples based on the assumption that the model cannot easily memorize noisy samples (Pleiss et al., 2020; Swayamdipta et al., 2020). However, applying these strategies to fine-tuning PLMs is suboptimal, as PLMs also present high confidence for noisy samples even in the early stage of fine-tuning, as",
+        "bbox": [
+            112,
+            677,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf296710eb0218084c85f1b93ca0b9a76fc3520c9515be4ea52901badfcdadf0.jpg",
+        "image_caption": [
+            "Figure 2: The detailed process of our framework LAFT. We perform two steps of separation to divide all training samples into three subsets with different degrees of label noise: Easy Clean Set  $\\mathcal{E}$ , Hard Clean Set  $\\mathcal{H}$ , and True Noisy Set  $\\mathcal{N}$ . We further propose three different losses to effectively learn from them:  $\\mathcal{L}_{\\mathcal{E}}$ ,  $\\mathcal{L}_{\\mathcal{H}}$ , and  $\\mathcal{L}_{\\mathcal{N}}$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            539,
+            84,
+            850,
+            309
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "shown in Fig. 1. To deal with this issue, we propose the utilization of external guidance, in the form of confidences generated by LLMs. Before delving into the specifics of our framework, we provide a formal definition of confidences. Denote the output of the final layer (i.e., the classifier) in PLMs for sample  $x_{i}$  as  $\\mathbf{z}(x_i)\\in \\mathbb{R}^N$ , where  $N$  is the number of classes. The confidence of  $x_{i}$  for the  $j$ -th class  $c_{j}$  can be represented as follows:",
+        "bbox": [
+            507,
+            412,
+            882,
+            557
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\widetilde {p} \\left(c _ {j}; x _ {i}\\right) = \\frac {\\exp \\left(z \\left(c _ {j} ; x _ {i}\\right)\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(z \\left(c _ {k} ; x _ {i}\\right)\\right)}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            561,
+            569,
+            882,
+            606
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $z(c_{j};x_{i})\\in \\mathbb{R}$  is the  $j$  -th value in  $\\mathbf{z}(x_i)$ . Notably, the confidences are obtained from  $\\mathbf{z}(x_i)$  after a softmax function and thus sum up to one.",
+        "bbox": [
+            507,
+            617,
+            882,
+            665
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Although LLM-generated confidences can provide external guidance, they are not completely accurate. Thus, We conduct a two-step sample separation process based on both LLM-generated and PLM-generated confidences, with the second step providing a more granular distinction.",
+        "bbox": [
+            507,
+            665,
+            882,
+            762
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.2 Coarse-Grained Separation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            769,
+            773,
+            784
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For the first step of separation, we aim to select samples that are easy to be identified as clean data with guidance from LLMs. Thus, we perform Coarse-grained Separation, utilizing confidences generated by LLMs with the raw text data included in the prompt. Here we provide an example of the prompt for querying the LLM to obtain the confidence value for each class:",
+        "bbox": [
+            507,
+            790,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12530",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Classify the following content:  $\\{input\\text{text}\\}$ . Select the label from  $\\{Class1\\}$ ,  $\\{Class2\\}$ , ...,  $\\{ClassN\\}$  and output a confidence value for each of them.",
+        "bbox": [
+            129,
+            87,
+            473,
+            151
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We denote the LLM-generated confidence for sample  $x_{i}$  regarding class  $c_{j}$  as  $\\widehat{p}(c_{j};x_{i})$ , where  $c_{j} \\in \\mathcal{C} = \\{c_{1},c_{2},\\ldots ,c_{N}\\}$ , and  $\\mathcal{C}$  is the class set. Then the label obtained by LLM can be represented as",
+        "bbox": [
+            112,
+            170,
+            489,
+            235
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\widehat {y} _ {i} = \\underset {j = 1, 2, \\dots , N} {\\operatorname {a r g m a x}} \\widehat {p} \\left(c _ {j}; x _ {i}\\right). \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            206,
+            247,
+            487,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To perform coarse-grained separation, we first provide an assumption as the justification:",
+        "bbox": [
+            112,
+            284,
+            487,
+            316
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Assumption 1. Samples whose assigned labels are the same as LLM-generated labels (i.e.,  $\\widehat{y}_i = \\widetilde{y}_i$ ) can be considered almost clean.",
+        "bbox": [
+            112,
+            321,
+            487,
+            368
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Note that Assumption 1 is empirically verified in Table 5 in Sec. 5.5. The underlying intuition is that the probability of an LLM-generated label being identical to the assigned label and inconsistent with the ground-truth label is significantly low and thus negligible. In concrete, this assumption allows us to segregate the training samples into two distinct subsets based on the concordance between the LLM-generated label  $\\widehat{y}_i$  and the assigned (potentially noisy) label  $\\widetilde{y}_i$  of  $x_i$ . More specifically, we define the resulting two subsets, the Easy Clean (EC) Set  $\\mathcal{E}$  and the Disagreed Set  $\\mathcal{D}$ , as follows:",
+        "bbox": [
+            112,
+            374,
+            487,
+            567
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {E} = \\left\\{x _ {i}, \\widetilde {y} _ {i} \\mid \\widehat {y} _ {i} = \\widetilde {y} _ {i} \\right\\}, \\mathcal {D} = \\left\\{x _ {i}, \\widetilde {y} _ {i} \\mid \\widehat {y} _ {i} \\neq \\widetilde {y} _ {i} \\right\\}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            124,
+            580,
+            487,
+            596
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It is naturally satisfied that  $\\mathcal{E} \\cup \\mathcal{D} = \\mathcal{D}_{tr}$  and  $\\mathcal{E} \\cap \\mathcal{D} = \\emptyset$ , where  $\\mathcal{D}_{tr}$  is the training set. Since the samples in  $\\mathcal{E}$  are already clean, we can directly fine-tune PLMs using LLM-generated labels  $\\widehat{y}_i$  based on the cross-entropy loss as follows:",
+        "bbox": [
+            112,
+            609,
+            487,
+            689
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathcal {E}} = - \\frac {1}{| \\mathcal {E} |} \\sum_ {i = 1} ^ {| \\mathcal {E} |} \\sum_ {j = 1} ^ {N} \\widehat {y} _ {i, j} \\log \\widetilde {p} \\left(c _ {j}; x _ {i}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            161,
+            701,
+            487,
+            747
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\widetilde{p}(c_j; x_i)$  is the PLM-generated confidence for  $x_i$  regarding the  $j$ -th class  $c_j$ . Here  $\\widehat{y}_{i,j} = 1$  if  $\\widehat{y}_i = c_j$ , and  $\\widehat{y}_{i,j} = 0$ , otherwise.",
+        "bbox": [
+            112,
+            758,
+            487,
+            808
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3 Fine-Grained Separation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            818,
+            357,
+            834
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It is noteworthy that, however, the samples in  $\\mathcal{D}$  are not completely noisy. This is because the LLM-generated labels are not perfectly correct, as shown in Table 5. Thus, the samples that are clean but incorrectly classified by LLMs will still",
+        "bbox": [
+            112,
+            838,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "be categorized into  $\\mathcal{D}$ . Therefore, although we can learn from samples in  $\\mathcal{E}$  directly with their LLM-generated label  $\\widetilde{y}_i$ , it is still challenging to learn from samples in  $\\mathcal{D}$ , which are only partially clean. Therefore, we further propose to separate the samples in the disagreed set  $\\mathcal{D}$  into two subsets: the Hard Clean (HC) Set  $\\mathcal{H}$  and the True Noisy (TN) Set  $\\mathcal{N}$ , referred to as fine-grained separation.",
+        "bbox": [
+            507,
+            84,
+            882,
+            212
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The intuition is that within the disagreed set  $\\mathcal{D}$ , the LLM-generated labels can be incorrect for specific hard samples with correct assigned labels, referred to as Hard Clean (HC) samples. Specifically, the ideal separation for  $\\mathcal{D}$  is as follows:",
+        "bbox": [
+            507,
+            212,
+            882,
+            292
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {H} ^ {*} = \\mathcal {D} \\cap \\left\\{x _ {i}, \\widetilde {y} _ {i} | \\widetilde {y} _ {i} = \\bar {y} _ {i} \\right\\}, \\\\ \\text {实} ^ {*} = \\mathcal {D} \\cap \\{\\dots \\widetilde {\\sim} | \\widetilde {\\sim} (\\dots) \\} \\end{array} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            586,
+            305,
+            880,
+            331
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {N} ^ {*} = \\mathcal {D} \\cap \\left\\{x _ {i}, \\widetilde {y} _ {i} | \\widetilde {y} _ {i} \\neq \\bar {y} _ {i} \\right\\},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            588,
+            326,
+            801,
+            342
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\overline{y}_i$  is the true label of  $x_i$ . Note that although this ideal separation can be completely precise for separating noisy samples in  $\\mathcal{D}$ , it is infeasible in practice, as the true labels are unknown. Therefore, to precisely separate the true noisy samples in  $\\mathcal{D}$ , we propose two thresholds for LLM-generated and PLM-generated confidences, respectively.",
+        "bbox": [
+            507,
+            353,
+            882,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to achieve more robust LLM-generated confidences distributed over the label space  $\\mathcal{C}$ , we adopt  $M$  different augmentations for each input sample to encourage input diversity while keeping the semantics immutable. Denote the augmented samples of  $x_{i}$  as  $v_{m}(x_{i})$ , where  $m = 1,2,\\ldots ,M$ , and  $M$  is the number of augmentations. We can obtain the  $M$  LLM-generated confidences as  $\\widehat{p} (c_j;v_m(x_i))$  for  $x_{i}$  regarding the  $j$ -th class  $c_{j}$  where  $m = 1,2,\\dots ,M$  and  $j = 1,2,\\dots ,N$ . We aggregate the LLM-generated confidences via the  $M$  augmentations as follows:",
+        "bbox": [
+            507,
+            467,
+            882,
+            659
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\widehat {p} _ {a} \\left(c _ {j}; x _ {i}\\right) = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\widehat {p} \\left(c _ {j}; v _ {m} \\left(x _ {i}\\right)\\right), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            559,
+            671,
+            882,
+            714
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $v_{m}(x_{i})$  is the input example after applying the  $m$ -th augmentation. As LLM-generated confidences remain fixed during fine-tuning PLMs, we adopt a fixed threshold  $\\widehat{\\tau}$  for fine-grained separation based on  $\\widehat{p}_a(c_j; x_i)$ .",
+        "bbox": [
+            507,
+            726,
+            882,
+            806
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "On the other hand, however, PLMs-generated confidences for each sample will change as the finetuning proceeds, which results in subpar separation performance if we adopt a fixed threshold to select high-confidence samples as clean ones. Intuitively, the confidences should be lower for HC samples at the beginning of fine-tuning, as the model cannot",
+        "bbox": [
+            507,
+            807,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12531",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "easily fit these hard samples within several epochs. Nonetheless, the noisy samples are relatively easier to achieve higher confidence, and thus their confidences will easily achieve higher at the beginning. Therefore, we propose an adaptive threshold  $\\widetilde{\\tau}(t)$  that will increase as fine-tuning proceeds while not reaching an excessively high value:",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\widetilde {\\tau} (t) = \\widetilde {\\tau} - \\exp (- \\lambda t), \\qquad (7)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            213,
+            208,
+            487,
+            225
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda$  and  $\\widetilde{\\tau}$  are hyper-parameters that control the value of threshold  $\\widetilde{\\tau}(t)$  as fine-tuning proceeds.  $t$  denotes the current number of fine-tuning epochs.",
+        "bbox": [
+            112,
+            236,
+            487,
+            282
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Combining the two thresholds, we can perform fine-grained separation by selecting the HC set  $\\mathcal{H}$  as follows:",
+        "bbox": [
+            112,
+            285,
+            485,
+            331
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {H} = \\mathcal {D} \\cap \\left\\{x _ {i}, \\widetilde {y} _ {i} \\mid \\max  _ {c \\in \\mathcal {C}} \\widehat {p} _ {a} (c; x _ {i}) <   \\widehat {\\tau} \\right\\} \\tag {8} \\\\ \\cap \\left\\{x _ {i}, \\widetilde {y} _ {i} \\mid \\max  _ {c \\in \\mathcal {C}} \\widetilde {p} (c; x _ {i}) <   \\widetilde {\\tau} (t) \\right\\}, \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            139,
+            338,
+            487,
+            388
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\widehat{\\tau}$  and  $\\widetilde{\\tau}(t)$  are the thresholds for LLM-generated confidence  $(\\widehat{p}_a(c_j; x_i))$  and PLMs-generated confidence  $(\\widetilde{p}(c_j; x_i))$ , respectively. In this manner, the process of separating HC and TN samples, i.e., fine-grained separation, can benefit from both the LLMs and PLMs. Then the remaining samples are categorized into the True Noisy (TN) set  $\\mathcal{N}$ :",
+        "bbox": [
+            112,
+            399,
+            489,
+            527
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {N} = \\mathcal {D} \\backslash \\mathcal {H}. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            248,
+            527,
+            485,
+            544
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.4 Learning from the Hard Clean (HC) Set",
+        "text_level": 1,
+        "bbox": [
+            112,
+            555,
+            475,
+            570
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Now we have divided the disagreed set  $\\mathcal{D}$  into  $\\mathcal{H}$  and  $\\mathcal{N}$ . Recall that ideally, samples in  $\\mathcal{H}$  are hard yet correct. Thus, for these samples, we can directly utilize their assigned labels  $\\widetilde{y}$  as training labels. Nevertheless, since the fine-grained separation of  $\\mathcal{H}$  and  $\\mathcal{N}$  cannot be perfect, the samples in  $\\mathcal{H}$  may still be noisy. As both LLMs and PLMs fail to provide a confident prediction for samples in  $\\mathcal{H}$ , we propose to prioritize the assigned label while also incorporating additional information from LLMs and PLMs. Specifically, we employ a weighted loss based on cross-entropy. The weight is enlarged if the summed confidence of the LLM and the PLM is high. We define the loss as follows:",
+        "bbox": [
+            112,
+            575,
+            487,
+            800
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathcal {H}} = - \\frac {1}{| \\mathcal {H} |} \\sum_ {i = 1} ^ {| \\mathcal {H} |} \\sum_ {j = 1} ^ {N} \\left(\\tilde {y} _ {i, j} + \\phi_ {i, j}\\right) \\log \\widetilde {p} \\left(c _ {j}; x _ {i}\\right), \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            810,
+            485,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\widetilde{p}(c_j; x_i)$  is the PLM-generated confidence for  $x_i$  regarding the  $j$ -th class  $c_j$ . Here  $\\widetilde{y}_{i,j} = 1$  if  $\\widetilde{y}_i = c_j$ , and  $\\widetilde{y}_{i,j} = 0$ , otherwise.  $\\phi_{i,j}$  acts as a",
+        "bbox": [
+            112,
+            870,
+            487,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "weight adjustment that increases when the sum of LLM confidence  $\\widehat{p}_a(c_j;x_i)$  and PLM confidence  $\\widetilde{p} (c_j;x_i)$  is larger. Intuitively, if the LLM and PLM are with high confidence regarding a specific class, then the information in their confidence can be useful as they are more likely to be correct. Therefore, we define  $\\phi_{i,j}$  as follows:",
+        "bbox": [
+            507,
+            84,
+            884,
+            198
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\phi_ {i, j} = \\max  \\left(\\widehat {p} _ {a} \\left(c _ {j}; x _ {i}\\right) + \\widetilde {p} \\left(c _ {j}; x _ {i}\\right) - \\alpha \\widetilde {\\tau} (t), 0\\right), \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            515,
+            209,
+            882,
+            241
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha > 1$  is a hyper-parameter that controls the threshold  $\\alpha \\widetilde{\\tau}(t)$  for  $\\phi_{i,j}$ . As such information can still be inaccurate, we subtract it by the threshold  $\\alpha \\widetilde{\\tau}(t)$  to control its magnitude, such that  $\\phi_{i,j}$  also acts as an adaptive loss weight for these samples.",
+        "bbox": [
+            507,
+            243,
+            882,
+            323
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.5 Learning from the True Noisy (TN) Set",
+        "text_level": 1,
+        "bbox": [
+            507,
+            334,
+            863,
+            350
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After the fine-grained separation, most samples in the True Noisy (TN) Set should be identified as noisy ones. Nevertheless, it is still challenging for PLMs to use their output as pseudo-labels for fine-tuning, as the prediction errors will accumulate and affect subsequent pseudo-labeling results. Fortunately, the confidence generated by LLMs can provide additional guidance to identify the potential labels for samples in the TN set. We first provide Remark 1 to justify the effectiveness of using LLM-generated labels for optimization.",
+        "bbox": [
+            507,
+            355,
+            882,
+            532
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Remark 1. LLM-generated labels on the True Noisy Set preserve the same accuracy as that on the whole dataset, as LLMs are not affected by noise.",
+        "bbox": [
+            507,
+            537,
+            882,
+            585
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Remark 1, as empirically verified in Sec. 5.5, demonstrates that even when we categorize most noisy samples into True Noisy Set, the LLM-generated labels can still provide decent guidance without sacrificing accuracy. Given the confidence provided by LLMs, we employ a loss that can enable the PLMs to benefit from them. Specifically,",
+        "bbox": [
+            507,
+            590,
+            882,
+            703
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathcal {N}} = - \\frac {1}{| \\mathcal {N} |} \\sum_ {i = 1} ^ {| \\mathcal {N} |} \\sum_ {j = 1} ^ {N} \\widehat {p} _ {a} (c _ {j}; x _ {i}) \\log \\widetilde {p} (c _ {j}; x _ {i}) \\\\ - \\frac {1}{| \\mathcal {N} |} \\sum_ {i = 1} ^ {| \\mathcal {N} |} \\delta \\left(x _ {i}\\right) \\cdot \\max  _ {c \\in \\mathcal {C}} \\widetilde {p} (c; x _ {i}) \\log \\max  _ {c \\in \\mathcal {C}} \\widetilde {p} (c; x _ {i}). \\tag {12} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            510,
+            715,
+            882,
+            821
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here in the first term, we leverage the confidences generated by LLMs to learn from potentially correct labels. This is because although LLMs cannot completely predict the correct labels, the confidences still preserve the potentially useful information in other incorrect but relevant labels. Such",
+        "bbox": [
+            507,
+            822,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12532",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "benefits cannot be provided by pseudo-labeling, which tends to output a definitive label. For the second term, we utilize the model predictions to exploit useful information from PLMs. The intuition is that, with the relevant label information provided by LLMs, the PLMs can learn accurate label information from the noisy samples in TN set. Consequently, if the model output tends to be confident on specific samples, we can utilize the prediction to further enhance the learning from it. Thus, we further set a threshold for this term, i.e.,  $\\delta(x_i)$ , defined as follows:",
+        "bbox": [
+            112,
+            84,
+            492,
+            277
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\delta \\left(x _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\quad \\max  _ {c \\in \\mathcal {C}} \\widetilde {p} (c; x _ {i}) > \\beta \\widetilde {\\tau} (t), \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            134,
+            287,
+            489,
+            331
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\widetilde{\\tau}(t)$  is computed by Eq. (7). To reduce the effect of confirmation bias, we multiply  $\\widetilde{\\tau}(t)$  by  $\\beta$ , a hyper-parameter that controls the final adaptive threshold, i.e.,  $\\beta \\widetilde{\\tau}(t)$ .",
+        "bbox": [
+            112,
+            341,
+            489,
+            405
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.6 Fine-tuning Objective",
+        "text_level": 1,
+        "bbox": [
+            112,
+            416,
+            332,
+            432
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "After we separate all training samples into  $\\mathcal{E}$ ,  $\\mathcal{H}$ , and  $\\mathcal{N}$ , we can combine the three individual losses for them for PLM fine-tuning. Our final fine-tuning objective can be represented as follows:",
+        "bbox": [
+            112,
+            437,
+            489,
+            501
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathcal {E}} + \\lambda_ {\\mathcal {H}} \\mathcal {L} _ {\\mathcal {H}} + \\lambda_ {\\mathcal {N}} \\mathcal {L} _ {\\mathcal {N}}, \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            512,
+            487,
+            529
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda_{\\mathcal{H}}$  and  $\\lambda_{\\mathcal{N}}$  are hyper-parameters that control the importance of  $\\mathcal{L}_{\\mathcal{H}}$  and  $\\mathcal{L}_{\\mathcal{N}}$ , respectively.",
+        "bbox": [
+            112,
+            539,
+            489,
+            571
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            583,
+            260,
+            600
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            609,
+            334,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets. To evaluate the performance of our framework, we first conduct experiments on two synthetic-noise datasets: 20Ng (Lang, 1995) and AGNews (Li and Roth, 2002; Zhang et al., 2015). Following existing works on learning from noisy labels (Patrini et al., 2017; Yao et al., 2020; Zhuang et al., 2023), we adopt three types of synthetic label noise: (1) Symmetric Noise (SN) uniformly changes labels to other classes (Han et al., 2018; Xia et al., 2021). (2) Asymmetric Noise (ASN) changes labels to other similar classes (Tanaka et al., 2018; Bae et al., 2022). (3) Instance-Dependent Noise (IDN) changes labels based on a probability in proportion to the specific sample features (Cheng et al., 2021; Yao et al., 2021). Moreover, we further conduct experiments on three real-world datasets with noisy labels: SemEval (Zhou et al., 2020), TREC (Awasthi et al., 2020), and",
+        "bbox": [
+            112,
+            629,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Hausa (Hedderich et al., 2020). More details about these datasets are provided in Appendix B.",
+        "bbox": [
+            507,
+            84,
+            880,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines. We compare our framework to state-of-the-art baselines for learning from noisy labels. In particular, we compare to (1) Base (Devlin et al., 2018) that performs fine-tuning with standard cross-entropy loss; (2) Regularization Methods: Mixup (Zhang et al., 2018) and GCE (Zhang and Sabuncu, 2018); (3) Sample-selection Methods: Co-teaching (Han et al., 2018), Co-teaching+ (Yu et al., 2019), JoCoR (Wei et al., 2020), CR (Zhou and Chen, 2021), and NPC (Bae et al., 2022). Additional details are provided in Appendix C.",
+        "bbox": [
+            507,
+            118,
+            884,
+            294
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. We use BERT (Devlin et al., 2018) as the text encoder for all datasets except Hausa, for which we use mBERT. The classifier is implemented as a fully-connected layer and randomly initialized at the beginning, while both the encoder and the classifier will be updated via gradient descent during fine-tuning. All experiments are evaluated on a clean test set, and the average accuracy along with the standard deviation over ten runs is reported for each dataset. We provide more details about the implementation in Appendix D, and our code is provided at https://github.com/SongW-SW/LAFT.",
+        "bbox": [
+            507,
+            297,
+            884,
+            507
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Comparison on Synthetic Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            524,
+            825,
+            539
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we compare our framework with other baselines on synthetic datasets  $20\\mathrm{Ng}$  and AGNews, considering different noise types and ratios. Specifically, for Symmetric Noise (SN), we conduct experiments on three different noise rates:  $20\\%$ ,  $40\\%$ , and  $60\\%$ . For the other two types of noise, i.e., Asymmetric Noise (AN) and Instance-Dependent Noise (IDN), we adopt two noise rates:  $20\\%$  and  $40\\%$ . We present the results in Table 1, 2, and 3. The key observations drawn from the outcomes are as follows: (1) Regardless of the noise type, our LAFT framework persistently surpasses other state-of-the-art baselines, thus showcasing its effectiveness in fine-tuning PLMs with noisy labels. (2) LAFT's performance improvement over other baselines is slightly more pronounced on the  $20\\mathrm{Ng}$  dataset compared to AGNews. This is attributable to  $20\\mathrm{Ng}$  containing a greater number of classes ( $N = 20$ ) as opposed to AGNews ( $N = 4$ ). Consequently, our strategy of utilizing LLM-generated confidence can capitalize on the similar labels within  $20\\mathrm{Ng}$  for PLM fine-tuning, even if LLM predictions are not entirely accurate.",
+        "bbox": [
+            507,
+            549,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12533",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/748719c58f03c945ab920b492655f390737b8b97572f49ed8faab95306bba765.jpg",
+        "table_caption": [
+            "Table 1: The overall performance of various models on synthetic noisy datasets  ${20}\\mathrm{{Ng}}$  and AGNews with Symmetric Noise (SN),where accuracy and standard deviation are reported in %,and the best results are in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td colspan=\"3\">20Ng</td><td colspan=\"3\">AGNews</td></tr><tr><td>SN Ratio</td><td>20%</td><td>40%</td><td>60%</td><td>20%</td><td>40%</td><td>60%</td></tr><tr><td>Base</td><td>79.15 ± 0.29</td><td>71.74 ± 0.07</td><td>61.87 ± 1.58</td><td>81.77 ± 0.33</td><td>78.33 ± 0.47</td><td>73.07 ± 1.49</td></tr><tr><td>Mixup</td><td>79.10 ± 0.21</td><td>70.58 ± 0.22</td><td>62.48 ± 2.04</td><td>81.25 ± 0.02</td><td>78.24 ± 0.77</td><td>73.12 ± 1.58</td></tr><tr><td>GCE</td><td>79.56 ± 0.39</td><td>70.46 ± 0.42</td><td>62.61 ± 1.01</td><td>82.98 ± 0.20</td><td>78.46 ± 0.20</td><td>72.75 ± 0.27</td></tr><tr><td>Co-teaching</td><td>76.68 ± 0.86</td><td>69.23 ± 0.62</td><td>61.39 ± 3.15</td><td>80.99 ± 0.88</td><td>76.82 ± 2.01</td><td>71.49 ± 2.74</td></tr><tr><td>Co-teaching+</td><td>78.78 ± 0.96</td><td>71.33 ± 2.31</td><td>62.90 ± 1.55</td><td>80.95 ± 0.86</td><td>77.98 ± 0.76</td><td>72.73 ± 0.98</td></tr><tr><td>JoCoR</td><td>80.11 ± 0.42</td><td>73.52 ± 0.88</td><td>63.52 ± 0.22</td><td>83.60 ± 0.54</td><td>79.45 ± 0.38</td><td>75.45 ± 0.30</td></tr><tr><td>CR</td><td>81.80 ± 0.48</td><td>74.16 ± 0.13</td><td>64.22 ± 1.25</td><td>84.09 ± 0.39</td><td>80.12 ± 0.95</td><td>75.06 ± 1.01</td></tr><tr><td>NPC</td><td>80.38 ± 0.35</td><td>72.93 ± 0.23</td><td>64.38 ± 0.38</td><td>84.28 ± 0.47</td><td>80.26 ± 1.39</td><td>73.81 ± 0.18</td></tr><tr><td>LAFT (Ours)</td><td>82.04 ± 0.11</td><td>76.93 ± 0.60</td><td>70.79 ± 1.98</td><td>90.86 ± 0.02</td><td>88.61 ± 0.25</td><td>80.79 ± 2.32</td></tr></table>",
+        "bbox": [
+            119,
+            116,
+            884,
+            302
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/ae70b3cb8c12cb6eec741e19318d513c52fa323d7f12002a69a6e7bbe03d9079.jpg",
+        "table_caption": [
+            "Table 2: The overall performance on  ${20}\\mathrm{{Ng}}$  and AGNews with Asymmetric Noise (AN)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td colspan=\"2\">20Ng</td><td colspan=\"2\">AGNews</td></tr><tr><td>AN Ratio</td><td>20%</td><td>40%</td><td>20%</td><td>40%</td></tr><tr><td>Base</td><td>75.34</td><td>59.65</td><td>82.24</td><td>76.35</td></tr><tr><td>Mixup</td><td>75.84</td><td>62.19</td><td>82.12</td><td>77.45</td></tr><tr><td>GCE</td><td>76.41</td><td>62.37</td><td>83.19</td><td>77.49</td></tr><tr><td>Co-teaching</td><td>77.49</td><td>64.55</td><td>84.15</td><td>79.16</td></tr><tr><td>Co-teaching+</td><td>77.98</td><td>63.48</td><td>86.12</td><td>81.02</td></tr><tr><td>JoCoR</td><td>81.27</td><td>70.99</td><td>87.63</td><td>82.79</td></tr><tr><td>CR</td><td>81.08</td><td>67.22</td><td>88.58</td><td>83.07</td></tr><tr><td>NPC</td><td>79.03</td><td>65.54</td><td>86.83</td><td>80.21</td></tr><tr><td>LAFT (Ours)</td><td>83.70</td><td>81.97</td><td>91.95</td><td>90.12</td></tr></table>",
+        "bbox": [
+            124,
+            349,
+            478,
+            533
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/24b8c987e9dd19e38aef317d07a7c7537b26e9125c6f18b43f419a9b8a045053.jpg",
+        "table_caption": [
+            "Table 3: The overall performance on  $20\\mathrm{Ng}$  and AGNews with Instance-Dependent Noise (IDN)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td colspan=\"2\">20Ng</td><td colspan=\"2\">AGNews</td></tr><tr><td>IDN Ratio</td><td>20%</td><td>40%</td><td>20%</td><td>40%</td></tr><tr><td>Base</td><td>78.23</td><td>70.28</td><td>85.40</td><td>78.66</td></tr><tr><td>Mixup</td><td>78.21</td><td>70.45</td><td>84.71</td><td>79.22</td></tr><tr><td>GCE</td><td>79.99</td><td>71.74</td><td>87.11</td><td>80.49</td></tr><tr><td>Co-teaching</td><td>77.43</td><td>70.76</td><td>84.09</td><td>78.43</td></tr><tr><td>Co-teaching+</td><td>79.89</td><td>71.52</td><td>88.21</td><td>82.30</td></tr><tr><td>JoCoR</td><td>81.83</td><td>74.04</td><td>88.24</td><td>82.66</td></tr><tr><td>CR</td><td>83.04</td><td>75.51</td><td>90.72</td><td>84.38</td></tr><tr><td>NPC</td><td>81.84</td><td>74.45</td><td>89.90</td><td>82.40</td></tr><tr><td>LAFT (Ours)</td><td>83.61</td><td>80.49</td><td>91.94</td><td>89.34</td></tr></table>",
+        "bbox": [
+            519,
+            349,
+            873,
+            533
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "(3) Certain regularization methods, such as Mixup and GCE, exhibit subpar performance when applied to datasets with a higher noise ratio of  $60\\%$ . This is because these methods rely on all training samples for PLM fine-tuning, making them more susceptible to strong label noises.",
+        "bbox": [
+            112,
+            558,
+            487,
+            653
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Comparison on Real-world Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            668,
+            445,
+            684
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we conduct experiments on real-world datasets: SemEval (Zhou et al., 2020), TREC (Awasthi et al., 2020), and Hausa (Hedderich et al., 2020). From the results in Table 4, we observe that LAFT outperforms other baseline methods on all three datasets. Moreover, the results of LAFT are noticeably competitive on TREC and Hausa with relatively higher noise ratios, which further demonstrates the strong robustness and generalization ability of LAFT to label noise.",
+        "bbox": [
+            112,
+            690,
+            489,
+            851
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            864,
+            280,
+            879
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we systematically remove specific components from our framework and ana",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "lyze the resulting impact on performance. We conduct experiments with the following variants: (1) LAFT\\C performs coarse-grained separation without LLMs and thus only separates the Easy Clean Set based on PLM predictions. (2) LAFT\\F performs fine-grained separation without LLMs, which means when separating Hard Clean Set, only PLM-generated confidence are used. (3) LAFT\\N removes our proposed loss  $\\mathcal{L}_{\\mathcal{N}}$  for True Noisy Set and replaces it with pseudo-labeling based on the most confident PLM prediction. From the results presented in Figure 3, we observe that LAFT outperforms all variants, which verifies the effectiveness of these designs in LAFT. Specifically, removing the learning loss  $\\mathcal{L}_{\\mathcal{N}}$  leads to significant performance degradation, indicating that such a design can effectively alleviate the adverse impact of noisy labels. Moreover, without the fine-grained separation strategy, the performance deteriorates rapidly when the noise ratio is larger, indicating the importance of fine-grained separation in the presence of higher noise ratios.",
+        "bbox": [
+            507,
+            565,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12534",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/f0ede45c0dd30cb5a4a09f6ce46916bd2220e933a8462cdcecc61a455b01cfee.jpg",
+        "table_caption": [
+            "Table 4: The overall performance of various models on three real-world noisy datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>SemEval</td><td>TREC</td><td>Hausa</td></tr><tr><td>Noise Ratio</td><td>16.00%</td><td>38.56%</td><td>50.37%</td></tr><tr><td>Base</td><td>70.61</td><td>67.42</td><td>47.80</td></tr><tr><td>Mixup</td><td>73.41</td><td>68.44</td><td>47.66</td></tr><tr><td>GCE</td><td>71.91</td><td>67.86</td><td>48.12</td></tr><tr><td>Co-teaching</td><td>72.13</td><td>68.19</td><td>47.56</td></tr><tr><td>Co-teaching+</td><td>72.24</td><td>69.83</td><td>48.25</td></tr><tr><td>JoCoR</td><td>70.11</td><td>66.62</td><td>46.48</td></tr><tr><td>CR</td><td>72.94</td><td>68.90</td><td>48.34</td></tr><tr><td>NPC</td><td>71.22</td><td>67.84</td><td>47.48</td></tr><tr><td>LAFT (Ours)</td><td>73.56</td><td>72.34</td><td>51.71</td></tr></table>",
+        "bbox": [
+            126,
+            116,
+            475,
+            302
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c31d82dbdf869e580a2f9c61eceef8f3d0843b796d5960c8a6a7325015caf233.jpg",
+        "image_caption": [
+            "Figure 3: Ablation study of our framework on  $20\\mathrm{Ng}$ . S-0.2 and S-0.4 (A, or ID) refer to the noise ratios of  $20\\%$  and  $40\\%$  for SN (AN, or IDN), respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            326,
+            473,
+            470
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.5 Evaluation of LLMs-generated Labels",
+        "text_level": 1,
+        "bbox": [
+            112,
+            555,
+            460,
+            570
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we evaluate the effectiveness of LLMs regarding the quality of generated confidence, presented in Table 5. Specifically, we report the accuracy of LLM-generated prediction on the training set in  $20\\mathrm{Ng}$  and AGNews. The statistics are from experiments on  $20\\mathrm{Ng}$  and AGNews with  $20\\%$  Symmetric Noise (SN). From the results, we can observe that: (1) For Easy Clean Set, the LLM-generated labels exhibit a remarkably high degree of correctness. This empirical finding provides the justification for Assumption 1, which allows for leveraging LLMs to perform coarse-grained separation in our framework. (2) For the TN set, LLM-generated labels are not entirely correct, while still preserving similar accuracy compared to that on all samples. This result empirically verifies Remark 1 and justifies our strategy of utilizing LMM-generated confidence for learning from the TN set. (3) Considering the overall result, LLM-generated labels generally exhibit lower accuracy than PLM-based baselines. That being said, LLMs cannot be",
+        "bbox": [
+            112,
+            581,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/16fa0df6b65518263846b1774ecd3f0298e14f8361b8f30839ca504f03ab6f90.jpg",
+        "table_caption": [
+            "Table 5: The LLM-generated prediction accuracy on the training set in  $20\\mathrm{Ng}$  and AGNews in the ideal separation and real separation (shown in gray) during fine-tuning."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>EC Set</td><td>HC Set</td><td>TN Set</td><td>Overall</td></tr><tr><td rowspan=\"2\">20Ng</td><td>99.46%</td><td>0%</td><td>76.47%</td><td>75.34%</td></tr><tr><td>99.46%</td><td>3.58%</td><td>69.44%</td><td>75.34%</td></tr><tr><td rowspan=\"2\">AGNews</td><td>99.03%</td><td>0%</td><td>79.35%</td><td>81.41%</td></tr><tr><td>99.03%</td><td>2.05%</td><td>76.07%</td><td>81.41%</td></tr></table>",
+        "bbox": [
+            521,
+            131,
+            868,
+            219
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4721ed0b99bd091e0337b9638f15f3c1c9508eb82d4f2f4bece63748b5ffd5a9.jpg",
+        "image_caption": [
+            "Figure 4: The results of our framework and the best baseline CR on  $20\\mathrm{Ng}$  with different noise ratios for three types of noise: SN, AN, and IDN."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            239,
+            868,
+            382
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "directly used to replace PLMs for the text classification task when the noise ratio is not extremely high. Nevertheless, our strategy can effectively leverage the LLM-generated confidences, despite their lack of complete correctness, to enhance the fine-tuning performance of PLMs with noisy labels.",
+        "bbox": [
+            507,
+            468,
+            882,
+            565
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.6 Results with Different Noise Ratios",
+        "text_level": 1,
+        "bbox": [
+            507,
+            575,
+            830,
+            589
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 presents the evaluation of our proposed framework across a spectrum of noise ratios on  $20\\mathrm{Ng}$ , ranging from 0.1 to 0.9, while considering three distinct types of noise: SN, AN, and IDN. From the results, several significant observations are discovered: (1) With increasing noise ratios, all baseline methods consistently show a deterioration in performance. This decline can be tied to the amplified difficulty of label noise during PLM fin-tuning. (2) Despite the escalating noise ratios, our framework maintains a relatively more robust performance. This is attributed to our sample separation strategy which adeptly discerns clean samples within the training set, thus alleviating the adverse effects of label noise. (3) Our framework demonstrates a slower performance decrease with AN noise compared to other types of noise. This can be credited to the adaptive nature of our approach that employs LLM-generated confidence to effectively learn from similar labels for each sample.",
+        "bbox": [
+            507,
+            596,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12535",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            247,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper delves into the issue of fine-tuning Pre-trained Language Models (PLMs) with noisy labels for text classification tasks. We address the detrimental effects of label noise by harnessing external guidance from Large Language Models (LLMs) in the form of confidence. Our approach entails a two-step separation strategy that accurately segregates all training samples into three distinct subsets, each treated with innovative noise-resilient strategies. In summary, our framework adeptly finetunes PLMs in the face of noisy samples, requiring minimal external guidance from LLMs.",
+        "bbox": [
+            115,
+            108,
+            489,
+            299
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            115,
+            312,
+            284,
+            328
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Song Wang and Jundong Li are supported by the National Science Foundation under grants (IIS-2006844, IIS-2144209, IIS-2223769, CNS2154962, and BCS-2228534), the Commonwealth Cyber Initiative awards (VV-1Q23-007, HV2Q23-003, and VV-1Q24-011), the JP Morgan Chase Faculty Research Award, the Cisco Faculty Research Award, the Jefferson Lab subcontract, and the UVA 4-VA collaborative research grant.",
+        "bbox": [
+            115,
+            336,
+            487,
+            480
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitation",
+        "text_level": 1,
+        "bbox": [
+            115,
+            492,
+            210,
+            507
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Despite the promising results, our work presents several limitations that we must acknowledge: (1) Dependence on Large Language Models (LLMs): Our framework leverages the guidance from LLMs in the form of confidences. This implies that the quality of these confidences and hence the effectiveness of our model are closely tied to the performance of the LLM. Unsatisfactory LLM performance could thus directly impact our framework's efficacy. (2) Noise Types and Ratios: The experiments in this paper mainly focus on three types of noise: Symmetric Noise (SN), Asymmetric Noise (AN), and Instance-Dependent Noise (IDN), with noise ratios up to  $60\\%$ . However, there may be other noise types or higher noise ratios in practice, and it remains to be investigated how well our framework performs under these circumstances. (3) Limitations in Sample Separation Strategies: The efficacy of our framework relies on accurately dividing the training samples into clean and noisy subsets. If this distinction is not clear or becomes more complex, our current approach may encounter difficulties. (4) Domain-Specific Text Classification: While the experiments are performed on specific text classification tasks, we do not investigate",
+        "bbox": [
+            115,
+            517,
+            487,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "the effectiveness of our approach in more domain-specific contexts where the nature of the noise could be different. (5) Computational Costs: Finally, our approach entails using LLMs which can be computationally expensive and could thus pose a challenge when applied to large datasets or in resource-constrained environments. In summary, future research could focus on overcoming these limitations and exploring the adaptability of our proposed framework to other noise types, higher noise ratios, more complex noise patterns, as well as different task domains.",
+        "bbox": [
+            512,
+            84,
+            882,
+            275
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            512,
+            294,
+            658,
+            309
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This research adheres to the principles of ethical conduct in artificial intelligence research. The development and utilization of our work are carried out with full recognition of the potential implications. While our method improves the performance of Pretrained Language Models (PLMs) with noisy labels in text classification tasks, we acknowledge that it could be potentially used in contexts that may involve ethical concerns, such as disinformation or manipulation of opinion on social platforms. Our experiments are performed on five publicly available datasets, namely  $20\\mathrm{Ng}$ , AGNews, SemEval, TREC, and Hausa. The datasets are already anonymized and do not contain any personally identifiable information. Moreover, we have complied with all guidelines and standards for the use of these datasets. Furthermore, the use of Large Language Models (LLMs) in our work is subject to the terms and conditions put forth by the creators and hosts of these models. We only empty them for the purpose of enhancing the fine-tuning performance of PLMs and do not exploit them for any inappropriate or unauthorized purposes. While our work focuses on improving the robustness of PLMs to label noise in data, we recognize the broader societal concerns related to the potential misuse of our framework. We advocate for the responsible use of our research findings and encourage continued discourse around the ethical use of machine learning and artificial intelligence technologies in real-world applications. Finally, it is crucial to note that the development of technologies and methodologies in machine learning should be done in conjunction with ongoing considerations about their ethical implications, potential misuse, and societal impacts. It is our responsibility to foster an environment of ethical vigilance in AI research and development.",
+        "bbox": [
+            512,
+            323,
+            882,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12536",
+        "bbox": [
+            478,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Christoph Alt, Aleksandra Gabryszak, and Leonhard Hennig. 2020. Tacred revisited: A thorough evaluation of the tacred relation extraction task. In ACL.",
+            "Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. 2017. A closer look at memorization in deep networks. In ICML.",
+            "Abhijeet Awasthi, Sabyasachi Ghosh, Rasna Goyal, and Sunita Sarawagi. 2020. Learning from rules generalizing labeled exemplars. In *ICLR*.",
+            "HeeSun Bae, Seungjae Shin, Byeonghu Na, JoonHo Jang, Kyungwoo Song, and Il-Chul Moon. 2022. From noisy prediction to true label: Noisy prediction calibration via generative model. In ICML.",
+            "David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. 2019. Mixmatch: A holistic approach to semi-supervised learning. NeurIPS.",
+            "Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. 2021. Learning with instance-dependent label noise: A sample sieve approach. In ICLR.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.",
+            "Erik Englesson and Hossein Azizpour. 2021. Generalized jensen-shannon divergence loss for learning with noisy labels. NeurIPS.",
+            "Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. 2018. Co-teaching: Robust training of deep neural networks with extremely noisy labels. NeurIPS.",
+            "Michael A Hedderich, David I Adelani, Dawei Zhu, Jesujoba Alabi, Udia Markus, and Dietrich Klakow. 2020. Transfer learning and distant supervision for multilingual transformer models: A study on african languages. In EMNLP.",
+            "Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. 2018. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML.",
+            "Nazmul Karim, Mamshad Nayeem Rizve, Nazanin Rahnavard, Ajmal Mian, and Mubarak Shah. 2022. Unicon: Combating label noise through uniform selection and contrastive learning. In CVPR.",
+            "Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In ICLR.",
+            "Ken Lang. 1995. Newsweeder: Learning to filter netnews. In Machine Learning Proceedings."
+        ],
+        "bbox": [
+            115,
+            107,
+            487,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. AI Open.",
+            "Junnan Li, Richard Socher, and Steven CH Hoi. 2020. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR.",
+            "Junnan Li, Caiming Xiong, and Steven CH Hoi. 2021. Learning from noisy data with robust representation learning. In ICCV.",
+            "Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. 2017. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862.",
+            "Xin Li and Dan Roth. 2002. Learning question classifiers. In COLING.",
+            "Yifan Li, Hu Han, Shiguang Shan, and Xilin Chen. 2023. Disc: Learning from noisy labels via dynamic instance-specific selection and correction. In CVPR.",
+            "Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. 2020. Early-learning regularization prevents memorization of noisy labels. NeurIPS.",
+            "Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. 2021. Augmentation strategies for learning with noisy labels. In CVPR.",
+            "OpenAI. 2023. Gpt-4 technical report.",
+            "Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. 2017. Making deep neural networks robust to label noise: A loss correction approach. In CVPR.",
+            "Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. 2020. Identifying unlabeled data using the area under the margin ranking. NeurIPS.",
+            "Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM.",
+            "Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. 2020. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS.",
+            "Hwanjun Song, Minseok Kim, and Jae-Gil Lee. 2019. Selfie: Refurbishing unclean samples for robust deep learning. In ICML.",
+            "Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In EMNLP."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12537",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. 2018. Joint optimization framework for learning with noisy labels. In CVPR.",
+            "Song Wang, Xingbo Fu, Kaize Ding, Chen Chen, Huiyuan Chen, and Jundong Li. 2023a. Federated few-shot learning. In SIGKDD.",
+            "Song Wang, Jing Ma, Lu Cheng, and Jundong Li. 2023b. Fair few-shot learning with auxiliary sets. In ECAI.",
+            "Yaqing Wang, Song Wang, Yanyan Li, and Dejing Dou. 2022. Recognizing medical search query intent by few-shot learning. In SIGIR.",
+            "Yaqing Wang, Song Wang, Quanming Yao, and Dejing Dou. 2021. Hierarchical heterogeneous graph representation learning for short text classification. In EMNLP.",
+            "Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. 2020. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR.",
+            "Jason Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. In EMNLP-IJCNLP.",
+            "Adina Williams, Nikita Nangia, and Samuel R Bowman. 2017. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426.",
+            "Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nanan Wang, Zongyuan Ge, and Yi Chang. 2021. Robust early-learning: Hinding the memorization of noisy labels. In ICLR.",
+            "Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. 2019. Unsupervised data augmentation for consistency training. arXiv preprint arXiv:1904.12848.",
+            "Yan Yan, Rómer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. 2014. Learning from multiple annotators with varying expertise. Machine Learning.",
+            "Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. 2021. Instance-dependent label-noise learning under a structural causal model. In NeurIPS.",
+            "Yu Yao, Tongliang Liu, Bo Han, Mingming Gong, Jiankang Deng, Gang Niu, and Masashi Sugiyama. 2020. Dual t: Reducing estimation error for transition matrix in label-noise learning. NeurIPS.",
+            "Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. 2019. How does disagreement help generalization against label corruption? In ICML.",
+            "Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. 2021a. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            916
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. 2018. mixup: Beyond empirical risk minimization. In ICLR.",
+            "Jieyu Zhang, Yue Yu, Yinghao Li, Yujing Wang, Yaming Yang, Mao Yang, and Alexander Ratner. 2021b. WRENCH: A comprehensive benchmark for weak supervision. In NeurIPS Datasets and Benchmarks Track.",
+            "Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. NIPS.",
+            "Yivan Zhang, Gang Niu, and Masashi Sugiyama. 2021c. Learning noise transition matrix from only noisy labels via total variation regularization. In ICML.",
+            "Zhilu Zhang and Mert Sabuncu. 2018. Generalized cross entropy loss for training deep neural networks with noisy labels. NeurIPS.",
+            "Ganlong Zhao, Guanbin Li, Yipeng Qin, Feng Liu, and Yizhou Yu. 2022. Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In ECCV.",
+            "Wenxuan Zhou and Muhao Chen. 2021. Learning from noisy labels for entity-centric information extraction. In EMNLP.",
+            "Wenxuan Zhou, Hongtao Lin, Bill Yuchen Lin, Ziqi Wang, Junyi Du, Leonardo Neves, and Xiang Ren. 2020. Nero: A neural rule grounding framework for label-efficient relation extraction. In TheWebConf.",
+            "Dawei Zhu, Michael Hedderich, Fangzhou Zhai, David Adelani, and Dietrich Klakow. 2022. Is bert robust to label noise? a study on learning with noisy labels in text classification. In Insights from Negative Results in NLP.",
+            "Yuchen Zhuang, Yue Yu, Lingkai Kong, Xiang Chen, and Chao Zhang. 2023. Dygen: Learning from noisy labels via dynamics-enhanced generative modeling. arXiv preprint arXiv:2305.19395."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            673
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Details of Three Subsets",
+        "text_level": 1,
+        "bbox": [
+            510,
+            688,
+            754,
+            703
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In this section, we further provide details about the three subsets of samples in our framework.",
+        "bbox": [
+            510,
+            714,
+            880,
+            745
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "1. Easy Clean (EC) Set. In this subset, the labels predicted by LLMs are the same as the assigned labels. Therefore, based on Assumption 1, we can infer that the labels of these samples are almost clean, which allows us to directly use them for model training. Note that a small portion of samples in this category can still be noisy when the LLM predictions happen to be the same as noisy labels. However, we empirically verify in Sec. 5.5 that this proportion is small.",
+        "bbox": [
+            510,
+            758,
+            882,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "12538",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "2. Hard Clean (HC) Set. In this (ideal) category, the labels predicted by LLMs deviate from the assigned labels, while the true labels are in accordance with the assigned labels. In this case, we are acknowledged that these samples are also clean. However, due to the semantic difficulties of these samples, the LLMs cannot easily predict the correct labels of them. Nevertheless, since the assigned labels are correct as true labels, we can use assigned labels for model training on these samples. It is noteworthy that in practice, it is infeasible to achieve a perfect separation of this subset. Therefore, in our framework, we propose to leverage LLM-generated and PLM-generated confidences to improve the separation performance.",
+        "bbox": [
+            112,
+            84,
+            490,
+            341
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "3. True Noisy (TN) Set. In this (ideal) category, the labels predicted by LLMs and the true labels are both different from the assigned labels. In other words, these samples are all noisy as their assigned labels are different from true labels. Since they are noisy, the labels predicted by LLMs are also likely to be different from the assigned labels. As these samples are noisy, we cannot directly use their labels for training. However, since we narrow down the potential range of noisy samples to this category, we can resort to specific techniques to learn from these noisy samples. In our framework, we utilize LLM-generated confidence as additional supervision information to effectively learn from these noisy samples.",
+        "bbox": [
+            112,
+            348,
+            490,
+            605
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Datasets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            619,
+            226,
+            634
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the details of the datasets used in our experiments. In particular,  $20\\mathrm{Ng}$  (Lang, 1995) and AGNews (Li and Roth, 2002; Zhang et al., 2015) are news topic classification datasets that are prevalently used for text classification tasks. We manually inject different types of noise (SN, AN, and IDN) into these two datasets for evaluation in our experiments. For real-world datasets, we utilize SemEval (Zhou et al., 2020), TREC (Awasthi et al., 2020), and Hausa (Hedderich et al., 2020). Specifically, SemEval is a relation extraction dataset, and we follow the process introduced in (Zhou et al., 2020) to obtain noisy labels. TREC is a question classification dataset in the weak supervision benchmark WRENCH (Zhang et al., 2021b). Moreover, Hausa is a text classification dataset in the language of",
+        "bbox": [
+            112,
+            645,
+            490,
+            919
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/22a5690cce3ecf236146ff727f20e1dd16b833e40c02fde2b178153260bb5557.jpg",
+        "table_caption": [
+            "Table 6: The detailed statistics of the five datasets used in our experiments."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td># Training</td><td># Validation</td><td># Test</td><td># Class</td></tr><tr><td>20Ng</td><td>9,051</td><td>2,263</td><td>7,532</td><td>20</td></tr><tr><td>AGNews</td><td>40,000</td><td>7,600</td><td>7,600</td><td>4</td></tr><tr><td>SemEval</td><td>1,749</td><td>200</td><td>692</td><td>9</td></tr><tr><td>TREC</td><td>4,965</td><td>500</td><td>500</td><td>6</td></tr><tr><td>Hausa</td><td>2,045</td><td>290</td><td>582</td><td>5</td></tr></table>",
+        "bbox": [
+            517,
+            116,
+            875,
+            219
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Hausa, the second most spoken indigenous language in Africa, with 40 million native speakers. For this dataset, gazetteers are used for automatic labeling, which results in feature-dependent label noise. Here we provide the detailed statistics of these datasets in Table 6.",
+        "bbox": [
+            507,
+            244,
+            884,
+            340
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Baselines",
+        "text_level": 1,
+        "bbox": [
+            509,
+            357,
+            628,
+            372
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In this section, we provide the details of the baselines used in our experiments.",
+        "bbox": [
+            507,
+            384,
+            882,
+            417
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Base (Devlin et al., 2018) is the BERT base model fine-tuned based on the standard cross-entropy loss. Note that for dataset Hausa, we utilize the multilingual BERT model.",
+            "- Mixup (Zhang et al., 2018) is a semi-supervised approach that performs a linear interpolation between clean and noisy samples.",
+            "- GCE (Zhang and Sabuncu, 2018) denotes Generalized Cross-Entropy loss, which can be regarded as a general loss that combines mean absolute error (MAE) and cross-entropy (CE).",
+            "- Co-teaching (Han et al., 2018) trains two different models and selects small-loss samples to feed them into each other for optimization.",
+            "- Co-teaching+ (Yu et al., 2019) is an updated version of Co-teaching that explicitly ensures the difference between the two models.",
+            "- JoCoR (Wei et al., 2020) trains two models and selects samples based on the sum of loss from the two models.",
+            "- CR (Zhou and Chen, 2021) also trains multiple models with a regularization strategy based on a soft target.",
+            "- NPC (Bae et al., 2022) utilizes a generative model to estimate the transition matrix from noisy predictions to the ground-truth labels of samples and uses it to correct noisy labels."
+        ],
+        "bbox": [
+            510,
+            432,
+            882,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "12539",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            356,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the implementation details for our experiments. We run the experiments on a single 48GB Nvidia GeForce RTX A6000 GPU. The experiments are run 10 times to achieve the average accuracy and the standard deviation.",
+        "bbox": [
+            112,
+            109,
+            489,
+            187
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "For the augmentations used in our framework, following existing works (Xie et al., 2019; Wei and Zou, 2019; Li et al., 2022), we adopt four typical text augmentations: Back Translation, Random Insertion, Random Deletion, and Random Swap. We set the threshold  $\\widehat{\\tau}$  for LLM-generated confidences as 0.8. We set the hyper-parameters  $\\lambda$  and  $\\widetilde{\\tau}$  for the adaptive threshold  $\\widetilde{\\tau}(t)$  for PLM-generated confidences as 0.7 and 0.8, respectively. We set  $\\alpha$  as 1.5 and  $\\beta$  as 0.9. For the loss weight, we set  $\\lambda_{\\mathcal{H}} = \\lambda_{\\mathcal{N}} = 1$ . We set the model fine-tuning learning rate as  $10^{-4}$  with a batch size of 32. The maximum length of texts is set as 256. The number of training epochs is set as 100 with early stopping based on the performance of the validation set. We optimize the model using the Adam (Kingma and Ba, 2015) optimization strategy. For the LLM, we utilize ChatGPT based on GPT4.",
+        "bbox": [
+            115,
+            190,
+            489,
+            476
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "In addition, we provide more detailed package requirements as listed below.",
+        "bbox": [
+            112,
+            479,
+            487,
+            511
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Python == 3.9.12",
+            "- torch == 2.0.1",
+            "- transformers == 4.29.2",
+            "- huggingface-hub == 0.15.1",
+            "- keras == 2.12.0",
+            "- numpy == 1.23.5",
+            "- scipy == 1.5.3",
+            "- CUDA == 11.6",
+            "networkx  $= = 2.5.1$",
+            "scikit-learn  $= = 0.24.1$"
+        ],
+        "bbox": [
+            115,
+            521,
+            332,
+            728
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "12540",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Noise-Robust Fine-Tuning of Pretrained Language Models via External Guidance/569fc499-7adf-4636-bebe-f243ca93393e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d7f3cf27438e563602aa8292e891260eaeb6b28207500980907a0cf9e764290
+size 609376

2023/Noise-Robust Fine-Tuning of Pretrained Language Models via External Guidance/full.md

@@ -0,0 +1,406 @@
+# Noise-Robust Fine-Tuning of Pretrained Language Models via External Guidance
+
+Song Wang
+
+University of Virginia
+
+sw3wv@virginia.edu
+
+Ruocheng Guo
+
+ByteDance Research
+
+ruocheng.guo@bytedance.com
+
+Zhen Tan
+
+Arizona State University
+
+ztan36@asu.edu
+
+Jundong Li
+
+University of Virginia
+
+jundong@virginia.edu
+
+# Abstract
+
+Adopting a two-stage paradigm of pretraining followed by fine-tuning, Pretrained Language Models (PLMs) have achieved substantial advancements in the field of natural language processing. However, in real-world scenarios, data labels are often noisy due to the complex annotation process, making it essential to develop strategies for fine-tuning PLMs with such noisy labels. To this end, we introduce an innovative approach for fine-tuning PLMs using noisy labels, which incorporates the guidance of Large Language Models (LLMs) like ChatGPT. This guidance assists in accurately distinguishing between clean and noisy samples and provides supplementary information beyond the noisy labels, thereby boosting the learning process during fine-tuning PLMs. Extensive experiments on synthetic and real-world noisy datasets further demonstrate the superiority of our framework over the state-of-the-art baselines.
+
+# 1 Introduction
+
+In recent years, the development of language models has significantly expanded the applications within the field of natural language processing (NLP). Fine-tuning Pretrained Language Models (PLMs) like BERT (Devlin et al., 2018) for specific downstream tasks has become an essential step in real-world implementations (Alt et al., 2020; Wang et al., 2021). In general, achieving significant performance gains in fine-tuning PLMs necessitates the availability of task-specific data for fine-tuning (Zhou and Chen, 2021; Wang et al., 2022). However, obtaining high-quality labeled datasets for this purpose poses significant challenges due to the expensive, complex, and labor-intensive nature of the annotation process (Yu et al., 2019; Bae et al., 2022). For example, large-scale datasets, often derived from web-crawling (Li et al., 2017; Song et al., 2019) or crowd-sourcing (Yan et al., 2014; Williams et al., 2017; Sakaguchi et al., 2021), frequently suffer from the presence of noisy labels.
+
+![](images/37f59962e010dae0084546f1130e1b2c56617463f7426f7f6bc29b1982450a4a.jpg)  
+Figure 1: The average confidence of clean and noisy samples during fine-tuning BERT on datasets  $20\mathrm{Ng}$  and AGNews with  $20\%$  noisy labels.
+
+Prior research (Arpit et al., 2017; Cheng et al., 2021; Zhang et al., 2021c; Wang et al., 2023a) has shown that PLMs are prone to overfitting and generally deliver subpar performance when fine-tuned on datasets containing label noise. Since fine-tuning involves incorporating supervision information to enhance performance on downstream tasks, the presence of noisy labels will mislead the training process and significantly impede the efficacy of PLMs (Zhu et al., 2022). Therefore, there is an immediate need to design effective algorithms for fine-tuning PLMs in the presence of noisy labels.
+
+In the context of learning from noisy labels, an intuitive approach is to separate clean samples from the noisy ones in the training set for model training (Han et al., 2018). For the remaining noisy samples, a prevalent strategy is to pseudo-label them based on model predictions to reduce the adverse impact of noise (Berthelot et al., 2019; Sohn et al., 2020). However, it remains challenging when applying this paradigm to PLMs. This is because PLMs, equipped with prior knowledge encoded in a large size of parameters, tend to easily memorize noisy samples early on during the fine-tuning process. As illustrated in Figure 1, PLMs exhibit similar prediction confidences for both clean and noisy samples, which will result in two challenges
+
+when utilizing existing methods: (1) Existing approaches often rely on confidence generated by models trained on potentially noisy samples (Li et al., 2020; Karim et al., 2022). However, as shown in Figure 1, it is challenging to accurately distinguish clean and noisy samples solely based on confidence. (2) Existing methods are susceptible to erroneous information during pseudo-labeling, as they directly utilize model predictions as pseudolabels, which can be detrimental when the predictions are inaccurate. PLMs, being capable of easily remembering noisy samples, further exacerbate the risk of capturing and amplifying erroneous information during pseudo-labeling. For example, in the  $20\mathrm{Ng}$  dataset (Lang, 1995), when an "autos" sample is assigned a wrong label "hardware", PLMs can easily memorize the erroneous information and thus cannot infer the correct labels in predictions, resulting in incorrect pseudo-labeling.
+
+To overcome these challenges, we propose a novel framework, named LAFT, that harnesses the power of Large Language Models (LLMs) for Fine-Tuning PLMs. We leverage LLMs trained on extensive corpora and fine-tuned with human instructions, such as GPT4 (OpenAI, 2023), to obtain external guidance in the form of confidence values. Our approach addresses the first challenge of separating reliable samples by utilizing LLM-generated confidence for each class of all training samples. By comparing these confidence with the assigned labels (i.e., the given noisy labels) of training samples, we categorize them into three disjoint subsets: the Easy Clean (EC) Set, the Hard Clean (HC) Set, and the True Noisy (TN) Set. Regarding the second challenge, we propose a novel method to incorporate LLM-generated confidence scores as robust supervision information for all samples to ensure that PLMs learn useful information from them. As LLM-generated confidence are not affected by label noise, they can provide potentially relevant labels that are useful even if they are not entirely accurate. In summary, our contributions are as follows:
+
+- We are the first to explore the potential of leveraging supervision information (i.e., confidence scores) generated by LLMs to tackle the noisy label problem in fine-tuning PLMs.  
+- We propose a novel framework LAFT that can effectively separate clean and noisy samples and learn from noisy labels, based on the LLMs-generated confidence.
+
+- We conduct extensive experiments on synthetic and real-world noisy datasets and demonstrate the superiority of our framework in fine-tuning PLMs with noisy labels.
+
+# 2 Related Work
+
+# 2.1 Learning from Noisy Labels
+
+Various strategies for learning from noisy labels have been proposed, falling primarily into three categories. The first category is Sample Selection methods, which typically employ loss or confidence to identify reliable samples for model optimization. These methods generally necessitate a predefined threshold (Li et al., 2021; Karim et al., 2022) or prior knowledge concerning the noise label rate (Han et al., 2018; Yu et al., 2019) to choose the instances. The second category, dubbed Label Transition methods, aims to learn (Tanaka et al., 2018) or generate pseudo-labels (Zhang et al., 2021c; Sohn et al., 2020) to replace the original noisy labels. Lastly, Regularization methods design robust loss functions (Liu et al., 2020; Englesson and Azizpour, 2021) or regularization techniques (Zhang and Sabuncu, 2018) that can effectively utilize all samples to enhance the model robustness against label noise. Nevertheless, these methods generally do not consider the scenario of fine-tuning a pretrained model with noisy labels.
+
+# 2.2 Confidence-Guided Sample Separation
+
+To separate noisy samples, existing works leverage the loss or confidence that provide insights into the model's prediction behavior as training proceeds (Yu et al., 2019; Karim et al., 2022). In the context of learning from noisy labels, the key concept is to leverage these dynamics as criteria for identifying and separating noisy samples. Several works propose to identify samples with lower training loss as the clean subset (Li et al., 2020; Han et al., 2018; Jiang et al., 2018; Zhao et al., 2022; Wang et al., 2023b), however, they are generally simplistic and inflexible, resulting in the selection of only easy samples. To address this limitation, alternative approaches have been proposed to effectively utilize the loss or confidence during training, as demonstrated in (Zhang et al., 2021a) and (Nishi et al., 2021). In contrast to existing methods that only rely on confidence generated by PLMs, we leverage LLMs as external guidance that provides more precise separations for fine-tuning PLMs.
+
+# 3 Problem Definition
+
+In this work, we study the problem of fine-tuning PLMs for text classification with noisy labels. Formally, consider a noisy training dataset containing  $n$  samples:  $\mathcal{D}_{tr} = \{(x_i,\widetilde{y}_i),i = 1,2,\ldots ,n\}$  where  $x_{i}$  is an input sample and  $\widetilde{y}_i$  denotes the assigned label of  $x_{i}$ . Note that  $\widetilde{y}_i$  is potentially corrupted, and we denote the true label of  $x_{i}$  as  $\overline{y}_i$  which is inaccessible during fine-tuning. Specifically, we aim to fine-tune PLMs with samples in  $\mathcal{D}_{tr}$  to achieve satisfactory prediction performance on test samples, while utilizing LLMs as external guidance. Notably, the LLMs remain fixed in our framework, which means we can also use blackbox LLMs. In practice, we implement PLMs for text classification by attaching an additional classifier, which will take the output of the PLM as input and generate class probabilities for classification.
+
+# 4 Methodology
+
+The overall framework of LAFT is illustrated in Fig. 2. In particular, we propose to divide all training samples into three subsets based on the accordance among LLM-generated confidence, PLM-generated confidence, and the assigned labels of samples. In particular, we query an LLM to provide confidence for each training sample, spanning all classes. Combining confidence obtained from both LLMs and PLMs, we perform two steps of separation to segregate the entire training set into three subsets: Easy Clean (EC) Set, Hard Clean (HC) Set, and True Noisy (TN) Set. Each of these subsets, displaying unique behaviors, is then subjected to specific noise-robust fine-tuning strategies.
+
+# 4.1 Confidences
+
+Existing studies establish a correlation between confidences and the degree to which deep models memorize specific samples during training (Arpit et al., 2017; Karim et al., 2022). It has been observed that as the model's memorization of a particular sample strengthens, the model tends to assign higher confidence for this sample (Li et al., 2023). Therefore, these methods generally employ confidences to distinguish between clean and noisy samples based on the assumption that the model cannot easily memorize noisy samples (Pleiss et al., 2020; Swayamdipta et al., 2020). However, applying these strategies to fine-tuning PLMs is suboptimal, as PLMs also present high confidence for noisy samples even in the early stage of fine-tuning, as
+
+![](images/bf296710eb0218084c85f1b93ca0b9a76fc3520c9515be4ea52901badfcdadf0.jpg)  
+Figure 2: The detailed process of our framework LAFT. We perform two steps of separation to divide all training samples into three subsets with different degrees of label noise: Easy Clean Set  $\mathcal{E}$ , Hard Clean Set  $\mathcal{H}$ , and True Noisy Set  $\mathcal{N}$ . We further propose three different losses to effectively learn from them:  $\mathcal{L}_{\mathcal{E}}$ ,  $\mathcal{L}_{\mathcal{H}}$ , and  $\mathcal{L}_{\mathcal{N}}$ .
+
+shown in Fig. 1. To deal with this issue, we propose the utilization of external guidance, in the form of confidences generated by LLMs. Before delving into the specifics of our framework, we provide a formal definition of confidences. Denote the output of the final layer (i.e., the classifier) in PLMs for sample  $x_{i}$  as  $\mathbf{z}(x_i)\in \mathbb{R}^N$ , where  $N$  is the number of classes. The confidence of  $x_{i}$  for the  $j$ -th class  $c_{j}$  can be represented as follows:
+
+$$
+\widetilde {p} \left(c _ {j}; x _ {i}\right) = \frac {\exp \left(z \left(c _ {j} ; x _ {i}\right)\right)}{\sum_ {k = 1} ^ {N} \exp \left(z \left(c _ {k} ; x _ {i}\right)\right)}, \tag {1}
+$$
+
+where  $z(c_{j};x_{i})\in \mathbb{R}$  is the  $j$  -th value in  $\mathbf{z}(x_i)$ . Notably, the confidences are obtained from  $\mathbf{z}(x_i)$  after a softmax function and thus sum up to one.
+
+Although LLM-generated confidences can provide external guidance, they are not completely accurate. Thus, We conduct a two-step sample separation process based on both LLM-generated and PLM-generated confidences, with the second step providing a more granular distinction.
+
+# 4.2 Coarse-Grained Separation
+
+For the first step of separation, we aim to select samples that are easy to be identified as clean data with guidance from LLMs. Thus, we perform Coarse-grained Separation, utilizing confidences generated by LLMs with the raw text data included in the prompt. Here we provide an example of the prompt for querying the LLM to obtain the confidence value for each class:
+
+Classify the following content:  $\{input\text{text}\}$ . Select the label from  $\{Class1\}$ ,  $\{Class2\}$ , ...,  $\{ClassN\}$  and output a confidence value for each of them.
+
+We denote the LLM-generated confidence for sample  $x_{i}$  regarding class  $c_{j}$  as  $\widehat{p}(c_{j};x_{i})$ , where  $c_{j} \in \mathcal{C} = \{c_{1},c_{2},\ldots ,c_{N}\}$ , and  $\mathcal{C}$  is the class set. Then the label obtained by LLM can be represented as
+
+$$
+\widehat {y} _ {i} = \underset {j = 1, 2, \dots , N} {\operatorname {a r g m a x}} \widehat {p} \left(c _ {j}; x _ {i}\right). \tag {2}
+$$
+
+To perform coarse-grained separation, we first provide an assumption as the justification:
+
+Assumption 1. Samples whose assigned labels are the same as LLM-generated labels (i.e.,  $\widehat{y}_i = \widetilde{y}_i$ ) can be considered almost clean.
+
+Note that Assumption 1 is empirically verified in Table 5 in Sec. 5.5. The underlying intuition is that the probability of an LLM-generated label being identical to the assigned label and inconsistent with the ground-truth label is significantly low and thus negligible. In concrete, this assumption allows us to segregate the training samples into two distinct subsets based on the concordance between the LLM-generated label  $\widehat{y}_i$  and the assigned (potentially noisy) label  $\widetilde{y}_i$  of  $x_i$ . More specifically, we define the resulting two subsets, the Easy Clean (EC) Set  $\mathcal{E}$  and the Disagreed Set  $\mathcal{D}$ , as follows:
+
+$$
+\mathcal {E} = \left\{x _ {i}, \widetilde {y} _ {i} \mid \widehat {y} _ {i} = \widetilde {y} _ {i} \right\}, \mathcal {D} = \left\{x _ {i}, \widetilde {y} _ {i} \mid \widehat {y} _ {i} \neq \widetilde {y} _ {i} \right\}. \tag {3}
+$$
+
+It is naturally satisfied that  $\mathcal{E} \cup \mathcal{D} = \mathcal{D}_{tr}$  and  $\mathcal{E} \cap \mathcal{D} = \emptyset$ , where  $\mathcal{D}_{tr}$  is the training set. Since the samples in  $\mathcal{E}$  are already clean, we can directly fine-tune PLMs using LLM-generated labels  $\widehat{y}_i$  based on the cross-entropy loss as follows:
+
+$$
+\mathcal {L} _ {\mathcal {E}} = - \frac {1}{| \mathcal {E} |} \sum_ {i = 1} ^ {| \mathcal {E} |} \sum_ {j = 1} ^ {N} \widehat {y} _ {i, j} \log \widetilde {p} \left(c _ {j}; x _ {i}\right), \tag {4}
+$$
+
+where  $\widetilde{p}(c_j; x_i)$  is the PLM-generated confidence for  $x_i$  regarding the  $j$ -th class  $c_j$ . Here  $\widehat{y}_{i,j} = 1$  if  $\widehat{y}_i = c_j$ , and  $\widehat{y}_{i,j} = 0$ , otherwise.
+
+# 4.3 Fine-Grained Separation
+
+It is noteworthy that, however, the samples in  $\mathcal{D}$  are not completely noisy. This is because the LLM-generated labels are not perfectly correct, as shown in Table 5. Thus, the samples that are clean but incorrectly classified by LLMs will still
+
+be categorized into  $\mathcal{D}$ . Therefore, although we can learn from samples in  $\mathcal{E}$  directly with their LLM-generated label  $\widetilde{y}_i$ , it is still challenging to learn from samples in  $\mathcal{D}$ , which are only partially clean. Therefore, we further propose to separate the samples in the disagreed set  $\mathcal{D}$  into two subsets: the Hard Clean (HC) Set  $\mathcal{H}$  and the True Noisy (TN) Set  $\mathcal{N}$ , referred to as fine-grained separation.
+
+The intuition is that within the disagreed set  $\mathcal{D}$ , the LLM-generated labels can be incorrect for specific hard samples with correct assigned labels, referred to as Hard Clean (HC) samples. Specifically, the ideal separation for  $\mathcal{D}$  is as follows:
+
+$$
+\begin{array}{l} \mathcal {H} ^ {*} = \mathcal {D} \cap \left\{x _ {i}, \widetilde {y} _ {i} | \widetilde {y} _ {i} = \bar {y} _ {i} \right\}, \\ \text {实} ^ {*} = \mathcal {D} \cap \{\dots \widetilde {\sim} | \widetilde {\sim} (\dots) \} \end{array} \tag {5}
+$$
+
+$$
+\mathcal {N} ^ {*} = \mathcal {D} \cap \left\{x _ {i}, \widetilde {y} _ {i} | \widetilde {y} _ {i} \neq \bar {y} _ {i} \right\},
+$$
+
+where  $\overline{y}_i$  is the true label of  $x_i$ . Note that although this ideal separation can be completely precise for separating noisy samples in  $\mathcal{D}$ , it is infeasible in practice, as the true labels are unknown. Therefore, to precisely separate the true noisy samples in  $\mathcal{D}$ , we propose two thresholds for LLM-generated and PLM-generated confidences, respectively.
+
+In order to achieve more robust LLM-generated confidences distributed over the label space  $\mathcal{C}$ , we adopt  $M$  different augmentations for each input sample to encourage input diversity while keeping the semantics immutable. Denote the augmented samples of  $x_{i}$  as  $v_{m}(x_{i})$ , where  $m = 1,2,\ldots ,M$ , and  $M$  is the number of augmentations. We can obtain the  $M$  LLM-generated confidences as  $\widehat{p} (c_j;v_m(x_i))$  for  $x_{i}$  regarding the  $j$ -th class  $c_{j}$  where  $m = 1,2,\dots ,M$  and  $j = 1,2,\dots ,N$ . We aggregate the LLM-generated confidences via the  $M$  augmentations as follows:
+
+$$
+\widehat {p} _ {a} \left(c _ {j}; x _ {i}\right) = \frac {1}{M} \sum_ {m = 1} ^ {M} \widehat {p} \left(c _ {j}; v _ {m} \left(x _ {i}\right)\right), \tag {6}
+$$
+
+where  $v_{m}(x_{i})$  is the input example after applying the  $m$ -th augmentation. As LLM-generated confidences remain fixed during fine-tuning PLMs, we adopt a fixed threshold  $\widehat{\tau}$  for fine-grained separation based on  $\widehat{p}_a(c_j; x_i)$ .
+
+On the other hand, however, PLMs-generated confidences for each sample will change as the finetuning proceeds, which results in subpar separation performance if we adopt a fixed threshold to select high-confidence samples as clean ones. Intuitively, the confidences should be lower for HC samples at the beginning of fine-tuning, as the model cannot
+
+easily fit these hard samples within several epochs. Nonetheless, the noisy samples are relatively easier to achieve higher confidence, and thus their confidences will easily achieve higher at the beginning. Therefore, we propose an adaptive threshold  $\widetilde{\tau}(t)$  that will increase as fine-tuning proceeds while not reaching an excessively high value:
+
+$$
+\widetilde {\tau} (t) = \widetilde {\tau} - \exp (- \lambda t), \qquad (7)
+$$
+
+where  $\lambda$  and  $\widetilde{\tau}$  are hyper-parameters that control the value of threshold  $\widetilde{\tau}(t)$  as fine-tuning proceeds.  $t$  denotes the current number of fine-tuning epochs.
+
+Combining the two thresholds, we can perform fine-grained separation by selecting the HC set  $\mathcal{H}$  as follows:
+
+$$
+\begin{array}{l} \mathcal {H} = \mathcal {D} \cap \left\{x _ {i}, \widetilde {y} _ {i} \mid \max  _ {c \in \mathcal {C}} \widehat {p} _ {a} (c; x _ {i}) <   \widehat {\tau} \right\} \tag {8} \\ \cap \left\{x _ {i}, \widetilde {y} _ {i} \mid \max  _ {c \in \mathcal {C}} \widetilde {p} (c; x _ {i}) <   \widetilde {\tau} (t) \right\}, \\ \end{array}
+$$
+
+where  $\widehat{\tau}$  and  $\widetilde{\tau}(t)$  are the thresholds for LLM-generated confidence  $(\widehat{p}_a(c_j; x_i))$  and PLMs-generated confidence  $(\widetilde{p}(c_j; x_i))$ , respectively. In this manner, the process of separating HC and TN samples, i.e., fine-grained separation, can benefit from both the LLMs and PLMs. Then the remaining samples are categorized into the True Noisy (TN) set  $\mathcal{N}$ :
+
+$$
+\mathcal {N} = \mathcal {D} \backslash \mathcal {H}. \tag {9}
+$$
+
+# 4.4 Learning from the Hard Clean (HC) Set
+
+Now we have divided the disagreed set  $\mathcal{D}$  into  $\mathcal{H}$  and  $\mathcal{N}$ . Recall that ideally, samples in  $\mathcal{H}$  are hard yet correct. Thus, for these samples, we can directly utilize their assigned labels  $\widetilde{y}$  as training labels. Nevertheless, since the fine-grained separation of  $\mathcal{H}$  and  $\mathcal{N}$  cannot be perfect, the samples in  $\mathcal{H}$  may still be noisy. As both LLMs and PLMs fail to provide a confident prediction for samples in  $\mathcal{H}$ , we propose to prioritize the assigned label while also incorporating additional information from LLMs and PLMs. Specifically, we employ a weighted loss based on cross-entropy. The weight is enlarged if the summed confidence of the LLM and the PLM is high. We define the loss as follows:
+
+$$
+\mathcal {L} _ {\mathcal {H}} = - \frac {1}{| \mathcal {H} |} \sum_ {i = 1} ^ {| \mathcal {H} |} \sum_ {j = 1} ^ {N} \left(\tilde {y} _ {i, j} + \phi_ {i, j}\right) \log \widetilde {p} \left(c _ {j}; x _ {i}\right), \tag {10}
+$$
+
+where  $\widetilde{p}(c_j; x_i)$  is the PLM-generated confidence for  $x_i$  regarding the  $j$ -th class  $c_j$ . Here  $\widetilde{y}_{i,j} = 1$  if  $\widetilde{y}_i = c_j$ , and  $\widetilde{y}_{i,j} = 0$ , otherwise.  $\phi_{i,j}$  acts as a
+
+weight adjustment that increases when the sum of LLM confidence  $\widehat{p}_a(c_j;x_i)$  and PLM confidence  $\widetilde{p} (c_j;x_i)$  is larger. Intuitively, if the LLM and PLM are with high confidence regarding a specific class, then the information in their confidence can be useful as they are more likely to be correct. Therefore, we define  $\phi_{i,j}$  as follows:
+
+$$
+\phi_ {i, j} = \max  \left(\widehat {p} _ {a} \left(c _ {j}; x _ {i}\right) + \widetilde {p} \left(c _ {j}; x _ {i}\right) - \alpha \widetilde {\tau} (t), 0\right), \tag {11}
+$$
+
+where  $\alpha > 1$  is a hyper-parameter that controls the threshold  $\alpha \widetilde{\tau}(t)$  for  $\phi_{i,j}$ . As such information can still be inaccurate, we subtract it by the threshold  $\alpha \widetilde{\tau}(t)$  to control its magnitude, such that  $\phi_{i,j}$  also acts as an adaptive loss weight for these samples.
+
+# 4.5 Learning from the True Noisy (TN) Set
+
+After the fine-grained separation, most samples in the True Noisy (TN) Set should be identified as noisy ones. Nevertheless, it is still challenging for PLMs to use their output as pseudo-labels for fine-tuning, as the prediction errors will accumulate and affect subsequent pseudo-labeling results. Fortunately, the confidence generated by LLMs can provide additional guidance to identify the potential labels for samples in the TN set. We first provide Remark 1 to justify the effectiveness of using LLM-generated labels for optimization.
+
+Remark 1. LLM-generated labels on the True Noisy Set preserve the same accuracy as that on the whole dataset, as LLMs are not affected by noise.
+
+Remark 1, as empirically verified in Sec. 5.5, demonstrates that even when we categorize most noisy samples into True Noisy Set, the LLM-generated labels can still provide decent guidance without sacrificing accuracy. Given the confidence provided by LLMs, we employ a loss that can enable the PLMs to benefit from them. Specifically,
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathcal {N}} = - \frac {1}{| \mathcal {N} |} \sum_ {i = 1} ^ {| \mathcal {N} |} \sum_ {j = 1} ^ {N} \widehat {p} _ {a} (c _ {j}; x _ {i}) \log \widetilde {p} (c _ {j}; x _ {i}) \\ - \frac {1}{| \mathcal {N} |} \sum_ {i = 1} ^ {| \mathcal {N} |} \delta \left(x _ {i}\right) \cdot \max  _ {c \in \mathcal {C}} \widetilde {p} (c; x _ {i}) \log \max  _ {c \in \mathcal {C}} \widetilde {p} (c; x _ {i}). \tag {12} \\ \end{array}
+$$
+
+Here in the first term, we leverage the confidences generated by LLMs to learn from potentially correct labels. This is because although LLMs cannot completely predict the correct labels, the confidences still preserve the potentially useful information in other incorrect but relevant labels. Such
+
+benefits cannot be provided by pseudo-labeling, which tends to output a definitive label. For the second term, we utilize the model predictions to exploit useful information from PLMs. The intuition is that, with the relevant label information provided by LLMs, the PLMs can learn accurate label information from the noisy samples in TN set. Consequently, if the model output tends to be confident on specific samples, we can utilize the prediction to further enhance the learning from it. Thus, we further set a threshold for this term, i.e.,  $\delta(x_i)$ , defined as follows:
+
+$$
+\delta \left(x _ {i}\right) = \left\{ \begin{array}{l l} 1, & \text {i f} \quad \max  _ {c \in \mathcal {C}} \widetilde {p} (c; x _ {i}) > \beta \widetilde {\tau} (t), \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {13}
+$$
+
+where  $\widetilde{\tau}(t)$  is computed by Eq. (7). To reduce the effect of confirmation bias, we multiply  $\widetilde{\tau}(t)$  by  $\beta$ , a hyper-parameter that controls the final adaptive threshold, i.e.,  $\beta \widetilde{\tau}(t)$ .
+
+# 4.6 Fine-tuning Objective
+
+After we separate all training samples into  $\mathcal{E}$ ,  $\mathcal{H}$ , and  $\mathcal{N}$ , we can combine the three individual losses for them for PLM fine-tuning. Our final fine-tuning objective can be represented as follows:
+
+$$
+\mathcal {L} = \mathcal {L} _ {\mathcal {E}} + \lambda_ {\mathcal {H}} \mathcal {L} _ {\mathcal {H}} + \lambda_ {\mathcal {N}} \mathcal {L} _ {\mathcal {N}}, \tag {14}
+$$
+
+where  $\lambda_{\mathcal{H}}$  and  $\lambda_{\mathcal{N}}$  are hyper-parameters that control the importance of  $\mathcal{L}_{\mathcal{H}}$  and  $\mathcal{L}_{\mathcal{N}}$ , respectively.
+
+# 5 Experiments
+
+# 5.1 Experimental Settings
+
+Datasets. To evaluate the performance of our framework, we first conduct experiments on two synthetic-noise datasets: 20Ng (Lang, 1995) and AGNews (Li and Roth, 2002; Zhang et al., 2015). Following existing works on learning from noisy labels (Patrini et al., 2017; Yao et al., 2020; Zhuang et al., 2023), we adopt three types of synthetic label noise: (1) Symmetric Noise (SN) uniformly changes labels to other classes (Han et al., 2018; Xia et al., 2021). (2) Asymmetric Noise (ASN) changes labels to other similar classes (Tanaka et al., 2018; Bae et al., 2022). (3) Instance-Dependent Noise (IDN) changes labels based on a probability in proportion to the specific sample features (Cheng et al., 2021; Yao et al., 2021). Moreover, we further conduct experiments on three real-world datasets with noisy labels: SemEval (Zhou et al., 2020), TREC (Awasthi et al., 2020), and
+
+Hausa (Hedderich et al., 2020). More details about these datasets are provided in Appendix B.
+
+Baselines. We compare our framework to state-of-the-art baselines for learning from noisy labels. In particular, we compare to (1) Base (Devlin et al., 2018) that performs fine-tuning with standard cross-entropy loss; (2) Regularization Methods: Mixup (Zhang et al., 2018) and GCE (Zhang and Sabuncu, 2018); (3) Sample-selection Methods: Co-teaching (Han et al., 2018), Co-teaching+ (Yu et al., 2019), JoCoR (Wei et al., 2020), CR (Zhou and Chen, 2021), and NPC (Bae et al., 2022). Additional details are provided in Appendix C.
+
+Implementation Details. We use BERT (Devlin et al., 2018) as the text encoder for all datasets except Hausa, for which we use mBERT. The classifier is implemented as a fully-connected layer and randomly initialized at the beginning, while both the encoder and the classifier will be updated via gradient descent during fine-tuning. All experiments are evaluated on a clean test set, and the average accuracy along with the standard deviation over ten runs is reported for each dataset. We provide more details about the implementation in Appendix D, and our code is provided at https://github.com/SongW-SW/LAFT.
+
+# 5.2 Comparison on Synthetic Datasets
+
+In this subsection, we compare our framework with other baselines on synthetic datasets  $20\mathrm{Ng}$  and AGNews, considering different noise types and ratios. Specifically, for Symmetric Noise (SN), we conduct experiments on three different noise rates:  $20\%$ ,  $40\%$ , and  $60\%$ . For the other two types of noise, i.e., Asymmetric Noise (AN) and Instance-Dependent Noise (IDN), we adopt two noise rates:  $20\%$  and  $40\%$ . We present the results in Table 1, 2, and 3. The key observations drawn from the outcomes are as follows: (1) Regardless of the noise type, our LAFT framework persistently surpasses other state-of-the-art baselines, thus showcasing its effectiveness in fine-tuning PLMs with noisy labels. (2) LAFT's performance improvement over other baselines is slightly more pronounced on the  $20\mathrm{Ng}$  dataset compared to AGNews. This is attributable to  $20\mathrm{Ng}$  containing a greater number of classes ( $N = 20$ ) as opposed to AGNews ( $N = 4$ ). Consequently, our strategy of utilizing LLM-generated confidence can capitalize on the similar labels within  $20\mathrm{Ng}$  for PLM fine-tuning, even if LLM predictions are not entirely accurate.
+
+Table 1: The overall performance of various models on synthetic noisy datasets  ${20}\mathrm{{Ng}}$  and AGNews with Symmetric Noise (SN),where accuracy and standard deviation are reported in %,and the best results are in bold.  
+
+<table><tr><td>Dataset</td><td colspan="3">20Ng</td><td colspan="3">AGNews</td></tr><tr><td>SN Ratio</td><td>20%</td><td>40%</td><td>60%</td><td>20%</td><td>40%</td><td>60%</td></tr><tr><td>Base</td><td>79.15 ± 0.29</td><td>71.74 ± 0.07</td><td>61.87 ± 1.58</td><td>81.77 ± 0.33</td><td>78.33 ± 0.47</td><td>73.07 ± 1.49</td></tr><tr><td>Mixup</td><td>79.10 ± 0.21</td><td>70.58 ± 0.22</td><td>62.48 ± 2.04</td><td>81.25 ± 0.02</td><td>78.24 ± 0.77</td><td>73.12 ± 1.58</td></tr><tr><td>GCE</td><td>79.56 ± 0.39</td><td>70.46 ± 0.42</td><td>62.61 ± 1.01</td><td>82.98 ± 0.20</td><td>78.46 ± 0.20</td><td>72.75 ± 0.27</td></tr><tr><td>Co-teaching</td><td>76.68 ± 0.86</td><td>69.23 ± 0.62</td><td>61.39 ± 3.15</td><td>80.99 ± 0.88</td><td>76.82 ± 2.01</td><td>71.49 ± 2.74</td></tr><tr><td>Co-teaching+</td><td>78.78 ± 0.96</td><td>71.33 ± 2.31</td><td>62.90 ± 1.55</td><td>80.95 ± 0.86</td><td>77.98 ± 0.76</td><td>72.73 ± 0.98</td></tr><tr><td>JoCoR</td><td>80.11 ± 0.42</td><td>73.52 ± 0.88</td><td>63.52 ± 0.22</td><td>83.60 ± 0.54</td><td>79.45 ± 0.38</td><td>75.45 ± 0.30</td></tr><tr><td>CR</td><td>81.80 ± 0.48</td><td>74.16 ± 0.13</td><td>64.22 ± 1.25</td><td>84.09 ± 0.39</td><td>80.12 ± 0.95</td><td>75.06 ± 1.01</td></tr><tr><td>NPC</td><td>80.38 ± 0.35</td><td>72.93 ± 0.23</td><td>64.38 ± 0.38</td><td>84.28 ± 0.47</td><td>80.26 ± 1.39</td><td>73.81 ± 0.18</td></tr><tr><td>LAFT (Ours)</td><td>82.04 ± 0.11</td><td>76.93 ± 0.60</td><td>70.79 ± 1.98</td><td>90.86 ± 0.02</td><td>88.61 ± 0.25</td><td>80.79 ± 2.32</td></tr></table>
+
+Table 2: The overall performance on  ${20}\mathrm{{Ng}}$  and AGNews with Asymmetric Noise (AN).  
+
+<table><tr><td>Dataset</td><td colspan="2">20Ng</td><td colspan="2">AGNews</td></tr><tr><td>AN Ratio</td><td>20%</td><td>40%</td><td>20%</td><td>40%</td></tr><tr><td>Base</td><td>75.34</td><td>59.65</td><td>82.24</td><td>76.35</td></tr><tr><td>Mixup</td><td>75.84</td><td>62.19</td><td>82.12</td><td>77.45</td></tr><tr><td>GCE</td><td>76.41</td><td>62.37</td><td>83.19</td><td>77.49</td></tr><tr><td>Co-teaching</td><td>77.49</td><td>64.55</td><td>84.15</td><td>79.16</td></tr><tr><td>Co-teaching+</td><td>77.98</td><td>63.48</td><td>86.12</td><td>81.02</td></tr><tr><td>JoCoR</td><td>81.27</td><td>70.99</td><td>87.63</td><td>82.79</td></tr><tr><td>CR</td><td>81.08</td><td>67.22</td><td>88.58</td><td>83.07</td></tr><tr><td>NPC</td><td>79.03</td><td>65.54</td><td>86.83</td><td>80.21</td></tr><tr><td>LAFT (Ours)</td><td>83.70</td><td>81.97</td><td>91.95</td><td>90.12</td></tr></table>
+
+Table 3: The overall performance on  $20\mathrm{Ng}$  and AGNews with Instance-Dependent Noise (IDN).  
+
+<table><tr><td>Dataset</td><td colspan="2">20Ng</td><td colspan="2">AGNews</td></tr><tr><td>IDN Ratio</td><td>20%</td><td>40%</td><td>20%</td><td>40%</td></tr><tr><td>Base</td><td>78.23</td><td>70.28</td><td>85.40</td><td>78.66</td></tr><tr><td>Mixup</td><td>78.21</td><td>70.45</td><td>84.71</td><td>79.22</td></tr><tr><td>GCE</td><td>79.99</td><td>71.74</td><td>87.11</td><td>80.49</td></tr><tr><td>Co-teaching</td><td>77.43</td><td>70.76</td><td>84.09</td><td>78.43</td></tr><tr><td>Co-teaching+</td><td>79.89</td><td>71.52</td><td>88.21</td><td>82.30</td></tr><tr><td>JoCoR</td><td>81.83</td><td>74.04</td><td>88.24</td><td>82.66</td></tr><tr><td>CR</td><td>83.04</td><td>75.51</td><td>90.72</td><td>84.38</td></tr><tr><td>NPC</td><td>81.84</td><td>74.45</td><td>89.90</td><td>82.40</td></tr><tr><td>LAFT (Ours)</td><td>83.61</td><td>80.49</td><td>91.94</td><td>89.34</td></tr></table>
+
+(3) Certain regularization methods, such as Mixup and GCE, exhibit subpar performance when applied to datasets with a higher noise ratio of  $60\%$ . This is because these methods rely on all training samples for PLM fine-tuning, making them more susceptible to strong label noises.
+
+# 5.3 Comparison on Real-world Datasets
+
+In this subsection, we conduct experiments on real-world datasets: SemEval (Zhou et al., 2020), TREC (Awasthi et al., 2020), and Hausa (Hedderich et al., 2020). From the results in Table 4, we observe that LAFT outperforms other baseline methods on all three datasets. Moreover, the results of LAFT are noticeably competitive on TREC and Hausa with relatively higher noise ratios, which further demonstrates the strong robustness and generalization ability of LAFT to label noise.
+
+# 5.4 Ablation Study
+
+In this subsection, we systematically remove specific components from our framework and ana
+
+lyze the resulting impact on performance. We conduct experiments with the following variants: (1) LAFT\C performs coarse-grained separation without LLMs and thus only separates the Easy Clean Set based on PLM predictions. (2) LAFT\F performs fine-grained separation without LLMs, which means when separating Hard Clean Set, only PLM-generated confidence are used. (3) LAFT\N removes our proposed loss  $\mathcal{L}_{\mathcal{N}}$  for True Noisy Set and replaces it with pseudo-labeling based on the most confident PLM prediction. From the results presented in Figure 3, we observe that LAFT outperforms all variants, which verifies the effectiveness of these designs in LAFT. Specifically, removing the learning loss  $\mathcal{L}_{\mathcal{N}}$  leads to significant performance degradation, indicating that such a design can effectively alleviate the adverse impact of noisy labels. Moreover, without the fine-grained separation strategy, the performance deteriorates rapidly when the noise ratio is larger, indicating the importance of fine-grained separation in the presence of higher noise ratios.
+
+Table 4: The overall performance of various models on three real-world noisy datasets.  
+
+<table><tr><td>Dataset</td><td>SemEval</td><td>TREC</td><td>Hausa</td></tr><tr><td>Noise Ratio</td><td>16.00%</td><td>38.56%</td><td>50.37%</td></tr><tr><td>Base</td><td>70.61</td><td>67.42</td><td>47.80</td></tr><tr><td>Mixup</td><td>73.41</td><td>68.44</td><td>47.66</td></tr><tr><td>GCE</td><td>71.91</td><td>67.86</td><td>48.12</td></tr><tr><td>Co-teaching</td><td>72.13</td><td>68.19</td><td>47.56</td></tr><tr><td>Co-teaching+</td><td>72.24</td><td>69.83</td><td>48.25</td></tr><tr><td>JoCoR</td><td>70.11</td><td>66.62</td><td>46.48</td></tr><tr><td>CR</td><td>72.94</td><td>68.90</td><td>48.34</td></tr><tr><td>NPC</td><td>71.22</td><td>67.84</td><td>47.48</td></tr><tr><td>LAFT (Ours)</td><td>73.56</td><td>72.34</td><td>51.71</td></tr></table>
+
+![](images/c31d82dbdf869e580a2f9c61eceef8f3d0843b796d5960c8a6a7325015caf233.jpg)  
+Figure 3: Ablation study of our framework on  $20\mathrm{Ng}$ . S-0.2 and S-0.4 (A, or ID) refer to the noise ratios of  $20\%$  and  $40\%$  for SN (AN, or IDN), respectively.
+
+# 5.5 Evaluation of LLMs-generated Labels
+
+In this subsection, we evaluate the effectiveness of LLMs regarding the quality of generated confidence, presented in Table 5. Specifically, we report the accuracy of LLM-generated prediction on the training set in  $20\mathrm{Ng}$  and AGNews. The statistics are from experiments on  $20\mathrm{Ng}$  and AGNews with  $20\%$  Symmetric Noise (SN). From the results, we can observe that: (1) For Easy Clean Set, the LLM-generated labels exhibit a remarkably high degree of correctness. This empirical finding provides the justification for Assumption 1, which allows for leveraging LLMs to perform coarse-grained separation in our framework. (2) For the TN set, LLM-generated labels are not entirely correct, while still preserving similar accuracy compared to that on all samples. This result empirically verifies Remark 1 and justifies our strategy of utilizing LMM-generated confidence for learning from the TN set. (3) Considering the overall result, LLM-generated labels generally exhibit lower accuracy than PLM-based baselines. That being said, LLMs cannot be
+
+Table 5: The LLM-generated prediction accuracy on the training set in  $20\mathrm{Ng}$  and AGNews in the ideal separation and real separation (shown in gray) during fine-tuning.  
+
+<table><tr><td>Dataset</td><td>EC Set</td><td>HC Set</td><td>TN Set</td><td>Overall</td></tr><tr><td rowspan="2">20Ng</td><td>99.46%</td><td>0%</td><td>76.47%</td><td>75.34%</td></tr><tr><td>99.46%</td><td>3.58%</td><td>69.44%</td><td>75.34%</td></tr><tr><td rowspan="2">AGNews</td><td>99.03%</td><td>0%</td><td>79.35%</td><td>81.41%</td></tr><tr><td>99.03%</td><td>2.05%</td><td>76.07%</td><td>81.41%</td></tr></table>
+
+![](images/4721ed0b99bd091e0337b9638f15f3c1c9508eb82d4f2f4bece63748b5ffd5a9.jpg)  
+Figure 4: The results of our framework and the best baseline CR on  $20\mathrm{Ng}$  with different noise ratios for three types of noise: SN, AN, and IDN.
+
+directly used to replace PLMs for the text classification task when the noise ratio is not extremely high. Nevertheless, our strategy can effectively leverage the LLM-generated confidences, despite their lack of complete correctness, to enhance the fine-tuning performance of PLMs with noisy labels.
+
+# 5.6 Results with Different Noise Ratios
+
+Figure 4 presents the evaluation of our proposed framework across a spectrum of noise ratios on  $20\mathrm{Ng}$ , ranging from 0.1 to 0.9, while considering three distinct types of noise: SN, AN, and IDN. From the results, several significant observations are discovered: (1) With increasing noise ratios, all baseline methods consistently show a deterioration in performance. This decline can be tied to the amplified difficulty of label noise during PLM fin-tuning. (2) Despite the escalating noise ratios, our framework maintains a relatively more robust performance. This is attributed to our sample separation strategy which adeptly discerns clean samples within the training set, thus alleviating the adverse effects of label noise. (3) Our framework demonstrates a slower performance decrease with AN noise compared to other types of noise. This can be credited to the adaptive nature of our approach that employs LLM-generated confidence to effectively learn from similar labels for each sample.
+
+# 6 Conclusion
+
+This paper delves into the issue of fine-tuning Pre-trained Language Models (PLMs) with noisy labels for text classification tasks. We address the detrimental effects of label noise by harnessing external guidance from Large Language Models (LLMs) in the form of confidence. Our approach entails a two-step separation strategy that accurately segregates all training samples into three distinct subsets, each treated with innovative noise-resilient strategies. In summary, our framework adeptly finetunes PLMs in the face of noisy samples, requiring minimal external guidance from LLMs.
+
+# Acknowledgements
+
+Song Wang and Jundong Li are supported by the National Science Foundation under grants (IIS-2006844, IIS-2144209, IIS-2223769, CNS2154962, and BCS-2228534), the Commonwealth Cyber Initiative awards (VV-1Q23-007, HV2Q23-003, and VV-1Q24-011), the JP Morgan Chase Faculty Research Award, the Cisco Faculty Research Award, the Jefferson Lab subcontract, and the UVA 4-VA collaborative research grant.
+
+# Limitation
+
+Despite the promising results, our work presents several limitations that we must acknowledge: (1) Dependence on Large Language Models (LLMs): Our framework leverages the guidance from LLMs in the form of confidences. This implies that the quality of these confidences and hence the effectiveness of our model are closely tied to the performance of the LLM. Unsatisfactory LLM performance could thus directly impact our framework's efficacy. (2) Noise Types and Ratios: The experiments in this paper mainly focus on three types of noise: Symmetric Noise (SN), Asymmetric Noise (AN), and Instance-Dependent Noise (IDN), with noise ratios up to  $60\%$ . However, there may be other noise types or higher noise ratios in practice, and it remains to be investigated how well our framework performs under these circumstances. (3) Limitations in Sample Separation Strategies: The efficacy of our framework relies on accurately dividing the training samples into clean and noisy subsets. If this distinction is not clear or becomes more complex, our current approach may encounter difficulties. (4) Domain-Specific Text Classification: While the experiments are performed on specific text classification tasks, we do not investigate
+
+the effectiveness of our approach in more domain-specific contexts where the nature of the noise could be different. (5) Computational Costs: Finally, our approach entails using LLMs which can be computationally expensive and could thus pose a challenge when applied to large datasets or in resource-constrained environments. In summary, future research could focus on overcoming these limitations and exploring the adaptability of our proposed framework to other noise types, higher noise ratios, more complex noise patterns, as well as different task domains.
+
+# Ethics Statement
+
+This research adheres to the principles of ethical conduct in artificial intelligence research. The development and utilization of our work are carried out with full recognition of the potential implications. While our method improves the performance of Pretrained Language Models (PLMs) with noisy labels in text classification tasks, we acknowledge that it could be potentially used in contexts that may involve ethical concerns, such as disinformation or manipulation of opinion on social platforms. Our experiments are performed on five publicly available datasets, namely  $20\mathrm{Ng}$ , AGNews, SemEval, TREC, and Hausa. The datasets are already anonymized and do not contain any personally identifiable information. Moreover, we have complied with all guidelines and standards for the use of these datasets. Furthermore, the use of Large Language Models (LLMs) in our work is subject to the terms and conditions put forth by the creators and hosts of these models. We only empty them for the purpose of enhancing the fine-tuning performance of PLMs and do not exploit them for any inappropriate or unauthorized purposes. While our work focuses on improving the robustness of PLMs to label noise in data, we recognize the broader societal concerns related to the potential misuse of our framework. We advocate for the responsible use of our research findings and encourage continued discourse around the ethical use of machine learning and artificial intelligence technologies in real-world applications. Finally, it is crucial to note that the development of technologies and methodologies in machine learning should be done in conjunction with ongoing considerations about their ethical implications, potential misuse, and societal impacts. It is our responsibility to foster an environment of ethical vigilance in AI research and development.
+
+# References
+
+Christoph Alt, Aleksandra Gabryszak, and Leonhard Hennig. 2020. Tacred revisited: A thorough evaluation of the tacred relation extraction task. In ACL.  
+Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. 2017. A closer look at memorization in deep networks. In ICML.  
+Abhijeet Awasthi, Sabyasachi Ghosh, Rasna Goyal, and Sunita Sarawagi. 2020. Learning from rules generalizing labeled exemplars. In *ICLR*.  
+HeeSun Bae, Seungjae Shin, Byeonghu Na, JoonHo Jang, Kyungwoo Song, and Il-Chul Moon. 2022. From noisy prediction to true label: Noisy prediction calibration via generative model. In ICML.  
+David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. 2019. Mixmatch: A holistic approach to semi-supervised learning. NeurIPS.  
+Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. 2021. Learning with instance-dependent label noise: A sample sieve approach. In ICLR.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.  
+Erik Englesson and Hossein Azizpour. 2021. Generalized jensen-shannon divergence loss for learning with noisy labels. NeurIPS.  
+Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. 2018. Co-teaching: Robust training of deep neural networks with extremely noisy labels. NeurIPS.  
+Michael A Hedderich, David I Adelani, Dawei Zhu, Jesujoba Alabi, Udia Markus, and Dietrich Klakow. 2020. Transfer learning and distant supervision for multilingual transformer models: A study on african languages. In EMNLP.  
+Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. 2018. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML.  
+Nazmul Karim, Mamshad Nayeem Rizve, Nazanin Rahnavard, Ajmal Mian, and Mubarak Shah. 2022. Unicon: Combating label noise through uniform selection and contrastive learning. In CVPR.  
+Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In ICLR.  
+Ken Lang. 1995. Newsweeder: Learning to filter netnews. In Machine Learning Proceedings.
+
+Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. AI Open.  
+Junnan Li, Richard Socher, and Steven CH Hoi. 2020. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR.  
+Junnan Li, Caiming Xiong, and Steven CH Hoi. 2021. Learning from noisy data with robust representation learning. In ICCV.  
+Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. 2017. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862.  
+Xin Li and Dan Roth. 2002. Learning question classifiers. In COLING.  
+Yifan Li, Hu Han, Shiguang Shan, and Xilin Chen. 2023. Disc: Learning from noisy labels via dynamic instance-specific selection and correction. In CVPR.  
+Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. 2020. Early-learning regularization prevents memorization of noisy labels. NeurIPS.  
+Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. 2021. Augmentation strategies for learning with noisy labels. In CVPR.  
+OpenAI. 2023. Gpt-4 technical report.  
+Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. 2017. Making deep neural networks robust to label noise: A loss correction approach. In CVPR.  
+Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. 2020. Identifying unlabeled data using the area under the margin ranking. NeurIPS.  
+Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM.  
+Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. 2020. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS.  
+Hwanjun Song, Minseok Kim, and Jae-Gil Lee. 2019. Selfie: Refurbishing unclean samples for robust deep learning. In ICML.  
+Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In EMNLP.
+
+Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. 2018. Joint optimization framework for learning with noisy labels. In CVPR.  
+Song Wang, Xingbo Fu, Kaize Ding, Chen Chen, Huiyuan Chen, and Jundong Li. 2023a. Federated few-shot learning. In SIGKDD.  
+Song Wang, Jing Ma, Lu Cheng, and Jundong Li. 2023b. Fair few-shot learning with auxiliary sets. In ECAI.  
+Yaqing Wang, Song Wang, Yanyan Li, and Dejing Dou. 2022. Recognizing medical search query intent by few-shot learning. In SIGIR.  
+Yaqing Wang, Song Wang, Quanming Yao, and Dejing Dou. 2021. Hierarchical heterogeneous graph representation learning for short text classification. In EMNLP.  
+Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. 2020. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR.  
+Jason Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. In EMNLP-IJCNLP.  
+Adina Williams, Nikita Nangia, and Samuel R Bowman. 2017. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426.  
+Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nanan Wang, Zongyuan Ge, and Yi Chang. 2021. Robust early-learning: Hinding the memorization of noisy labels. In ICLR.  
+Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. 2019. Unsupervised data augmentation for consistency training. arXiv preprint arXiv:1904.12848.  
+Yan Yan, Rómer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. 2014. Learning from multiple annotators with varying expertise. Machine Learning.  
+Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. 2021. Instance-dependent label-noise learning under a structural causal model. In NeurIPS.  
+Yu Yao, Tongliang Liu, Bo Han, Mingming Gong, Jiankang Deng, Gang Niu, and Masashi Sugiyama. 2020. Dual t: Reducing estimation error for transition matrix in label-noise learning. NeurIPS.  
+Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. 2019. How does disagreement help generalization against label corruption? In ICML.  
+Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. 2021a. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS.
+
+Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. 2018. mixup: Beyond empirical risk minimization. In ICLR.  
+Jieyu Zhang, Yue Yu, Yinghao Li, Yujing Wang, Yaming Yang, Mao Yang, and Alexander Ratner. 2021b. WRENCH: A comprehensive benchmark for weak supervision. In NeurIPS Datasets and Benchmarks Track.  
+Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. NIPS.  
+Yivan Zhang, Gang Niu, and Masashi Sugiyama. 2021c. Learning noise transition matrix from only noisy labels via total variation regularization. In ICML.  
+Zhilu Zhang and Mert Sabuncu. 2018. Generalized cross entropy loss for training deep neural networks with noisy labels. NeurIPS.  
+Ganlong Zhao, Guanbin Li, Yipeng Qin, Feng Liu, and Yizhou Yu. 2022. Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In ECCV.  
+Wenxuan Zhou and Muhao Chen. 2021. Learning from noisy labels for entity-centric information extraction. In EMNLP.  
+Wenxuan Zhou, Hongtao Lin, Bill Yuchen Lin, Ziqi Wang, Junyi Du, Leonardo Neves, and Xiang Ren. 2020. Nero: A neural rule grounding framework for label-efficient relation extraction. In TheWebConf.  
+Dawei Zhu, Michael Hedderich, Fangzhou Zhai, David Adelani, and Dietrich Klakow. 2022. Is bert robust to label noise? a study on learning with noisy labels in text classification. In Insights from Negative Results in NLP.  
+Yuchen Zhuang, Yue Yu, Lingkai Kong, Xiang Chen, and Chao Zhang. 2023. Dygen: Learning from noisy labels via dynamics-enhanced generative modeling. arXiv preprint arXiv:2305.19395.
+
+# A Details of Three Subsets
+
+In this section, we further provide details about the three subsets of samples in our framework.
+
+1. Easy Clean (EC) Set. In this subset, the labels predicted by LLMs are the same as the assigned labels. Therefore, based on Assumption 1, we can infer that the labels of these samples are almost clean, which allows us to directly use them for model training. Note that a small portion of samples in this category can still be noisy when the LLM predictions happen to be the same as noisy labels. However, we empirically verify in Sec. 5.5 that this proportion is small.
+
+2. Hard Clean (HC) Set. In this (ideal) category, the labels predicted by LLMs deviate from the assigned labels, while the true labels are in accordance with the assigned labels. In this case, we are acknowledged that these samples are also clean. However, due to the semantic difficulties of these samples, the LLMs cannot easily predict the correct labels of them. Nevertheless, since the assigned labels are correct as true labels, we can use assigned labels for model training on these samples. It is noteworthy that in practice, it is infeasible to achieve a perfect separation of this subset. Therefore, in our framework, we propose to leverage LLM-generated and PLM-generated confidences to improve the separation performance.
+
+3. True Noisy (TN) Set. In this (ideal) category, the labels predicted by LLMs and the true labels are both different from the assigned labels. In other words, these samples are all noisy as their assigned labels are different from true labels. Since they are noisy, the labels predicted by LLMs are also likely to be different from the assigned labels. As these samples are noisy, we cannot directly use their labels for training. However, since we narrow down the potential range of noisy samples to this category, we can resort to specific techniques to learn from these noisy samples. In our framework, we utilize LLM-generated confidence as additional supervision information to effectively learn from these noisy samples.
+
+# B Datasets
+
+In this section, we introduce the details of the datasets used in our experiments. In particular,  $20\mathrm{Ng}$  (Lang, 1995) and AGNews (Li and Roth, 2002; Zhang et al., 2015) are news topic classification datasets that are prevalently used for text classification tasks. We manually inject different types of noise (SN, AN, and IDN) into these two datasets for evaluation in our experiments. For real-world datasets, we utilize SemEval (Zhou et al., 2020), TREC (Awasthi et al., 2020), and Hausa (Hedderich et al., 2020). Specifically, SemEval is a relation extraction dataset, and we follow the process introduced in (Zhou et al., 2020) to obtain noisy labels. TREC is a question classification dataset in the weak supervision benchmark WRENCH (Zhang et al., 2021b). Moreover, Hausa is a text classification dataset in the language of
+
+Table 6: The detailed statistics of the five datasets used in our experiments.  
+
+<table><tr><td>Dataset</td><td># Training</td><td># Validation</td><td># Test</td><td># Class</td></tr><tr><td>20Ng</td><td>9,051</td><td>2,263</td><td>7,532</td><td>20</td></tr><tr><td>AGNews</td><td>40,000</td><td>7,600</td><td>7,600</td><td>4</td></tr><tr><td>SemEval</td><td>1,749</td><td>200</td><td>692</td><td>9</td></tr><tr><td>TREC</td><td>4,965</td><td>500</td><td>500</td><td>6</td></tr><tr><td>Hausa</td><td>2,045</td><td>290</td><td>582</td><td>5</td></tr></table>
+
+Hausa, the second most spoken indigenous language in Africa, with 40 million native speakers. For this dataset, gazetteers are used for automatic labeling, which results in feature-dependent label noise. Here we provide the detailed statistics of these datasets in Table 6.
+
+# C Baselines
+
+In this section, we provide the details of the baselines used in our experiments.
+
+- Base (Devlin et al., 2018) is the BERT base model fine-tuned based on the standard cross-entropy loss. Note that for dataset Hausa, we utilize the multilingual BERT model.  
+- Mixup (Zhang et al., 2018) is a semi-supervised approach that performs a linear interpolation between clean and noisy samples.  
+- GCE (Zhang and Sabuncu, 2018) denotes Generalized Cross-Entropy loss, which can be regarded as a general loss that combines mean absolute error (MAE) and cross-entropy (CE).  
+- Co-teaching (Han et al., 2018) trains two different models and selects small-loss samples to feed them into each other for optimization.  
+- Co-teaching+ (Yu et al., 2019) is an updated version of Co-teaching that explicitly ensures the difference between the two models.  
+- JoCoR (Wei et al., 2020) trains two models and selects samples based on the sum of loss from the two models.  
+- CR (Zhou and Chen, 2021) also trains multiple models with a regularization strategy based on a soft target.  
+- NPC (Bae et al., 2022) utilizes a generative model to estimate the transition matrix from noisy predictions to the ground-truth labels of samples and uses it to correct noisy labels.
+
+# D Implementation Details
+
+In this section, we introduce the implementation details for our experiments. We run the experiments on a single 48GB Nvidia GeForce RTX A6000 GPU. The experiments are run 10 times to achieve the average accuracy and the standard deviation.
+
+For the augmentations used in our framework, following existing works (Xie et al., 2019; Wei and Zou, 2019; Li et al., 2022), we adopt four typical text augmentations: Back Translation, Random Insertion, Random Deletion, and Random Swap. We set the threshold  $\widehat{\tau}$  for LLM-generated confidences as 0.8. We set the hyper-parameters  $\lambda$  and  $\widetilde{\tau}$  for the adaptive threshold  $\widetilde{\tau}(t)$  for PLM-generated confidences as 0.7 and 0.8, respectively. We set  $\alpha$  as 1.5 and  $\beta$  as 0.9. For the loss weight, we set  $\lambda_{\mathcal{H}} = \lambda_{\mathcal{N}} = 1$ . We set the model fine-tuning learning rate as  $10^{-4}$  with a batch size of 32. The maximum length of texts is set as 256. The number of training epochs is set as 100 with early stopping based on the performance of the validation set. We optimize the model using the Adam (Kingma and Ba, 2015) optimization strategy. For the LLM, we utilize ChatGPT based on GPT4.
+
+In addition, we provide more detailed package requirements as listed below.
+
+- Python == 3.9.12  
+- torch == 2.0.1  
+- transformers == 4.29.2  
+- huggingface-hub == 0.15.1  
+- keras == 2.12.0  
+- numpy == 1.23.5  
+- scipy == 1.5.3  
+- CUDA == 11.6  
+networkx  $= = 2.5.1$  
+scikit-learn  $= = 0.24.1$

2023/Noise-Robust Fine-Tuning of Pretrained Language Models via External Guidance/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7cd1fb401924758e546e67db3a092cbfc7651ee3b7c9f3faf6d3f0c37319a40
+size 502706

2023/Noise-Robust Semi-Supervised Learning for Distantly Supervised Relation Extraction/fc076ec4-b157-42d2-a4a8-4199f3a5df0a_content_list.json

@@ -0,0 +1,2023 @@
+[
+    {
+        "type": "text",
+        "text": "Noise-Robust Semi-Supervised Learning for Distantly Supervised Relation Extraction",
+        "text_level": 1,
+        "bbox": [
+            115,
+            87,
+            880,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xin Sun $^{1,2}$ , Qiang Liu $^{2,3*}$ , Shu Wu $^{2,3}$ , Zilei Wang $^{1}$ , Liang Wang $^{2,3}$",
+        "bbox": [
+            221,
+            140,
+            776,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>University of Science and Technology of China",
+        "bbox": [
+            304,
+            159,
+            697,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ CRIPAC & MAIS, Institute of Automation, Chinese Academy of Sciences",
+        "bbox": [
+            196,
+            175,
+            805,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>3</sup>University of Chinese Academy of Sciences",
+        "bbox": [
+            317,
+            193,
+            682,
+            209
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "sunxin000@mail.ustc.edu.cn, qiang.liu@nlpr ia.ac.cn",
+        "bbox": [
+            310,
+            212,
+            690,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Distantly supervised relation extraction (DSRE) aims to extract relational facts from texts but suffers from noisy instances. To mitigate the influence of noisy labels, current methods typically use the Multi-Instance-Learning framework to extract relations for each bag. However, these approaches are not capable of extracting relation labels for individual sentences. Several studies have focused on sentence-level DSRE to solve the above problem. These studies primarily aim to develop methods for identifying noisy samples and filtering them out to mitigate the impact of noise. However, discarding noisy samples directly leads to the loss of useful information. To this end, we propose SSLRE, a novel Semi-Supervised-Learning Relation Extraction framework for sentence-level DSRE. We discard only the labels of the noisy samples and utilize these instances without labels as unlabeled samples. Our SSLRE framework utilizes a weighted K-NN graph to select confident samples as labeled data and the rest as unlabeled. We then design a robust semi-supervised learning framework that can efficiently handle remaining label noise present in the labeled dataset, while also making effective use of unlabeled samples. Based on our experiments on two real-world datasets, the SSLRE framework we proposed has achieved significant enhancements in sentence-level relation extraction performance compared to the existing state-of-the-art methods. Moreover, it has also attained a state-of-the-art level of performance in bag-level relation extraction with ONE aggregation strategy.",
+        "bbox": [
+            144,
+            281,
+            460,
+            793
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            806,
+            258,
+            821
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Relation extraction (RE) is a fundamental process for constructing knowledge graphs, as it aims to predict the relationship between entities based on their context. However, most supervised RE",
+        "bbox": [
+            112,
+            832,
+            489,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/487a4bbf26088f7c7dc661d596f7a5e371d0031f089d0ba355fbb643dfe7cbc3.jpg",
+        "image_caption": [
+            "Figure 1: Bag-level RE maps a bag of sentences to bag labels. Sentence-level RE maps each sentence to a specific relation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            255,
+            884,
+            357
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "techniques require extensive labeled training data, which can be difficult to obtain manually. To address this issue, Distant Supervision (DS) was proposed (Mintz et al., 2009) to automatically generate labeled text corpus by aligning plain texts with knowledge bases (KB). For instance, if a sentence contains both the subject  $(s)$  and object  $(o)$  of a triple  $\\langle s,r,o\\rangle$  ( $\\langle \\text{subject}, \\text{relation}, \\text{object} \\rangle$ ), then the DS method considers  $\\langle s,r,o\\rangle$  as a valid sample for that sentence. Conversely, if no relational triples apply, then the sentence is labeled as \"NA\".",
+        "bbox": [
+            507,
+            450,
+            884,
+            626
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Distantly supervised datasets usually face high label noise in training data due to the annotation process. To mitigate the impact of noisy labels caused by distant supervision, contemporary techniques (Lin et al., 2016; Alt et al., 2019; Chen et al., 2021b; Li et al., 2022; Dai et al., 2022) usually employ multi-instance-learning (MIL) framework or modify MIL to train the relation extraction model.",
+        "bbox": [
+            507,
+            627,
+            882,
+            756
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Although MIL-based techniques can identify bag relation labels, they are not proficient in precisely mapping each sentence in a bag with explicit sentence labels (Feng et al., 2018; Jia et al., 2019; Ma et al., 2021). Several studies have focused on improving sentence-level DSRE and have empirically demonstrated the inadequacy of bag-level methods on sentence-level evaluation. (Feng et al., 2018; Qin et al., 2018) apply reinforcement learning to train a sample selector. (Jia et al., 2019) iden",
+        "bbox": [
+            507,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding author.",
+        "bbox": [
+            141,
+            904,
+            289,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "13145",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13145-13157 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            208,
+            945,
+            786,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tify confident samples by frequent patterns. (Ma et al., 2021) utilizes negative training to separate noisy data from the training data.",
+        "bbox": [
+            112,
+            84,
+            487,
+            131
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, these methods have two main issues: (1) These works simply discard all noisy samples and train a relation extraction model with selected samples. However, filtering out noisy samples directly results in the loss of useful information. (Gao et al., 2021a) notes that the DSRE dataset has a noisereate exceeding  $50\\%$ . Discarding all these samples would result in a significant loss of information. (2) The confident samples selection procedure is not impeccable, and there may still exist a small amount of noise in the chosen confident samples. Directly training a classifier in the presence of label noise is known to result in noise memorization.",
+        "bbox": [
+            115,
+            134,
+            489,
+            341
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the two issues, this work proposes a novel semi-supervised-learning framework for sentence-level DSRE, First, we construct a K-NN graph for all samples using the hidden features. Then, we identify confident samples from the graph structure and consider the remaining samples as noisy. For issue (1): Our method discards only the noisy labels and treats corresponding samples as unlabeled data. We then utilize this unlabeled data by pseudo labeling within our robust semi-supervised learning framework to learn a better feature representation for relation. For issue (2): Despite our initial selection of confident samples, there may still be noise in the labeled dataset. we have developed a noise-robust semi-supervised learning framework that leverages mixup supervised contrastive learning to learn from the labeled dataset and curriculum pseudo labeling to learn from the unlabeled dataset.",
+        "bbox": [
+            115,
+            344,
+            489,
+            646
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize the contribution of this work:",
+        "bbox": [
+            132,
+            650,
+            460,
+            665
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a noise-robust Semi-Supervised-Learning framework SSLRE for DSRE task, which effectively mitigate the impact of noisy labels.",
+            "- We propose to use graph structure information (weighted K-NN) to identify the confident samples and effectively convert noisy samples as useful training data by utilizing pseudo labeling.",
+            "- The proposed framework achieves significant improvement over previous methods in terms of both sent-level and bag-level relation extraction performance."
+        ],
+        "bbox": [
+            136,
+            681,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            660,
+            98
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Distantly Supervised Relation Extraction",
+        "text_level": 1,
+        "bbox": [
+            507,
+            109,
+            877,
+            124
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Relation extraction (RE) is a fundamental pro- cess for constructing knowledge graphs(Zhang et al., 2023a; Xia et al., 2022; Zhang et al., 2023b). To generate large-scale auto-labeled data without human effort, (Mintz et al., 2009) first use Distant Supervision to label sentences mentioning two entities with their relations in KGs, which inevitably brings wrongly labeled instances. To tackle the predicament of noise, most of the existing studies on DSRE are founded on the multi-instance learning framework. This approach is leveraged to handle noisy sentences in each bag and train models by capitalizing on the constructed trustworthy baglevel representations. Usually, these methods employ attention mechanisms to assign less weights to the probable noisy sentences in the bag(Lin et al., 2016; Han et al., 2018b; Alt et al., 2019; Shen et al., 2019; Ye and Ling, 2019; Chen et al., 2021a; Li et al., 2022), apply adversarial training or reinforcement learning to remove the noisy sentences from the bag (Zeng et al., 2015; Qin et al., 2018; Shang and Wei, 2019; Chen et al., 2020; Hao et al., 2021). However, the studies (Feng et al., 2018; Jia et al., 2019; Ma et al., 2021; Gao et al., 2021a) indicate that the bag-level DSRE methods are ineffective for sentence-level relation extraction.",
+        "bbox": [
+            507,
+            130,
+            885,
+            546
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This work focuses on extracting relations at the sentence level, (Feng et al., 2018) applied reinforcement learning to identify confident instances based on the reward of noisy labels. (Jia et al., 2019) involve building initial reliable sentences based on several manually defined frequent relation patterns. (Ma et al., 2021) assigning complementary labels that cooperate with negative training to filter out noisy instances. Unlike previous studies, our method only discard noisy labels and keep the unlabeled samples. We use pseudo labeling to effectively utilize unlabeled samples, which helps to learn better representation.",
+        "bbox": [
+            507,
+            549,
+            885,
+            758
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Semi-Supervised-Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            769,
+            766,
+            784
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In SSL, a portion of training dataset is labeled and the remaining portion is unlabeled. SSL has seen great progress in recent years. Since (Bachman et al., 2014) proposed a consistency regularization based method, many approaches have migrated it into the semi-supervised learning field. MixMatch (Berthelot et al., 2019) proposes to combine consistency regularization with entropy minimization.",
+        "bbox": [
+            507,
+            790,
+            885,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "13146",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Mean Teacher (Tarvainen and Valpola, 2017) and Dual Student (Ke et al., 2019) are also based on consistent learning, aiming for the same outputs for different networks. Recently, FixMatch (Sohn et al., 2020) provides a simple yet effective weak-to-strong consistency regularization framework. Flexmatch (Zhang et al., 2021) provides curriculum pseudo Labels learning to combat the imbalance of pseudo labels.",
+        "bbox": [
+            112,
+            84,
+            489,
+            227
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our SSLRE framework differs from these frameworks in two main ways. Firstly, our labeled dataset still contains a small amount of noise due to the fact that confident sample identification cannot achieve perfection. Therefore, we utilize mixup contrastive supervised learning to combat this noise. Secondly, current SSL methods generate and utilize pseudo labels with the same head, which causes error accumulation during the training stage. To address this issue, we propose utilizing a pseudo classifier head, which decouples the generation and utilization of pseudo labels by two parameter-independent heads to avoid error accumulation.",
+        "bbox": [
+            115,
+            230,
+            489,
+            439
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Learning with Noisy Data",
+        "text_level": 1,
+        "bbox": [
+            112,
+            451,
+            363,
+            467
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In both computer vision and natural language processing, learning with noisy data is a widely discussed problem. Existing approaches include but not limit to estimating the noise transition matrix (Chen and Gupta, 2015; Goldberger and Ben-Reuven, 2016), leveraging a robust loss function (Lyu and Tsang, 2019; Ghosh et al., 2017; Liu and Guo, 2019), introducing regularization (Liu et al., 2020; Iscen et al., 2022), selecting noisy samples by multi-network learning or multi-round learning (Han et al., 2018a; Wu et al., 2020), re-weighting examples (Liu and Tao, 2014), generating pseudo labels (Li et al., 2020; Han et al., 2019), and so on. In addition, some advanced state-of-the-art methods combine serveral techniques, e.g., Dividemix (Li et al., 2020) and  $\\mathrm{ELR + }$  (Liu et al., 2020).",
+        "bbox": [
+            112,
+            474,
+            489,
+            731
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we address the issue of noisy labels in distant relation extraction. Our approach first involves constructing a K-NN graph to identify confident samples based on their graph structure, and then use noise-robust mixup supervised contrastive learning to train with the labeled samples.",
+        "bbox": [
+            112,
+            733,
+            489,
+            829
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            859,
+            263,
+            876
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To achieve sentence-level relation extraction in DSRE, we propose a framework called SSLRE,",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "which consists of two main steps. Firstly, we select confident samples from the distantly supervised dataset using a weighted K-NN approach built by all sample representations. We use the selected samples as labeled data and the remaining samples as unlabeled data (as detailed in Section 3.1). Secondly, we employ our robust Semi-Supervised Learning framework to learn from the Semi-Supervised datasets (as described in Section 3.2). Appendix A delineates the full algorithm.",
+        "bbox": [
+            507,
+            84,
+            884,
+            243
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we denote the original dataset in this task as  $\\tilde{\\mathcal{D}} = \\{(s_1,\\tilde{y}_1),(s_2,\\tilde{y}_2),\\dots ,(s_N,\\tilde{y}_N)\\}$  where  $\\tilde{y}_i\\in \\{1,\\dots ,C\\}$  is the label of the ith input sentence  $s_i$  . The labeled dataset (identified confident samples) is denoted as  $\\mathcal{X} = \\{(x_{1},y_{1}),(x_{2},y_{2}),\\dots ,(x_{n},y_{n})\\}$  and the unlabeled dataset (the noisy samples without labels) is denoted as  $\\mathcal{U} = \\{(u_1,u_2,\\dots ,u_m\\}$  where  $m + n = N$",
+        "bbox": [
+            507,
+            244,
+            884,
+            405
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Confident Samples Identification with Weighted K-NN",
+        "text_level": 1,
+        "bbox": [
+            507,
+            416,
+            853,
+            447
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our Semi-Supervised-Learning module requires us to divide the noisy dataset into a labeled dataset and an unlabeled dataset. Inspired by (Lee et al., 2019; Bahri et al., 2020), we utilize neighborhood information of the hidden feature spaces to identify confident samples. We employ supervised contrastive learning to warm up our model and obtain the representations for all instances. It is noteworthy that deep neural networks tend to initially fit the training data with clean labels during an early training phase, prior to ultimately memorizing the examples with false labels (Arpit et al., 2017; Liu et al., 2020). Consequently, we only warm up our model for a single epoch. Given two sentences  $s_i$  and  $s_j$ , we can obtain their low-dimensional representations as  $z_i = \\theta(s_i)$  and  $z_j = \\theta(s_j)$ , where  $\\theta$  is the sentence encoder. We then calculate their representation similarity using the cosine distance",
+        "bbox": [
+            507,
+            451,
+            884,
+            741
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd \\left(z _ {i}, z _ {j}\\right) = \\frac {z _ {i} z _ {j} ^ {T}}{\\| z _ {i} \\| \\| z _ {j} \\|.} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            611,
+            747,
+            882,
+            785
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Then, we build a weighted K-NN graph for all samples based on the consine distance. To quantify the agreement between  $s_i$  and  $\\tilde{y}_i$ , We first use the label distribution in the  $K$ -neighborhood to approximate clean posterior probabilities,",
+        "bbox": [
+            507,
+            789,
+            884,
+            870
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {q} _ {c} \\left(s _ {i}\\right) = \\frac {1}{\\sum_ {s _ {k} \\in \\mathcal {N} _ {i}} d \\left(z _ {i} , z _ {k}\\right)} \\sum_ {\\boldsymbol {s} _ {k} \\in \\mathcal {N} _ {i}} d \\left(z _ {i}, z _ {k}\\right) \\cdot \\mathbb {1} \\left(\\tilde {y} _ {k} = c\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            517,
+            875,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "13147",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/043c5beee27f8eacbec8f6f0812c2decc0325b1f7dc16853e2793881e21d1da4.jpg",
+        "image_caption": [
+            "Figure 2: An overview of the proposed framework, SSLRE.  $\\hat{D},\\mathcal{X},\\mathcal{U}$  denote the original noisy dataset, labeled dataset and unlabeled dataset.  $\\theta$  indicates the encoder.  $\\theta_w$  and  $\\theta_s$  mean forward with lower and higher dropout rate, respectively.  $\\phi$  and  $\\psi$  are classifier head and pseudo classifier head.  $\\mathcal{L}_s$  is the mixup supervised contrastive loss defined in Eq. (12), and  $\\mathcal{L}_{u,t}$  is the unsupervised loss defined in Eq. (6)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            139,
+            80,
+            850,
+            268
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{N}_i$  represents the set of  $K$  closest neighbors to  $s_i$ . We then use the cross-entropy loss  $\\ell$  to calculate the disagreement between  $\\hat{\\mathbf{q}}_c(\\boldsymbol{s}_i)$  and  $\\tilde{y}_i$ . Denoting the set of confident examples belonging to the  $c$ -th class as  $\\mathcal{X}_c$ , we have",
+        "bbox": [
+            112,
+            362,
+            487,
+            441
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {X} _ {c} = \\left\\{s _ {i}, \\tilde {y} _ {i} \\mid \\ell \\left(\\hat {\\mathbf {q}} \\left(s _ {i}\\right), \\tilde {y} _ {i}\\right) <   \\gamma_ {c}, \\tilde {y} _ {i} = c \\right\\}, c \\in [ C ], \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            115,
+            454,
+            485,
+            486
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\gamma_{c}$  is a threshold for the  $c$ -th class, which is dynamically defined to ensure a class-balanced set of identified confident examples. To achieve this goal, we use the  $\\alpha$  fractile of per-class agreements between the original label  $\\tilde{y}_i$  and  $\\max_c\\hat{\\mathbf{q}}_c(\\boldsymbol{s}_i)$  across all classes to determine how many examples should be selected for each class, i.e.  $\\sum_{i=1}^{n} = \\mathbb{1}(\\max_{c'}\\hat{\\mathbf{q}}_{c'}(\\boldsymbol{s}_i) = \\tilde{y}_i) \\cdot (\\tilde{y}_i = c), c \\in [C]$ . Finally, we can get the labeled set and unlabeled set as",
+        "bbox": [
+            112,
+            488,
+            487,
+            632
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {X} = \\cup_ {c = 1} ^ {C} \\mathcal {X} _ {c} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            642,
+            487,
+            668
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {U} = \\left\\{s _ {i} \\mid \\left(s _ {i}, \\tilde {y} _ {i}\\right) \\in \\tilde {\\mathcal {D}} \\backslash \\mathcal {X} \\right\\}.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            665,
+            408,
+            683
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Noise-Robust Semi-Supervised learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            693,
+            470,
+            709
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Despite selecting confident samples from the distantly supervised dataset, there still remains a small amount of noise in the labeled dataset. Naively training a classifier in the presence of label noise leads to noise memorization (Liu et al., 2020), which degrades the performance. We propose a noise-robust semi-supervised learning framework to mitigate the influence of remaining noise.",
+        "bbox": [
+            112,
+            714,
+            487,
+            843
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Data Augmentation with Dropout",
+        "text_level": 1,
+        "bbox": [
+            112,
+            852,
+            438,
+            866
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inspired by SimCSE (Gao et al., 2021b), we augment training samples by embedding processing. In particular, We obtain different embeddings of a",
+        "bbox": [
+            112,
+            870,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "sentence by applying dropout during the forward process. Additionally, we propose using a high dropout rate for strong augmentation and a low dropout rate for weak augmentation. The sentence encoder is denoted as  $\\theta$ , with forward propagation using a high dropout rate denoted as  $\\theta_{s}$  and forward propagation using a low dropout rate denoted as  $\\theta_{w}$ .",
+        "bbox": [
+            507,
+            362,
+            882,
+            491
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Unsupervised Learning with Pseudo Labeling",
+        "text_level": 1,
+        "bbox": [
+            507,
+            507,
+            853,
+            539
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this part, we propose two modules to learn from the unlabeled dataset: (1) To generate and utilize pseudo labels independently, we propose pseudo classifier head. (2) Utilize Curriculum Pseudo Labeling to perform consistent learning while combatting the unbalance of the generated pseudo labels.",
+        "bbox": [
+            507,
+            546,
+            882,
+            643
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Pseudo classifier head: pseudo labeling is one of the prevalent techniques in semi-supervised learning. The existing approaches generate and utilize pseudo labels with the same head. However, this may cause training bias, ultimately amplifying the model's errors as self-training continues. (Wang et al., 2022). To reduce this bias when using pseudo labels, we propose utilizing a two-classifier model consisting of an encoder  $\\theta$  with both a classifier head  $\\phi$  and a pseudo classifier head  $\\psi$ . We optimize the classifier head  $\\phi$  using only labeled samples and without any unreliable pseudo labels from unlabeled samples. Unlabeled samples are used solely for updating encoder  $\\theta$  and pseudo classifier head  $\\psi$ . In particular, the classifier head  $\\phi$  generates pseudo labels  $(\\phi \\circ \\theta_w)(u_b)$  for unlabeled samples (which have",
+        "bbox": [
+            507,
+            646,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "13148",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "no gradient), the loss of unlabeled samples is calculated by  $\\ell ((\\psi \\circ \\theta_s)(u_b),(\\phi \\circ \\theta_w)(u_b))$  , where  $\\ell$  denotes cross entropy loss. This decouples the generation and utilization of pseudo labels by two parameter-independent heads to mitigate error accumulation.",
+        "bbox": [
+            112,
+            84,
+            489,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Curriculum Pseudo Labeling: Due to the highly unbalanced dataset, using a constant cut-off  $\\tau$  for all classes in Pseudo labeling results in almost all selected samples (those with confidence greater than the cut-off) being labeled as 'NA', which is the dominant class.",
+        "bbox": [
+            112,
+            180,
+            487,
+            275
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inspired by Flexmatch (Zhang et al., 2021), we use Curriculum Pseudo Labeling (CPL) to combat unbalanced pseudo labels. We first generate pseudo labels for iteration  $t$",
+        "bbox": [
+            112,
+            277,
+            487,
+            340
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {p}} _ {b} = \\left(\\phi_ {t} \\circ \\theta_ {w, t}\\right) \\left(u _ {b}\\right), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            218,
+            350,
+            487,
+            367
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "These labels are then used as the target of strongly-augmented data. The unsupervised loss term has the form as",
+        "bbox": [
+            112,
+            375,
+            489,
+            420
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {u, t} = \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} \\mathbb {1} (\\max  (\\hat {\\mathbf {p}} _ {b}) > \\mathcal {T} _ {t} (\\arg \\max  (\\hat {\\mathbf {p}} _ {b}))) \\tag {6} \\\\ \\cdot \\ell \\left(\\left(\\psi_ {t} \\circ \\theta_ {s, t}\\right) (u _ {b}), \\hat {\\mathbf {p}} _ {b}\\right), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            428,
+            487,
+            480
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where",
+        "bbox": [
+            112,
+            487,
+            164,
+            499
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {T} _ {t} (c) = \\frac {\\sigma_ {t} (c)}{\\max  _ {c} \\sigma_ {t}} \\tau , \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            226,
+            497,
+            487,
+            530
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "and  $\\sigma_t(c)$  represents the numbers of the samples whose predictions fall into class  $c$  and above the threshold, formulated as",
+        "bbox": [
+            112,
+            533,
+            487,
+            580
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\sigma_ {t} (c) = \\sum_ {n = 1} ^ {N} \\mathbb {1} (\\max  \\hat {\\mathbf {p}} _ {b} > \\tau) \\cdot \\mathbb {1} (\\arg \\max  \\hat {\\mathbf {p}} _ {b} = c). \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            588,
+            487,
+            644
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.3 Mixup Supervised Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            652,
+            487,
+            668
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We target learning robust relation representation in the presence of remaining label noise. In particular, we adopt the contrastive learning approach and randomly sample  $N$  sentences and inference the sentences twice with same dropout rate to get two view. Then we normalize the embedding by  $L_{2}$  normalization. The resulting minibatch  $\\{z_i,y_i\\}_{i = 1}^{2N}$  consists of  $2N$  normalized sentence embedding and corresponding labels. We perform supervised contrastive learning on labeled samples",
+        "bbox": [
+            112,
+            671,
+            487,
+            832
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {i} \\left(z _ {i}, y _ {i}\\right) = \\frac {1}{2 N _ {y _ {i}} - 1} \\sum_ {j = 1} ^ {2 N} \\mathbb {1} _ {i \\neq j} \\mathbb {1} _ {y _ {i} = y _ {j}} \\tag {9} \\\\ \\cdot \\log \\frac {\\exp (z _ {i} \\cdot z _ {j} / \\tau)}{\\sum_ {r = 1} ^ {2 N} \\mathbb {1} _ {r \\neq i} \\exp (z _ {i} \\cdot z _ {r} / \\tau)}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            152,
+            839,
+            485,
+            922
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To make representation learning robust, we add Mixup (Berthelot et al., 2019) to supervised contrastive learning. Mixup strategies have demonstrated excellent performance in classification frameworks and have further shown promising results to prevent label noise memorization. Inspired by this success, we propose mixup supervised contrastive learning, a novel adaptation of mixup data augmentation for supervised contrastive learning. Mixup performs convex combinations of pairs of samples as",
+        "bbox": [
+            507,
+            84,
+            884,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx _ {i} = \\lambda x _ {a} + (1 - \\lambda) x _ {b}, \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            606,
+            273,
+            882,
+            288
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda \\in [0,1] \\sim \\text{Beta}(\\alpha_m, \\alpha_m)$  and  $x_i$  denotes the training example that combines two mini-batch examples  $x_a$  and  $x_b$ . A linear relation in the contrastive loss is imposed as",
+        "bbox": [
+            507,
+            300,
+            882,
+            365
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {i} ^ {M I X} = \\lambda \\mathcal {L} _ {a} \\left(z _ {i}\\right) + (1 - \\lambda) \\mathcal {L} _ {b} \\left(z _ {i}\\right), \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            546,
+            374,
+            882,
+            393
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_a$  and  $\\mathcal{L}_b$  have the same form as  $\\mathcal{L}_i$  in Eq. (9).",
+        "bbox": [
+            507,
+            404,
+            880,
+            434
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The supervised loss is the sum of Eq. (11) for each mixed instance:",
+        "bbox": [
+            507,
+            436,
+            880,
+            466
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {i} ^ {2 N} \\mathcal {L} _ {i} ^ {M I X}. \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            628,
+            475,
+            882,
+            517
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Mixup supervised contrastive learning helps to learn a robust representation for relations, but it cannot map the representation to a class. To learn the map function from the learned representation to relation class, classification learning using cross entropy loss is also employed as",
+        "bbox": [
+            507,
+            527,
+            882,
+            624
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} ^ {C L S} = \\sum_ {\\left(x _ {i}, y _ {i}\\right) \\in \\mathcal {X}} \\mathcal {L} _ {i} ^ {c l s} \\left(x _ {i}\\right) = \\sum_ {\\left(x _ {i}, y _ {i}\\right) \\in \\mathcal {X}} \\ell \\left(\\left(\\phi \\circ \\theta\\right) \\left(x _ {i}\\right), y _ {i}\\right). \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            510,
+            632,
+            882,
+            675
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Training Objective",
+        "text_level": 1,
+        "bbox": [
+            507,
+            684,
+            705,
+            700
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Combining the above analyses, the total objective loss is",
+        "bbox": [
+            507,
+            705,
+            880,
+            734
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda_ {u} \\mathcal {L} _ {u, t} + \\lambda_ {c} \\mathcal {L} ^ {C L S}. \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            583,
+            746,
+            882,
+            765
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            776,
+            655,
+            793
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            801,
+            621,
+            815
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our SSLRE framework on three DSRE datasets, including NYT10 (Riedel et al., 2010), NYT10m (Gao et al., 2021a), and wiki20m (Gao et al., 2021a).",
+        "bbox": [
+            507,
+            822,
+            882,
+            885
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "NYT10 dataset is created by aligning information from FreeBase with the New York Times (NYT)",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "13149",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "corpus. However, (1) it contains many duplicated instances in the dataset; (2) there is no public validation set; (3) The distantly supervised test set is quite noisy since the annotated errors. (Gao et al., 2021a) notes that  $53\\%$  samples of the NYT10 test set are wrongly labeled. We only use it to perform held-out evaluation experiment in Table 5 with some strong baselines.",
+        "bbox": [
+            112,
+            84,
+            487,
+            212
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "NYT10m build a manually annotated test set for NYT10 for more accurate evaluation. Furthermore, it removes all the duplicated instances and create a new relation ontology by merging semantically similar relations and delete the relations that only show up in train set or the test set.",
+        "bbox": [
+            112,
+            212,
+            487,
+            309
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "wiki20m is a processed version of Wiki20 (Han et al., 2020), which is constructed by aligning the English Wikipedia corpus with Wikidata (Vrandecic and Krötzsch, 2014). It has a manually annotated test set for evaluation.",
+        "bbox": [
+            112,
+            309,
+            487,
+            388
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Evaluation and Parameter Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            401,
+            435,
+            416
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To guarantee the fairness of evaluation. We take both sentence-level evaluation and bag-level evaluation in our experiments. Further details of the evaluation methods are available in the appendix C. To achieve bag-level evaluation under sentence-level training, we use at-least-one (ONE) aggregation stragy (Zeng et al., 2015), which first predicts relation scores for each sentence in the bag, and then takes the highest score for each relation. The details of the hyper-parameters are available in the appendix D.",
+        "bbox": [
+            112,
+            420,
+            487,
+            598
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            609,
+            233,
+            621
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In order to prove the effectiveness of the SSLRE, we compare our method with state-of-the-art methods sentence-level DSRE framework and bag-level DSRE framework.",
+        "bbox": [
+            112,
+            629,
+            487,
+            690
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For bag-level methods baselines, RESIDE (Vashisht et al., 2018) exploits the information of entity type and relation alias to add a soft limitation for DSRE. DISTRE (Alt et al., 2019) combines the selective attention to its Transformer-based model. CIL (Chen et al., 2021a) utilize contrastive instance learning under MIL framework, HiCLRE (Li et al., 2022) propose hierarchical contrastive learning framework, PARE (Rathore et al., 2022) propose concatenate all sentences in the bag to attend every token in the bag. Besides, we combine Bert with different aggregation strategies: ONE, which is mentioned in section 4.2; AVG averages the representations of all the sentences in the bag;",
+        "bbox": [
+            112,
+            694,
+            487,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ATT (Lin et al., 2016) produces bag-level representation as a weighted average over embeddings of sentences and determines weights by attention scores between sentences and relations.",
+        "bbox": [
+            507,
+            84,
+            882,
+            148
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For sentence-level baselines, RL-DSRE (Feng et al., 2018) apply reinforcement learning to train sample selector by feedback from the manually designed reward function. ARNOR (Jia et al., 2019) selects the reliable instances based on reward of attention score on the selected patterns. SENT (Ma et al., 2021) filters noisy instances based on negative training.",
+        "bbox": [
+            507,
+            151,
+            884,
+            279
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            297,
+            613,
+            311
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We first evaluate our SSLRE framework in the NYT10m and WIKI20m dataset. Table 1 shows the overall performance in terms of sentence-level evaluation. From the results, we can observe that (1) Our SSLRE framework demonstrates superior performance on both datasets, surpassing all strong baseline models significantly in terms of F1 score. In comparison to the most robust baseline models in two distantly supervised datasets, SSLRE displays a significant enhancement in performance (i.e.,  $+6.3\\%$  F1 and  $+3.4\\%$  F1). (2) The current sentence-level DSRE models (SENT, ARNOR) fail to outperform the state-of-the-art MIL-based techniques in terms of F1 score on the aforementioned datasets. This could be attributed to the loss of information resulting from the elimination of samples. Unlike the MIL-based methods that employ all samples for training, these models only utilize selected samples. Moreover, the selection procedure may not always be reliable. (3) The performance of state-of-the-art MIL-based methods is not substantially superior to that of the Bert baseline. This suggests that the MIL modules, which are specifically crafted for this task, do not exhibit significant effectiveness when evaluated at the sentence level.",
+        "bbox": [
+            507,
+            321,
+            884,
+            738
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2 presents the results of bag-level evaluation of SSLRE with ONE strategy on NYT10m and WIKI20m datasets. We compared our SSLRE framework with state-of-the-art methods for bag-level relation extraction, and found that our approach outperformed all strong baselines. Specifically, SSLRE achieved a  $5.5\\%$  improvement in AUC compared to the best baseline model PARE, and a  $10.1\\%$  improvement in micro F1 score compared with best baseline model HICLRE on the NYT10m dataset. Despite being trained without a",
+        "bbox": [
+            507,
+            741,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "13150",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/ef195fcc9bb32aa5a8f1a41531db20909bee3044146a8db1e5c37db79099b993.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"3\">NYT10m</td><td colspan=\"3\">wiki20m</td></tr><tr><td>μPrec.</td><td>μRec.</td><td>μF1</td><td>μPrec.</td><td>μRec.</td><td>μF1</td></tr><tr><td>Bert-ATT</td><td>49.1</td><td>58.2</td><td>53.3</td><td>71.3</td><td>77.8</td><td>74.4</td></tr><tr><td>Bert-ONE</td><td>51.0</td><td>60.7</td><td>55.5</td><td>72.8</td><td>73.2</td><td>73.0</td></tr><tr><td>Bert-AVG</td><td>52.0</td><td>55.8</td><td>53.9</td><td>80.7</td><td>76.9</td><td>78.7</td></tr><tr><td>RESIDE</td><td>45.5</td><td>48.7</td><td>47.0</td><td>69.9</td><td>72.3</td><td>70.9</td></tr><tr><td>DISTRE</td><td>55.1</td><td>51.0</td><td>52.9</td><td>80.3</td><td>73.6</td><td>76.8</td></tr><tr><td>CIL</td><td>58.1</td><td>49.3</td><td>53.4</td><td>81.8</td><td>73.1</td><td>77.0</td></tr><tr><td>HiCLRE</td><td>54.6</td><td>60.7</td><td>57.5</td><td>81.6</td><td>74.6</td><td>77.9</td></tr><tr><td>PARE</td><td>53.1</td><td>58.1</td><td>55.4</td><td>77.3</td><td>74.2</td><td>75.7</td></tr><tr><td>RL-DSRE</td><td>48.6</td><td>57.1</td><td>52.5</td><td>69.9</td><td>74.3</td><td>72.0</td></tr><tr><td>ARNOR</td><td>53.2</td><td>59.1</td><td>55.9</td><td>76.8</td><td>76.9</td><td>76.8</td></tr><tr><td>SENT</td><td>57.2</td><td>56.3</td><td>56.7</td><td>79.8</td><td>78.2</td><td>78.9</td></tr><tr><td>SSLRE</td><td>57.4</td><td>72.0</td><td>63.8</td><td>81.9</td><td>81.0</td><td>81.5</td></tr></table>",
+        "bbox": [
+            260,
+            80,
+            736,
+            342
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/c8ac870f293ece271811214c90a59d38eaadc8c7941a947d82ac37bc16eacb76.jpg",
+        "table_caption": [
+            "Table 1: Sentence-level evaluation results on NYT10m and wiki20m. Bold and underline indicate the best and the second best scores."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"3\">NYT10m</td><td colspan=\"3\">wiki20m</td></tr><tr><td>AUC</td><td>μF1</td><td>Macro_F1</td><td>AUC</td><td>μF1</td><td>Macro_F1</td></tr><tr><td>Bert-ATT</td><td>49.5</td><td>52.9</td><td>24.5</td><td>88.0</td><td>80.9</td><td>80.7</td></tr><tr><td>Bert-ONE</td><td>56.7</td><td>54.1</td><td>35.7</td><td>89.9</td><td>81.3</td><td>82.0</td></tr><tr><td>Bert-AVG</td><td>57.1</td><td>56.2</td><td>33.9</td><td>88.9</td><td>82.6</td><td>81.1</td></tr><tr><td>RESIDE</td><td>36.8</td><td>44.2</td><td>11.2</td><td>80.5</td><td>75.1</td><td>74.2</td></tr><tr><td>HiCLRE</td><td>61.0</td><td>61.2</td><td>32.0</td><td>89.1</td><td>82.3</td><td>81.1</td></tr><tr><td>CIL</td><td>57.4</td><td>59.6</td><td>29.4</td><td>89.3</td><td>81.8</td><td>82.4</td></tr><tr><td>PARE</td><td>61.1</td><td>59.8</td><td>37.2</td><td>90.3</td><td>83.2</td><td>82.6</td></tr><tr><td>SSLRE-ONE</td><td>66.5</td><td>71.3</td><td>36.9</td><td>91.6</td><td>83.3</td><td>84.1</td></tr></table>",
+        "bbox": [
+            164,
+            395,
+            835,
+            586
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Bag-level evaluation results on nyt10m and wiki20m. SSLRE-ONE represents the SSLRE with ONE aggregation strategy",
+        "bbox": [
+            112,
+            595,
+            882,
+            624
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "MIL framework, our SSLRE framework achieves state-of-the-art performance on bag-level relation extraction. This finding suggests that sentence-level training can also yield excellent results on bag-label prediction. This observation is also consistent with (Gao et al., 2021a; Amin et al., 2020).On the wiki20m dataset, we note a consistent improvement on as well, although it is not as evident as in the case of NYT10m. We surmise that this could be attributed to the fact that the wiki20m dataset is relatively less noisy when compared to NYT10m.",
+        "bbox": [
+            112,
+            649,
+            487,
+            826
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We also compared our framework to several strong baselines using held-out evaluation on the NYT10 dataset, which is detailed in appendix B.",
+        "bbox": [
+            112,
+            871,
+            487,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.5 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            650,
+            673,
+            665
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conducted ablation study experiments on the NYT10m dataset to assess the effectiveness of different modules in our SSLRE framework. We specifically removed each of the argued contributions one by one to evaluate their effectiveness. For unsupervised learning part, we remove the pseudo classifier head and CPL one at a time. For supervised learning part, we switch from mixup supervised learning to supervised contrastive learning and cross entropy as our new objective. In terms of confident samples identifications methods, we alternate between using random (randomly selection) and NLI-based selection instead of our K-NN method. The NLI method involves performing zero-shot relation extraction through Natural Lan",
+        "bbox": [
+            507,
+            677,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "13151",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "guage Inference (NLI)(Sainz et al., 2021), then identify the confident samples based on the level of agreement between the distant label and NLI soft label.",
+        "bbox": [
+            112,
+            84,
+            489,
+            147
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/050094c5e5ca8189b2c099082b28595d7d537911239dfb6ab5f2f246ea4b3616.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Methods</td><td>Prec.</td><td>Rec.</td><td>F1</td></tr><tr><td colspan=\"2\">SSLRE</td><td>57.4</td><td>72.0</td><td>63.8</td></tr><tr><td rowspan=\"2\">Unsupervised Learning</td><td>w/o pseudo head</td><td>56.2</td><td>70.6</td><td>62.5</td></tr><tr><td>w/o CPL</td><td>50.8</td><td>69.6</td><td>58.7</td></tr><tr><td rowspan=\"2\">Supervised Learning</td><td>SupCon</td><td>55.4</td><td>67.5</td><td>60.9</td></tr><tr><td>CE</td><td>51.8</td><td>70.3</td><td>59.6</td></tr><tr><td rowspan=\"2\">Conf-Samples Identification</td><td>random</td><td>58.6</td><td>56.0</td><td>57.3</td></tr><tr><td>NLI</td><td>56.4</td><td>65.3</td><td>60.5</td></tr></table>",
+        "bbox": [
+            115,
+            158,
+            510,
+            338
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table. 3 shows the ablation study results. We conclude that (1) Unsupervised learning apart effectively utilize the unlabeled samples. Removing pseudo classifier head and CPL leads to a decrease of  $1.3\\%$  and  $5.1\\%$  on micro-F1, respectively. (2) When dealing with noisy labeled data in supervised learning, Mixup Contrastive Supervised Learning proves to be more robust than both Supervised Contrastive Learning  $(-2.9\\%)$  and Cross Entropy  $(-4.2\\%)$ . (3) Our K-NN-based confident samples identification method outperforms the random method by  $6.5\\%$  and the NLI method by  $3.3\\%$ . This indicates that our K-NN method can effectively select confident samples.",
+        "bbox": [
+            112,
+            384,
+            490,
+            609
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6 Analysis on KNN",
+        "text_level": 1,
+        "bbox": [
+            112,
+            623,
+            295,
+            638
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conducted an evaluation of the performance of weighted k-nearest neighbors (kNN) in terms of its ability to select confident samples. To elaborate, we intentionally corrupted the labels of instances in the nyt10m test set with a random probability of  $20\\%$ ,  $40\\%$ , and  $60\\%$ . Our objective was to assess whether our weighted kNN method could effectively identify the uncorrupted (confident) instances. We trained our model on the corrupted nyt10m test set for 10 epochs, considering its relatively smaller size compared to the training set, which required more epochs to converge. In order to evaluate the ability of the weighted kNN in identifying confident samples, we reported the recall and precision metrics. The results are shown as 4:",
+        "bbox": [
+            112,
+            644,
+            489,
+            885
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "It is worth noting that precision is the more important metric because our goal is to make the la",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/5da5dbbcaa99de161c3f919fc95996c475a8f092e90c0171a182b7bce86671cf.jpg",
+        "table_caption": [
+            "Table 3: Ablation study of SSLRE on NYT10m"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Pre.</td><td>Rec.</td></tr><tr><td>20%</td><td>0.9851</td><td>0.8970</td></tr><tr><td>30%</td><td>0.9732</td><td>0.8758</td></tr><tr><td>40%</td><td>0.9531</td><td>0.8723</td></tr><tr><td>50%</td><td>0.9254</td><td>0.8655</td></tr><tr><td>60%</td><td>0.8735</td><td>0.8172</td></tr></table>",
+        "bbox": [
+            596,
+            82,
+            796,
+            195
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: The effect of KNN.",
+        "bbox": [
+            596,
+            204,
+            793,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "beled data as clean as possible after selecting confident samples. Even with a noisy label rate of  $60\\%$ , our weighted kNN method still achieves a precision of  $87.35\\%$  for the identified confident samples. This indicates that only  $12.65\\%$  of the labeled data is noisy, which is significantly smaller than the  $60\\%$  noise rate. Additionally, our weighted kNN method demonstrates a recall of over  $80\\%$  for confident samples. Although a few clean samples may not be selected, they can still be utilized through pseudo-labeling techniques.",
+        "bbox": [
+            507,
+            249,
+            885,
+            426
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.7 Analysis on Dropout rate",
+        "text_level": 1,
+        "bbox": [
+            507,
+            444,
+            754,
+            460
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 shows the performance of SSLRE under different dropout rate for strong augmentation. we can observe that: (1) Increasing the dropout rate appropriately improves the model's performance, with SSLRE achieving the best result (63.8 on F1) when the strong augmentation dropout rate is set to 0.4. (2) Augmentation using a very high dropout rate harms the performance, as it results in a loss of a significant amount of information. However, this does not significantly degrade the performance since we only use strong augmentation to predict pseudo-labels, which only affects  $\\mathcal{L}_u$ .",
+        "bbox": [
+            505,
+            469,
+            882,
+            663
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5e2020bfef393a62a9c4c224524f58a374e2d81abbaa8a70041a84603a717d80.jpg",
+        "image_caption": [
+            "Figure 3: Strong augmentation with different dropout rate"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            544,
+            695,
+            835,
+            848
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "13152",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/541d54363fbb9e6436cbf6a01bd13b47558a76a192fdd9615d7b903a17a55f86.jpg",
+        "image_caption": [
+            "(a) Sup"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            144,
+            95,
+            278,
+            167
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/38b6b7280c8481f2508f8a7b80b1fded57bc9ec9c951158b99bede98ab498c8e.jpg",
+        "image_caption": [
+            "(b) SSL",
+            "Figure 4: t-SNE visualization of representations with Pseudo labeling(SSL) and without(Sup). SSL achieves a better cluster results comparing with Sup."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            327,
+            95,
+            463,
+            167
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4.8 t-SNE analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            279,
+            275,
+            294
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "To demonstrate that preserving unlabeled samples can facilitate the learning of a superior representation compared to discarding them, we utilized sentence representations obtained from theta as the input to conduct dimension reduction via t-SNE and acquire two-dimensional representations. We focused on four primary categories of relation classes, which are \"/location/location/contains\", \"/business/person/company\", \"/location/administrative_division/country\", and \"/people/person/nationality\". As depicted in Figure 4, leveraging unlabeled samples via Pseudo labeling enhances the clustering of identical relation data points and effectively separates distinct classes from one another. Appendix F shows the t-SNE results of all classes.",
+        "bbox": [
+            112,
+            302,
+            489,
+            543
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            558,
+            245,
+            571
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose SSLRE, a novel sentence-level framework that is grounded in Semi-Supervised Learning for the DSRE task. Our SSLRE framework employs mixup supervised contrastive learning to tackle the noise present in selected samples, and it leverages unlabeled samples through Pseudo Labeling, which effectively utilize the information contained within noisy samples. Experimental results demonstrate that SSLRE outperforms strong baseline methods in both sentence-level and bag-level relation extraction.",
+        "bbox": [
+            112,
+            586,
+            489,
+            763
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            778,
+            218,
+            793
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In order to augment textual instances, we leverage dropout during forward propagation. This necessitates propagating each instance twice to generate the augmented sentence embeddings. However, the demand for GPU resources is higher compared to previous methods. Furthermore, we adjust the dropout rate to regulate the augmentation intensity",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "for semi-supervised learning and show its effectiveness through the performance results. Nonetheless, we have not conducted explicit experiments to investigate the interpretability, which needs further investigation.",
+        "bbox": [
+            507,
+            84,
+            884,
+            165
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgment",
+        "text_level": 1,
+        "bbox": [
+            509,
+            177,
+            665,
+            193
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is sponsored by National Natural Science Foundation of China (NSFC) (62206291, 62141608, and 62236010).",
+        "bbox": [
+            507,
+            203,
+            885,
+            250
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            279,
+            608,
+            293
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Christoph Alt, Marc Hubner, and Leonhard Hennig. 2019. Fine-tuning pre-trained transformer language models to distantly supervised relation extraction. ArXiv, abs/1906.08646.",
+            "Saadullah Amin, Katherine Ann Dunfield, Anna Vechkaeva, and Guenter Neumann. 2020. A data-driven approach for noise reduction in distantly supervised biomedical relation extraction. In Proceedings of the 19th SIGBioMed Workshop on Biomedical Language Processing, pages 187-194, Online. Association for Computational Linguistics.",
+            "Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. 2017. A closer look at memorization in deep networks. In International conference on machine learning, pages 233-242. PMLR.",
+            "Philip Bachman, Ouais Alsharif, and Doina Precup. 2014. Learning with pseudo-ensembles. ArXiv, abs/1412.4864.",
+            "Dara Bahri, Heinrich Jiang, and Maya Gupta. 2020. Deep k-nn for noisy labels. In International Conference on Machine Learning, pages 540-550. PMLR.",
+            "David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. 2019. Mixmatch: A holistic approach to semi-supervised learning. *ArXiv*, abs/1905.02249.",
+            "Jhih-Wei Chen, Tsu-Jui Fu, Chen-Kang Lee, and WeiYun Ma. 2020. H-fnd: Hierarchical false-negative denoising for distant supervision relation extraction. In Findings.",
+            "Tao Chen, Haizhou Shi, Siliang Tang, Zhigang Chen, Fei Wu, and Yue Ting Zhuang. 2021a. Cil: Contrastive instance learning framework for distantly supervised relation extraction. ArXiv, abs/2106.10855.",
+            "Tao Chen, Haochen Shi, Liyuan Liu, Siliang Tang, Jian Shao, Zhigang Chen, and Yueting Zhuang. 2021b. Empower distantly supervised relation extraction with collaborative adversarial training. In AAAI Conference on Artificial Intelligence."
+        ],
+        "bbox": [
+            509,
+            302,
+            885,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "13153",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xinlei Chen and Abhinav Kumar Gupta. 2015. Webly supervised learning of convolutional networks. 2015 IEEE International Conference on Computer Vision (ICCV), pages 1431-1439.",
+            "Qin Dai, Benjamin Heinzerling, and Kentaro Inui. 2022. Cross-stitching text and knowledge graph encoders for distantly supervised relation extraction. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6947-6958, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Jun Feng, Minlie Huang, Li Zhao, Yang Yang, and Xiaoyan Zhu. 2018. Reinforcement learning for relation classification from noisy data. In Proceedings of the aai conference on artificial intelligence, volume 32.",
+            "Tianyu Gao, Xu Han, Keyue Qiu, Yuzhuo Bai, Zhiyu Xie, Yankai Lin, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. 2021a. Manual evaluation matters: Reviewing test protocols of distantly supervised relation extraction. In Findings.",
+            "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021b. SimCSE: Simple contrastive learning of sentence embeddings. In Empirical Methods in Natural Language Processing (EMNLP).",
+            "Aritra Ghosh, Himanshu Kumar, and P. Shanti Sastry. 2017. Robust loss functions under label noise for deep neural networks. In AAAI Conference on Artificial Intelligence.",
+            "Jacob Goldberger and Ehud Ben-Reuven. 2016. Training deep neural-networks using a noise adaptation layer. In International Conference on Learning Representations.",
+            "Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Wai-Hung Tsang, and Masashi Sugiyama. 2018a. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In Neural Information Processing Systems.",
+            "Jiangfan Han, Ping Luo, and Xiaogang Wang. 2019. Deep self-learning from noisy labels. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5137-5146.",
+            "Xu Han, Tianyu Gao, Yankai Lin, Hao Peng, Yaoliang Yang, Chaojun Xiao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. 2020. More data, more relations, more context and more openness: A review and outlook for relation extraction. In AACL."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xu Han, Zhiyuan Liu, and Maosong Sun. 2018b. Neural knowledge acquisition via mutual attention between knowledge graph and text. In AAAI Conference on Artificial Intelligence.",
+            "Kailong Hao, Botao Yu, and Wei Hu. 2021. Knowing false negatives: An adversarial training method for distantly supervised relation extraction. In Conference on Empirical Methods in Natural Language Processing.",
+            "Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. 2022. Learning with neighbor consistency for noisy labels. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4662-4671.",
+            "Wei Jia, Dai Dai, Xinyan Xiao, and Hua Wu. 2019. Arnor: Attention regularization based noise reduction for distant supervision relation classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1399-1408.",
+            "Zhanghan Ke, Daoye Wang, Qiong Yan, Jimmy S. J. Ren, and Rynson W. H. Lau. 2019. Dual student: Breaking the limits of the teacher in semi-supervised learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6727-6735.",
+            "Diederik P. Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. CoRR, abs/1412.6980.",
+            "Kimin Lee, Sukmin Yun, Kibok Lee, Honglak Lee, Bo Li, and Jinwoo Shin. 2019. Robust inference via generative classifiers for handling noisy labels. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 3763-3772. PMLR.",
+            "Dongyang Li, Taolin Zhang, Nan Hu, Chengyu Wang, and Xiaofeng He. 2022. Hic Ire: A hierarchical contrastive learning framework for distantly supervised relation extraction. In Findings.",
+            "Junnan Li, Richard Socher, and Steven C. H. Hoi. 2020. Dividemix: Learning with noisy labels as semi-supervised learning. ArXiv, abs/2002.07394.",
+            "Yankai Lin, Shiqi Shen, Zhiyuan Liu, Huanbo Luan, and Maosong Sun. 2016. Neural relation extraction with selective attention over instances. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2124-2133.",
+            "Ruri Liu, Shasha Mo, Jianwei Niu, and Shengda Fan. 2022. CETA: A consensus enhanced training approach for denoising in distantly supervised relation extraction. In Proceedings of the 29th International Conference on Computational Linguistics, pages 2247-2258, Gyeongju, Republic of Korea. International Committee on Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "13154",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. 2020. Early-learning regularization prevents memorization of noisy labels. ArXiv, abs/2007.00151.",
+            "Tongliang Liu and Dacheng Tao. 2014. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38:447-461.",
+            "Yang Liu and Hongyi Guo. 2019. Peer loss functions: Learning from noisy labels without knowing noise rates. ArXiv, abs/1910.03231.",
+            "Yueming Lyu and Ivor Wai-Hung Tsang. 2019. Curriculum loss: Robust learning and generalization against label corruption. ArXiv, abs/1905.10045.",
+            "Ruotian Ma, Tao Gui, Linyang Li, Qi Zhang, Xuan-Jing Huang, and Yaqian Zhou. 2021. Sent: Sentence-level distant relation extraction via negative training. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6201-6213.",
+            "Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1003-1011.",
+            "Pengda Qin, Weiran Xu, and William Yang Wang. 2018. Robust distant supervision relation extraction via deep reinforcement learning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2137-2147.",
+            "Vipul Rathore, Kartikeya Badola, Parag Singla, and Mausam. 2022. PARE: A simple and strong baseline for monolingual and multilingual distantly supervised relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 340-354, Dublin, Ireland. Association for Computational Linguistics.",
+            "Sebastian Riedel, Limin Yao, and Andrew McCallum. 2010. Modeling relations and their mentions without labeled text. In *Joint European Conference on Machine Learning and Knowledge Discovery in Databases*, pages 148-163. Springer.",
+            "Oscar Sainz, Oier Lopez de Lacalle, Gorka Labaka, Ander Barrena, and Eneko Agirre. 2021. Label verbalization and entailment for effective zero and few-shot relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1199-1212.",
+            "Yu-Ming Shang and Wei Wei. 2019. Are noisy sentences useless for distant supervised relation extraction? In AAAI Conference on Artificial Intelligence."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Tao Shen, Guodong Long, Tao Shen, Tianyi Zhou, Lina Yao, Huan Huo, and Jing Jiang. 2019. Self-attention enhanced selective gate with entity-aware embedding for distantly supervised relation extraction. ArXiv, abs/1911.11899.",
+            "Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin Dogus Cubuk, Alexey Kurakin, Han Zhang, and Colin Raffel. 2020. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. ArXiv, abs/2001.07685.",
+            "Antti Tarvainen and Harri Valpola. 2017. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NIPS.",
+            "Shikhar Vashishth, Rishabh Joshi, Sai Suman Prayaga, Chiranjib Bhattacharyya, and Partha Talukdar. 2018. RESIDE: Improving distantly-supervised neural relation extraction using side information. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1257-1266, Brussels, Belgium. Association for Computational Linguistics.",
+            "Denny Vrandecic and Markus Krötzsch. 2014. Wiki-data: A free collaborative knowledge base. Communications of the ACM, 57:78-85.",
+            "Xudong Wang, Zhi-Li Wu, Long Lian, and Stella X. Yu. 2022. Debiased learning from naturally imbalanced pseudo-labels. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14627-14637.",
+            "Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris N. Metaxas, and Chao Chen. 2020. A topological filter for learning with label noise. ArXiv, abs/2012.04835.",
+            "Yuwei Xia, Mengqi Zhang, Qiang Liu, Shu Wu, and Xiao-Yu Zhang. 2022. MetaTKG: Learning evolutionary meta-knowledge for temporal knowledge graph reasoning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7230-7240, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Zhiquan Ye and Zhenhua Ling. 2019. Distant supervision relation extraction with intra-bag and inter-bag attentions. ArXiv, abs/1904.00143.",
+            "Daojian Zeng, Kang Liu, Yubo Chen, and Jun Zhao. 2015. Distant supervision for relation extraction via piecewise convolutional neural networks. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 1753-1762.",
+            "Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. 2021. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. ArXiv, abs/2110.08263."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "13155",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Mengqi Zhang, Yuwei Xia, Qiang Liu, Shu Wu, and Liang Wang. 2023a. Learning latent relations for temporal knowledge graph reasoning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12617-12631, Toronto, Canada. Association for Computational Linguistics.",
+        "bbox": [
+            114,
+            85,
+            489,
+            178
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Mengqi Zhang, Yuwei Xia, Qiang Liu, Shu Wu, and Liang Wang. 2023b. Learning long- and short-term representations for temporal knowledge graph reasoning. In Proceedings of the ACM Web Conference 2023, WWW '23, page 2412-2422, New York, NY, USA. Association for Computing Machinery.",
+        "bbox": [
+            114,
+            187,
+            489,
+            268
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Algorithm",
+        "text_level": 1,
+        "bbox": [
+            114,
+            281,
+            243,
+            298
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1 provides the pseudo-code of the overall framework.",
+        "bbox": [
+            112,
+            307,
+            489,
+            338
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Held-out evaluation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            351,
+            322,
+            368
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/7a8b7c1905a667d11c18f4632242e1163cc65900aa4d006823446f637add0d28.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Models</td><td>Prec.</td><td>Rec.</td><td>F1</td></tr><tr><td>SENT</td><td>45.3</td><td>50.1</td><td>47.5</td></tr><tr><td>HiCLRE</td><td>44.3</td><td>50.6</td><td>47.2</td></tr><tr><td>SSLRE(Ours)</td><td>49.2</td><td>54.3</td><td>51.6</td></tr></table>",
+        "bbox": [
+            154,
+            387,
+            445,
+            475
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Held-out evaluation on NYT10",
+        "bbox": [
+            163,
+            485,
+            436,
+            499
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 5 shows the held-out evaluation results on NYT10 dataset.",
+        "bbox": [
+            112,
+            519,
+            487,
+            550
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Evaluation Settings",
+        "text_level": 1,
+        "bbox": [
+            114,
+            564,
+            321,
+            580
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Sentence-level evaluation: Different from bag-level evaluation used by MIL-based model, a sentence-level (or instance-level) evaluation accesses model performance directly on all of the individual instances in the dataset, which require the model to accurately predict the relation for each sentence. Following (Jia et al., 2019; Ma et al., 2021; Liu et al., 2022), we report micro-Precision  $(\\mu \\mathrm{Prec.})$ , micro-Recall  $(\\mu \\mathrm{Rec.})$  and microF1  $(\\mu \\mathrm{F1})$  for sentence-level evaluation.",
+        "bbox": [
+            112,
+            590,
+            489,
+            750
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Parameter Settings",
+        "text_level": 1,
+        "bbox": [
+            114,
+            764,
+            319,
+            781
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Bag-level evaluation: Bag evaluation accesses the performance of bag relation label extraction. Since manually annotated data are at the sentence-level, following (Gao et al., 2021a), we construct bag-level annotations in the following way: For each bag, if one sentence in the bag has a human-labeled relation, this bag is labeled with this relation; if no sentence in the bag is annotated with any relation,",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/5c3ebe4b63ac4922d423f49ab0d5a254d00dddaf4d5513ea2820c6496cfe9cca.jpg",
+        "image_caption": [
+            "Figure 5: PR-curve"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            536,
+            101,
+            858,
+            253
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "this bag is labeled as N/A. We report AUC, MicroF1 and Macro-F1 for bag level evaluation.",
+        "bbox": [
+            507,
+            313,
+            884,
+            346
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The underlying encoder for sentence are implemented by BERT-base (Devlin et al., 2019), which generates 768 hidden units for each token's context-aware representation. During the training stage, we set the learning rate of the model to  $2 \\times 10^{-5}$  and the batch size to 64, which was determined through a grid search over batch sizes in  $\\{16, 32, 64\\}$  and learning rates in  $\\{1e-5, 2e-5, 5e-5\\}$ . We train the model for 5 epochs and use Adam (Kingma and Ba, 2014) as the optimizer. The Mixup parameter  $\\alpha_{m}$  is set to 1, the classifier loss weight  $\\lambda_{c}$  is set to 0.2, the fractile alpha is set to 0.8, the unsupervised loss weight  $\\lambda_{u}$  is set to 1, and the CPL threshold  $\\tau$  is set to 0.95. We set the dropout rate for weak augmentation to 0.2 and the dropout rate for strong augmentation to 0.4. Further analysis on the strong augmentation dropout rate is presented in Section 4.7.",
+        "bbox": [
+            507,
+            347,
+            884,
+            634
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "E PR-curve",
+        "text_level": 1,
+        "bbox": [
+            509,
+            649,
+            630,
+            664
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We report the P-R curve on NYT10m dataset as Figure 5:",
+        "bbox": [
+            507,
+            676,
+            880,
+            708
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/12367125ff4214d54131c8e6e0d41a0628383e3d0d06356bfb984c7c98f87f95.jpg",
+        "image_caption": [
+            "(a) Sup"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            539,
+            741,
+            675,
+            816
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/9f09bc1cc074ae9ac9fb9e7e46befa4478d5c663ee676a4c0d56d43bb9d00be7.jpg",
+        "image_caption": [
+            "(b) SSL",
+            "Figure 6: t-SNE visualization of representations with Pseudo labeling(SSL) and without(Sup). SSL achieves a better cluster results comparing with Sup, especially on color green and light purple."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            724,
+            743,
+            858,
+            814
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "13156",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "F Additional t-SNE analysis",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            374,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Figure 6 shows the t-SNE results on all classes of the sentence representation.",
+        "bbox": [
+            112,
+            108,
+            489,
+            142
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1: SSLRE Algorithm"
+        ],
+        "code_body": "input: Noisy Dataset  $\\tilde{\\mathcal{D}}$  output:   \n1 Warm up  $\\theta$  and  $\\phi$  for one epoch using supervised contrastive learning and get  $\\theta^{\\prime}$    \n2 Gain features:  $z_{i} = \\theta^{\\prime}(x_{i})$    \n3 Build weighted K-NN graph using Eq. (1)   \n4 Gain  $\\mathcal{X}$  and  $\\mathcal{U}$  using Eq. (2-4)   \n5 Reinitialize  $\\phi$  and  $\\theta$    \n6 while not reach the maximum iteration do   \n7 for  $c\\gets 1$  to  $C$  do   \n8 Calculate  $\\mathcal{T}(c)$  using Eq. (7) {Determine the flexible threshold for class c.}   \n9 end   \n10 for  $t\\gets 1$  to num_iters do   \n11 From  $\\mathcal{X}$  , draw a mini-batch  $X^{t} = \\{(x_{b},y_{b});b\\in (1,\\dots ,B)\\}$    \n12 From  $\\mathcal{U}$  , draw a mini-batch  $U^{t} = \\{u_{b};b\\in (1,\\dots ,B)\\}$  /\\* Learning from labeled dataset X \\*/   \n13  $Z^{t} = \\theta_{w}(X^{t})\\cup \\theta_{w}(X^{t}) / /$  Get two augmented embedding views by dropout inference twice.   \n14 Calculate  $L^{CLS}$  using Eq. (13).   \n15 Mixup embedding using Eq. (10), and gain  $Z_{mix}^{t}$    \n16 Calculate  $L_{s}$  using Eq. (11) /\\* Learning from unlabeled dataset U \\*/   \n17 Generate pseudo labels using Eq. (5).   \n18 Calculate unsupervised loss using Eq. (6) with  $\\mathcal{T}(c)$    \n19 Calculate the overall loss  $L$  using Eq. (14).   \n20 end   \n21 end",
+        "bbox": [
+            115,
+            175,
+            485,
+            829
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "13157",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Noise-Robust Semi-Supervised Learning for Distantly Supervised Relation Extraction/fc076ec4-b157-42d2-a4a8-4199f3a5df0a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86d9740643c53d16fd7e713aa5d842efeea5ab58470e2a97e7523d9e825718f1
+size 1060001

2023/Noise-Robust Semi-Supervised Learning for Distantly Supervised Relation Extraction/full.md

@@ -0,0 +1,421 @@
+# Noise-Robust Semi-Supervised Learning for Distantly Supervised Relation Extraction
+
+Xin Sun $^{1,2}$ , Qiang Liu $^{2,3*}$ , Shu Wu $^{2,3}$ , Zilei Wang $^{1}$ , Liang Wang $^{2,3}$
+
+<sup>1</sup>University of Science and Technology of China
+
+$^{2}$ CRIPAC & MAIS, Institute of Automation, Chinese Academy of Sciences
+
+<sup>3</sup>University of Chinese Academy of Sciences
+
+sunxin000@mail.ustc.edu.cn, qiang.liu@nlpr ia.ac.cn
+
+# Abstract
+
+Distantly supervised relation extraction (DSRE) aims to extract relational facts from texts but suffers from noisy instances. To mitigate the influence of noisy labels, current methods typically use the Multi-Instance-Learning framework to extract relations for each bag. However, these approaches are not capable of extracting relation labels for individual sentences. Several studies have focused on sentence-level DSRE to solve the above problem. These studies primarily aim to develop methods for identifying noisy samples and filtering them out to mitigate the impact of noise. However, discarding noisy samples directly leads to the loss of useful information. To this end, we propose SSLRE, a novel Semi-Supervised-Learning Relation Extraction framework for sentence-level DSRE. We discard only the labels of the noisy samples and utilize these instances without labels as unlabeled samples. Our SSLRE framework utilizes a weighted K-NN graph to select confident samples as labeled data and the rest as unlabeled. We then design a robust semi-supervised learning framework that can efficiently handle remaining label noise present in the labeled dataset, while also making effective use of unlabeled samples. Based on our experiments on two real-world datasets, the SSLRE framework we proposed has achieved significant enhancements in sentence-level relation extraction performance compared to the existing state-of-the-art methods. Moreover, it has also attained a state-of-the-art level of performance in bag-level relation extraction with ONE aggregation strategy.
+
+# 1 Introduction
+
+Relation extraction (RE) is a fundamental process for constructing knowledge graphs, as it aims to predict the relationship between entities based on their context. However, most supervised RE
+
+![](images/487a4bbf26088f7c7dc661d596f7a5e371d0031f089d0ba355fbb643dfe7cbc3.jpg)  
+Figure 1: Bag-level RE maps a bag of sentences to bag labels. Sentence-level RE maps each sentence to a specific relation.
+
+techniques require extensive labeled training data, which can be difficult to obtain manually. To address this issue, Distant Supervision (DS) was proposed (Mintz et al., 2009) to automatically generate labeled text corpus by aligning plain texts with knowledge bases (KB). For instance, if a sentence contains both the subject  $(s)$  and object  $(o)$  of a triple  $\langle s,r,o\rangle$  ( $\langle \text{subject}, \text{relation}, \text{object} \rangle$ ), then the DS method considers  $\langle s,r,o\rangle$  as a valid sample for that sentence. Conversely, if no relational triples apply, then the sentence is labeled as "NA".
+
+Distantly supervised datasets usually face high label noise in training data due to the annotation process. To mitigate the impact of noisy labels caused by distant supervision, contemporary techniques (Lin et al., 2016; Alt et al., 2019; Chen et al., 2021b; Li et al., 2022; Dai et al., 2022) usually employ multi-instance-learning (MIL) framework or modify MIL to train the relation extraction model.
+
+Although MIL-based techniques can identify bag relation labels, they are not proficient in precisely mapping each sentence in a bag with explicit sentence labels (Feng et al., 2018; Jia et al., 2019; Ma et al., 2021). Several studies have focused on improving sentence-level DSRE and have empirically demonstrated the inadequacy of bag-level methods on sentence-level evaluation. (Feng et al., 2018; Qin et al., 2018) apply reinforcement learning to train a sample selector. (Jia et al., 2019) iden
+
+tify confident samples by frequent patterns. (Ma et al., 2021) utilizes negative training to separate noisy data from the training data.
+
+However, these methods have two main issues: (1) These works simply discard all noisy samples and train a relation extraction model with selected samples. However, filtering out noisy samples directly results in the loss of useful information. (Gao et al., 2021a) notes that the DSRE dataset has a noisereate exceeding  $50\%$ . Discarding all these samples would result in a significant loss of information. (2) The confident samples selection procedure is not impeccable, and there may still exist a small amount of noise in the chosen confident samples. Directly training a classifier in the presence of label noise is known to result in noise memorization.
+
+To address the two issues, this work proposes a novel semi-supervised-learning framework for sentence-level DSRE, First, we construct a K-NN graph for all samples using the hidden features. Then, we identify confident samples from the graph structure and consider the remaining samples as noisy. For issue (1): Our method discards only the noisy labels and treats corresponding samples as unlabeled data. We then utilize this unlabeled data by pseudo labeling within our robust semi-supervised learning framework to learn a better feature representation for relation. For issue (2): Despite our initial selection of confident samples, there may still be noise in the labeled dataset. we have developed a noise-robust semi-supervised learning framework that leverages mixup supervised contrastive learning to learn from the labeled dataset and curriculum pseudo labeling to learn from the unlabeled dataset.
+
+To summarize the contribution of this work:
+
+- We propose a noise-robust Semi-Supervised-Learning framework SSLRE for DSRE task, which effectively mitigate the impact of noisy labels.  
+- We propose to use graph structure information (weighted K-NN) to identify the confident samples and effectively convert noisy samples as useful training data by utilizing pseudo labeling.  
+- The proposed framework achieves significant improvement over previous methods in terms of both sent-level and bag-level relation extraction performance.
+
+# 2 Related work
+
+# 2.1 Distantly Supervised Relation Extraction
+
+Relation extraction (RE) is a fundamental pro- cess for constructing knowledge graphs(Zhang et al., 2023a; Xia et al., 2022; Zhang et al., 2023b). To generate large-scale auto-labeled data without human effort, (Mintz et al., 2009) first use Distant Supervision to label sentences mentioning two entities with their relations in KGs, which inevitably brings wrongly labeled instances. To tackle the predicament of noise, most of the existing studies on DSRE are founded on the multi-instance learning framework. This approach is leveraged to handle noisy sentences in each bag and train models by capitalizing on the constructed trustworthy baglevel representations. Usually, these methods employ attention mechanisms to assign less weights to the probable noisy sentences in the bag(Lin et al., 2016; Han et al., 2018b; Alt et al., 2019; Shen et al., 2019; Ye and Ling, 2019; Chen et al., 2021a; Li et al., 2022), apply adversarial training or reinforcement learning to remove the noisy sentences from the bag (Zeng et al., 2015; Qin et al., 2018; Shang and Wei, 2019; Chen et al., 2020; Hao et al., 2021). However, the studies (Feng et al., 2018; Jia et al., 2019; Ma et al., 2021; Gao et al., 2021a) indicate that the bag-level DSRE methods are ineffective for sentence-level relation extraction.
+
+This work focuses on extracting relations at the sentence level, (Feng et al., 2018) applied reinforcement learning to identify confident instances based on the reward of noisy labels. (Jia et al., 2019) involve building initial reliable sentences based on several manually defined frequent relation patterns. (Ma et al., 2021) assigning complementary labels that cooperate with negative training to filter out noisy instances. Unlike previous studies, our method only discard noisy labels and keep the unlabeled samples. We use pseudo labeling to effectively utilize unlabeled samples, which helps to learn better representation.
+
+# 2.2 Semi-Supervised-Learning
+
+In SSL, a portion of training dataset is labeled and the remaining portion is unlabeled. SSL has seen great progress in recent years. Since (Bachman et al., 2014) proposed a consistency regularization based method, many approaches have migrated it into the semi-supervised learning field. MixMatch (Berthelot et al., 2019) proposes to combine consistency regularization with entropy minimization.
+
+Mean Teacher (Tarvainen and Valpola, 2017) and Dual Student (Ke et al., 2019) are also based on consistent learning, aiming for the same outputs for different networks. Recently, FixMatch (Sohn et al., 2020) provides a simple yet effective weak-to-strong consistency regularization framework. Flexmatch (Zhang et al., 2021) provides curriculum pseudo Labels learning to combat the imbalance of pseudo labels.
+
+Our SSLRE framework differs from these frameworks in two main ways. Firstly, our labeled dataset still contains a small amount of noise due to the fact that confident sample identification cannot achieve perfection. Therefore, we utilize mixup contrastive supervised learning to combat this noise. Secondly, current SSL methods generate and utilize pseudo labels with the same head, which causes error accumulation during the training stage. To address this issue, we propose utilizing a pseudo classifier head, which decouples the generation and utilization of pseudo labels by two parameter-independent heads to avoid error accumulation.
+
+# 2.3 Learning with Noisy Data
+
+In both computer vision and natural language processing, learning with noisy data is a widely discussed problem. Existing approaches include but not limit to estimating the noise transition matrix (Chen and Gupta, 2015; Goldberger and Ben-Reuven, 2016), leveraging a robust loss function (Lyu and Tsang, 2019; Ghosh et al., 2017; Liu and Guo, 2019), introducing regularization (Liu et al., 2020; Iscen et al., 2022), selecting noisy samples by multi-network learning or multi-round learning (Han et al., 2018a; Wu et al., 2020), re-weighting examples (Liu and Tao, 2014), generating pseudo labels (Li et al., 2020; Han et al., 2019), and so on. In addition, some advanced state-of-the-art methods combine serveral techniques, e.g., Dividemix (Li et al., 2020) and  $\mathrm{ELR + }$  (Liu et al., 2020).
+
+In this paper, we address the issue of noisy labels in distant relation extraction. Our approach first involves constructing a K-NN graph to identify confident samples based on their graph structure, and then use noise-robust mixup supervised contrastive learning to train with the labeled samples.
+
+# 3 Methodology
+
+To achieve sentence-level relation extraction in DSRE, we propose a framework called SSLRE,
+
+which consists of two main steps. Firstly, we select confident samples from the distantly supervised dataset using a weighted K-NN approach built by all sample representations. We use the selected samples as labeled data and the remaining samples as unlabeled data (as detailed in Section 3.1). Secondly, we employ our robust Semi-Supervised Learning framework to learn from the Semi-Supervised datasets (as described in Section 3.2). Appendix A delineates the full algorithm.
+
+Specifically, we denote the original dataset in this task as  $\tilde{\mathcal{D}} = \{(s_1,\tilde{y}_1),(s_2,\tilde{y}_2),\dots ,(s_N,\tilde{y}_N)\}$  where  $\tilde{y}_i\in \{1,\dots ,C\}$  is the label of the ith input sentence  $s_i$  . The labeled dataset (identified confident samples) is denoted as  $\mathcal{X} = \{(x_{1},y_{1}),(x_{2},y_{2}),\dots ,(x_{n},y_{n})\}$  and the unlabeled dataset (the noisy samples without labels) is denoted as  $\mathcal{U} = \{(u_1,u_2,\dots ,u_m\}$  where  $m + n = N$
+
+# 3.1 Confident Samples Identification with Weighted K-NN
+
+Our Semi-Supervised-Learning module requires us to divide the noisy dataset into a labeled dataset and an unlabeled dataset. Inspired by (Lee et al., 2019; Bahri et al., 2020), we utilize neighborhood information of the hidden feature spaces to identify confident samples. We employ supervised contrastive learning to warm up our model and obtain the representations for all instances. It is noteworthy that deep neural networks tend to initially fit the training data with clean labels during an early training phase, prior to ultimately memorizing the examples with false labels (Arpit et al., 2017; Liu et al., 2020). Consequently, we only warm up our model for a single epoch. Given two sentences  $s_i$  and  $s_j$ , we can obtain their low-dimensional representations as  $z_i = \theta(s_i)$  and  $z_j = \theta(s_j)$ , where  $\theta$  is the sentence encoder. We then calculate their representation similarity using the cosine distance
+
+$$
+d \left(z _ {i}, z _ {j}\right) = \frac {z _ {i} z _ {j} ^ {T}}{\| z _ {i} \| \| z _ {j} \|.} \tag {1}
+$$
+
+Then, we build a weighted K-NN graph for all samples based on the consine distance. To quantify the agreement between  $s_i$  and  $\tilde{y}_i$ , We first use the label distribution in the  $K$ -neighborhood to approximate clean posterior probabilities,
+
+$$
+\hat {q} _ {c} \left(s _ {i}\right) = \frac {1}{\sum_ {s _ {k} \in \mathcal {N} _ {i}} d \left(z _ {i} , z _ {k}\right)} \sum_ {\boldsymbol {s} _ {k} \in \mathcal {N} _ {i}} d \left(z _ {i}, z _ {k}\right) \cdot \mathbb {1} \left(\tilde {y} _ {k} = c\right) \tag {2}
+$$
+
+![](images/043c5beee27f8eacbec8f6f0812c2decc0325b1f7dc16853e2793881e21d1da4.jpg)  
+Figure 2: An overview of the proposed framework, SSLRE.  $\hat{D},\mathcal{X},\mathcal{U}$  denote the original noisy dataset, labeled dataset and unlabeled dataset.  $\theta$  indicates the encoder.  $\theta_w$  and  $\theta_s$  mean forward with lower and higher dropout rate, respectively.  $\phi$  and  $\psi$  are classifier head and pseudo classifier head.  $\mathcal{L}_s$  is the mixup supervised contrastive loss defined in Eq. (12), and  $\mathcal{L}_{u,t}$  is the unsupervised loss defined in Eq. (6).
+
+where  $\mathcal{N}_i$  represents the set of  $K$  closest neighbors to  $s_i$ . We then use the cross-entropy loss  $\ell$  to calculate the disagreement between  $\hat{\mathbf{q}}_c(\boldsymbol{s}_i)$  and  $\tilde{y}_i$ . Denoting the set of confident examples belonging to the  $c$ -th class as  $\mathcal{X}_c$ , we have
+
+$$
+\mathcal {X} _ {c} = \left\{s _ {i}, \tilde {y} _ {i} \mid \ell \left(\hat {\mathbf {q}} \left(s _ {i}\right), \tilde {y} _ {i}\right) <   \gamma_ {c}, \tilde {y} _ {i} = c \right\}, c \in [ C ], \tag {3}
+$$
+
+where  $\gamma_{c}$  is a threshold for the  $c$ -th class, which is dynamically defined to ensure a class-balanced set of identified confident examples. To achieve this goal, we use the  $\alpha$  fractile of per-class agreements between the original label  $\tilde{y}_i$  and  $\max_c\hat{\mathbf{q}}_c(\boldsymbol{s}_i)$  across all classes to determine how many examples should be selected for each class, i.e.  $\sum_{i=1}^{n} = \mathbb{1}(\max_{c'}\hat{\mathbf{q}}_{c'}(\boldsymbol{s}_i) = \tilde{y}_i) \cdot (\tilde{y}_i = c), c \in [C]$ . Finally, we can get the labeled set and unlabeled set as
+
+$$
+\mathcal {X} = \cup_ {c = 1} ^ {C} \mathcal {X} _ {c} \tag {4}
+$$
+
+$$
+\mathcal {U} = \left\{s _ {i} \mid \left(s _ {i}, \tilde {y} _ {i}\right) \in \tilde {\mathcal {D}} \backslash \mathcal {X} \right\}.
+$$
+
+# 3.2 Noise-Robust Semi-Supervised learning
+
+Despite selecting confident samples from the distantly supervised dataset, there still remains a small amount of noise in the labeled dataset. Naively training a classifier in the presence of label noise leads to noise memorization (Liu et al., 2020), which degrades the performance. We propose a noise-robust semi-supervised learning framework to mitigate the influence of remaining noise.
+
+# 3.2.1 Data Augmentation with Dropout
+
+Inspired by SimCSE (Gao et al., 2021b), we augment training samples by embedding processing. In particular, We obtain different embeddings of a
+
+sentence by applying dropout during the forward process. Additionally, we propose using a high dropout rate for strong augmentation and a low dropout rate for weak augmentation. The sentence encoder is denoted as  $\theta$ , with forward propagation using a high dropout rate denoted as  $\theta_{s}$  and forward propagation using a low dropout rate denoted as  $\theta_{w}$ .
+
+# 3.2.2 Unsupervised Learning with Pseudo Labeling
+
+In this part, we propose two modules to learn from the unlabeled dataset: (1) To generate and utilize pseudo labels independently, we propose pseudo classifier head. (2) Utilize Curriculum Pseudo Labeling to perform consistent learning while combatting the unbalance of the generated pseudo labels.
+
+Pseudo classifier head: pseudo labeling is one of the prevalent techniques in semi-supervised learning. The existing approaches generate and utilize pseudo labels with the same head. However, this may cause training bias, ultimately amplifying the model's errors as self-training continues. (Wang et al., 2022). To reduce this bias when using pseudo labels, we propose utilizing a two-classifier model consisting of an encoder  $\theta$  with both a classifier head  $\phi$  and a pseudo classifier head  $\psi$ . We optimize the classifier head  $\phi$  using only labeled samples and without any unreliable pseudo labels from unlabeled samples. Unlabeled samples are used solely for updating encoder  $\theta$  and pseudo classifier head  $\psi$ . In particular, the classifier head  $\phi$  generates pseudo labels  $(\phi \circ \theta_w)(u_b)$  for unlabeled samples (which have
+
+no gradient), the loss of unlabeled samples is calculated by  $\ell ((\psi \circ \theta_s)(u_b),(\phi \circ \theta_w)(u_b))$  , where  $\ell$  denotes cross entropy loss. This decouples the generation and utilization of pseudo labels by two parameter-independent heads to mitigate error accumulation.
+
+Curriculum Pseudo Labeling: Due to the highly unbalanced dataset, using a constant cut-off  $\tau$  for all classes in Pseudo labeling results in almost all selected samples (those with confidence greater than the cut-off) being labeled as 'NA', which is the dominant class.
+
+Inspired by Flexmatch (Zhang et al., 2021), we use Curriculum Pseudo Labeling (CPL) to combat unbalanced pseudo labels. We first generate pseudo labels for iteration  $t$
+
+$$
+\hat {\mathbf {p}} _ {b} = \left(\phi_ {t} \circ \theta_ {w, t}\right) \left(u _ {b}\right), \tag {5}
+$$
+
+These labels are then used as the target of strongly-augmented data. The unsupervised loss term has the form as
+
+$$
+\begin{array}{l} \mathcal {L} _ {u, t} = \frac {1}{\mu B} \sum_ {b = 1} ^ {\mu B} \mathbb {1} (\max  (\hat {\mathbf {p}} _ {b}) > \mathcal {T} _ {t} (\arg \max  (\hat {\mathbf {p}} _ {b}))) \tag {6} \\ \cdot \ell \left(\left(\psi_ {t} \circ \theta_ {s, t}\right) (u _ {b}), \hat {\mathbf {p}} _ {b}\right), \\ \end{array}
+$$
+
+where
+
+$$
+\mathcal {T} _ {t} (c) = \frac {\sigma_ {t} (c)}{\max  _ {c} \sigma_ {t}} \tau , \tag {7}
+$$
+
+and  $\sigma_t(c)$  represents the numbers of the samples whose predictions fall into class  $c$  and above the threshold, formulated as
+
+$$
+\sigma_ {t} (c) = \sum_ {n = 1} ^ {N} \mathbb {1} (\max  \hat {\mathbf {p}} _ {b} > \tau) \cdot \mathbb {1} (\arg \max  \hat {\mathbf {p}} _ {b} = c). \tag {8}
+$$
+
+# 3.2.3 Mixup Supervised Contrastive Learning
+
+We target learning robust relation representation in the presence of remaining label noise. In particular, we adopt the contrastive learning approach and randomly sample  $N$  sentences and inference the sentences twice with same dropout rate to get two view. Then we normalize the embedding by  $L_{2}$  normalization. The resulting minibatch  $\{z_i,y_i\}_{i = 1}^{2N}$  consists of  $2N$  normalized sentence embedding and corresponding labels. We perform supervised contrastive learning on labeled samples
+
+$$
+\begin{array}{l} \mathcal {L} _ {i} \left(z _ {i}, y _ {i}\right) = \frac {1}{2 N _ {y _ {i}} - 1} \sum_ {j = 1} ^ {2 N} \mathbb {1} _ {i \neq j} \mathbb {1} _ {y _ {i} = y _ {j}} \tag {9} \\ \cdot \log \frac {\exp (z _ {i} \cdot z _ {j} / \tau)}{\sum_ {r = 1} ^ {2 N} \mathbb {1} _ {r \neq i} \exp (z _ {i} \cdot z _ {r} / \tau)}. \\ \end{array}
+$$
+
+To make representation learning robust, we add Mixup (Berthelot et al., 2019) to supervised contrastive learning. Mixup strategies have demonstrated excellent performance in classification frameworks and have further shown promising results to prevent label noise memorization. Inspired by this success, we propose mixup supervised contrastive learning, a novel adaptation of mixup data augmentation for supervised contrastive learning. Mixup performs convex combinations of pairs of samples as
+
+$$
+x _ {i} = \lambda x _ {a} + (1 - \lambda) x _ {b}, \tag {10}
+$$
+
+where  $\lambda \in [0,1] \sim \text{Beta}(\alpha_m, \alpha_m)$  and  $x_i$  denotes the training example that combines two mini-batch examples  $x_a$  and  $x_b$ . A linear relation in the contrastive loss is imposed as
+
+$$
+\mathcal {L} _ {i} ^ {M I X} = \lambda \mathcal {L} _ {a} \left(z _ {i}\right) + (1 - \lambda) \mathcal {L} _ {b} \left(z _ {i}\right), \tag {11}
+$$
+
+where  $\mathcal{L}_a$  and  $\mathcal{L}_b$  have the same form as  $\mathcal{L}_i$  in Eq. (9).
+
+The supervised loss is the sum of Eq. (11) for each mixed instance:
+
+$$
+\mathcal {L} _ {s} = \sum_ {i} ^ {2 N} \mathcal {L} _ {i} ^ {M I X}. \tag {12}
+$$
+
+Mixup supervised contrastive learning helps to learn a robust representation for relations, but it cannot map the representation to a class. To learn the map function from the learned representation to relation class, classification learning using cross entropy loss is also employed as
+
+$$
+\mathcal {L} ^ {C L S} = \sum_ {\left(x _ {i}, y _ {i}\right) \in \mathcal {X}} \mathcal {L} _ {i} ^ {c l s} \left(x _ {i}\right) = \sum_ {\left(x _ {i}, y _ {i}\right) \in \mathcal {X}} \ell \left(\left(\phi \circ \theta\right) \left(x _ {i}\right), y _ {i}\right). \tag {13}
+$$
+
+# 3.3 Training Objective
+
+Combining the above analyses, the total objective loss is
+
+$$
+\mathcal {L} = \mathcal {L} _ {s} + \lambda_ {u} \mathcal {L} _ {u, t} + \lambda_ {c} \mathcal {L} ^ {C L S}. \tag {14}
+$$
+
+# 4 Experiments
+
+# 4.1 Datasets
+
+We evaluate our SSLRE framework on three DSRE datasets, including NYT10 (Riedel et al., 2010), NYT10m (Gao et al., 2021a), and wiki20m (Gao et al., 2021a).
+
+NYT10 dataset is created by aligning information from FreeBase with the New York Times (NYT)
+
+corpus. However, (1) it contains many duplicated instances in the dataset; (2) there is no public validation set; (3) The distantly supervised test set is quite noisy since the annotated errors. (Gao et al., 2021a) notes that  $53\%$  samples of the NYT10 test set are wrongly labeled. We only use it to perform held-out evaluation experiment in Table 5 with some strong baselines.
+
+NYT10m build a manually annotated test set for NYT10 for more accurate evaluation. Furthermore, it removes all the duplicated instances and create a new relation ontology by merging semantically similar relations and delete the relations that only show up in train set or the test set.
+
+wiki20m is a processed version of Wiki20 (Han et al., 2020), which is constructed by aligning the English Wikipedia corpus with Wikidata (Vrandecic and Krötzsch, 2014). It has a manually annotated test set for evaluation.
+
+# 4.2 Evaluation and Parameter Settings
+
+To guarantee the fairness of evaluation. We take both sentence-level evaluation and bag-level evaluation in our experiments. Further details of the evaluation methods are available in the appendix C. To achieve bag-level evaluation under sentence-level training, we use at-least-one (ONE) aggregation stragy (Zeng et al., 2015), which first predicts relation scores for each sentence in the bag, and then takes the highest score for each relation. The details of the hyper-parameters are available in the appendix D.
+
+# 4.3 Baselines
+
+In order to prove the effectiveness of the SSLRE, we compare our method with state-of-the-art methods sentence-level DSRE framework and bag-level DSRE framework.
+
+For bag-level methods baselines, RESIDE (Vashisht et al., 2018) exploits the information of entity type and relation alias to add a soft limitation for DSRE. DISTRE (Alt et al., 2019) combines the selective attention to its Transformer-based model. CIL (Chen et al., 2021a) utilize contrastive instance learning under MIL framework, HiCLRE (Li et al., 2022) propose hierarchical contrastive learning framework, PARE (Rathore et al., 2022) propose concatenate all sentences in the bag to attend every token in the bag. Besides, we combine Bert with different aggregation strategies: ONE, which is mentioned in section 4.2; AVG averages the representations of all the sentences in the bag;
+
+ATT (Lin et al., 2016) produces bag-level representation as a weighted average over embeddings of sentences and determines weights by attention scores between sentences and relations.
+
+For sentence-level baselines, RL-DSRE (Feng et al., 2018) apply reinforcement learning to train sample selector by feedback from the manually designed reward function. ARNOR (Jia et al., 2019) selects the reliable instances based on reward of attention score on the selected patterns. SENT (Ma et al., 2021) filters noisy instances based on negative training.
+
+# 4.4 Results
+
+We first evaluate our SSLRE framework in the NYT10m and WIKI20m dataset. Table 1 shows the overall performance in terms of sentence-level evaluation. From the results, we can observe that (1) Our SSLRE framework demonstrates superior performance on both datasets, surpassing all strong baseline models significantly in terms of F1 score. In comparison to the most robust baseline models in two distantly supervised datasets, SSLRE displays a significant enhancement in performance (i.e.,  $+6.3\%$  F1 and  $+3.4\%$  F1). (2) The current sentence-level DSRE models (SENT, ARNOR) fail to outperform the state-of-the-art MIL-based techniques in terms of F1 score on the aforementioned datasets. This could be attributed to the loss of information resulting from the elimination of samples. Unlike the MIL-based methods that employ all samples for training, these models only utilize selected samples. Moreover, the selection procedure may not always be reliable. (3) The performance of state-of-the-art MIL-based methods is not substantially superior to that of the Bert baseline. This suggests that the MIL modules, which are specifically crafted for this task, do not exhibit significant effectiveness when evaluated at the sentence level.
+
+Table 2 presents the results of bag-level evaluation of SSLRE with ONE strategy on NYT10m and WIKI20m datasets. We compared our SSLRE framework with state-of-the-art methods for bag-level relation extraction, and found that our approach outperformed all strong baselines. Specifically, SSLRE achieved a  $5.5\%$  improvement in AUC compared to the best baseline model PARE, and a  $10.1\%$  improvement in micro F1 score compared with best baseline model HICLRE on the NYT10m dataset. Despite being trained without a
+
+<table><tr><td rowspan="2">Models</td><td colspan="3">NYT10m</td><td colspan="3">wiki20m</td></tr><tr><td>μPrec.</td><td>μRec.</td><td>μF1</td><td>μPrec.</td><td>μRec.</td><td>μF1</td></tr><tr><td>Bert-ATT</td><td>49.1</td><td>58.2</td><td>53.3</td><td>71.3</td><td>77.8</td><td>74.4</td></tr><tr><td>Bert-ONE</td><td>51.0</td><td>60.7</td><td>55.5</td><td>72.8</td><td>73.2</td><td>73.0</td></tr><tr><td>Bert-AVG</td><td>52.0</td><td>55.8</td><td>53.9</td><td>80.7</td><td>76.9</td><td>78.7</td></tr><tr><td>RESIDE</td><td>45.5</td><td>48.7</td><td>47.0</td><td>69.9</td><td>72.3</td><td>70.9</td></tr><tr><td>DISTRE</td><td>55.1</td><td>51.0</td><td>52.9</td><td>80.3</td><td>73.6</td><td>76.8</td></tr><tr><td>CIL</td><td>58.1</td><td>49.3</td><td>53.4</td><td>81.8</td><td>73.1</td><td>77.0</td></tr><tr><td>HiCLRE</td><td>54.6</td><td>60.7</td><td>57.5</td><td>81.6</td><td>74.6</td><td>77.9</td></tr><tr><td>PARE</td><td>53.1</td><td>58.1</td><td>55.4</td><td>77.3</td><td>74.2</td><td>75.7</td></tr><tr><td>RL-DSRE</td><td>48.6</td><td>57.1</td><td>52.5</td><td>69.9</td><td>74.3</td><td>72.0</td></tr><tr><td>ARNOR</td><td>53.2</td><td>59.1</td><td>55.9</td><td>76.8</td><td>76.9</td><td>76.8</td></tr><tr><td>SENT</td><td>57.2</td><td>56.3</td><td>56.7</td><td>79.8</td><td>78.2</td><td>78.9</td></tr><tr><td>SSLRE</td><td>57.4</td><td>72.0</td><td>63.8</td><td>81.9</td><td>81.0</td><td>81.5</td></tr></table>
+
+Table 1: Sentence-level evaluation results on NYT10m and wiki20m. Bold and underline indicate the best and the second best scores.  
+
+<table><tr><td rowspan="2">Models</td><td colspan="3">NYT10m</td><td colspan="3">wiki20m</td></tr><tr><td>AUC</td><td>μF1</td><td>Macro_F1</td><td>AUC</td><td>μF1</td><td>Macro_F1</td></tr><tr><td>Bert-ATT</td><td>49.5</td><td>52.9</td><td>24.5</td><td>88.0</td><td>80.9</td><td>80.7</td></tr><tr><td>Bert-ONE</td><td>56.7</td><td>54.1</td><td>35.7</td><td>89.9</td><td>81.3</td><td>82.0</td></tr><tr><td>Bert-AVG</td><td>57.1</td><td>56.2</td><td>33.9</td><td>88.9</td><td>82.6</td><td>81.1</td></tr><tr><td>RESIDE</td><td>36.8</td><td>44.2</td><td>11.2</td><td>80.5</td><td>75.1</td><td>74.2</td></tr><tr><td>HiCLRE</td><td>61.0</td><td>61.2</td><td>32.0</td><td>89.1</td><td>82.3</td><td>81.1</td></tr><tr><td>CIL</td><td>57.4</td><td>59.6</td><td>29.4</td><td>89.3</td><td>81.8</td><td>82.4</td></tr><tr><td>PARE</td><td>61.1</td><td>59.8</td><td>37.2</td><td>90.3</td><td>83.2</td><td>82.6</td></tr><tr><td>SSLRE-ONE</td><td>66.5</td><td>71.3</td><td>36.9</td><td>91.6</td><td>83.3</td><td>84.1</td></tr></table>
+
+Table 2: Bag-level evaluation results on nyt10m and wiki20m. SSLRE-ONE represents the SSLRE with ONE aggregation strategy
+
+MIL framework, our SSLRE framework achieves state-of-the-art performance on bag-level relation extraction. This finding suggests that sentence-level training can also yield excellent results on bag-label prediction. This observation is also consistent with (Gao et al., 2021a; Amin et al., 2020).On the wiki20m dataset, we note a consistent improvement on as well, although it is not as evident as in the case of NYT10m. We surmise that this could be attributed to the fact that the wiki20m dataset is relatively less noisy when compared to NYT10m.
+
+We also compared our framework to several strong baselines using held-out evaluation on the NYT10 dataset, which is detailed in appendix B.
+
+# 4.5 Ablation Study
+
+We conducted ablation study experiments on the NYT10m dataset to assess the effectiveness of different modules in our SSLRE framework. We specifically removed each of the argued contributions one by one to evaluate their effectiveness. For unsupervised learning part, we remove the pseudo classifier head and CPL one at a time. For supervised learning part, we switch from mixup supervised learning to supervised contrastive learning and cross entropy as our new objective. In terms of confident samples identifications methods, we alternate between using random (randomly selection) and NLI-based selection instead of our K-NN method. The NLI method involves performing zero-shot relation extraction through Natural Lan
+
+guage Inference (NLI)(Sainz et al., 2021), then identify the confident samples based on the level of agreement between the distant label and NLI soft label.
+
+<table><tr><td colspan="2">Methods</td><td>Prec.</td><td>Rec.</td><td>F1</td></tr><tr><td colspan="2">SSLRE</td><td>57.4</td><td>72.0</td><td>63.8</td></tr><tr><td rowspan="2">Unsupervised Learning</td><td>w/o pseudo head</td><td>56.2</td><td>70.6</td><td>62.5</td></tr><tr><td>w/o CPL</td><td>50.8</td><td>69.6</td><td>58.7</td></tr><tr><td rowspan="2">Supervised Learning</td><td>SupCon</td><td>55.4</td><td>67.5</td><td>60.9</td></tr><tr><td>CE</td><td>51.8</td><td>70.3</td><td>59.6</td></tr><tr><td rowspan="2">Conf-Samples Identification</td><td>random</td><td>58.6</td><td>56.0</td><td>57.3</td></tr><tr><td>NLI</td><td>56.4</td><td>65.3</td><td>60.5</td></tr></table>
+
+Table. 3 shows the ablation study results. We conclude that (1) Unsupervised learning apart effectively utilize the unlabeled samples. Removing pseudo classifier head and CPL leads to a decrease of  $1.3\%$  and  $5.1\%$  on micro-F1, respectively. (2) When dealing with noisy labeled data in supervised learning, Mixup Contrastive Supervised Learning proves to be more robust than both Supervised Contrastive Learning  $(-2.9\%)$  and Cross Entropy  $(-4.2\%)$ . (3) Our K-NN-based confident samples identification method outperforms the random method by  $6.5\%$  and the NLI method by  $3.3\%$ . This indicates that our K-NN method can effectively select confident samples.
+
+# 4.6 Analysis on KNN
+
+We conducted an evaluation of the performance of weighted k-nearest neighbors (kNN) in terms of its ability to select confident samples. To elaborate, we intentionally corrupted the labels of instances in the nyt10m test set with a random probability of  $20\%$ ,  $40\%$ , and  $60\%$ . Our objective was to assess whether our weighted kNN method could effectively identify the uncorrupted (confident) instances. We trained our model on the corrupted nyt10m test set for 10 epochs, considering its relatively smaller size compared to the training set, which required more epochs to converge. In order to evaluate the ability of the weighted kNN in identifying confident samples, we reported the recall and precision metrics. The results are shown as 4:
+
+It is worth noting that precision is the more important metric because our goal is to make the la
+
+Table 3: Ablation study of SSLRE on NYT10m  
+
+<table><tr><td></td><td>Pre.</td><td>Rec.</td></tr><tr><td>20%</td><td>0.9851</td><td>0.8970</td></tr><tr><td>30%</td><td>0.9732</td><td>0.8758</td></tr><tr><td>40%</td><td>0.9531</td><td>0.8723</td></tr><tr><td>50%</td><td>0.9254</td><td>0.8655</td></tr><tr><td>60%</td><td>0.8735</td><td>0.8172</td></tr></table>
+
+Table 4: The effect of KNN.
+
+beled data as clean as possible after selecting confident samples. Even with a noisy label rate of  $60\%$ , our weighted kNN method still achieves a precision of  $87.35\%$  for the identified confident samples. This indicates that only  $12.65\%$  of the labeled data is noisy, which is significantly smaller than the  $60\%$  noise rate. Additionally, our weighted kNN method demonstrates a recall of over  $80\%$  for confident samples. Although a few clean samples may not be selected, they can still be utilized through pseudo-labeling techniques.
+
+# 4.7 Analysis on Dropout rate
+
+Figure 3 shows the performance of SSLRE under different dropout rate for strong augmentation. we can observe that: (1) Increasing the dropout rate appropriately improves the model's performance, with SSLRE achieving the best result (63.8 on F1) when the strong augmentation dropout rate is set to 0.4. (2) Augmentation using a very high dropout rate harms the performance, as it results in a loss of a significant amount of information. However, this does not significantly degrade the performance since we only use strong augmentation to predict pseudo-labels, which only affects  $\mathcal{L}_u$ .
+
+![](images/5e2020bfef393a62a9c4c224524f58a374e2d81abbaa8a70041a84603a717d80.jpg)  
+Figure 3: Strong augmentation with different dropout rate
+
+![](images/541d54363fbb9e6436cbf6a01bd13b47558a76a192fdd9615d7b903a17a55f86.jpg)  
+(a) Sup
+
+![](images/38b6b7280c8481f2508f8a7b80b1fded57bc9ec9c951158b99bede98ab498c8e.jpg)  
+(b) SSL  
+Figure 4: t-SNE visualization of representations with Pseudo labeling(SSL) and without(Sup). SSL achieves a better cluster results comparing with Sup.
+
+# 4.8 t-SNE analysis
+
+To demonstrate that preserving unlabeled samples can facilitate the learning of a superior representation compared to discarding them, we utilized sentence representations obtained from theta as the input to conduct dimension reduction via t-SNE and acquire two-dimensional representations. We focused on four primary categories of relation classes, which are "/location/location/contains", "/business/person/company", "/location/administrative_division/country", and "/people/person/nationality". As depicted in Figure 4, leveraging unlabeled samples via Pseudo labeling enhances the clustering of identical relation data points and effectively separates distinct classes from one another. Appendix F shows the t-SNE results of all classes.
+
+# 5 Conclusion
+
+In this paper, we propose SSLRE, a novel sentence-level framework that is grounded in Semi-Supervised Learning for the DSRE task. Our SSLRE framework employs mixup supervised contrastive learning to tackle the noise present in selected samples, and it leverages unlabeled samples through Pseudo Labeling, which effectively utilize the information contained within noisy samples. Experimental results demonstrate that SSLRE outperforms strong baseline methods in both sentence-level and bag-level relation extraction.
+
+# Limitations
+
+In order to augment textual instances, we leverage dropout during forward propagation. This necessitates propagating each instance twice to generate the augmented sentence embeddings. However, the demand for GPU resources is higher compared to previous methods. Furthermore, we adjust the dropout rate to regulate the augmentation intensity
+
+for semi-supervised learning and show its effectiveness through the performance results. Nonetheless, we have not conducted explicit experiments to investigate the interpretability, which needs further investigation.
+
+# Acknowledgment
+
+This work is sponsored by National Natural Science Foundation of China (NSFC) (62206291, 62141608, and 62236010).
+
+# References
+
+Christoph Alt, Marc Hubner, and Leonhard Hennig. 2019. Fine-tuning pre-trained transformer language models to distantly supervised relation extraction. ArXiv, abs/1906.08646.  
+Saadullah Amin, Katherine Ann Dunfield, Anna Vechkaeva, and Guenter Neumann. 2020. A data-driven approach for noise reduction in distantly supervised biomedical relation extraction. In Proceedings of the 19th SIGBioMed Workshop on Biomedical Language Processing, pages 187-194, Online. Association for Computational Linguistics.  
+Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. 2017. A closer look at memorization in deep networks. In International conference on machine learning, pages 233-242. PMLR.  
+Philip Bachman, Ouais Alsharif, and Doina Precup. 2014. Learning with pseudo-ensembles. ArXiv, abs/1412.4864.  
+Dara Bahri, Heinrich Jiang, and Maya Gupta. 2020. Deep k-nn for noisy labels. In International Conference on Machine Learning, pages 540-550. PMLR.  
+David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. 2019. Mixmatch: A holistic approach to semi-supervised learning. *ArXiv*, abs/1905.02249.  
+Jhih-Wei Chen, Tsu-Jui Fu, Chen-Kang Lee, and WeiYun Ma. 2020. H-fnd: Hierarchical false-negative denoising for distant supervision relation extraction. In Findings.  
+Tao Chen, Haizhou Shi, Siliang Tang, Zhigang Chen, Fei Wu, and Yue Ting Zhuang. 2021a. Cil: Contrastive instance learning framework for distantly supervised relation extraction. ArXiv, abs/2106.10855.  
+Tao Chen, Haochen Shi, Liyuan Liu, Siliang Tang, Jian Shao, Zhigang Chen, and Yueting Zhuang. 2021b. Empower distantly supervised relation extraction with collaborative adversarial training. In AAAI Conference on Artificial Intelligence.
+
+Xinlei Chen and Abhinav Kumar Gupta. 2015. Webly supervised learning of convolutional networks. 2015 IEEE International Conference on Computer Vision (ICCV), pages 1431-1439.  
+Qin Dai, Benjamin Heinzerling, and Kentaro Inui. 2022. Cross-stitching text and knowledge graph encoders for distantly supervised relation extraction. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6947-6958, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Jun Feng, Minlie Huang, Li Zhao, Yang Yang, and Xiaoyan Zhu. 2018. Reinforcement learning for relation classification from noisy data. In Proceedings of the aai conference on artificial intelligence, volume 32.  
+Tianyu Gao, Xu Han, Keyue Qiu, Yuzhuo Bai, Zhiyu Xie, Yankai Lin, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. 2021a. Manual evaluation matters: Reviewing test protocols of distantly supervised relation extraction. In Findings.  
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021b. SimCSE: Simple contrastive learning of sentence embeddings. In Empirical Methods in Natural Language Processing (EMNLP).  
+Aritra Ghosh, Himanshu Kumar, and P. Shanti Sastry. 2017. Robust loss functions under label noise for deep neural networks. In AAAI Conference on Artificial Intelligence.  
+Jacob Goldberger and Ehud Ben-Reuven. 2016. Training deep neural-networks using a noise adaptation layer. In International Conference on Learning Representations.  
+Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Wai-Hung Tsang, and Masashi Sugiyama. 2018a. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In Neural Information Processing Systems.  
+Jiangfan Han, Ping Luo, and Xiaogang Wang. 2019. Deep self-learning from noisy labels. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5137-5146.  
+Xu Han, Tianyu Gao, Yankai Lin, Hao Peng, Yaoliang Yang, Chaojun Xiao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. 2020. More data, more relations, more context and more openness: A review and outlook for relation extraction. In AACL.
+
+Xu Han, Zhiyuan Liu, and Maosong Sun. 2018b. Neural knowledge acquisition via mutual attention between knowledge graph and text. In AAAI Conference on Artificial Intelligence.  
+Kailong Hao, Botao Yu, and Wei Hu. 2021. Knowing false negatives: An adversarial training method for distantly supervised relation extraction. In Conference on Empirical Methods in Natural Language Processing.  
+Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. 2022. Learning with neighbor consistency for noisy labels. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4662-4671.  
+Wei Jia, Dai Dai, Xinyan Xiao, and Hua Wu. 2019. Arnor: Attention regularization based noise reduction for distant supervision relation classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1399-1408.  
+Zhanghan Ke, Daoye Wang, Qiong Yan, Jimmy S. J. Ren, and Rynson W. H. Lau. 2019. Dual student: Breaking the limits of the teacher in semi-supervised learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6727-6735.  
+Diederik P. Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. CoRR, abs/1412.6980.  
+Kimin Lee, Sukmin Yun, Kibok Lee, Honglak Lee, Bo Li, and Jinwoo Shin. 2019. Robust inference via generative classifiers for handling noisy labels. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 3763-3772. PMLR.  
+Dongyang Li, Taolin Zhang, Nan Hu, Chengyu Wang, and Xiaofeng He. 2022. Hic Ire: A hierarchical contrastive learning framework for distantly supervised relation extraction. In Findings.  
+Junnan Li, Richard Socher, and Steven C. H. Hoi. 2020. Dividemix: Learning with noisy labels as semi-supervised learning. ArXiv, abs/2002.07394.  
+Yankai Lin, Shiqi Shen, Zhiyuan Liu, Huanbo Luan, and Maosong Sun. 2016. Neural relation extraction with selective attention over instances. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2124-2133.  
+Ruri Liu, Shasha Mo, Jianwei Niu, and Shengda Fan. 2022. CETA: A consensus enhanced training approach for denoising in distantly supervised relation extraction. In Proceedings of the 29th International Conference on Computational Linguistics, pages 2247-2258, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.
+
+Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. 2020. Early-learning regularization prevents memorization of noisy labels. ArXiv, abs/2007.00151.  
+Tongliang Liu and Dacheng Tao. 2014. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38:447-461.  
+Yang Liu and Hongyi Guo. 2019. Peer loss functions: Learning from noisy labels without knowing noise rates. ArXiv, abs/1910.03231.  
+Yueming Lyu and Ivor Wai-Hung Tsang. 2019. Curriculum loss: Robust learning and generalization against label corruption. ArXiv, abs/1905.10045.  
+Ruotian Ma, Tao Gui, Linyang Li, Qi Zhang, Xuan-Jing Huang, and Yaqian Zhou. 2021. Sent: Sentence-level distant relation extraction via negative training. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6201-6213.  
+Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1003-1011.  
+Pengda Qin, Weiran Xu, and William Yang Wang. 2018. Robust distant supervision relation extraction via deep reinforcement learning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2137-2147.  
+Vipul Rathore, Kartikeya Badola, Parag Singla, and Mausam. 2022. PARE: A simple and strong baseline for monolingual and multilingual distantly supervised relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 340-354, Dublin, Ireland. Association for Computational Linguistics.  
+Sebastian Riedel, Limin Yao, and Andrew McCallum. 2010. Modeling relations and their mentions without labeled text. In *Joint European Conference on Machine Learning and Knowledge Discovery in Databases*, pages 148-163. Springer.  
+Oscar Sainz, Oier Lopez de Lacalle, Gorka Labaka, Ander Barrena, and Eneko Agirre. 2021. Label verbalization and entailment for effective zero and few-shot relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1199-1212.  
+Yu-Ming Shang and Wei Wei. 2019. Are noisy sentences useless for distant supervised relation extraction? In AAAI Conference on Artificial Intelligence.
+
+Tao Shen, Guodong Long, Tao Shen, Tianyi Zhou, Lina Yao, Huan Huo, and Jing Jiang. 2019. Self-attention enhanced selective gate with entity-aware embedding for distantly supervised relation extraction. ArXiv, abs/1911.11899.  
+Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin Dogus Cubuk, Alexey Kurakin, Han Zhang, and Colin Raffel. 2020. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. ArXiv, abs/2001.07685.  
+Antti Tarvainen and Harri Valpola. 2017. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NIPS.  
+Shikhar Vashishth, Rishabh Joshi, Sai Suman Prayaga, Chiranjib Bhattacharyya, and Partha Talukdar. 2018. RESIDE: Improving distantly-supervised neural relation extraction using side information. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1257-1266, Brussels, Belgium. Association for Computational Linguistics.  
+Denny Vrandecic and Markus Krötzsch. 2014. Wiki-data: A free collaborative knowledge base. Communications of the ACM, 57:78-85.  
+Xudong Wang, Zhi-Li Wu, Long Lian, and Stella X. Yu. 2022. Debiased learning from naturally imbalanced pseudo-labels. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14627-14637.  
+Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris N. Metaxas, and Chao Chen. 2020. A topological filter for learning with label noise. ArXiv, abs/2012.04835.  
+Yuwei Xia, Mengqi Zhang, Qiang Liu, Shu Wu, and Xiao-Yu Zhang. 2022. MetaTKG: Learning evolutionary meta-knowledge for temporal knowledge graph reasoning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7230-7240, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Zhiquan Ye and Zhenhua Ling. 2019. Distant supervision relation extraction with intra-bag and inter-bag attentions. ArXiv, abs/1904.00143.  
+Daojian Zeng, Kang Liu, Yubo Chen, and Jun Zhao. 2015. Distant supervision for relation extraction via piecewise convolutional neural networks. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 1753-1762.  
+Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. 2021. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. ArXiv, abs/2110.08263.
+
+Mengqi Zhang, Yuwei Xia, Qiang Liu, Shu Wu, and Liang Wang. 2023a. Learning latent relations for temporal knowledge graph reasoning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12617-12631, Toronto, Canada. Association for Computational Linguistics.
+
+Mengqi Zhang, Yuwei Xia, Qiang Liu, Shu Wu, and Liang Wang. 2023b. Learning long- and short-term representations for temporal knowledge graph reasoning. In Proceedings of the ACM Web Conference 2023, WWW '23, page 2412-2422, New York, NY, USA. Association for Computing Machinery.
+
+# A Algorithm
+
+Algorithm 1 provides the pseudo-code of the overall framework.
+
+# B Held-out evaluation
+
+<table><tr><td>Models</td><td>Prec.</td><td>Rec.</td><td>F1</td></tr><tr><td>SENT</td><td>45.3</td><td>50.1</td><td>47.5</td></tr><tr><td>HiCLRE</td><td>44.3</td><td>50.6</td><td>47.2</td></tr><tr><td>SSLRE(Ours)</td><td>49.2</td><td>54.3</td><td>51.6</td></tr></table>
+
+Table 5: Held-out evaluation on NYT10
+
+Table 5 shows the held-out evaluation results on NYT10 dataset.
+
+# C Evaluation Settings
+
+Sentence-level evaluation: Different from bag-level evaluation used by MIL-based model, a sentence-level (or instance-level) evaluation accesses model performance directly on all of the individual instances in the dataset, which require the model to accurately predict the relation for each sentence. Following (Jia et al., 2019; Ma et al., 2021; Liu et al., 2022), we report micro-Precision  $(\mu \mathrm{Prec.})$ , micro-Recall  $(\mu \mathrm{Rec.})$  and microF1  $(\mu \mathrm{F1})$  for sentence-level evaluation.
+
+# D Parameter Settings
+
+Bag-level evaluation: Bag evaluation accesses the performance of bag relation label extraction. Since manually annotated data are at the sentence-level, following (Gao et al., 2021a), we construct bag-level annotations in the following way: For each bag, if one sentence in the bag has a human-labeled relation, this bag is labeled with this relation; if no sentence in the bag is annotated with any relation,
+
+![](images/5c3ebe4b63ac4922d423f49ab0d5a254d00dddaf4d5513ea2820c6496cfe9cca.jpg)  
+Figure 5: PR-curve
+
+this bag is labeled as N/A. We report AUC, MicroF1 and Macro-F1 for bag level evaluation.
+
+The underlying encoder for sentence are implemented by BERT-base (Devlin et al., 2019), which generates 768 hidden units for each token's context-aware representation. During the training stage, we set the learning rate of the model to  $2 \times 10^{-5}$  and the batch size to 64, which was determined through a grid search over batch sizes in  $\{16, 32, 64\}$  and learning rates in  $\{1e-5, 2e-5, 5e-5\}$ . We train the model for 5 epochs and use Adam (Kingma and Ba, 2014) as the optimizer. The Mixup parameter  $\alpha_{m}$  is set to 1, the classifier loss weight  $\lambda_{c}$  is set to 0.2, the fractile alpha is set to 0.8, the unsupervised loss weight  $\lambda_{u}$  is set to 1, and the CPL threshold  $\tau$  is set to 0.95. We set the dropout rate for weak augmentation to 0.2 and the dropout rate for strong augmentation to 0.4. Further analysis on the strong augmentation dropout rate is presented in Section 4.7.
+
+# E PR-curve
+
+We report the P-R curve on NYT10m dataset as Figure 5:
+
+![](images/12367125ff4214d54131c8e6e0d41a0628383e3d0d06356bfb984c7c98f87f95.jpg)  
+(a) Sup
+
+![](images/9f09bc1cc074ae9ac9fb9e7e46befa4478d5c663ee676a4c0d56d43bb9d00be7.jpg)  
+(b) SSL  
+Figure 6: t-SNE visualization of representations with Pseudo labeling(SSL) and without(Sup). SSL achieves a better cluster results comparing with Sup, especially on color green and light purple.
+
+# F Additional t-SNE analysis
+
+Figure 6 shows the t-SNE results on all classes of the sentence representation.
+
+Algorithm 1: SSLRE Algorithm  
+input: Noisy Dataset  $\tilde{\mathcal{D}}$  output:   
+1 Warm up  $\theta$  and  $\phi$  for one epoch using supervised contrastive learning and get  $\theta^{\prime}$    
+2 Gain features:  $z_{i} = \theta^{\prime}(x_{i})$    
+3 Build weighted K-NN graph using Eq. (1)   
+4 Gain  $\mathcal{X}$  and  $\mathcal{U}$  using Eq. (2-4)   
+5 Reinitialize  $\phi$  and  $\theta$    
+6 while not reach the maximum iteration do   
+7 for  $c\gets 1$  to  $C$  do   
+8 Calculate  $\mathcal{T}(c)$  using Eq. (7) {Determine the flexible threshold for class c.}   
+9 end   
+10 for  $t\gets 1$  to num_iters do   
+11 From  $\mathcal{X}$  , draw a mini-batch  $X^{t} = \{(x_{b},y_{b});b\in (1,\dots ,B)\}$    
+12 From  $\mathcal{U}$  , draw a mini-batch  $U^{t} = \{u_{b};b\in (1,\dots ,B)\}$  /\* Learning from labeled dataset X \*/   
+13  $Z^{t} = \theta_{w}(X^{t})\cup \theta_{w}(X^{t}) / /$  Get two augmented embedding views by dropout inference twice.   
+14 Calculate  $L^{CLS}$  using Eq. (13).   
+15 Mixup embedding using Eq. (10), and gain  $Z_{mix}^{t}$    
+16 Calculate  $L_{s}$  using Eq. (11) /\* Learning from unlabeled dataset U \*/   
+17 Generate pseudo labels using Eq. (5).   
+18 Calculate unsupervised loss using Eq. (6) with  $\mathcal{T}(c)$    
+19 Calculate the overall loss  $L$  using Eq. (14).   
+20 end   
+21 end

2023/Noise-Robust Semi-Supervised Learning for Distantly Supervised Relation Extraction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:020e00c36dda24176fac7a08d86b25bb3431f58da9de4f3b6ed52dabf08eb872
+size 442400

2023/Noisy Pair Corrector for Dense Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_content_list.json

@@ -0,0 +1,1978 @@
+[
+    {
+        "type": "text",
+        "text": "Noisy Pair Corrector for Dense Retrieval",
+        "text_level": 1,
+        "bbox": [
+            285,
+            90,
+            712,
+            110
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hang Zhang $^{1}$ , Yeyun Gong, Xingwei He $^{†2}$ , Dayiheng Liu $^{1}$ , Daya Guo $^{3}$ , Jiancheng Lv $^{†1}$ , Jian Guo $^{†4}$",
+        "bbox": [
+            240,
+            124,
+            754,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup> College of Computer Science, Sichuan University",
+        "bbox": [
+            294,
+            158,
+            709,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "<sup>1</sup> Engineering Research Center of Machine Learning and Industry Intelligence",
+            "<sup>2</sup> The University of Hong Kong <sup>3</sup> Sun Yat-sen University <sup>4</sup> IDEA Research, China hangzhang_scu@foxmail.com, hexingwei15@gmail.com"
+        ],
+        "bbox": [
+            169,
+            174,
+            831,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Most dense retrieval models contain an implicit assumption: the training query-document pairs are exactly matched. Since it is expensive to annotate the corpus manually, training pairs in real-world applications are usually collected automatically, which inevitably introduces mismatched-pair noise. In this paper, we explore an interesting and challenging problem in dense retrieval, how to train an effective model with mismatched-pair noise. To solve this problem, we propose a novel approach called Noisy Pair Corrector (NPC), which consists of a detection module and a correction module. The detection module estimates noise pairs by calculating the perplexity between annotated positive and easy negative documents. The correction module utilizes an exponential moving average (EMA) model to provide a soft supervised signal, aiding in mitigating the effects of noise. We conduct experiments on text-retrieval benchmarks Natural Question and TriviaQA, code-search benchmarks StaQC and SO-DS. Experimental results show that NPC achieves excellent performance in handling both synthetic and realistic noise.",
+        "bbox": [
+            141,
+            280,
+            460,
+            636
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            649,
+            258,
+            664
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "With the advancements in pre-trained language models (Devlin et al., 2019; Liu et al., 2019), dense retrieval has developed rapidly in recent years. It is essential to many applications including search engine (Brickley et al., 2019), open-domain question answering (Karpukhin et al., 2020a; Zhang et al., 2021), and code intelligence (Guo et al., 2021). A typical dense retrieval model maps both queries and documents into a low-dimensional vector space and measures the relevance between them by the similarity between their respective representations (Shen et al., 2014). During training, the model utilizes query-document pairs as labelled",
+        "bbox": [
+            112,
+            675,
+            489,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/08f56f5cf93fcf9e6b5d29530e4fe73cb11044c730ce1906f44e21f465f88049.jpg",
+        "image_caption": [
+            "Figure 1: Two examples from StaQC training set. In the bottom example, the given code is mismatched with the query, since it can not answer the query."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            252,
+            875,
+            460
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "training data (Xiong et al., 2021) and samples negative documents for each pair. Then the model learns to minimize the contrastive loss for obtaining a good representation ability (Zhang et al., 2022b; Qu et al., 2021).",
+        "bbox": [
+            507,
+            532,
+            882,
+            611
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent studies on dense retrieval have achieved promising results with hard negative mining (Xiong et al., 2021), pretraining (Gao and Callan, 2021a), distillation (Yang and Seo, 2020), and adversarial training (Zhang et al., 2022a). All methods contain an implicit assumption: each query is precisely aligned with the positive documents in the training set. In practical applications, this assumption becomes challenging to satisfy, particularly when the corpora is automatically collected from the internet. In such scenarios, it is inevitable that the training data will contain mismatched pairs, incorporating instances such as user mis-click noise in search engines or low-quality reply noise in Q&A communities. As shown in Fig. 1, the examples are from StaQC benchmark (Yao et al., 2018), which is automatically collected from StackOverflow. The document, i.e., code solution, can not answer the query but is incorrectly annotated as a positive doc",
+        "bbox": [
+            507,
+            613,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Work is done during internship at IDEA Research",
+        "bbox": [
+            134,
+            891,
+            447,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "†Corresponding author",
+        "bbox": [
+            136,
+            904,
+            278,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "11439",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 11439-11451",
+        "bbox": [
+            210,
+            945,
+            784,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            277,
+            958,
+            719,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/cebbd4ae4812b6f292e3700c18d2d75a0c56e784203664521ebd64d599a63c65.jpg",
+        "image_caption": [
+            "Figure 2: Effect of matched & mismatched pair for training. Green objects refer to annotated pairs, while pentagram and triangle are actually aligned pairs. In the left case, retrieval models are required to push the query with true-positive document (TP Doc) together and pull the query with true-negative documents (TN Doc) apart. In the right case, the retrieval models are misled by the mismatched data pair, where the false-positive document (FP Doc) and the false-negative document (FN Doc) are wrongly pulled and pushed, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            84,
+            480,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "uments. Such noisy pairs are widely present in automatically constructed datasets, which ultimately impact the performance of dense retrievers.",
+        "bbox": [
+            112,
+            416,
+            489,
+            463
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To train robust dense retrievers, previous works have explored addressing various types of noise. For example, RocketQA (Qu et al., 2021) and AR2 (Zhang et al., 2022a) mitigate the false-negative noise with a cross-encoder filter and distillation, respectively; coCondenser (Gao and Callan, 2021b) reduce the noise during fine-tuning with pre-training technique; RoDR (Chen et al., 2022) deal with query spelling noise with local ranking alignment. However, mismatched-pair noise (false positive problem) in dense retrieval has not been well studied. As shown in Fig. 2, mismatched-pair noise will mislead the retriever to update in the opposite direction.",
+        "bbox": [
+            112,
+            464,
+            489,
+            690
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Based on these observations, we propose a Noisy Pair Corrector (NPC) framework to solve the false-positive problem. NPC consists of noise detection and correction modules. At each epoch, the detection module estimates whether a query-document pair is mismatched by the perplexity between the annotated document and easy negative documents. Then the correction module provides a soft supervised signal for both estimated noisy data and clean data via an exponential moving average (EMA) model. Both modules are plug-and-play, which means NPC is a general training paradigm that can be easily applied to almost all retrieval models.",
+        "bbox": [
+            112,
+            692,
+            489,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The contributions of this paper are as follows: (1)",
+        "bbox": [
+            131,
+            903,
+            487,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We reveal and extensively explore a long-neglected problem in dense retrieval, i.e., mismatched-pair noise, which is ubiquitous in the real world. (2) We propose a simple yet effective method for training dense retrievers with mismatched-pair noise. (3) Extensive experiments on four datasets and comprehensive analyses verify the effectiveness of our method against synthetic and realistic noise. Code is available at https://github.com/hangzhang-nlp/NPC.",
+        "bbox": [
+            507,
+            84,
+            884,
+            247
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Preliminary",
+        "text_level": 1,
+        "bbox": [
+            507,
+            256,
+            650,
+            273
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Before describing our model in detail, we first introduce the basic elements of dense retrieval, including problem definition, model architecture, and model training.",
+        "bbox": [
+            507,
+            282,
+            884,
+            344
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given a query  $q$ , and a document collection  $\\mathbb{D}$ , dense retrieval aims to find document  $d^{+}$  relevant to  $q$  from  $\\mathbb{D}$ . The training set consists of a collection of query-document pairs, donated as  $C = \\{(q_{1}, d_{1}^{+}), \\dots, (q_{N}, d_{N}^{+})\\}$ , where  $N$  is the data size. Typical dense retrieval models adopt a dual encoder architecture to map queries and documents into a dense representation space. Then the relevance score  $f(q, d)$  of query  $q$  and document  $d$  can be calculated with their dense representations:",
+        "bbox": [
+            507,
+            347,
+            884,
+            507
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf _ {\\theta} (q, d) = \\operatorname {s i m} \\left(E (q; \\theta), E (d; \\theta)\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            520,
+            880,
+            537
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $E(\\cdot ;\\theta)$  denotes the encoder module parameterized with  $\\theta$  and sim is the similarity function, e.g., euclidean distance, cosine distance, inner-product. Existing methods generally leverage the approximate nearest neighbor technique (ANN) (Johnson et al., 2019) for efficient search.",
+        "bbox": [
+            507,
+            548,
+            884,
+            645
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For training dense retrievers, conventional approaches leverage contrastive learning techniques (Karpukhin et al., 2020a; Zhang et al., 2022b). Given a training pair  $(q_{i},d_{i}^{+})\\in C$ , these methods sample  $m$  negative documents  $\\{d_{i,1}^{-},\\dots,d_{i,m}^{-}\\}$  from a large document collection  $\\mathbb{D}$ . The retriever's objective is to minimize the contrastive loss, pushing the similarity of positive pairs higher than negative pairs. Previous work (Xiong et al., 2021) has verified the effectiveness of the negative sampling strategy. Two commonly employed strategies are \"In-Batch Negative\" and \"Hard Negative\" (Karpukhin et al., 2020a; Qu et al., 2021).",
+        "bbox": [
+            507,
+            646,
+            885,
+            853
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The above training paradigm assumes that the query-document pair  $(q_{i},d_{i}^{+})$  in training set  $C$  is correctly aligned. However, this assumption is difficult to satisfy in real-world applications (Qu et al.,",
+        "bbox": [
+            507,
+            854,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "11440",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/cbd09a5537eb88688a21601df295fdc6babb3d9c3c18cd09702026ac1f3b093c.jpg",
+        "image_caption": [
+            "(a) Noise Detection",
+            "(b) Noise Correction",
+            "Figure 3: Overview of noise detection and noise correction. (a) Procedure of Noise Detection. At each epoch, we first calculate the perplexity of all training query-document pairs using the retriever  $\\theta$ ; next fit the perplexity distribution with Gaussian Mixture Model to get the correctly matched probability of each pair; finally estimate the flag set  $\\{\\hat{y}_i\\}_{i=1}^N$  by setting the threshold. (b) Framework of Noise Correction. Given a batch of data pairs, where  $d_{i,1}^{-}$  is the hard negative of  $q_i$  and  $\\{q_3, d_3^{+}\\}$  is the estimated noisy pair, the retriever  $\\theta$  and teacher  $\\theta^*$  compute similarity matrices  $S_\\theta$  and  $S_{\\theta^*}$  for all queries and documents, respectively. The retriever learns to minimize (1)  $L_{cont}$ : the negative likelihood probability of true positive documents; (2)  $L_{cons}$ : the KL divergence between  $S_\\theta$  and the rectified soft label  $S_{\\theta^*}$  after normalization."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            131,
+            91,
+            868,
+            223
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2021; Li et al., 2022; Wang et al., 2022). In practice, most training data pairs are collected automatically without manual inspection, such as inevitably leading to the inclusion of some mismatched pairs.",
+        "bbox": [
+            112,
+            397,
+            489,
+            462
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            114,
+            473,
+            216,
+            487
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We propose NPC framework to learn retrievers with mismatched-pair noise. As shown in Fig. 3, NPC consists of two parts: (a) the noise detection module as described in Sec. 3.1, and (b) the noise correction module as described in Sec. 3.2.",
+        "bbox": [
+            112,
+            497,
+            489,
+            577
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Noise Detection",
+        "text_level": 1,
+        "bbox": [
+            112,
+            589,
+            282,
+            602
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The noise detection module is meant to detect mismatched pairs in the training set. We hypothesize that: dense retrievers will first learn to distinguish correctly matched pairs from easy negatives, and then gradually overfit the mismatched pairs. Therefore, we determine whether a training pair is mismatched by the perplexity between the annotated document and easy negative documents.",
+        "bbox": [
+            112,
+            609,
+            489,
+            736
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Specifically, given a retriever  $\\theta$  equipped with preliminary retrieval capabilities, and an uncertain pair  $(q_{i},d_{i})$ , we calculate the perplexity as follows:",
+        "bbox": [
+            112,
+            738,
+            487,
+            785
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP P L _ {\\left(q _ {i}, d _ {i}, \\theta\\right)} = - \\log \\frac {e ^ {\\tau f _ {\\theta} \\left(q _ {i} , d _ {i}\\right)}}{e ^ {\\tau f _ {\\theta} \\left(q _ {i} , d _ {i}\\right)} + \\sum_ {j = 1} ^ {m} e ^ {\\tau f _ {\\theta} \\left(q _ {i} , d _ {i , j} ^ {-}\\right)}}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            808,
+            485,
+            853
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tau$  is a hyper-parameter,  $d_{i,j}^{-}$  is the negative document randomly sampled from the document collection  $\\mathbb{D}$ . Note that  $d_{i,j}^{-}$  is a randomly selected negative document, not a hard negative. We discuss",
+        "bbox": [
+            112,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "this further in Appendix C. In practice, we adopt the \"In-Batch Negative\" strategy for efficiency.",
+        "bbox": [
+            507,
+            397,
+            880,
+            429
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "After obtaining the perplexity of each pair, an automated method is necessary to differentiate between the noise and the clean data. We note that there is a bimodal effect between the distribution of clean samples and the distribution of noisy samples. An example can be seen in Figure 4(b). Motivated by this, we fit the perplexity distribution over all training pairs with a two-component Gaussian Mixture Model (GMM):",
+        "bbox": [
+            507,
+            430,
+            884,
+            573
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np (P P L \\mid \\theta) = \\sum_ {k = 1} ^ {K} \\pi_ {k} \\phi (P P L \\mid k), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            564,
+            583,
+            880,
+            627
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\pi_{k}$  and  $\\phi(PPL \\mid k)$  are the mixture coefficient and the probability density of the  $k$ -th component, respectively. We optimize the GMM with the Expectation-Maximization algorithm (Dempster et al., 1977).",
+        "bbox": [
+            507,
+            638,
+            882,
+            717
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Based on the above hypothesis, we treat training pairs with higher  $PPL$  as noise and those with lower  $PPL$  as clean data. So the estimated clean flag can be calculated as follows:",
+        "bbox": [
+            507,
+            719,
+            880,
+            781
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {y} _ {i} = \\mathbb {I} \\left(p (\\kappa \\mid P P L _ {\\left(q _ {i}, d _ {i}, \\theta\\right)}) > \\lambda\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            569,
+            797,
+            880,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{y}_i\\in \\{1,0\\}$  denotes whether we estimate the pair  $(q_i,d_i)$  to be correctly matched or not,  $\\kappa$  is the GMM component with the lower mean,  $\\lambda$  is the threshold.  $p(\\kappa \\mid PPL_{(q_i,d_i,\\theta)})$  is the posterior probability over the component  $\\kappa$ , which can be intuitively understood as the correctly annotated con",
+        "bbox": [
+            507,
+            822,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "11441",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "fidence. We set  $\\lambda$  to 0.5 in all experiments. Note that before noise detection, the retriever should equip with preliminary retrieval capabilities. This can be achieved by initializing it with a strong unsupervised retriever or by pre-training it on the entire noise dataset.",
+        "bbox": [
+            112,
+            84,
+            489,
+            179
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Noise Correction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            191,
+            294,
+            205
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Next, we will introduce how to reduce the impact of noise pairs after obtaining the estimated flag set  $\\{\\hat{y}_i\\}_{i=1}^N$ . One quick fix is to discard the noise data directly, which is sub-optimal since it wastes the query data in noisy pairs. In this work, we adopt a self-ensemble teacher to provide rectified soft labels for noisy pairs. The teacher is an exponential moving average (EMA) of the retriever, and the retriever is trained with a weight-averaged consistency target on noisy data.",
+        "bbox": [
+            112,
+            212,
+            489,
+            372
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, given a retriever  $\\theta$ , the teacher  $\\theta^{*}$  is updated with an exponential moving average strategy as follows:",
+        "bbox": [
+            112,
+            373,
+            487,
+            420
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {t} ^ {*} = \\alpha \\theta_ {t - 1} ^ {\\star} + (1 - \\alpha) \\theta_ {t}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            203,
+            432,
+            485,
+            451
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha$  is a momentum coefficient. Only the parameters  $\\theta$  are updated by back-propagation.",
+        "bbox": [
+            112,
+            463,
+            489,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For a query  $q_{i}$  and the candidate document set  $D_{q_i}$ , where  $D_{q_i} = \\{d_{i,j}\\}_{j = 1}^m$  could consist of annotated documents, hard negatives and in-batch negatives, we first get teacher's and retriever's similarity scores, respectively. Then, the retriever  $\\theta$  is expected to keep consistent with its smooth teacher  $\\theta^{*}$ . To achieve this goal, we update the retriever  $\\theta$  by minimizing the KL divergence between the student's distribution and the teacher's distribution.",
+        "bbox": [
+            112,
+            495,
+            489,
+            639
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To be concrete, the similarity scores between  $q_{i}$  and  $D_{q_i}$  are normalized into the following distributions:",
+        "bbox": [
+            112,
+            640,
+            487,
+            686
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\phi} \\left(d _ {i, j} \\mid q _ {i}; D _ {q _ {i}}\\right) = \\frac {e ^ {\\tau f _ {\\phi} \\left(q _ {i} , d _ {i , j}\\right)}}{\\sum_ {j = 1} ^ {m} e ^ {\\tau f _ {\\phi} \\left(q _ {i} , d _ {i , j}\\right)}}, \\phi \\in \\left\\{\\theta , \\theta^ {*} \\right\\}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            696,
+            489,
+            747
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Then, the consistency loss  $L_{cons}$  can be written as:",
+        "bbox": [
+            112,
+            749,
+            487,
+            764
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {c o n s} = K L (p _ {\\theta} (. | q _ {i}; D _ {q _ {i}}), p _ {\\theta^ {*}} (. | q _ {i}; D _ {q _ {i}})), \\quad (7)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            131,
+            793,
+            487,
+            810
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $KL(\\cdot)$  is the KL divergence,  $p_{\\theta}(.|q_i;D_{q_i})$  and  $p_{\\theta^*}(.|q_i;D_{q_i})$  denote the conditional probabilities of candidate documents  $D_{q_i}$  by the retriever  $\\theta$  and the teacher  $\\theta^*$ , respectively.",
+        "bbox": [
+            112,
+            822,
+            487,
+            885
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For the estimated noisy pair, the teacher corrects the supervised signal into a soft label. For the",
+        "bbox": [
+            112,
+            887,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1 Noisy Pair Corrector (NPC)",
+        "text_level": 1,
+        "bbox": [
+            510,
+            83,
+            818,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Require: Retriever  $\\theta$ ; Noisy Training dataset  $C$ .",
+        "bbox": [
+            510,
+            103,
+            873,
+            118
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1: Warm up the retriever  $\\theta$",
+            "2: Initial EMA model  $\\theta^{*}$  with  $\\theta$",
+            "3: for  $i = 1 :$  num_epoch do",
+            "4: Calculate PPL of training pairs with random negatives using Eq.2;",
+            "5: Fit PPL distribution with GMM;",
+            "6: Get the estimated flag set  $\\{\\hat{y}_i\\}$  using Eq.4;",
+            "7: for  $i = 1 :$  num_batch do",
+            "8: Sample negatives with \"In-Batch Negative\" or \"Hard Negative\" strategy;",
+            "9: Calculate rectified soft labels with EMA model  $\\theta^{*}$",
+            "10: Train  $\\theta$  by optimizing Eq.8;",
+            "11: Update EMA model  $\\theta^{*}$  using Eq.5;",
+            "12: end for",
+            "13: end for"
+        ],
+        "bbox": [
+            514,
+            120,
+            884,
+            374
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "estimated clean pair, we calculate the contrastive loss and consistency loss. So the overall loss is formalized:",
+        "bbox": [
+            507,
+            403,
+            882,
+            450
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL = \\hat {y} _ {i} L _ {\\text {c o n t}} + L _ {\\text {c o n s}}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            611,
+            466,
+            880,
+            482
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{y}_i\\in \\{1,0\\}$  is estimated by the noise detection module.",
+        "bbox": [
+            507,
+            495,
+            882,
+            526
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Overall Procedure",
+        "text_level": 1,
+        "bbox": [
+            507,
+            539,
+            700,
+            554
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "NPC is a general training framework that can be easily applied to most retrieval methods. Under the classical training process of dense retrieval, We first warmup the retriever with the typical contrastive learning method to provide it with basic retrieval abilities, and then add the noise detection module before training each epoch and the noise correction module during training. The detail is presented in Algorithm 1.",
+        "bbox": [
+            507,
+            562,
+            882,
+            706
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            719,
+            655,
+            734
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            746,
+            623,
+            759
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To verify the effectiveness of NPC in robust dense retrieval, we conduct experiments on four commonly-used benchmarks, including Natural Questions (Kwiatkowski et al., 2019), Trivia QA (Joshi et al., 2017), StaQC (Yao et al., 2018) and SO-DS (Heyman and Van Cutsem, 2020).",
+        "bbox": [
+            507,
+            766,
+            882,
+            863
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "StaQC is a large dataset that collects real querycode pairs from Stack Overflow*. The dataset has",
+        "bbox": [
+            507,
+            864,
+            882,
+            896
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "\\*https://stackoverflow.com/",
+        "bbox": [
+            529,
+            904,
+            776,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "11442",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/fa72da3b4f438baa439f55cc23d4578fc68d5c88c479adada14949b5f8d94c77.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"3\">StaQC</td><td colspan=\"3\">SO-DS</td></tr><tr><td>R@3</td><td>R@10</td><td>MRR</td><td>R@3</td><td>R@10</td><td>MRR</td></tr><tr><td>BM25 desc (Heyman and Van Cutsem, 2020)</td><td>8.0</td><td>13.3</td><td>7.5</td><td>23.8</td><td>32.3</td><td>21.6</td></tr><tr><td>NBOW (Heyman and Van Cutsem, 2020)</td><td>10.9</td><td>16.6</td><td>9.5</td><td>27.7</td><td>38.0</td><td>24.7</td></tr><tr><td>USE (Heyman and Van Cutsem, 2020)</td><td>12.8</td><td>20.3</td><td>11.7</td><td>33.3</td><td>48.5</td><td>30.4</td></tr><tr><td>CodeBERT (Feng et al., 2020)</td><td>-</td><td>-</td><td>23.4</td><td>-</td><td>-</td><td>23.1</td></tr><tr><td>GraphCodeBERT (Guo et al., 2021)</td><td>-</td><td>-</td><td>24.1</td><td>-</td><td>-</td><td>25.2</td></tr><tr><td>CodeRetriever (In-Batch Negative) (Li et al., 2022)</td><td>-</td><td>-</td><td>25.5</td><td>-</td><td>-</td><td>27.1</td></tr><tr><td>CodeRetriever (Hard Negative) (Li et al., 2022)</td><td>-</td><td>-</td><td>24.6</td><td>-</td><td>-</td><td>31.8</td></tr><tr><td>UniXcoder (In-Batch Negative) (Guo et al., 2022)</td><td>29.98</td><td>47.47</td><td>28.04</td><td>31.90</td><td>51.21</td><td>28.29</td></tr><tr><td>UniXcoder (Hard Negative) (Guo et al., 2022)</td><td>31.18</td><td>48.38</td><td>28.63</td><td>33.42</td><td>53.37</td><td>29.97</td></tr><tr><td>NPC (In-Batch Negative)</td><td>33.07</td><td>50.35</td><td>30.39</td><td>35.58</td><td>54.54</td><td>30.96</td></tr><tr><td>NPC (Hard Negative)</td><td>34.38</td><td>52.20</td><td>31.36</td><td>38.00</td><td>56.51</td><td>32.49</td></tr></table>",
+        "bbox": [
+            137,
+            80,
+            858,
+            296
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Retrieval performance on StaQC and SO-DS, which are realistic-noisy datasets. The results of the first block are mainly borrowed from published papers (Heyman and Van Cutsem, 2020; Li et al., 2022). If the results are not provided, we mark them as “-”.",
+        "bbox": [
+            112,
+            306,
+            882,
+            350
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "been widely used on code summarization (Ped-damail et al., 2018) and code search (Heyman and Van Cutsem, 2020). SO-DS mines query-code pairs from the most upvoted Stack Overflow posts, mainly focuses on the data science domain. Following previous works (Heyman and Van Cutsem, 2020; Li et al., 2022), we resort to Recall of topk  $(\\mathbb{R}@\\mathbb{k})$  and Mean Reciprocal Rank (MRR) as the evaluation metric. StaQC and SO-DS are constructed automatically without human annotation. Therefore, there are numerous mismatched pairs in training data.",
+        "bbox": [
+            110,
+            367,
+            489,
+            558
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Natural Questions (NQ) collects real queries from the Google search engine. Each question is paired with an answer span and golden passages from the Wikipedia pages. Trivia QA (TQ) is a reading comprehension dataset authored by trivia enthusiasts. During the retrieval stage of both datasets, the objective is to identify positive passages from a large collection. Positive pairs in these datasets are assessed based on strict rule, i.e., whether passages contain answers or not (Karpukhin et al., 2020a). Consequently, we consider these datasets to be of high quality. Thus, we leverage them for simulation experiments to quantitatively analyze the impact of varying proportions of noise. Drawing inspiration from the setting in the noisy classification task (Han et al., 2018), we simulate the mismatched-pair noise by randomly pairing queries with unrelated documents.",
+        "bbox": [
+            110,
+            561,
+            489,
+            851
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            864,
+            346,
+            879
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "NPC is a general training paradigm that can be directly applied to almost all retrieval models. For",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "StaQC and SO-DS, we adopt UniXcoder (Guo et al., 2022) as our backbone, which is the SoTA model for code representation. Following Guo et al. (2022), we adopt the cosine distance as a similarity function and set temperature  $\\lambda$  to 20. We update model parameters using the Adam optimizer and perform early stopping on the development set. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 256, 5, and 10, respectively. In the \"Hard Negative\" setting, we adopt the same strategy as Li et al. (2022).",
+        "bbox": [
+            507,
+            367,
+            884,
+            543
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For NQ and TQ, we adopt BERT (Devlin et al., 2019) as our initial model. Following Karpukhin et al. (2020a), we adopt inner-product as the similarity function and set temperature  $\\lambda$  to 1. The max sequence length is 16 for query and 128 for passage. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 512, 10, and 40, respectively. We adopt \"BM25 Negative\" and \"Hard Negative\" strategies as described in the DPR toolkit. For a fair comparison, we implement DPR (Karpukhin et al., 2020a) with the same hyperparameters. All experiments are run on 8 NVIDIA Tesla A100 GPUs. The implementation of NPC is based on Huggingface (Wolf et al., 2020).",
+        "bbox": [
+            507,
+            545,
+            884,
+            770
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            781,
+            613,
+            796
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Results on StaQC and SO-DS: Table 1 shows the results on the realistic-noisy datasets StaQC and SO-DS. Both datasets contain a large number of real noise pairs. The first block shows the results of previous SoTA methods. BM25 desc is a traditional",
+        "bbox": [
+            507,
+            802,
+            882,
+            883
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$\\dagger$ https://github.com/facebookresearch/DPR",
+        "bbox": [
+            509,
+            891,
+            867,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "11443",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/508d1a78b29f8358f760df53799576b37b81fd7f214792d73f9106efa010ff5a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Noisy</td><td rowspan=\"2\">Methods</td><td colspan=\"4\">Natural Questions</td><td colspan=\"4\">Trivia QA</td></tr><tr><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td rowspan=\"2\"></td><td>BM25*</td><td>-</td><td>-</td><td>59.1</td><td>73.7</td><td>-</td><td>-</td><td>66.9</td><td>76.7</td></tr><tr><td>DPR*</td><td>-</td><td>-</td><td>78.4</td><td>85.4</td><td>-</td><td>-</td><td>79.4</td><td>85.0</td></tr><tr><td rowspan=\"11\">20</td><td>DPR (BM25 Negative)</td><td>27.07</td><td>47.79</td><td>63.36</td><td>75.69</td><td>35.73</td><td>52.88</td><td>64.05</td><td>74.16</td></tr><tr><td>coCondenser (BM25 Negative)</td><td>29.12</td><td>51.02</td><td>67.45</td><td>77.93</td><td>39.41</td><td>56.72</td><td>67.34</td><td>76.04</td></tr><tr><td>Co-teaching (BM25 Negative)</td><td>26.02</td><td>52.48</td><td>63.46</td><td>76.11</td><td>28.65</td><td>53.01</td><td>64.99</td><td>74.05</td></tr><tr><td>DPR-C (BM25 Negative)</td><td>43.69</td><td>66.62</td><td>79.07</td><td>86.12</td><td>52.10</td><td>70.52</td><td>79.05</td><td>85.08</td></tr><tr><td>NPC (BM25 Negative)</td><td>45.22</td><td>68.42</td><td>79.76</td><td>86.56</td><td>52.34</td><td>70.22</td><td>79.10</td><td>84.86</td></tr><tr><td>DPR (Hard Negative)</td><td>37.61</td><td>60.73</td><td>71.68</td><td>79.56</td><td>43.39</td><td>60.67</td><td>70.34</td><td>77.88</td></tr><tr><td>coCondenser (Hard Negative)</td><td>40.71</td><td>63.41</td><td>74.33</td><td>81.22</td><td>47.42</td><td>64.80</td><td>73.38</td><td>80.07</td></tr><tr><td>RocketQA (Hard Negative)</td><td>43.32</td><td>64.25</td><td>74.96</td><td>81.42</td><td>49.90</td><td>65.72</td><td>74.04</td><td>80.39</td></tr><tr><td>Co-teaching (Hard Negative)</td><td>31.78</td><td>56.32</td><td>66.12</td><td>77.56</td><td>33.28</td><td>57.29</td><td>66.50</td><td>75.62</td></tr><tr><td>DPR-C (Hard Negative)</td><td>51.66</td><td>72.40</td><td>81.50</td><td>87.80</td><td>55.35</td><td>72.36</td><td>80.33</td><td>85.34</td></tr><tr><td>NPC (Hard Negative)</td><td>51.85</td><td>73.06</td><td>82.47</td><td>87.80</td><td>56.03</td><td>72.54</td><td>80.59</td><td>85.58</td></tr><tr><td rowspan=\"11\">50</td><td>DPR (BM25 Negative)</td><td>16.12</td><td>33.88</td><td>49.70</td><td>63.38</td><td>20.09</td><td>34.63</td><td>47.42</td><td>61.04</td></tr><tr><td>coCondenser (BM25 Negative)</td><td>18.28</td><td>36.37</td><td>52.01</td><td>65.92</td><td>22.80</td><td>38.01</td><td>51.00</td><td>63.79</td></tr><tr><td>Co-teaching (BM25 Negative)</td><td>23.72</td><td>50.32</td><td>64.86</td><td>74.92</td><td>26.56</td><td>51.22</td><td>63.78</td><td>73.77</td></tr><tr><td>DPR-C (BM25 Negative)</td><td>41.29</td><td>65.21</td><td>78.48</td><td>85.70</td><td>49.61</td><td>68.81</td><td>78.00</td><td>84.23</td></tr><tr><td>NPC (BM25 Negative)</td><td>42.87</td><td>65.65</td><td>78.37</td><td>85.76</td><td>50.80</td><td>68.98</td><td>78.21</td><td>84.43</td></tr><tr><td>DPR (Hard Negative)</td><td>23.87</td><td>42.34</td><td>55.12</td><td>67.06</td><td>28.47</td><td>45.12</td><td>56.88</td><td>67.62</td></tr><tr><td>coCondenser (Hard Negative)</td><td>24.55</td><td>44.16</td><td>56.69</td><td>68.72</td><td>31.05</td><td>47.81</td><td>59.48</td><td>70.14</td></tr><tr><td>RocketQA (Hard Negative)</td><td>26.83</td><td>45.72</td><td>57.32</td><td>69.24</td><td>33.67</td><td>49.28</td><td>60.32</td><td>70.46</td></tr><tr><td>Co-teaching (Hard Negative)</td><td>30.12</td><td>55.94</td><td>65.81</td><td>76.90</td><td>31.85</td><td>55.37</td><td>65.29</td><td>75.02</td></tr><tr><td>DPR-C (Hard Negative)</td><td>48.87</td><td>70.52</td><td>81.44</td><td>87.17</td><td>53.07</td><td>70.36</td><td>79.02</td><td>84.69</td></tr><tr><td>NPC (Hard Negative)</td><td>48.81</td><td>70.60</td><td>81.17</td><td>87.20</td><td>53.09</td><td>70.27</td><td>79.31</td><td>84.96</td></tr></table>",
+        "bbox": [
+            137,
+            80,
+            858,
+            461
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Retrieval performance on Natural Questions and Trivia QA under the noise ratio of  $20\\%$ , and  $50\\%$ , respectively. The results of BM25* and DPR* are borrowed from Karpukhin et al. (2020a). If the results are not provided, we mark them as “-”.",
+        "bbox": [
+            112,
+            470,
+            882,
+            516
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "sparse retriever based on the exact term matching of queries and code descriptions. NBOW is an unsupervised retriever that leverages pre-trained word embedding of queries and code descriptions for retrieval. USE is a simple dense retriever based on transformer. CodeBERT, GraphCodeBERT are pre-trained models for code understanding using large-scale code corpus. CodeRetriever is a pretrained model dedicated to code retrieval, which is pre-trained with unimodal and bimodal contrastive learning on a large-scale corpus. UniXcoder is also a pretrained model that utilizes multi-modal data, including code, comment, and AST, for better code representation. The results are implemented by ourselves for a fair comparison with NPC. The bottom block shows the results of NPC using two negative sampling strategies.",
+        "bbox": [
+            110,
+            533,
+            487,
+            804
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "From the results, we can see that our proposed NPC consistently performs better than the evaluated models across all metrics. Compared with the strong baseline UniXcoder which ignores the mismatched-pair problem, NPC achieves a significant improvement with both \"in-batch negative\" and \"hard negative\" sampling strategies. It indi",
+        "bbox": [
+            112,
+            807,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "cates that the mismatched-pair noise greatly limits the performance of dense retrieval models, and NPC can mitigate this negative effect. We also show some noisy examples detected by NPC in Appendix A.",
+        "bbox": [
+            507,
+            533,
+            882,
+            612
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results on NQ and TQ: Table 2 shows the results on the synthetic-noisy datasets NQ and TQ under the noise ratio of  $20\\%$ , and  $50\\%$ . We compare NPC with BM25 (Yang et al., 2017) and DPR (Karpukhin et al., 2020a). BM25 is an unsupervised sparse retriever that is not affected by noisy data. DPR (Karpukhin et al., 2020a) is a widely used method for training dense retrievers. coCondenser (Gao and Callan, 2021b) leverage pre-training to enhance models' robustness. RocketQA (Qu et al., 2021) adopts a cross-encoder to filter false negatives in the \"Hard Negative\" strategy. Co-teaching (Han et al., 2018) uses the samples with small loss to iteratively train two networks, which is widely used in the noisy label classification task. We implement baselines using two negative sampling strategies. Besides, we evaluate DPR on clean datasets by discarding the synthetic-noisy pairs, denoted by DPR-C. DPR-C is a strong",
+        "bbox": [
+            507,
+            613,
+            884,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "11444",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/3ddc2d12852bd0861ed6e90ae8e0114ab6cb391746342161c0979ddeb51d9892.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"3\">Methods</td><td colspan=\"2\">NQ</td><td colspan=\"2\">StaQC</td></tr><tr><td>De</td><td>Co</td><td>HN</td><td>R@20</td><td>R@100</td><td>R@1</td><td>R@3</td></tr><tr><td>-</td><td>-</td><td>-</td><td>48.22</td><td>62.31</td><td>18.08</td><td>31.09</td></tr><tr><td>-</td><td>✓</td><td>-</td><td>55.90</td><td>69.33</td><td>18.51</td><td>31.01</td></tr><tr><td>✓</td><td>-</td><td>-</td><td>75.19</td><td>83.31</td><td>20.05</td><td>32.71</td></tr><tr><td>✓</td><td>✓</td><td>-</td><td>77.50</td><td>84.79</td><td>20.70</td><td>33.55</td></tr><tr><td>-</td><td>-</td><td>✓</td><td>54.63</td><td>65.54</td><td>18.66</td><td>31.74</td></tr><tr><td>-</td><td>✓</td><td>✓</td><td>58.63</td><td>69.06</td><td>19.35</td><td>32.09</td></tr><tr><td>✓</td><td>-</td><td>✓</td><td>77.59</td><td>85.03</td><td>20.93</td><td>33.55</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>80.07</td><td>85.89</td><td>21.93</td><td>34.51</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            497,
+            247
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "baseline that is not affected by mismatched pairs.",
+        "bbox": [
+            112,
+            313,
+            478,
+            330
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We can observe that (1) As the noise ratio increases, DPR, coCondenser, and RocketQA experience a significant decrease in performance. At a noise rate of  $50\\%$ , they perform worse than unsupervised BM25. (2) Despite Co-teaching having good noise resistance, its performance is still low. This indicates that methods for dealing with label noise in classification are not effective for retrieval. (3) NPC outperforms baselines by a large margin, with only a slight performance drop when the noise increases. Even comparing DPR-C, NPC still achieves competitive results.",
+        "bbox": [
+            112,
+            332,
+            489,
+            524
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            539,
+            226,
+            555
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Ablations of Noise Detection and Noise Correction: To get a better insight into NPC, we conduct ablation studies on the realistic-noisy dataset StaQC and the synthetic-noisy dataset NQ under the noise ratio of  $50\\%$ . The results are shown in Table 3. \"De\" and \"Co\" refer to noise detection and noise correction, respectively. \"HN\" indicates whether to perform \"Hard Negative\" strategy. For both synthetic noise and realistic noise, we can see that the noise detection module brings a significant gain, no matter which negative sampling strategy is used. Correction also enhances the robustness of the retriever since it provides rectified soft labels which can lead the model output to be smoother. The results show that combining the two obtains better performance compared with only using the detection module or correction module.",
+        "bbox": [
+            112,
+            563,
+            489,
+            835
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of Warmup Epoch: According to the foregoing, NPC starts by warming up. In Table 4, we pre-training the retriever on the noisy dataset for warming up, and show the performance of NPC with different various epoch numbers  $n$ . In this",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/aef375c9798e8d2b240b54c3c4ac14f24aa9a154520145f957590cbe3e522781.jpg",
+        "table_caption": [
+            "Table 3: Ablation studies on StaQC dev set and NQ dev set under noise ratio of  $50\\%$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Setting</td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td>n=1</td><td>50.58</td><td>69.93</td><td>79.87</td><td>84.96</td></tr><tr><td>n=5</td><td>50.03</td><td>69.64</td><td>80.17</td><td>85.76</td></tr><tr><td>n=10</td><td>50.07</td><td>69.93</td><td>80.07</td><td>85.89</td></tr><tr><td>n=20</td><td>38.09</td><td>60.31</td><td>72.00</td><td>80.07</td></tr></table>",
+        "bbox": [
+            527,
+            80,
+            867,
+            178
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Performance of NPC on NQ dev set with different warmup epoch number  $n$  .",
+        "bbox": [
+            507,
+            187,
+            882,
+            217
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/dcbc0024eec1373ac363c3fc24a44b0459ee0f3fd2ed17ccacebaf9d1af0612c.jpg",
+        "image_caption": [
+            "(a) Before warmup"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            221,
+            695,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/ff0d9d101b4a5bbe04c616c134f92ae47d66262ae98438926fa3ec8100d901c4.jpg",
+        "image_caption": [
+            "(b) After warmup"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            221,
+            870,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/baa431921633f57acfa52b1ff1f2b97d511463f29dc689800c8de230bbdbfa4c.jpg",
+        "image_caption": [
+            "(c) DPR"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            337,
+            695,
+            429
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/55b0dadef6128a055152c36cf1a1411946ef6047a5e8dac813c3bc5943bf141c.jpg",
+        "image_caption": [
+            "(d) NPC",
+            "Figure 4: Perplexity distribution of training pairs under different settings."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            337,
+            870,
+            429
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "experiment, we adopt \"Hard Negative\" sampling strategy. We find that NPC achieves good results when the warmup epoch is relatively small (1 - 10). However, when the warmup epoch is too large, the performance will degrade. We believe that a prolonged warmup causes overfitting to noise samples.",
+        "bbox": [
+            507,
+            510,
+            884,
+            607
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of Iterative Detection: In the training of NPC, we perform iterative noise detection every epoch. A straightforward approach is to detect the noise only once after warmup and fix the estimated flag set  $\\{\\hat{y}_i\\}$ . To study the effectiveness of iterative detection, we conducted an ablation study. The results are shown in Table 5. We can see that the model performance degrades after removing iterative detection.",
+        "bbox": [
+            507,
+            609,
+            882,
+            752
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Ablations of PPL: We distinguish noise pairs according to the perplexity between the annotated positive document and easy negatives. When calculating the perplexity, \"Hard Negative\" will cause trouble for detection. We construct ablation experiments to verify this, and the results are shown in Table 5. We can see that the perplexity with \"Hard Negative\" results in performance degradation.",
+        "bbox": [
+            507,
+            756,
+            884,
+            884
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Visualization of Perplexity Distribution: In Fig. 4, we illustrate the perplexity distribution of",
+        "bbox": [
+            507,
+            887,
+            882,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "11445",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/e88108902edeec047ff0d601d8f3bb3885458a2b1af5f05dc6f864df4aaf0dd9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Setting</td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td>NPC</td><td>50.07</td><td>69.93</td><td>80.07</td><td>85.89</td></tr><tr><td>-w/o iterative detection</td><td>47.29</td><td>68.39</td><td>78.79</td><td>85.38</td></tr><tr><td>-ppl with HN</td><td>42.81</td><td>65.06</td><td>75.22</td><td>83.09</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            482,
+            147
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Ablation studies of iterative noise detection and perplexity variants",
+        "bbox": [
+            112,
+            155,
+            485,
+            185
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/536d70a8d2cb88df20639cfd6d5ca8cc0333dd50b240e02dd21b4c40ea60155c.jpg",
+        "image_caption": [
+            "(a) w/o Hard Negative"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            193,
+            295,
+            282
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/83a8a9353639f66c54f9cffd67434fb4498a65290355222e85117a1823ac5622.jpg",
+        "image_caption": [
+            "(b) w/ Hard Negative",
+            "Figure 5: Retrieval performance of DPR and NPC on NQ dev set under different noise ratios."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            305,
+            192,
+            478,
+            282
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "training pairs before and after warmup, after training with DPR, and after training with NPC. The experiment is on NQ under the noise ratio of  $50\\%$ . We can see that the perplexity of most noisy pairs is larger than the clean pairs after warmup, which verifies our hypothesis in Sec. 3.1. Comparing Fig. 4(c) and Fig. 4(d), we find that the retriever trained with DPR will overfit the noise pairs. However, NPC enables the retriever to correctly distinguish clean and noisy pairs because it avoids the dominant effect of noise during network optimization.",
+        "bbox": [
+            112,
+            356,
+            487,
+            531
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Analysis of Generalizability Fig. 5 shows the performance of DPR and NPC under the noise ratio ranging from  $0\\%$  to  $80\\%$ . We can see that as the noise ratio increases, the performance degradation of DPR is much larger than that of NPC, which demonstrates the generalizability of NPC. Furthermore, even though NPC is designed to deal with mismatched-pair noise, it achieves competitive results when used in a noise-free setting.",
+        "bbox": [
+            112,
+            533,
+            489,
+            678
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            692,
+            270,
+            707
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.1 Dense Retrieval",
+        "text_level": 1,
+        "bbox": [
+            112,
+            720,
+            284,
+            734
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Dense retrieval has shown better performance than traditional sparse retrieval methods (Lee et al., 2019; Karpukhin et al., 2020a). The studies of dense retrieval can be divided into two categories, (1) unsupervised pre-training to get better initialization and (2) more effective fine-tuning on labeled data. In the first category, some researchers focus on how to generate contrastive pairs automatically from a large unsupervised corpus (Lee et al., 2019; Chang et al., 2019; Ma et al., 2022; Li et al., 2022). Another line of research enforces",
+        "bbox": [
+            112,
+            741,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the model to produce an information-rich CLS representation (Gao and Callan, 2021a,b; Lu et al., 2021). As for effective fine-tuning strategies (He et al., 2022b), recent studies show that negative sampling techniques are critical to the performance of dense retrievers. DPR (Karpukhin et al., 2020b) adopts in-batch negatives and BM25 negatives; ANCE (Xiong et al., 2021), RocketQA (Qu et al., 2021), and AR2 (Zhang et al., 2022a) improve the hard negative sampling by iterative replacement, denoising, and adversarial framework, respectively. Several works distill knowledge from ranker to retriever (Izacard and Grave, 2020; Yang and Seo, 2020; Ren et al., 2021; Zeng et al., 2022). Some studies incorporate lexical-aware sparse retrievers to convey lexical-related knowledge to dense retrievers, thereby enhancing the dense retriever's ability to recognize lexical matches (Shen et al., 2023; Zhang et al., 2023).",
+        "bbox": [
+            507,
+            84,
+            884,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Although the above methods have achieved promising results, they are highly dependent on correctly matched data, which is difficult to satisfy in real scenes. The mismatched-pair noise problem has seldom been considered. Besides, some studies utilize large-sized generative models (He et al., 2023) to guide retrievers, which achieve impressive performance without paired data (Sachan et al., 2022, 2021; Gao et al., 2022; He et al., 2022a). Although these models exhibit some robustness to noisy data, their success depends on the availability of strong generative models. Moreover, their applicability will be limited in domains where generative models do not perform well.",
+        "bbox": [
+            507,
+            392,
+            884,
+            617
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2 Denoising Techniques",
+        "text_level": 1,
+        "bbox": [
+            507,
+            634,
+            726,
+            650
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "One related task to our work is Noisy Label. Numerous methods have been proposed to solve this problem, and most of them focus on the classification task (Han et al., 2020). Some works design robust loss functions to mitigate label noise (Ghosh et al., 2017; Ma et al., 2020). Another line of work aims to identify noise from the training set with the memorization effect of neural networks (Silva et al., 2022; Liang et al., 2022; Bai et al., 2021).",
+        "bbox": [
+            507,
+            659,
+            882,
+            804
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "These studies mainly focus on classification. NPC studies the mismatched noise problem in dense retrieval rather than the noise in category annotations, which is more complex to handle. Several pre-training approaches noticed the problem of mismatched noisy pairs. ALIGN (Jia et al., 2021) and CLIP (Radford et al., 2021) claim that utilizing",
+        "bbox": [
+            507,
+            806,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "11446",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "large-scale image-text pairs can ignore the existence of noise. E5 (Wang et al., 2022) employs a consistency-based rule to filter the pre-training data. Although they slightly realized the existence of noisy pairs during pre-train, none of them give a specialized solution to solve it and extensively explored the characteristics of noisy text pairs. Some recent works (Huang et al., 2021; Han et al., 2023) study the noisy correspondence problem in cross-modal retrieval. Although the \"mismatched-pair noisy\" problem in cross-modal retrieval and text retrieval shares similarities, the specific settings and methods used in these two areas are notably distinct. It is challenging to directly apply these cross-modal retrieval works to document and code retrieval. Our NPC is the first systematic work to explore mismatched-pair noise in document/code retrieval.",
+        "bbox": [
+            112,
+            84,
+            492,
+            374
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            388,
+            247,
+            404
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper explores a neglected problem in dense retrieval, i.e., mismatched-pair noise. To solve this problem, we propose a generalized Noisy Pair Corrector(NPC) framework, which iteratively detects noisy pairs per epoch based on the perplexity and then provides rectified soft labels via an EMA model. The experimental results and analysis demonstrate the effectiveness of NPC in effectively handling both synthetic and realistic mismatched-pair noise.",
+        "bbox": [
+            112,
+            414,
+            489,
+            575
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            590,
+            220,
+            606
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work mainly focuses on training the dense retrieval models with mismatched noise. There may be two possible limitations in our study.",
+        "bbox": [
+            112,
+            618,
+            487,
+            665
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1) Due to the limited computing infrastructure, we only verified the robustness performance of NPC based on the classical retriever training framework. We leave experiments to combine NPC with more effective retriever training methods such as distillation (Ren et al., 2021), AR2 (Zhang et al., 2022a), as future work.",
+            "2) Mismatched-pair noise may also exist in other tasks, such as recommender systems. We will consider extending NPC to more tasks."
+        ],
+        "bbox": [
+            112,
+            668,
+            489,
+            829
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            112,
+            844,
+            278,
+            860
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by the Fundamental Research Funds for the Central Universities under Grant 1082204112364 and the Key Program of the",
+        "bbox": [
+            112,
+            871,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "National Science Foundation of China under Grant 61836006.",
+        "bbox": [
+            507,
+            84,
+            882,
+            115
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            143,
+            608,
+            156
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yingbin Bai, Erkun Yang, Bo Han, Yanhua Yang, Jia-tong Li, Yinian Mao, Gang Niu, and Tongliang Liu. 2021. Understanding and improving early stopping for learning with noisy labels. In NIPS.",
+            "Dan Brickley, Matthew Burgess, and Natasha Noy. 2019. Google dataset search: Building a search engine for datasets in an open web ecosystem. In WWW.",
+            "Wei-Cheng Chang, X Yu Felix, Yin-Wen Chang, Yiming Yang, and Sanjiv Kumar. 2019. Pre-training tasks for embedding-based large-scale retrieval. In International Conference on Learning Representations.",
+            "Xuanang Chen, Jian Luo, Ben He, Le Sun, and Yingfei Sun. 2022. Towards robust dense retrieval via local ranking alignment. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI, pages 1980-1986.",
+            "Arthur P Dempster, Nan M Laird, and Donald B Rubin. 1977. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers).",
+            "Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. Codebert: A pre-trained model for programming and natural languages. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, Online Event, 16-20 November 2020.",
+            "Luyu Gao and Jamie Callan. 2021a. Is your language model ready for dense representation fine-tuning? arXiv preprint arXiv:2104.08253.",
+            "Luyu Gao and Jamie Callan. 2021b. Unsupervised corpus aware language model pre-training for dense passage retrieval. arXiv preprint arXiv:2108.05540.",
+            "Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2022. Precise zero-shot dense retrieval without relevance labels. CoRR.",
+            "Aritra Ghosh, Himanshu Kumar, and P Shanti Sastry. 2017. Robust loss functions under label noise for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence.",
+            "Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin. 2022. Unixcoder: Unified cross-modal pre-training for code representation. In ACL."
+        ],
+        "bbox": [
+            509,
+            164,
+            885,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "11447",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Daya Guo, Shuo Ren, Shuai Lu, Zhangyin Feng, Duyu Tang, Shujie Liu, Long Zhou, Nan Duan, Alexey Svyatkovskiy, Shengyu Fu, Michele Tufano, Shao Kun Deng, Colin B. Clement, Dawn Drain, Neel Sundaresan, Jian Yin, Daxin Jiang, and Ming Zhou. 2021. Graphcodebert: Pre-training code representations with data flow. In International Conference on Learning Representations.",
+            "Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W Tsang, James T Kwok, and Masashi Sugiyama. 2020. A survey of label-noise representation learning: Past, present and future. arXiv preprint arXiv:2011.04406.",
+            "Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. 2018. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NIPS.",
+            "Haochen Han, Kaiyao Miao, Qinghua Zheng, and Minnan Luo. 2023. Noisy correspondence learning with meta similarity correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7517-7526.",
+            "Xingwei He, Yeyun Gong, A Jin, Weizhen Qi, Hang Zhang, Jian Jiao, Bartuer Zhou, Biao Cheng, Siu Ming Yiu, Nan Duan, et al. 2022a. Metric-guided distillation: Distilling knowledge from the metric to ranker and retriever for generative commonsense reasoning. arXiv preprint arXiv:2210.11708.",
+            "Xingwei He, Yeyun Gong, A Jin, Hang Zhang, Anlei Dong, Jian Jiao, Siu Ming Yiu, Nan Duan, et al. 2022b. Curriculum sampling for dense retrieval with document expansion. arXiv preprint arXiv:2212.09114.",
+            "Xingwei He, Zhenghao Lin, Yeyun Gong, Alex Jin, Hang Zhang, Chen Lin, Jian Jiao, Siu Ming Yiu, Nan Duan, Weizhu Chen, et al. 2023. Annollm: Making large language models to be better crowdsourced annotators. arXiv preprint arXiv:2303.16854.",
+            "Geert Heyman and Tom Van Cutsem. 2020. Neural code search revisited: Enhancing code snippet retrieval through natural language intent. arXiv preprint arXiv:2008.12193.",
+            "Zhenyu Huang, Guocheng Niu, Xiao Liu, Wenbiao Ding, Xinyan Xiao, Hua Wu, and Xi Peng. 2021. Learning with noisy correspondence for cross-modal matching. Advances in Neural Information Processing Systems, 34:29406-29419.",
+            "Gautier Izacard and Edouard Grave. 2020. Distilling knowledge from reader to retriever for question answering. arXiv preprint arXiv:2012.04584.",
+            "Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. 2021. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547.",
+            "Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In ACL.",
+            "Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020a. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906.",
+            "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick S. H. Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020b. Dense passage retrieval for open-domain question answering. In EMNLP.",
+            "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur P. Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: a benchmark for question answering research. Trans. Assoc. Comput. Linguistics, 7:452-466.",
+            "Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.",
+            "Xiaonan Li, Yeyun Gong, Yelong Shen, Xipeng Qiu, Hang Zhang, Bolun Yao, Weizhen Qi, Daxin Jiang, Weizhu Chen, and Nan Duan. 2022. Coderetriever: Unimodal and bimodal contrastive learning. arXiv preprint arXiv:2201.10866.",
+            "Kevin J Liang, Samrudhdhi B. Rangrej, Vladan Petrovic, and Tal Hassner. 2022. Few-shot learning with noisy labels. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9079-9088.",
+            "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.",
+            "Shuqi Lu, Di He, Chenyan Xiong, Guolin Ke, Waleed Malik, Zhicheng Dou, Paul Bennett, Tie-Yan Liu, and Arnold Overwijk. 2021. Less is more: Pretrain a strong siamese encoder for dense text retrieval using a weak decoder. In Empirical Methods in Natural Language Processing.",
+            "Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. 2020. Normalized loss functions for deep learning with noisy labels. In ICML."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "11448",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xinyu Ma, Jiafeng Guo, Ruqing Zhang, Yixing Fan, and Xueqi Cheng. 2022. Pre-train a discriminative text encoder for dense retrieval via contrastive span prediction. In SIGIR '22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, July 11 - 15, 2022.",
+            "Jayavardhan Reddy Peddamail, Ziyu Yao, Zhen Wang, and Huan Sun. 2018. A comprehensive study of staqc for deep code summarization. In KDD.",
+            "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. Rocketqa: An optimized training approach to dense passage retrieval for open-domain question answering. In NAACL-HLT.",
+            "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In ICML.",
+            "Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Qiaoqiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021. Rocketqav2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2825-2835.",
+            "Devendra Singh Sachan, Mike Lewis, Dani Yogatama, Luke Zettlemoyer, Joelle Pineau, and Manzil Zaheer. 2022. Questions are all you need to train a dense passage retriever. CoRR.",
+            "Devendra Singh Sachan, Mostofa Patwary, Mohammad Shoeybi, Neel Kant, Wei Ping, William L. Hamilton, and Bryan Catanzaro. 2021. End-to-end training of neural retrievers for open-domain question answering. In ACL/IJCNLP.",
+            "Tao Shen, Xiubo Geng, Chongyang Tao, Can Xu, Guodong Long, Kai Zhang, and Daxin Jiang. 2023. Unifier: A unified retriever for large-scale retrieval. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4787-4799.",
+            "Yelong Shen, Xiaodong He, Jianfeng Gao, Li Deng, and Grégoire Mesnil. 2014. Learning semantic representations using convolutional neural networks for web search. In Proceedings of the 23rd international conference on world wide web, pages 373-374.",
+            "Amila Silva, Ling Luo, Shanika Karunasekera, and Christopher Leckie. 2022. Noise-robust learning from multiple unsupervised sources of inferred labels. In AAAI Conference on Artificial Intelligence.",
+            "Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. ArXiv, abs/2212.03533."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In EMNLP.",
+            "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations.",
+            "Peilin Yang, Hui Fang, and Jimmy Lin. 2017. Anserini: Enabling the use of lucene for information retrieval research. In Proceedings of the 40th international ACM SIGIR conference on research and development in information retrieval.",
+            "Sohee Yang and Minjoon Seo. 2020. Is retriever merely an approximator of reader? arXiv preprint arXiv:2010.10999.",
+            "Ziyu Yao, Daniel S Weld, Wei-Peng Chen, and Huan Sun. 2018. Staqc: A systematically mined questioncode dataset from stack overflow. In WWW.",
+            "Hansi Zeng, Hamed Zamani, and Vishwa Vinay. 2022. Curriculum learning for dense retrieval distillation. In SIGIR '22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, July 11 - 15, 2022.",
+            "Hang Zhang, Yeyun Gong, Yelong Shen, Weisheng Li, Jiancheng Lv, Nan Duan, and Weizhu Chen. 2021. Poolingformer: Long document modeling with pooling attention. In International Conference on Machine Learning, pages 12437-12446. PMLR.",
+            "Hang Zhang, Yeyun Gong, Yelong Shen, Jiancheng Lv, Nan Duan, and Weizhu Chen. 2022a. Adversarial retriever-ranker for dense text retrieval. In International Conference on Learning Representations.",
+            "Kai Zhang, Chongyang Tao, Tao Shen, Can Xu, Xiubo Geng, Binxing Jiao, and Daxin Jiang. 2023. Led: Lexicon-enlightened dense retriever for large-scale retrieval. In Proceedings of the ACM Web Conference 2023, pages 3203-3213.",
+            "Shunyu Zhang, Yaobo Liang, Ming Gong, Daxin Jiang, and Nan Duan. 2022b. Multi-view document representation learning for open-domain dense retrieval. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            846
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11449",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Qualitative Analysis",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            327,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 9 lists some mismatched pairs detected by NPC in StaQC training set. We can see that these mismatched pairs are almost irrelevant and can be correctly detected by NPC. These examples are not well aligned, mainly due to the low-quality answers of the open community (cases 2 and 4), inappropriate data preprocessing in the collection phase (cases 2 and 3), and other reasons. It is well known that collecting and cleaning training data is expensive and complex work. Automatically constructed datasets in real-world applications often contain such mismatched-pair noise. Our method can mitigate the impact caused by such noise during training.",
+        "bbox": [
+            112,
+            116,
+            489,
+            341
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Statistics of Datasets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            362,
+            331,
+            378
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The statistics of datasets are shown in Table 6.",
+        "bbox": [
+            112,
+            395,
+            458,
+            410
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/7887df5e9a08c0bb7e0248bca0fe09db2d200870ac8ff813537c70607f42de1e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Train</td><td>Dev</td><td>Test</td><td>Corpus size</td></tr><tr><td>StaQC</td><td>203.7K</td><td>2.6K</td><td>2.7K</td><td>14.6K</td></tr><tr><td>SO-DS</td><td>12.1K</td><td>0.9K</td><td>1.1K</td><td>12.1K</td></tr><tr><td>NQ</td><td>79.2K</td><td>8,8K</td><td>3.6K</td><td>21 M</td></tr><tr><td>TQ</td><td>78.8K</td><td>8.8k</td><td>11.3K</td><td>21 M</td></tr></table>",
+        "bbox": [
+            131,
+            425,
+            470,
+            513
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Discussion about Perplexity",
+        "text_level": 1,
+        "bbox": [
+            114,
+            596,
+            391,
+            614
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We calculate the perplexity between the annotated document and easy negative documents during noise detection. We emphasize that the negative documents are randomly selected from the document collection  $\\mathbb{D}$ . It is not suitable to adopt \"Hard Negative\" sampling strategy when calculating the perplexity. Although hard negatives are important to train a strong dense retriever, they will cause trouble during noise detection. Specifically, it is expected that the retriever is confused only between false positive and negative documents and can confidently distinguish true positive and negative documents. But if we adopt \"Hard Negative\" when calculating the perplexity, the retriever will also be confused between true positive and hard negative documents, which will affect noise detection. We construct ablation experiments to verify this, and the results are shown in Table 5.",
+        "bbox": [
+            112,
+            629,
+            489,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Integration with stronger methods",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            845,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We conducted comprehensive experiments that integrated NPC into coCondenser and RocketQAv2. The subsequent experiments were conducted on the NQ dataset with a  $50\\%$  noise ratio. We first combined NPC with coCondenser which is a pretrained model specialized for dense retrieval tasks. The results are shown in Table 7",
+        "bbox": [
+            507,
+            109,
+            885,
+            219
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/6f0713bfef298a91afdd6eede55bf7315a19f503cab5861c447e3ec519d4d314.jpg",
+        "table_caption": [
+            "Table 6: The statistics of datasets. Corpus size means the size of document corpus for evaluation."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td colspan=\"5\">BM25 Negative</td></tr><tr><td>coCondenser</td><td>18.28</td><td>36.37</td><td>52.01</td><td>65.92</td></tr><tr><td>coCondenser-C</td><td>44.75</td><td>68.91</td><td>80.89</td><td>87.31</td></tr><tr><td>coCondenser+NPC</td><td>47.31</td><td>70.38</td><td>81.58</td><td>87.38</td></tr><tr><td colspan=\"5\">Hard Negative</td></tr><tr><td>coCondenser</td><td>24.55</td><td>44.16</td><td>56.69</td><td>68.72</td></tr><tr><td>coCondenser-C</td><td>49.31</td><td>71.99</td><td>82.41</td><td>88.38</td></tr><tr><td>coCondenser+NPC</td><td>50.66</td><td>72.42</td><td>82.64</td><td>88.31</td></tr></table>",
+        "bbox": [
+            514,
+            229,
+            884,
+            357
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "It's evident that NPC significantly enhances the robustness of coCondenser against noise associated with mismatched pairs. This observation underscores the compatibility between NPC and pretrained dense retrievers.",
+        "bbox": [
+            507,
+            414,
+            882,
+            494
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we combined NPC with RocketQAv2 which adopted a cross-encoder as a teacher and dynamically distilled knowledge to the dense retriever. The results are shown in Table 8:",
+        "bbox": [
+            507,
+            495,
+            882,
+            558
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/c14a2b6e50fa00f68f3a969a6eada738eb54059077ab275b005b8c7a1a445dd7.jpg",
+        "table_caption": [
+            "Table 7: Retrieval performance on NQ after combining NPC with coCondenser."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td>RocketQAv2</td><td>32.30</td><td>51.37</td><td>62.19</td><td>70.79</td></tr><tr><td>RocketQAv2-C</td><td>52.63</td><td>73.51</td><td>83.21</td><td>88.71</td></tr><tr><td>RocketQAv2+NPC</td><td>52.59</td><td>73.83</td><td>83.32</td><td>88.69</td></tr></table>",
+        "bbox": [
+            514,
+            567,
+            880,
+            627
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Retrieval performance on NQ after combining NPC with RocketQAv2.",
+        "bbox": [
+            507,
+            636,
+            880,
+            665
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To combine NPC with RocketQAv2, we integrate the noise detection and the correction modules in each training epoch of RocketQAv2. From the table, we can find that although RocketQAv2 uses a powerful cross-encoder as a teacher, it is still limited by the noise of the training data and shows low performance. NPC can effectively harmonize with RocketQAv2 to mitigate the problems caused by mismatched pair noise. We will add these experiment results to the next version.",
+        "bbox": [
+            507,
+            684,
+            882,
+            844
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "11450",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/013248572b5114b398641d8ed44af23db091529f116f4c49f6b702063a6f4128.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Question</td><td>Code</td></tr><tr><td>1</td><td>Split words in a nested list into letters</td><td>» [list(l[0]) for l in mylist]</td></tr><tr><td>2</td><td>Dictionary in python problem</td><td>» s = problem.getSuccessors( getStartState())</td></tr><tr><td>3</td><td>Find the Common first name from Django Auth user Model</td><td>» import operator</td></tr><tr><td>4</td><td>Find all text files not containing some text string</td><td>» lst = [1,2,4,6,3,8,0,5] \n» for n in lst[:]: \n» if n % 2 == 0: \n»&gt;&gt;&gt; lst.remove(n) \n» lst</td></tr></table>",
+        "bbox": [
+            114,
+            395,
+            882,
+            576
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 9: Some noisy pairs detected by NPC in StaQC training set.",
+        "bbox": [
+            273,
+            585,
+            722,
+            600
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "11451",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Noisy Pair Corrector for Dense Retrieval/ff6504db-4532-48dc-8468-050b9abe58d7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:921869deba6db20b6db9ea4928326374e8e8b483585df1e9ef38de967e8314c9
+size 1539262

2023/Noisy Pair Corrector for Dense Retrieval/full.md

@@ -0,0 +1,388 @@
+# Noisy Pair Corrector for Dense Retrieval
+
+Hang Zhang $^{1}$ , Yeyun Gong, Xingwei He $^{†2}$ , Dayiheng Liu $^{1}$ , Daya Guo $^{3}$ , Jiancheng Lv $^{†1}$ , Jian Guo $^{†4}$
+
+<sup>1</sup> College of Computer Science, Sichuan University
+
+<sup>1</sup> Engineering Research Center of Machine Learning and Industry Intelligence  
+<sup>2</sup> The University of Hong Kong <sup>3</sup> Sun Yat-sen University <sup>4</sup> IDEA Research, China hangzhang_scu@foxmail.com, hexingwei15@gmail.com
+
+# Abstract
+
+Most dense retrieval models contain an implicit assumption: the training query-document pairs are exactly matched. Since it is expensive to annotate the corpus manually, training pairs in real-world applications are usually collected automatically, which inevitably introduces mismatched-pair noise. In this paper, we explore an interesting and challenging problem in dense retrieval, how to train an effective model with mismatched-pair noise. To solve this problem, we propose a novel approach called Noisy Pair Corrector (NPC), which consists of a detection module and a correction module. The detection module estimates noise pairs by calculating the perplexity between annotated positive and easy negative documents. The correction module utilizes an exponential moving average (EMA) model to provide a soft supervised signal, aiding in mitigating the effects of noise. We conduct experiments on text-retrieval benchmarks Natural Question and TriviaQA, code-search benchmarks StaQC and SO-DS. Experimental results show that NPC achieves excellent performance in handling both synthetic and realistic noise.
+
+# 1 Introduction
+
+With the advancements in pre-trained language models (Devlin et al., 2019; Liu et al., 2019), dense retrieval has developed rapidly in recent years. It is essential to many applications including search engine (Brickley et al., 2019), open-domain question answering (Karpukhin et al., 2020a; Zhang et al., 2021), and code intelligence (Guo et al., 2021). A typical dense retrieval model maps both queries and documents into a low-dimensional vector space and measures the relevance between them by the similarity between their respective representations (Shen et al., 2014). During training, the model utilizes query-document pairs as labelled
+
+![](images/08f56f5cf93fcf9e6b5d29530e4fe73cb11044c730ce1906f44e21f465f88049.jpg)  
+Figure 1: Two examples from StaQC training set. In the bottom example, the given code is mismatched with the query, since it can not answer the query.
+
+training data (Xiong et al., 2021) and samples negative documents for each pair. Then the model learns to minimize the contrastive loss for obtaining a good representation ability (Zhang et al., 2022b; Qu et al., 2021).
+
+Recent studies on dense retrieval have achieved promising results with hard negative mining (Xiong et al., 2021), pretraining (Gao and Callan, 2021a), distillation (Yang and Seo, 2020), and adversarial training (Zhang et al., 2022a). All methods contain an implicit assumption: each query is precisely aligned with the positive documents in the training set. In practical applications, this assumption becomes challenging to satisfy, particularly when the corpora is automatically collected from the internet. In such scenarios, it is inevitable that the training data will contain mismatched pairs, incorporating instances such as user mis-click noise in search engines or low-quality reply noise in Q&A communities. As shown in Fig. 1, the examples are from StaQC benchmark (Yao et al., 2018), which is automatically collected from StackOverflow. The document, i.e., code solution, can not answer the query but is incorrectly annotated as a positive doc
+
+![](images/cebbd4ae4812b6f292e3700c18d2d75a0c56e784203664521ebd64d599a63c65.jpg)  
+Figure 2: Effect of matched & mismatched pair for training. Green objects refer to annotated pairs, while pentagram and triangle are actually aligned pairs. In the left case, retrieval models are required to push the query with true-positive document (TP Doc) together and pull the query with true-negative documents (TN Doc) apart. In the right case, the retrieval models are misled by the mismatched data pair, where the false-positive document (FP Doc) and the false-negative document (FN Doc) are wrongly pulled and pushed, respectively.
+
+uments. Such noisy pairs are widely present in automatically constructed datasets, which ultimately impact the performance of dense retrievers.
+
+To train robust dense retrievers, previous works have explored addressing various types of noise. For example, RocketQA (Qu et al., 2021) and AR2 (Zhang et al., 2022a) mitigate the false-negative noise with a cross-encoder filter and distillation, respectively; coCondenser (Gao and Callan, 2021b) reduce the noise during fine-tuning with pre-training technique; RoDR (Chen et al., 2022) deal with query spelling noise with local ranking alignment. However, mismatched-pair noise (false positive problem) in dense retrieval has not been well studied. As shown in Fig. 2, mismatched-pair noise will mislead the retriever to update in the opposite direction.
+
+Based on these observations, we propose a Noisy Pair Corrector (NPC) framework to solve the false-positive problem. NPC consists of noise detection and correction modules. At each epoch, the detection module estimates whether a query-document pair is mismatched by the perplexity between the annotated document and easy negative documents. Then the correction module provides a soft supervised signal for both estimated noisy data and clean data via an exponential moving average (EMA) model. Both modules are plug-and-play, which means NPC is a general training paradigm that can be easily applied to almost all retrieval models.
+
+The contributions of this paper are as follows: (1)
+
+We reveal and extensively explore a long-neglected problem in dense retrieval, i.e., mismatched-pair noise, which is ubiquitous in the real world. (2) We propose a simple yet effective method for training dense retrievers with mismatched-pair noise. (3) Extensive experiments on four datasets and comprehensive analyses verify the effectiveness of our method against synthetic and realistic noise. Code is available at https://github.com/hangzhang-nlp/NPC.
+
+# 2 Preliminary
+
+Before describing our model in detail, we first introduce the basic elements of dense retrieval, including problem definition, model architecture, and model training.
+
+Given a query  $q$ , and a document collection  $\mathbb{D}$ , dense retrieval aims to find document  $d^{+}$  relevant to  $q$  from  $\mathbb{D}$ . The training set consists of a collection of query-document pairs, donated as  $C = \{(q_{1}, d_{1}^{+}), \dots, (q_{N}, d_{N}^{+})\}$ , where  $N$  is the data size. Typical dense retrieval models adopt a dual encoder architecture to map queries and documents into a dense representation space. Then the relevance score  $f(q, d)$  of query  $q$  and document  $d$  can be calculated with their dense representations:
+
+$$
+f _ {\theta} (q, d) = \operatorname {s i m} \left(E (q; \theta), E (d; \theta)\right), \tag {1}
+$$
+
+where  $E(\cdot ;\theta)$  denotes the encoder module parameterized with  $\theta$  and sim is the similarity function, e.g., euclidean distance, cosine distance, inner-product. Existing methods generally leverage the approximate nearest neighbor technique (ANN) (Johnson et al., 2019) for efficient search.
+
+For training dense retrievers, conventional approaches leverage contrastive learning techniques (Karpukhin et al., 2020a; Zhang et al., 2022b). Given a training pair  $(q_{i},d_{i}^{+})\in C$ , these methods sample  $m$  negative documents  $\{d_{i,1}^{-},\dots,d_{i,m}^{-}\}$  from a large document collection  $\mathbb{D}$ . The retriever's objective is to minimize the contrastive loss, pushing the similarity of positive pairs higher than negative pairs. Previous work (Xiong et al., 2021) has verified the effectiveness of the negative sampling strategy. Two commonly employed strategies are "In-Batch Negative" and "Hard Negative" (Karpukhin et al., 2020a; Qu et al., 2021).
+
+The above training paradigm assumes that the query-document pair  $(q_{i},d_{i}^{+})$  in training set  $C$  is correctly aligned. However, this assumption is difficult to satisfy in real-world applications (Qu et al.,
+
+![](images/cbd09a5537eb88688a21601df295fdc6babb3d9c3c18cd09702026ac1f3b093c.jpg)  
+(a) Noise Detection  
+(b) Noise Correction  
+Figure 3: Overview of noise detection and noise correction. (a) Procedure of Noise Detection. At each epoch, we first calculate the perplexity of all training query-document pairs using the retriever  $\theta$ ; next fit the perplexity distribution with Gaussian Mixture Model to get the correctly matched probability of each pair; finally estimate the flag set  $\{\hat{y}_i\}_{i=1}^N$  by setting the threshold. (b) Framework of Noise Correction. Given a batch of data pairs, where  $d_{i,1}^{-}$  is the hard negative of  $q_i$  and  $\{q_3, d_3^{+}\}$  is the estimated noisy pair, the retriever  $\theta$  and teacher  $\theta^*$  compute similarity matrices  $S_\theta$  and  $S_{\theta^*}$  for all queries and documents, respectively. The retriever learns to minimize (1)  $L_{cont}$ : the negative likelihood probability of true positive documents; (2)  $L_{cons}$ : the KL divergence between  $S_\theta$  and the rectified soft label  $S_{\theta^*}$  after normalization.
+
+2021; Li et al., 2022; Wang et al., 2022). In practice, most training data pairs are collected automatically without manual inspection, such as inevitably leading to the inclusion of some mismatched pairs.
+
+# 3 Method
+
+We propose NPC framework to learn retrievers with mismatched-pair noise. As shown in Fig. 3, NPC consists of two parts: (a) the noise detection module as described in Sec. 3.1, and (b) the noise correction module as described in Sec. 3.2.
+
+# 3.1 Noise Detection
+
+The noise detection module is meant to detect mismatched pairs in the training set. We hypothesize that: dense retrievers will first learn to distinguish correctly matched pairs from easy negatives, and then gradually overfit the mismatched pairs. Therefore, we determine whether a training pair is mismatched by the perplexity between the annotated document and easy negative documents.
+
+Specifically, given a retriever  $\theta$  equipped with preliminary retrieval capabilities, and an uncertain pair  $(q_{i},d_{i})$ , we calculate the perplexity as follows:
+
+$$
+P P L _ {\left(q _ {i}, d _ {i}, \theta\right)} = - \log \frac {e ^ {\tau f _ {\theta} \left(q _ {i} , d _ {i}\right)}}{e ^ {\tau f _ {\theta} \left(q _ {i} , d _ {i}\right)} + \sum_ {j = 1} ^ {m} e ^ {\tau f _ {\theta} \left(q _ {i} , d _ {i , j} ^ {-}\right)}}, \tag {2}
+$$
+
+where  $\tau$  is a hyper-parameter,  $d_{i,j}^{-}$  is the negative document randomly sampled from the document collection  $\mathbb{D}$ . Note that  $d_{i,j}^{-}$  is a randomly selected negative document, not a hard negative. We discuss
+
+this further in Appendix C. In practice, we adopt the "In-Batch Negative" strategy for efficiency.
+
+After obtaining the perplexity of each pair, an automated method is necessary to differentiate between the noise and the clean data. We note that there is a bimodal effect between the distribution of clean samples and the distribution of noisy samples. An example can be seen in Figure 4(b). Motivated by this, we fit the perplexity distribution over all training pairs with a two-component Gaussian Mixture Model (GMM):
+
+$$
+p (P P L \mid \theta) = \sum_ {k = 1} ^ {K} \pi_ {k} \phi (P P L \mid k), \tag {3}
+$$
+
+where  $\pi_{k}$  and  $\phi(PPL \mid k)$  are the mixture coefficient and the probability density of the  $k$ -th component, respectively. We optimize the GMM with the Expectation-Maximization algorithm (Dempster et al., 1977).
+
+Based on the above hypothesis, we treat training pairs with higher  $PPL$  as noise and those with lower  $PPL$  as clean data. So the estimated clean flag can be calculated as follows:
+
+$$
+\hat {y} _ {i} = \mathbb {I} \left(p (\kappa \mid P P L _ {\left(q _ {i}, d _ {i}, \theta\right)}) > \lambda\right), \tag {4}
+$$
+
+where  $\hat{y}_i\in \{1,0\}$  denotes whether we estimate the pair  $(q_i,d_i)$  to be correctly matched or not,  $\kappa$  is the GMM component with the lower mean,  $\lambda$  is the threshold.  $p(\kappa \mid PPL_{(q_i,d_i,\theta)})$  is the posterior probability over the component  $\kappa$ , which can be intuitively understood as the correctly annotated con
+
+fidence. We set  $\lambda$  to 0.5 in all experiments. Note that before noise detection, the retriever should equip with preliminary retrieval capabilities. This can be achieved by initializing it with a strong unsupervised retriever or by pre-training it on the entire noise dataset.
+
+# 3.2 Noise Correction
+
+Next, we will introduce how to reduce the impact of noise pairs after obtaining the estimated flag set  $\{\hat{y}_i\}_{i=1}^N$ . One quick fix is to discard the noise data directly, which is sub-optimal since it wastes the query data in noisy pairs. In this work, we adopt a self-ensemble teacher to provide rectified soft labels for noisy pairs. The teacher is an exponential moving average (EMA) of the retriever, and the retriever is trained with a weight-averaged consistency target on noisy data.
+
+Specifically, given a retriever  $\theta$ , the teacher  $\theta^{*}$  is updated with an exponential moving average strategy as follows:
+
+$$
+\theta_ {t} ^ {*} = \alpha \theta_ {t - 1} ^ {\star} + (1 - \alpha) \theta_ {t}, \tag {5}
+$$
+
+where  $\alpha$  is a momentum coefficient. Only the parameters  $\theta$  are updated by back-propagation.
+
+For a query  $q_{i}$  and the candidate document set  $D_{q_i}$ , where  $D_{q_i} = \{d_{i,j}\}_{j = 1}^m$  could consist of annotated documents, hard negatives and in-batch negatives, we first get teacher's and retriever's similarity scores, respectively. Then, the retriever  $\theta$  is expected to keep consistent with its smooth teacher  $\theta^{*}$ . To achieve this goal, we update the retriever  $\theta$  by minimizing the KL divergence between the student's distribution and the teacher's distribution.
+
+To be concrete, the similarity scores between  $q_{i}$  and  $D_{q_i}$  are normalized into the following distributions:
+
+$$
+p _ {\phi} \left(d _ {i, j} \mid q _ {i}; D _ {q _ {i}}\right) = \frac {e ^ {\tau f _ {\phi} \left(q _ {i} , d _ {i , j}\right)}}{\sum_ {j = 1} ^ {m} e ^ {\tau f _ {\phi} \left(q _ {i} , d _ {i , j}\right)}}, \phi \in \left\{\theta , \theta^ {*} \right\}, \tag {6}
+$$
+
+Then, the consistency loss  $L_{cons}$  can be written as:
+
+$$
+L _ {c o n s} = K L (p _ {\theta} (. | q _ {i}; D _ {q _ {i}}), p _ {\theta^ {*}} (. | q _ {i}; D _ {q _ {i}})), \quad (7)
+$$
+
+where  $KL(\cdot)$  is the KL divergence,  $p_{\theta}(.|q_i;D_{q_i})$  and  $p_{\theta^*}(.|q_i;D_{q_i})$  denote the conditional probabilities of candidate documents  $D_{q_i}$  by the retriever  $\theta$  and the teacher  $\theta^*$ , respectively.
+
+For the estimated noisy pair, the teacher corrects the supervised signal into a soft label. For the
+
+# Algorithm 1 Noisy Pair Corrector (NPC)
+
+Require: Retriever  $\theta$ ; Noisy Training dataset  $C$ .
+
+1: Warm up the retriever  $\theta$  
+2: Initial EMA model  $\theta^{*}$  with  $\theta$  
+3: for  $i = 1 :$  num_epoch do  
+4: Calculate PPL of training pairs with random negatives using Eq.2;  
+5: Fit PPL distribution with GMM;  
+6: Get the estimated flag set  $\{\hat{y}_i\}$  using Eq.4;  
+7: for  $i = 1 :$  num_batch do  
+8: Sample negatives with "In-Batch Negative" or "Hard Negative" strategy;  
+9: Calculate rectified soft labels with EMA model  $\theta^{*}$  
+10: Train  $\theta$  by optimizing Eq.8;  
+11: Update EMA model  $\theta^{*}$  using Eq.5;  
+12: end for  
+13: end for
+
+estimated clean pair, we calculate the contrastive loss and consistency loss. So the overall loss is formalized:
+
+$$
+L = \hat {y} _ {i} L _ {\text {c o n t}} + L _ {\text {c o n s}}, \tag {8}
+$$
+
+where  $\hat{y}_i\in \{1,0\}$  is estimated by the noise detection module.
+
+# 3.3 Overall Procedure
+
+NPC is a general training framework that can be easily applied to most retrieval methods. Under the classical training process of dense retrieval, We first warmup the retriever with the typical contrastive learning method to provide it with basic retrieval abilities, and then add the noise detection module before training each epoch and the noise correction module during training. The detail is presented in Algorithm 1.
+
+# 4 Experiments
+
+# 4.1 Datasets
+
+To verify the effectiveness of NPC in robust dense retrieval, we conduct experiments on four commonly-used benchmarks, including Natural Questions (Kwiatkowski et al., 2019), Trivia QA (Joshi et al., 2017), StaQC (Yao et al., 2018) and SO-DS (Heyman and Van Cutsem, 2020).
+
+StaQC is a large dataset that collects real querycode pairs from Stack Overflow*. The dataset has
+
+<table><tr><td rowspan="2">Methods</td><td colspan="3">StaQC</td><td colspan="3">SO-DS</td></tr><tr><td>R@3</td><td>R@10</td><td>MRR</td><td>R@3</td><td>R@10</td><td>MRR</td></tr><tr><td>BM25 desc (Heyman and Van Cutsem, 2020)</td><td>8.0</td><td>13.3</td><td>7.5</td><td>23.8</td><td>32.3</td><td>21.6</td></tr><tr><td>NBOW (Heyman and Van Cutsem, 2020)</td><td>10.9</td><td>16.6</td><td>9.5</td><td>27.7</td><td>38.0</td><td>24.7</td></tr><tr><td>USE (Heyman and Van Cutsem, 2020)</td><td>12.8</td><td>20.3</td><td>11.7</td><td>33.3</td><td>48.5</td><td>30.4</td></tr><tr><td>CodeBERT (Feng et al., 2020)</td><td>-</td><td>-</td><td>23.4</td><td>-</td><td>-</td><td>23.1</td></tr><tr><td>GraphCodeBERT (Guo et al., 2021)</td><td>-</td><td>-</td><td>24.1</td><td>-</td><td>-</td><td>25.2</td></tr><tr><td>CodeRetriever (In-Batch Negative) (Li et al., 2022)</td><td>-</td><td>-</td><td>25.5</td><td>-</td><td>-</td><td>27.1</td></tr><tr><td>CodeRetriever (Hard Negative) (Li et al., 2022)</td><td>-</td><td>-</td><td>24.6</td><td>-</td><td>-</td><td>31.8</td></tr><tr><td>UniXcoder (In-Batch Negative) (Guo et al., 2022)</td><td>29.98</td><td>47.47</td><td>28.04</td><td>31.90</td><td>51.21</td><td>28.29</td></tr><tr><td>UniXcoder (Hard Negative) (Guo et al., 2022)</td><td>31.18</td><td>48.38</td><td>28.63</td><td>33.42</td><td>53.37</td><td>29.97</td></tr><tr><td>NPC (In-Batch Negative)</td><td>33.07</td><td>50.35</td><td>30.39</td><td>35.58</td><td>54.54</td><td>30.96</td></tr><tr><td>NPC (Hard Negative)</td><td>34.38</td><td>52.20</td><td>31.36</td><td>38.00</td><td>56.51</td><td>32.49</td></tr></table>
+
+Table 1: Retrieval performance on StaQC and SO-DS, which are realistic-noisy datasets. The results of the first block are mainly borrowed from published papers (Heyman and Van Cutsem, 2020; Li et al., 2022). If the results are not provided, we mark them as “-”.
+
+been widely used on code summarization (Ped-damail et al., 2018) and code search (Heyman and Van Cutsem, 2020). SO-DS mines query-code pairs from the most upvoted Stack Overflow posts, mainly focuses on the data science domain. Following previous works (Heyman and Van Cutsem, 2020; Li et al., 2022), we resort to Recall of topk  $(\mathbb{R}@\mathbb{k})$  and Mean Reciprocal Rank (MRR) as the evaluation metric. StaQC and SO-DS are constructed automatically without human annotation. Therefore, there are numerous mismatched pairs in training data.
+
+Natural Questions (NQ) collects real queries from the Google search engine. Each question is paired with an answer span and golden passages from the Wikipedia pages. Trivia QA (TQ) is a reading comprehension dataset authored by trivia enthusiasts. During the retrieval stage of both datasets, the objective is to identify positive passages from a large collection. Positive pairs in these datasets are assessed based on strict rule, i.e., whether passages contain answers or not (Karpukhin et al., 2020a). Consequently, we consider these datasets to be of high quality. Thus, we leverage them for simulation experiments to quantitatively analyze the impact of varying proportions of noise. Drawing inspiration from the setting in the noisy classification task (Han et al., 2018), we simulate the mismatched-pair noise by randomly pairing queries with unrelated documents.
+
+# 4.2 Implementation Details
+
+NPC is a general training paradigm that can be directly applied to almost all retrieval models. For
+
+StaQC and SO-DS, we adopt UniXcoder (Guo et al., 2022) as our backbone, which is the SoTA model for code representation. Following Guo et al. (2022), we adopt the cosine distance as a similarity function and set temperature  $\lambda$  to 20. We update model parameters using the Adam optimizer and perform early stopping on the development set. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 256, 5, and 10, respectively. In the "Hard Negative" setting, we adopt the same strategy as Li et al. (2022).
+
+For NQ and TQ, we adopt BERT (Devlin et al., 2019) as our initial model. Following Karpukhin et al. (2020a), we adopt inner-product as the similarity function and set temperature  $\lambda$  to 1. The max sequence length is 16 for query and 128 for passage. The learning rate, batch size, warmup epoch, and training epoch are set to 2e-5, 512, 10, and 40, respectively. We adopt "BM25 Negative" and "Hard Negative" strategies as described in the DPR toolkit. For a fair comparison, we implement DPR (Karpukhin et al., 2020a) with the same hyperparameters. All experiments are run on 8 NVIDIA Tesla A100 GPUs. The implementation of NPC is based on Huggingface (Wolf et al., 2020).
+
+# 4.3 Results
+
+Results on StaQC and SO-DS: Table 1 shows the results on the realistic-noisy datasets StaQC and SO-DS. Both datasets contain a large number of real noise pairs. The first block shows the results of previous SoTA methods. BM25 desc is a traditional
+
+<table><tr><td rowspan="2">Noisy</td><td rowspan="2">Methods</td><td colspan="4">Natural Questions</td><td colspan="4">Trivia QA</td></tr><tr><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td rowspan="2"></td><td>BM25*</td><td>-</td><td>-</td><td>59.1</td><td>73.7</td><td>-</td><td>-</td><td>66.9</td><td>76.7</td></tr><tr><td>DPR*</td><td>-</td><td>-</td><td>78.4</td><td>85.4</td><td>-</td><td>-</td><td>79.4</td><td>85.0</td></tr><tr><td rowspan="11">20</td><td>DPR (BM25 Negative)</td><td>27.07</td><td>47.79</td><td>63.36</td><td>75.69</td><td>35.73</td><td>52.88</td><td>64.05</td><td>74.16</td></tr><tr><td>coCondenser (BM25 Negative)</td><td>29.12</td><td>51.02</td><td>67.45</td><td>77.93</td><td>39.41</td><td>56.72</td><td>67.34</td><td>76.04</td></tr><tr><td>Co-teaching (BM25 Negative)</td><td>26.02</td><td>52.48</td><td>63.46</td><td>76.11</td><td>28.65</td><td>53.01</td><td>64.99</td><td>74.05</td></tr><tr><td>DPR-C (BM25 Negative)</td><td>43.69</td><td>66.62</td><td>79.07</td><td>86.12</td><td>52.10</td><td>70.52</td><td>79.05</td><td>85.08</td></tr><tr><td>NPC (BM25 Negative)</td><td>45.22</td><td>68.42</td><td>79.76</td><td>86.56</td><td>52.34</td><td>70.22</td><td>79.10</td><td>84.86</td></tr><tr><td>DPR (Hard Negative)</td><td>37.61</td><td>60.73</td><td>71.68</td><td>79.56</td><td>43.39</td><td>60.67</td><td>70.34</td><td>77.88</td></tr><tr><td>coCondenser (Hard Negative)</td><td>40.71</td><td>63.41</td><td>74.33</td><td>81.22</td><td>47.42</td><td>64.80</td><td>73.38</td><td>80.07</td></tr><tr><td>RocketQA (Hard Negative)</td><td>43.32</td><td>64.25</td><td>74.96</td><td>81.42</td><td>49.90</td><td>65.72</td><td>74.04</td><td>80.39</td></tr><tr><td>Co-teaching (Hard Negative)</td><td>31.78</td><td>56.32</td><td>66.12</td><td>77.56</td><td>33.28</td><td>57.29</td><td>66.50</td><td>75.62</td></tr><tr><td>DPR-C (Hard Negative)</td><td>51.66</td><td>72.40</td><td>81.50</td><td>87.80</td><td>55.35</td><td>72.36</td><td>80.33</td><td>85.34</td></tr><tr><td>NPC (Hard Negative)</td><td>51.85</td><td>73.06</td><td>82.47</td><td>87.80</td><td>56.03</td><td>72.54</td><td>80.59</td><td>85.58</td></tr><tr><td rowspan="11">50</td><td>DPR (BM25 Negative)</td><td>16.12</td><td>33.88</td><td>49.70</td><td>63.38</td><td>20.09</td><td>34.63</td><td>47.42</td><td>61.04</td></tr><tr><td>coCondenser (BM25 Negative)</td><td>18.28</td><td>36.37</td><td>52.01</td><td>65.92</td><td>22.80</td><td>38.01</td><td>51.00</td><td>63.79</td></tr><tr><td>Co-teaching (BM25 Negative)</td><td>23.72</td><td>50.32</td><td>64.86</td><td>74.92</td><td>26.56</td><td>51.22</td><td>63.78</td><td>73.77</td></tr><tr><td>DPR-C (BM25 Negative)</td><td>41.29</td><td>65.21</td><td>78.48</td><td>85.70</td><td>49.61</td><td>68.81</td><td>78.00</td><td>84.23</td></tr><tr><td>NPC (BM25 Negative)</td><td>42.87</td><td>65.65</td><td>78.37</td><td>85.76</td><td>50.80</td><td>68.98</td><td>78.21</td><td>84.43</td></tr><tr><td>DPR (Hard Negative)</td><td>23.87</td><td>42.34</td><td>55.12</td><td>67.06</td><td>28.47</td><td>45.12</td><td>56.88</td><td>67.62</td></tr><tr><td>coCondenser (Hard Negative)</td><td>24.55</td><td>44.16</td><td>56.69</td><td>68.72</td><td>31.05</td><td>47.81</td><td>59.48</td><td>70.14</td></tr><tr><td>RocketQA (Hard Negative)</td><td>26.83</td><td>45.72</td><td>57.32</td><td>69.24</td><td>33.67</td><td>49.28</td><td>60.32</td><td>70.46</td></tr><tr><td>Co-teaching (Hard Negative)</td><td>30.12</td><td>55.94</td><td>65.81</td><td>76.90</td><td>31.85</td><td>55.37</td><td>65.29</td><td>75.02</td></tr><tr><td>DPR-C (Hard Negative)</td><td>48.87</td><td>70.52</td><td>81.44</td><td>87.17</td><td>53.07</td><td>70.36</td><td>79.02</td><td>84.69</td></tr><tr><td>NPC (Hard Negative)</td><td>48.81</td><td>70.60</td><td>81.17</td><td>87.20</td><td>53.09</td><td>70.27</td><td>79.31</td><td>84.96</td></tr></table>
+
+Table 2: Retrieval performance on Natural Questions and Trivia QA under the noise ratio of  $20\%$ , and  $50\%$ , respectively. The results of BM25* and DPR* are borrowed from Karpukhin et al. (2020a). If the results are not provided, we mark them as “-”.
+
+sparse retriever based on the exact term matching of queries and code descriptions. NBOW is an unsupervised retriever that leverages pre-trained word embedding of queries and code descriptions for retrieval. USE is a simple dense retriever based on transformer. CodeBERT, GraphCodeBERT are pre-trained models for code understanding using large-scale code corpus. CodeRetriever is a pretrained model dedicated to code retrieval, which is pre-trained with unimodal and bimodal contrastive learning on a large-scale corpus. UniXcoder is also a pretrained model that utilizes multi-modal data, including code, comment, and AST, for better code representation. The results are implemented by ourselves for a fair comparison with NPC. The bottom block shows the results of NPC using two negative sampling strategies.
+
+From the results, we can see that our proposed NPC consistently performs better than the evaluated models across all metrics. Compared with the strong baseline UniXcoder which ignores the mismatched-pair problem, NPC achieves a significant improvement with both "in-batch negative" and "hard negative" sampling strategies. It indi
+
+cates that the mismatched-pair noise greatly limits the performance of dense retrieval models, and NPC can mitigate this negative effect. We also show some noisy examples detected by NPC in Appendix A.
+
+Results on NQ and TQ: Table 2 shows the results on the synthetic-noisy datasets NQ and TQ under the noise ratio of  $20\%$ , and  $50\%$ . We compare NPC with BM25 (Yang et al., 2017) and DPR (Karpukhin et al., 2020a). BM25 is an unsupervised sparse retriever that is not affected by noisy data. DPR (Karpukhin et al., 2020a) is a widely used method for training dense retrievers. coCondenser (Gao and Callan, 2021b) leverage pre-training to enhance models' robustness. RocketQA (Qu et al., 2021) adopts a cross-encoder to filter false negatives in the "Hard Negative" strategy. Co-teaching (Han et al., 2018) uses the samples with small loss to iteratively train two networks, which is widely used in the noisy label classification task. We implement baselines using two negative sampling strategies. Besides, we evaluate DPR on clean datasets by discarding the synthetic-noisy pairs, denoted by DPR-C. DPR-C is a strong
+
+<table><tr><td colspan="3">Methods</td><td colspan="2">NQ</td><td colspan="2">StaQC</td></tr><tr><td>De</td><td>Co</td><td>HN</td><td>R@20</td><td>R@100</td><td>R@1</td><td>R@3</td></tr><tr><td>-</td><td>-</td><td>-</td><td>48.22</td><td>62.31</td><td>18.08</td><td>31.09</td></tr><tr><td>-</td><td>✓</td><td>-</td><td>55.90</td><td>69.33</td><td>18.51</td><td>31.01</td></tr><tr><td>✓</td><td>-</td><td>-</td><td>75.19</td><td>83.31</td><td>20.05</td><td>32.71</td></tr><tr><td>✓</td><td>✓</td><td>-</td><td>77.50</td><td>84.79</td><td>20.70</td><td>33.55</td></tr><tr><td>-</td><td>-</td><td>✓</td><td>54.63</td><td>65.54</td><td>18.66</td><td>31.74</td></tr><tr><td>-</td><td>✓</td><td>✓</td><td>58.63</td><td>69.06</td><td>19.35</td><td>32.09</td></tr><tr><td>✓</td><td>-</td><td>✓</td><td>77.59</td><td>85.03</td><td>20.93</td><td>33.55</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>80.07</td><td>85.89</td><td>21.93</td><td>34.51</td></tr></table>
+
+baseline that is not affected by mismatched pairs.
+
+We can observe that (1) As the noise ratio increases, DPR, coCondenser, and RocketQA experience a significant decrease in performance. At a noise rate of  $50\%$ , they perform worse than unsupervised BM25. (2) Despite Co-teaching having good noise resistance, its performance is still low. This indicates that methods for dealing with label noise in classification are not effective for retrieval. (3) NPC outperforms baselines by a large margin, with only a slight performance drop when the noise increases. Even comparing DPR-C, NPC still achieves competitive results.
+
+# 4.4 Analysis
+
+Ablations of Noise Detection and Noise Correction: To get a better insight into NPC, we conduct ablation studies on the realistic-noisy dataset StaQC and the synthetic-noisy dataset NQ under the noise ratio of  $50\%$ . The results are shown in Table 3. "De" and "Co" refer to noise detection and noise correction, respectively. "HN" indicates whether to perform "Hard Negative" strategy. For both synthetic noise and realistic noise, we can see that the noise detection module brings a significant gain, no matter which negative sampling strategy is used. Correction also enhances the robustness of the retriever since it provides rectified soft labels which can lead the model output to be smoother. The results show that combining the two obtains better performance compared with only using the detection module or correction module.
+
+Impact of Warmup Epoch: According to the foregoing, NPC starts by warming up. In Table 4, we pre-training the retriever on the noisy dataset for warming up, and show the performance of NPC with different various epoch numbers  $n$ . In this
+
+Table 3: Ablation studies on StaQC dev set and NQ dev set under noise ratio of  $50\%$ .  
+
+<table><tr><td>Setting</td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td>n=1</td><td>50.58</td><td>69.93</td><td>79.87</td><td>84.96</td></tr><tr><td>n=5</td><td>50.03</td><td>69.64</td><td>80.17</td><td>85.76</td></tr><tr><td>n=10</td><td>50.07</td><td>69.93</td><td>80.07</td><td>85.89</td></tr><tr><td>n=20</td><td>38.09</td><td>60.31</td><td>72.00</td><td>80.07</td></tr></table>
+
+Table 4: Performance of NPC on NQ dev set with different warmup epoch number  $n$  .
+
+![](images/dcbc0024eec1373ac363c3fc24a44b0459ee0f3fd2ed17ccacebaf9d1af0612c.jpg)  
+(a) Before warmup
+
+![](images/ff0d9d101b4a5bbe04c616c134f92ae47d66262ae98438926fa3ec8100d901c4.jpg)  
+(b) After warmup
+
+![](images/baa431921633f57acfa52b1ff1f2b97d511463f29dc689800c8de230bbdbfa4c.jpg)  
+(c) DPR
+
+![](images/55b0dadef6128a055152c36cf1a1411946ef6047a5e8dac813c3bc5943bf141c.jpg)  
+(d) NPC  
+Figure 4: Perplexity distribution of training pairs under different settings.
+
+experiment, we adopt "Hard Negative" sampling strategy. We find that NPC achieves good results when the warmup epoch is relatively small (1 - 10). However, when the warmup epoch is too large, the performance will degrade. We believe that a prolonged warmup causes overfitting to noise samples.
+
+Impact of Iterative Detection: In the training of NPC, we perform iterative noise detection every epoch. A straightforward approach is to detect the noise only once after warmup and fix the estimated flag set  $\{\hat{y}_i\}$ . To study the effectiveness of iterative detection, we conducted an ablation study. The results are shown in Table 5. We can see that the model performance degrades after removing iterative detection.
+
+Ablations of PPL: We distinguish noise pairs according to the perplexity between the annotated positive document and easy negatives. When calculating the perplexity, "Hard Negative" will cause trouble for detection. We construct ablation experiments to verify this, and the results are shown in Table 5. We can see that the perplexity with "Hard Negative" results in performance degradation.
+
+Visualization of Perplexity Distribution: In Fig. 4, we illustrate the perplexity distribution of
+
+<table><tr><td>Setting</td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td>NPC</td><td>50.07</td><td>69.93</td><td>80.07</td><td>85.89</td></tr><tr><td>-w/o iterative detection</td><td>47.29</td><td>68.39</td><td>78.79</td><td>85.38</td></tr><tr><td>-ppl with HN</td><td>42.81</td><td>65.06</td><td>75.22</td><td>83.09</td></tr></table>
+
+Table 5: Ablation studies of iterative noise detection and perplexity variants
+
+![](images/536d70a8d2cb88df20639cfd6d5ca8cc0333dd50b240e02dd21b4c40ea60155c.jpg)  
+(a) w/o Hard Negative
+
+![](images/83a8a9353639f66c54f9cffd67434fb4498a65290355222e85117a1823ac5622.jpg)  
+(b) w/ Hard Negative  
+Figure 5: Retrieval performance of DPR and NPC on NQ dev set under different noise ratios.
+
+training pairs before and after warmup, after training with DPR, and after training with NPC. The experiment is on NQ under the noise ratio of  $50\%$ . We can see that the perplexity of most noisy pairs is larger than the clean pairs after warmup, which verifies our hypothesis in Sec. 3.1. Comparing Fig. 4(c) and Fig. 4(d), we find that the retriever trained with DPR will overfit the noise pairs. However, NPC enables the retriever to correctly distinguish clean and noisy pairs because it avoids the dominant effect of noise during network optimization.
+
+Analysis of Generalizability Fig. 5 shows the performance of DPR and NPC under the noise ratio ranging from  $0\%$  to  $80\%$ . We can see that as the noise ratio increases, the performance degradation of DPR is much larger than that of NPC, which demonstrates the generalizability of NPC. Furthermore, even though NPC is designed to deal with mismatched-pair noise, it achieves competitive results when used in a noise-free setting.
+
+# 5 Related Work
+
+# 5.1 Dense Retrieval
+
+Dense retrieval has shown better performance than traditional sparse retrieval methods (Lee et al., 2019; Karpukhin et al., 2020a). The studies of dense retrieval can be divided into two categories, (1) unsupervised pre-training to get better initialization and (2) more effective fine-tuning on labeled data. In the first category, some researchers focus on how to generate contrastive pairs automatically from a large unsupervised corpus (Lee et al., 2019; Chang et al., 2019; Ma et al., 2022; Li et al., 2022). Another line of research enforces
+
+the model to produce an information-rich CLS representation (Gao and Callan, 2021a,b; Lu et al., 2021). As for effective fine-tuning strategies (He et al., 2022b), recent studies show that negative sampling techniques are critical to the performance of dense retrievers. DPR (Karpukhin et al., 2020b) adopts in-batch negatives and BM25 negatives; ANCE (Xiong et al., 2021), RocketQA (Qu et al., 2021), and AR2 (Zhang et al., 2022a) improve the hard negative sampling by iterative replacement, denoising, and adversarial framework, respectively. Several works distill knowledge from ranker to retriever (Izacard and Grave, 2020; Yang and Seo, 2020; Ren et al., 2021; Zeng et al., 2022). Some studies incorporate lexical-aware sparse retrievers to convey lexical-related knowledge to dense retrievers, thereby enhancing the dense retriever's ability to recognize lexical matches (Shen et al., 2023; Zhang et al., 2023).
+
+Although the above methods have achieved promising results, they are highly dependent on correctly matched data, which is difficult to satisfy in real scenes. The mismatched-pair noise problem has seldom been considered. Besides, some studies utilize large-sized generative models (He et al., 2023) to guide retrievers, which achieve impressive performance without paired data (Sachan et al., 2022, 2021; Gao et al., 2022; He et al., 2022a). Although these models exhibit some robustness to noisy data, their success depends on the availability of strong generative models. Moreover, their applicability will be limited in domains where generative models do not perform well.
+
+# 5.2 Denoising Techniques
+
+One related task to our work is Noisy Label. Numerous methods have been proposed to solve this problem, and most of them focus on the classification task (Han et al., 2020). Some works design robust loss functions to mitigate label noise (Ghosh et al., 2017; Ma et al., 2020). Another line of work aims to identify noise from the training set with the memorization effect of neural networks (Silva et al., 2022; Liang et al., 2022; Bai et al., 2021).
+
+These studies mainly focus on classification. NPC studies the mismatched noise problem in dense retrieval rather than the noise in category annotations, which is more complex to handle. Several pre-training approaches noticed the problem of mismatched noisy pairs. ALIGN (Jia et al., 2021) and CLIP (Radford et al., 2021) claim that utilizing
+
+large-scale image-text pairs can ignore the existence of noise. E5 (Wang et al., 2022) employs a consistency-based rule to filter the pre-training data. Although they slightly realized the existence of noisy pairs during pre-train, none of them give a specialized solution to solve it and extensively explored the characteristics of noisy text pairs. Some recent works (Huang et al., 2021; Han et al., 2023) study the noisy correspondence problem in cross-modal retrieval. Although the "mismatched-pair noisy" problem in cross-modal retrieval and text retrieval shares similarities, the specific settings and methods used in these two areas are notably distinct. It is challenging to directly apply these cross-modal retrieval works to document and code retrieval. Our NPC is the first systematic work to explore mismatched-pair noise in document/code retrieval.
+
+# 6 Conclusion
+
+This paper explores a neglected problem in dense retrieval, i.e., mismatched-pair noise. To solve this problem, we propose a generalized Noisy Pair Corrector(NPC) framework, which iteratively detects noisy pairs per epoch based on the perplexity and then provides rectified soft labels via an EMA model. The experimental results and analysis demonstrate the effectiveness of NPC in effectively handling both synthetic and realistic mismatched-pair noise.
+
+# Limitations
+
+This work mainly focuses on training the dense retrieval models with mismatched noise. There may be two possible limitations in our study.
+
+1) Due to the limited computing infrastructure, we only verified the robustness performance of NPC based on the classical retriever training framework. We leave experiments to combine NPC with more effective retriever training methods such as distillation (Ren et al., 2021), AR2 (Zhang et al., 2022a), as future work.  
+2) Mismatched-pair noise may also exist in other tasks, such as recommender systems. We will consider extending NPC to more tasks.
+
+# Acknowledgement
+
+This work is supported by the Fundamental Research Funds for the Central Universities under Grant 1082204112364 and the Key Program of the
+
+National Science Foundation of China under Grant 61836006.
+
+# References
+
+Yingbin Bai, Erkun Yang, Bo Han, Yanhua Yang, Jia-tong Li, Yinian Mao, Gang Niu, and Tongliang Liu. 2021. Understanding and improving early stopping for learning with noisy labels. In NIPS.  
+Dan Brickley, Matthew Burgess, and Natasha Noy. 2019. Google dataset search: Building a search engine for datasets in an open web ecosystem. In WWW.  
+Wei-Cheng Chang, X Yu Felix, Yin-Wen Chang, Yiming Yang, and Sanjiv Kumar. 2019. Pre-training tasks for embedding-based large-scale retrieval. In International Conference on Learning Representations.  
+Xuanang Chen, Jian Luo, Ben He, Le Sun, and Yingfei Sun. 2022. Towards robust dense retrieval via local ranking alignment. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI, pages 1980-1986.  
+Arthur P Dempster, Nan M Laird, and Donald B Rubin. 1977. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers).  
+Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. Codebert: A pre-trained model for programming and natural languages. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, Online Event, 16-20 November 2020.  
+Luyu Gao and Jamie Callan. 2021a. Is your language model ready for dense representation fine-tuning? arXiv preprint arXiv:2104.08253.  
+Luyu Gao and Jamie Callan. 2021b. Unsupervised corpus aware language model pre-training for dense passage retrieval. arXiv preprint arXiv:2108.05540.  
+Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2022. Precise zero-shot dense retrieval without relevance labels. CoRR.  
+Aritra Ghosh, Himanshu Kumar, and P Shanti Sastry. 2017. Robust loss functions under label noise for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence.  
+Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin. 2022. Unixcoder: Unified cross-modal pre-training for code representation. In ACL.
+
+Daya Guo, Shuo Ren, Shuai Lu, Zhangyin Feng, Duyu Tang, Shujie Liu, Long Zhou, Nan Duan, Alexey Svyatkovskiy, Shengyu Fu, Michele Tufano, Shao Kun Deng, Colin B. Clement, Dawn Drain, Neel Sundaresan, Jian Yin, Daxin Jiang, and Ming Zhou. 2021. Graphcodebert: Pre-training code representations with data flow. In International Conference on Learning Representations.  
+Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W Tsang, James T Kwok, and Masashi Sugiyama. 2020. A survey of label-noise representation learning: Past, present and future. arXiv preprint arXiv:2011.04406.  
+Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. 2018. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NIPS.  
+Haochen Han, Kaiyao Miao, Qinghua Zheng, and Minnan Luo. 2023. Noisy correspondence learning with meta similarity correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7517-7526.  
+Xingwei He, Yeyun Gong, A Jin, Weizhen Qi, Hang Zhang, Jian Jiao, Bartuer Zhou, Biao Cheng, Siu Ming Yiu, Nan Duan, et al. 2022a. Metric-guided distillation: Distilling knowledge from the metric to ranker and retriever for generative commonsense reasoning. arXiv preprint arXiv:2210.11708.  
+Xingwei He, Yeyun Gong, A Jin, Hang Zhang, Anlei Dong, Jian Jiao, Siu Ming Yiu, Nan Duan, et al. 2022b. Curriculum sampling for dense retrieval with document expansion. arXiv preprint arXiv:2212.09114.  
+Xingwei He, Zhenghao Lin, Yeyun Gong, Alex Jin, Hang Zhang, Chen Lin, Jian Jiao, Siu Ming Yiu, Nan Duan, Weizhu Chen, et al. 2023. Annollm: Making large language models to be better crowdsourced annotators. arXiv preprint arXiv:2303.16854.  
+Geert Heyman and Tom Van Cutsem. 2020. Neural code search revisited: Enhancing code snippet retrieval through natural language intent. arXiv preprint arXiv:2008.12193.  
+Zhenyu Huang, Guocheng Niu, Xiao Liu, Wenbiao Ding, Xinyan Xiao, Hua Wu, and Xi Peng. 2021. Learning with noisy correspondence for cross-modal matching. Advances in Neural Information Processing Systems, 34:29406-29419.  
+Gautier Izacard and Edouard Grave. 2020. Distilling knowledge from reader to retriever for question answering. arXiv preprint arXiv:2012.04584.  
+Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. 2021. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning.
+
+Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547.  
+Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In ACL.  
+Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020a. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906.  
+Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick S. H. Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020b. Dense passage retrieval for open-domain question answering. In EMNLP.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur P. Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: a benchmark for question answering research. Trans. Assoc. Comput. Linguistics, 7:452-466.  
+Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.  
+Xiaonan Li, Yeyun Gong, Yelong Shen, Xipeng Qiu, Hang Zhang, Bolun Yao, Weizhen Qi, Daxin Jiang, Weizhu Chen, and Nan Duan. 2022. Coderetriever: Unimodal and bimodal contrastive learning. arXiv preprint arXiv:2201.10866.  
+Kevin J Liang, Samrudhdhi B. Rangrej, Vladan Petrovic, and Tal Hassner. 2022. Few-shot learning with noisy labels. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9079-9088.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.  
+Shuqi Lu, Di He, Chenyan Xiong, Guolin Ke, Waleed Malik, Zhicheng Dou, Paul Bennett, Tie-Yan Liu, and Arnold Overwijk. 2021. Less is more: Pretrain a strong siamese encoder for dense text retrieval using a weak decoder. In Empirical Methods in Natural Language Processing.  
+Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. 2020. Normalized loss functions for deep learning with noisy labels. In ICML.
+
+Xinyu Ma, Jiafeng Guo, Ruqing Zhang, Yixing Fan, and Xueqi Cheng. 2022. Pre-train a discriminative text encoder for dense retrieval via contrastive span prediction. In SIGIR '22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, July 11 - 15, 2022.  
+Jayavardhan Reddy Peddamail, Ziyu Yao, Zhen Wang, and Huan Sun. 2018. A comprehensive study of staqc for deep code summarization. In KDD.  
+Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. Rocketqa: An optimized training approach to dense passage retrieval for open-domain question answering. In NAACL-HLT.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In ICML.  
+Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Qiaoqiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021. Rocketqav2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2825-2835.  
+Devendra Singh Sachan, Mike Lewis, Dani Yogatama, Luke Zettlemoyer, Joelle Pineau, and Manzil Zaheer. 2022. Questions are all you need to train a dense passage retriever. CoRR.  
+Devendra Singh Sachan, Mostofa Patwary, Mohammad Shoeybi, Neel Kant, Wei Ping, William L. Hamilton, and Bryan Catanzaro. 2021. End-to-end training of neural retrievers for open-domain question answering. In ACL/IJCNLP.  
+Tao Shen, Xiubo Geng, Chongyang Tao, Can Xu, Guodong Long, Kai Zhang, and Daxin Jiang. 2023. Unifier: A unified retriever for large-scale retrieval. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4787-4799.  
+Yelong Shen, Xiaodong He, Jianfeng Gao, Li Deng, and Grégoire Mesnil. 2014. Learning semantic representations using convolutional neural networks for web search. In Proceedings of the 23rd international conference on world wide web, pages 373-374.  
+Amila Silva, Ling Luo, Shanika Karunasekera, and Christopher Leckie. 2022. Noise-robust learning from multiple unsupervised sources of inferred labels. In AAAI Conference on Artificial Intelligence.  
+Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. ArXiv, abs/2212.03533.
+
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In EMNLP.  
+Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations.  
+Peilin Yang, Hui Fang, and Jimmy Lin. 2017. Anserini: Enabling the use of lucene for information retrieval research. In Proceedings of the 40th international ACM SIGIR conference on research and development in information retrieval.  
+Sohee Yang and Minjoon Seo. 2020. Is retriever merely an approximator of reader? arXiv preprint arXiv:2010.10999.  
+Ziyu Yao, Daniel S Weld, Wei-Peng Chen, and Huan Sun. 2018. Staqc: A systematically mined questioncode dataset from stack overflow. In WWW.  
+Hansi Zeng, Hamed Zamani, and Vishwa Vinay. 2022. Curriculum learning for dense retrieval distillation. In SIGIR '22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, July 11 - 15, 2022.  
+Hang Zhang, Yeyun Gong, Yelong Shen, Weisheng Li, Jiancheng Lv, Nan Duan, and Weizhu Chen. 2021. Poolingformer: Long document modeling with pooling attention. In International Conference on Machine Learning, pages 12437-12446. PMLR.  
+Hang Zhang, Yeyun Gong, Yelong Shen, Jiancheng Lv, Nan Duan, and Weizhu Chen. 2022a. Adversarial retriever-ranker for dense text retrieval. In International Conference on Learning Representations.  
+Kai Zhang, Chongyang Tao, Tao Shen, Can Xu, Xiubo Geng, Binxing Jiao, and Daxin Jiang. 2023. Led: Lexicon-enlightened dense retriever for large-scale retrieval. In Proceedings of the ACM Web Conference 2023, pages 3203-3213.  
+Shunyu Zhang, Yaobo Liang, Ming Gong, Daxin Jiang, and Nan Duan. 2022b. Multi-view document representation learning for open-domain dense retrieval. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).
+
+# A Qualitative Analysis
+
+Table 9 lists some mismatched pairs detected by NPC in StaQC training set. We can see that these mismatched pairs are almost irrelevant and can be correctly detected by NPC. These examples are not well aligned, mainly due to the low-quality answers of the open community (cases 2 and 4), inappropriate data preprocessing in the collection phase (cases 2 and 3), and other reasons. It is well known that collecting and cleaning training data is expensive and complex work. Automatically constructed datasets in real-world applications often contain such mismatched-pair noise. Our method can mitigate the impact caused by such noise during training.
+
+# B Statistics of Datasets
+
+The statistics of datasets are shown in Table 6.
+
+<table><tr><td>Dataset</td><td>Train</td><td>Dev</td><td>Test</td><td>Corpus size</td></tr><tr><td>StaQC</td><td>203.7K</td><td>2.6K</td><td>2.7K</td><td>14.6K</td></tr><tr><td>SO-DS</td><td>12.1K</td><td>0.9K</td><td>1.1K</td><td>12.1K</td></tr><tr><td>NQ</td><td>79.2K</td><td>8,8K</td><td>3.6K</td><td>21 M</td></tr><tr><td>TQ</td><td>78.8K</td><td>8.8k</td><td>11.3K</td><td>21 M</td></tr></table>
+
+# C Discussion about Perplexity
+
+We calculate the perplexity between the annotated document and easy negative documents during noise detection. We emphasize that the negative documents are randomly selected from the document collection  $\mathbb{D}$ . It is not suitable to adopt "Hard Negative" sampling strategy when calculating the perplexity. Although hard negatives are important to train a strong dense retriever, they will cause trouble during noise detection. Specifically, it is expected that the retriever is confused only between false positive and negative documents and can confidently distinguish true positive and negative documents. But if we adopt "Hard Negative" when calculating the perplexity, the retriever will also be confused between true positive and hard negative documents, which will affect noise detection. We construct ablation experiments to verify this, and the results are shown in Table 5.
+
+# D Integration with stronger methods
+
+We conducted comprehensive experiments that integrated NPC into coCondenser and RocketQAv2. The subsequent experiments were conducted on the NQ dataset with a  $50\%$  noise ratio. We first combined NPC with coCondenser which is a pretrained model specialized for dense retrieval tasks. The results are shown in Table 7
+
+Table 6: The statistics of datasets. Corpus size means the size of document corpus for evaluation.  
+
+<table><tr><td></td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td colspan="5">BM25 Negative</td></tr><tr><td>coCondenser</td><td>18.28</td><td>36.37</td><td>52.01</td><td>65.92</td></tr><tr><td>coCondenser-C</td><td>44.75</td><td>68.91</td><td>80.89</td><td>87.31</td></tr><tr><td>coCondenser+NPC</td><td>47.31</td><td>70.38</td><td>81.58</td><td>87.38</td></tr><tr><td colspan="5">Hard Negative</td></tr><tr><td>coCondenser</td><td>24.55</td><td>44.16</td><td>56.69</td><td>68.72</td></tr><tr><td>coCondenser-C</td><td>49.31</td><td>71.99</td><td>82.41</td><td>88.38</td></tr><tr><td>coCondenser+NPC</td><td>50.66</td><td>72.42</td><td>82.64</td><td>88.31</td></tr></table>
+
+It's evident that NPC significantly enhances the robustness of coCondenser against noise associated with mismatched pairs. This observation underscores the compatibility between NPC and pretrained dense retrievers.
+
+Furthermore, we combined NPC with RocketQAv2 which adopted a cross-encoder as a teacher and dynamically distilled knowledge to the dense retriever. The results are shown in Table 8:
+
+Table 7: Retrieval performance on NQ after combining NPC with coCondenser.  
+
+<table><tr><td></td><td>R@1</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td>RocketQAv2</td><td>32.30</td><td>51.37</td><td>62.19</td><td>70.79</td></tr><tr><td>RocketQAv2-C</td><td>52.63</td><td>73.51</td><td>83.21</td><td>88.71</td></tr><tr><td>RocketQAv2+NPC</td><td>52.59</td><td>73.83</td><td>83.32</td><td>88.69</td></tr></table>
+
+Table 8: Retrieval performance on NQ after combining NPC with RocketQAv2.
+
+To combine NPC with RocketQAv2, we integrate the noise detection and the correction modules in each training epoch of RocketQAv2. From the table, we can find that although RocketQAv2 uses a powerful cross-encoder as a teacher, it is still limited by the noise of the training data and shows low performance. NPC can effectively harmonize with RocketQAv2 to mitigate the problems caused by mismatched pair noise. We will add these experiment results to the next version.
+
+<table><tr><td></td><td>Question</td><td>Code</td></tr><tr><td>1</td><td>Split words in a nested list into letters</td><td>» [list(l[0]) for l in mylist]</td></tr><tr><td>2</td><td>Dictionary in python problem</td><td>» s = problem.getSuccessors( getStartState())</td></tr><tr><td>3</td><td>Find the Common first name from Django Auth user Model</td><td>» import operator</td></tr><tr><td>4</td><td>Find all text files not containing some text string</td><td>» lst = [1,2,4,6,3,8,0,5] 
+» for n in lst[:]: 
+» if n % 2 == 0: 
+»&gt;&gt;&gt; lst.remove(n) 
+» lst</td></tr></table>
+
+Table 9: Some noisy pairs detected by NPC in StaQC training set.

2023/Noisy Pair Corrector for Dense Retrieval/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94cb83dba96e64b17ba5ed2b2bb4a7573acdc1d14841a0d265f201c487e5443d
+size 668527

2023/Noisy Self-Training with Synthetic Queries for Dense Retrieval/260bb510-850d-4d78-a651-3f79c3ed0b9d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52f0571bb1c3359e795529bbde48e4a77863148edb9aaca32152ea7b574d97fa
+size 638796

2023/Noisy Self-Training with Synthetic Queries for Dense Retrieval/full.md

@@ -0,0 +1,391 @@
+# Noisy Self-Training with Synthetic Queries for Dense Retrieval
+
+Fan Jiang and Tom Drummond and Trevor Cohn *
+
+School of Computing and Information Systems
+
+The University of Melbourne, Victoria, Australia
+
+fan.jiang1@student.unimelb.edu.au
+
+{tom.drummond, trevor.cohn}@unimelb.edu.au
+
+# Abstract
+
+Although existing neural retrieval models reveal promising results when training data is abundant and the performance keeps improving as training data increases, collecting high-quality annotated data is prohibitively costly. To this end, we introduce a novel noisy self-training framework combined with synthetic queries, showing that neural retrievers can be improved in a self-evolution manner with no reliance on any external models. Experimental results show that our method improves consistently over existing methods on both general-domain (e.g., MS-MARCO) and out-of-domain (i.e., BEIR) retrieval benchmarks. Extra analysis on low-resource settings reveals that our method is data efficient and outperforms competitive baselines, with as little as  $30\%$  of labelled training data. Further extending the framework for reranker training demonstrates that the proposed method is general and yields additional gains on tasks of diverse domains. $^{1}$
+
+# 1 Introduction
+
+As an important task, information retrieval (IR) refers to the task of finding relevant texts from a large collection of passages or documents to satisfy the specific information needs of users. The information need is usually expressed as a short textual query, and the task is formulated as retrieving texts that are most relevant to the given query.
+
+Recently, impressive achievements have been made in neural retrieval models through adopting large-scale pre-trained models (Devlin et al., 2019). Dual-encoders typically serve as the backbone architecture, which enables retrieving relevant knowledge from collections with millions or billions of passages in a fraction of time (Karpukhin et al., 2020). Unlike traditional term-matching-based lexical retrievers, such as TF-IDF (Manning et al.,
+
+2008) or BM25 (Robertson and Zaragoza, 2009), which can be applied without any training, neural retrieval methods normally require training on a sufficient number of human-labelled query-passage pairs to work well. Nevertheless, due to the high cost of human annotations, the number of available query-passage pairs is way smaller compared to the size of passage collections (500k at most (Bajaj et al., 2016) v.s. 21M passages (Kwiatkowski et al., 2019)). Moreover, the situation becomes increasingly worse in out-of-domain applications, where only a few or no training examples are available. Directly applying neural retrieval models trained on high-resource datasets typically achieves low out-of-domain performance and lags behind their lexical counterparts (Thakur et al., 2021).
+
+In this work, we aim to improve the performance of state-of-the-art neural retrieval models on both general domain and out-of-domain datasets, by using automatically synthesised queries. For this purpose, we use a query generator to automatically synthesise queries for each passage in the target dataset, which we then use for pre-training, based on a self-training objective (Scudder, 1965). More specifically, instead of directly training the model on synthetic query-passage pairs (Ma et al., 2021), we use a neural retrieval model that was trained on the labelled dataset as the teacher to generate soft labels for each synthetic query, providing a supervision signal which is more robust to noise in the generated data. This ameliorates issues with the synthetic queries which are often generic or ambiguous, meaning that the query is only weakly related to its originating passage, and accordingly, other passages may match the synthetic query equally well or better. Table 1 shows a motivating example, where the synthetic query is spuriously-related to its originating passage with only the keyword achievement matched. However, the top-2 sampled "negatives" are highly related to the query in topics and semantics, according to both human judgment
+
+Synthetic Query: what was the major achievement of the Manhattan project?  
+
+<table><tr><td>Originating Passage</td><td>Logits</td><td>Hard</td></tr><tr><td>...The only cloud hanging over the impressive achievement of the atomic researchers and engineers is what their success truly meant; hundreds of thousands of innocent lives obliterated.</td><td>0.14</td><td>1</td></tr><tr><td colspan="3">Sampled Negatives</td></tr><tr><td>...success of the 20th century was the Manhattan Project. The Manhattan Project assimilated concepts and leaders from all scientific fields and engineering disciplines to construct the first two atomic bombs.</td><td>1.0</td><td>0</td></tr><tr><td>The Manhattan Project was an epic, secret, wartime effort to design and build the world&#x27;s first nuclear weapon. ... the $20 billion project resulted in the production of the first uranium and plutonium bombs.</td><td>0.92</td><td>0</td></tr><tr><td>The Manhattan Project is an American film, released in 1986. The plot revolves around a gifted high school student who decides to construct an atomic bomb for a national science fair.</td><td>0.4</td><td>0</td></tr></table>
+
+Table 1: A motivating example for the issue of weak-relation between synthetic query and originating passages, where top-2 sampled ‘negatives’ are way more relevant to the query than its originating passage. Negatives and associated Logits are from a dense model fine-tuned on MS-MARCO and are normalised for clarity.
+
+and model predictions. By contrast, simply taking the originating passage as positive and equally treating all negatives with hard labels regardless of their relevance to the query will completely mislead the model, leading to poor semantic matching patterns between query and passages.
+
+Furthermore, to prevent the student from blindly imitating the teacher's behaviour, we pollute its inputs either in the query or passage by injecting noise. This ensures that the student has a different view on inputs compared to the teacher, encouraging it to learn generalised signals from the teacher. Moreover, by polluting the inputs (e.g., shuffling), the student is encouraged to capture salient phrases in addition to semantic matching, being more robust to possible perturbed inputs (Figure 4). After completing pre-training on synthetic queries, we further finetune the model on labelled data. Through iterating the pre-training and finetuning steps by using the latest model to relabel synthetic data, the resulting model can significantly outperform the one trained only using labelled data. In summary, the contributions are as follows:
+
+1. We proposed a novel self-training framework to make use of automatically generated data more wisely for neural retrieval models (§3).  
+2. Our experimental results on general-domain benchmarks show that the proposed framework can not only significantly boost the performance of state-of-the-art neural retrievers, but also yield superior results in low-resource settings. We further show our framework is general and can be extended to improve more powerful cross-encoder-based rerankers (§4).  
+3. Further experiments are conducted on the out-of-domain BEIR benchmark (Thakur et al.,
+
+2021), and both neural retrievers and rerankers surpass a series of strong models (§4.6).
+
+# 2 Preliminaries
+
+# 2.1 Task Definition
+
+We focus on the task of short-passage retrieval (IR) in this work. Given a query in the form of a short text, the task requires retrieving a small set of passages that can satisfy the information needs, from a collection of passages in million or billion scale. Formally, suppose we have a passage collection  $\mathcal{P} = \{p_1,p_2,\dots ,p_n\}$ , the retriever is required to fetch top- $k$  passages  $\mathcal{P}_q = \{p_1,p_2,\dots ,p_k\}$  from  $\mathcal{P}$  that are relevant to a specific query  $q$ .
+
+# 2.2 Dense Passage Retrieval
+
+In contrast to traditional IR methods, such as BM25 (Robertson and Zaragoza, 2009), which represent texts in high dimensional and sparse vectors with inverted index, dense retrieval methods alternatively adopt neural models to encode texts (queries or passages) in dense latent vectors with much smaller dimensions. A dense passage retrieval model (Karpukhin et al., 2020) typically adopts the dual-encoder architecture, where neural models are used to encode the query and passage into dense vectors separately. The relevance is measured by the dot product between their embeddings:
+
+$$
+s (q, p; \theta) = \mathbf {E} _ {q} (q; \theta) ^ {\top} \cdot \mathbf {E} _ {p} (p; \theta) \tag {1}
+$$
+
+where  $\mathbf{E}(\cdot ;\theta)$  is an encoder parameterised by  $\theta$  The adoption of this form of 'dual-encoder' architecture decouples the encoding of query and passage. At inference, all passages in  $\mathcal{P}$  can be encoded offline. When a query  $q$  comes in, efficient
+
+nearest neighbour search (Johnson et al., 2021) can be performed to fetch the top- $k$  passages.
+
+Contrastive learning is applied to train the dual-encoder. Given a query  $q$ , we have a positive passage  $p^+$  and a set of  $n$  negative passages  $\mathcal{P}_q^- = \{p_i^-\}_{i=1}^n$ . The model is being optimised by minimising the negative log likelihood of the positive passage:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {C L}} = - \log D (p ^ {+} | q, \mathcal {P}, \theta) \\ = - \log \frac {s (q , p ^ {+} ; \theta)}{\sum_ {p \in \{p ^ {+} \} \cup \mathcal {P} _ {q} ^ {-}} s (q , p ; \theta)} \tag {2} \\ \end{array}
+$$
+
+$\mathcal{P}_q^-$  is the set of irrelevant passages constructed from in-batch negatives (Chen et al., 2020) (i.e. positive passages of other queries in the same minibatch) and mined hard negatives from existing retrievers (Karpukhin et al., 2020; Xiong et al., 2021).
+
+# 3 Method
+
+# 3.1 Self-Training with Synthetic Queries
+
+Self-training (Yarowsky, 1995) has been shown effective in improving model performance through using unlabelled or synthetic data. It first use labelled data to train a good teacher, then the teacher is used to generate pseudo labels on unlabelled data. Finally, an identical student model is first pre-trained on unlabelled data with soft labels and then finetuned on labelled data. Recently, it has been shown to work well for a variety of tasks, including image classification (Xie et al., 2020) and neural machine translation (He et al., 2020). However, its effectiveness has not yet been evaluated for dense passage retrieval, especially with automatically synthesised queries.
+
+Suppose we have a well-trained query generator, which is used to generate a set of synthetic query-passage pairs  $\mathcal{T}^{\prime} = \{(q^{\prime},p^{\prime +})\}$ . Moreover, we assume that a teacher dense retrieval model that is fully trained on labelled data is available. The student retriever has the same architecture as the teacher, but it only accesses the soft labels produced by the teacher during training. This ensures that all involved negatives are not treated equally but with different soft labels, and more relevant passages are more likely to be penalised less. Moreover, we assume the use of teacher supervision will make the supervision signal for problematic queries (e.g., general or ambiguous queries) much more diverse, as the teacher will have little clue and a corresponding high entropy predictive distribution. Formally,
+
+the learning process is conducted via KL divergence by closing the distribution distance between the student and teacher:
+
+$$
+\mathcal {L} _ {\mathrm {S T}} = \operatorname {K L} \left(T \left(\cdot | q ^ {\prime}, \mathcal {P} _ {q ^ {\prime}} ^ {\prime}, \theta_ {T}\right), S \left(\cdot | q ^ {\prime}, \mathcal {P} _ {q ^ {\prime}} ^ {\prime}, \theta_ {S}\right)\right) \tag {3}
+$$
+
+where  $\mathcal{P}_{q'}' = \{p'+\} \cup \mathcal{P}_{q'}'$ , which is the union of the pseudo positive passage of synthetic query  $q'$  and the sampled negatives as in §2.2.  $T(\cdot | q', \mathcal{P}_{q'}', \theta_T)$  and  $S(\cdot | q', \mathcal{P}_{q'}', \theta_S)$  are the distributions from the teacher and student, respectively.
+
+We further inject noises into student's inputs, hoping that it can generate more robust embeddings. Through injecting noises, the student needs to imitate the behaviour of the teacher with perturbed inputs (i.e. different views), being encouraged to learn generalised signals from the teacher (He et al., 2020). Inspired by Wu et al. (2019), we use following strategies for noise injection:
+
+1. Shuffle: randomly choose some words in the query or passage as candidates for shuffling, then randomly shuffle these candidates.  
+2. Delete: randomly delete some words in the query or passage.  
+3. Mask: randomly mask some words in the query or passage with a [MASK] token.
+
+Empirically, we found that applying them sequentially to both queries and passages with a probability 0.1 achieves the best results. Thus, the self-training loss of noised version becomes:
+
+$$
+\mathcal {L} _ {\mathrm {S T}} = \mathrm {K L} (T (\cdot | q ^ {\prime}, \mathcal {P} _ {q ^ {\prime}} ^ {\prime}, \theta_ {T}), S (\cdot | \widetilde {q} ^ {\prime}, \widetilde {\mathcal {P}} _ {q ^ {\prime}} ^ {\prime}, \theta_ {S})) \quad (4)
+$$
+
+where  $\widetilde{q'}$  and  $\widetilde{\mathcal{P}}_{q'}'$  are noised query and passages.
+
+# 3.2 Training Pipeline
+
+Algorithm 1 summarises the training pipeline, with an overview shown in Figure 1. More specifically:
+
+Teacher Preparation (line 3 in Alg. 1) We first train a teacher model in labelled data using a two-stage training similar to (Gao and Callan, 2022). In the first stage, the retriever is trained with hard negatives sampled from a BM25 retriever. Then, the retriever trained in the first stage is used to discover hard negatives, which are later used to train a second-stage retriever. The resulted retriever serves as the teacher  $\theta_T$  in our algorithm.
+
+Index Building & Hard Negative Mining (lines 4-5 in Alg. 1) The teacher model  $\theta_T$  encodes all passages on  $\mathcal{P}$  into dense vectors, which are later used to build the ANN search index with FAISS.
+
+![](images/3c6b17e4d3c5e8be0e583df10e122a1f3d2cc65a1ca8d555bd834746ece7bedb.jpg)  
+Figure 1: The overview of noisy self-training algorithm. Line numbers in Algorithm 1 are above each arrow type.
+
+# Algorithm 1 Noisy Self-Training with Synthetic Queries
+
+Require: Gold query-passage pairs  $\mathcal{T} = \{(q,p^{+})\}$
+
+Require: Passage collection  $\mathcal{P}$
+
+1: Train a query generator  $G$  on gold pairs  $\mathcal{T}$  
+2: Generate queries from  $\mathcal{P}$  and construct synthetic query-passage pairs  $\mathcal{T}^{\prime} = \{(q^{\prime},p^{\prime +})\}_{1}^{\| \mathcal{P}\|\}$ .  
+3: Train a dual-encoder retriever  $\theta_T$  on gold pairs  $\mathcal{T}$  
+4: Use  $\theta_T$  to build ANN index on  $\mathcal{P}$  
+5: Retrieve negatives  $\mathcal{P}_q^-$  for each  $(q,p^{+})\in \mathcal{T}$  and  $\mathcal{P}_{q'}^{-}$  for each  $(q',p'^{+})\in \mathcal{T}'$  
+6: Use  $\theta_T$  as the teacher to generate soft-labels on synthetic queries  $\mathcal{T}' = \{(q',p'^{+},\mathcal{P}_{q'}^{-})\}$  
+7: Pre-train a student retriever  $\theta_S$  on soft-labelled synthetic queries  $\mathcal{T}'$ , with noises injected to  $\theta_S$ 's inputs  
+8: Finetune  $\theta_S$  on gold pairs  $\mathcal{T} = \{(q,p^{+},\mathcal{P}_q^{-})\}$  
+9: Take  $\theta_S$  as new teacher:  $\theta_T \gets \theta_S$  and go back to line 6
+
+Hard negatives for both labelled data  $\{(q,p^{+})\}$  and synthetic data  $\{(q^{\prime},p^{\prime +})\}$  are retrieved using the built index, by taking  $k$  -nearest neighbours while excluding the gold passage.
+
+Noisy Self Training & Fine-tuning (lines 6-8 in Alg. 1) A student model  $\theta_{S}$  is first pre-trained on the synthetic queries with the noised self-training objective (Eq. 4), after which it is fine-tuned on labelled data according to Eq. 2.
+
+Iterative Training (line 9 in Alg. 1) The teacher model can be replaced by the student to generate new pseudo labels and do noisy self-training and fine-tuning all over again.
+
+# 4 General Domain Experiments
+
+# 4.1 Datasets
+
+We evaluate our method on two general-domain datasets: MS-MARCO passage ranking (Bajaj et al., 2016) and Natural Questions (NQ) (Kwiatkowski et al., 2019). MRR@10, Recall@50, Recall@1000 on the MS-MARCO dev set are reported, where MRR represents the Mean Reciprocal Rank, which is calculated as the sum of the reciprocal rank of the first retrieved relevant
+
+passage for each query; and Recall@k is the proportion of relevant passages that appear in the top-k retrievals. Recall@k (k=5, 20, 100) is reported on Natural Questions, which is the proportion of top-k retrieved passages that contain the answer string to each query. The evaluation script provided by Py-serini (Lin et al., 2021) is used for all experiments.
+
+# 4.2 Experimental Settings
+
+We replicate the coCondenser model (Gao and Callan, 2022) using PyTorch (Paszke et al., 2019) and treat it as our baseline.
+
+For self-training, the coCondenser model that was fully trained on labelled data serves as the teacher, which we use to train the student retriever on synthetic queries. After self-training, the student is further finetuned on labelled data, with hyperparameters following Gao and Callan (2022). The last checkpoint is selected for evaluation on test set for all experiments. More details are in Appendix A.1.
+
+# 4.3 Main Results
+
+Table 2 shows the results of our model compared with a range of lexical and neural retrievers. We observe that the performance of our replicated coCondenser is competitive with that of Gao and Callan (2022), achieving slightly better results on Natural Questions but slightly worse on MS-MARCO. By applying our proposed noisy self-training framework, the student coCondenser outperforms the state-of-the-art results on both datasets, resulting in significant improvements over coCondenser (1.2% MRR@10 and R@5 on MS-MARCO and Natural Questions, respectively). This also shows that although coCondenser has already been pre-trained on the target corpus, continuing pre-training it on synthetic queries with noisy self-training can still lead to further performance boost.
+
+To test if the proposed method is general, we also combine it with various pre-trained models, includ
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">MS-MARCO</td><td colspan="3">Natural Questions</td></tr><tr><td>MRR@10</td><td>R@50</td><td>R@1k</td><td>R@5</td><td>R@20</td><td>R@100</td></tr><tr><td>BM25 (Robertson and Zaragoza, 2009)*</td><td>18.7</td><td>59.2</td><td>85.7</td><td>-</td><td>59.1</td><td></td></tr><tr><td>DeepCT (Dai and Callan, 2020)*</td><td>24.3</td><td>69.0</td><td>91.0</td><td>-</td><td>-</td><td>-</td></tr><tr><td>docT5query (Nogueira and Lin, 2019)*</td><td>27.7</td><td>75.6</td><td>94.7</td><td>-</td><td>-</td><td>-</td></tr><tr><td>GAR (Mao et al., 2021)*</td><td>-</td><td>-</td><td>-</td><td>60.9</td><td>74.4</td><td>85.3</td></tr><tr><td>DPR (Karpukhin et al., 2020)*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>74.4</td><td>85.3</td></tr><tr><td>ANCE (Xiong et al., 2021)*</td><td>33.0</td><td>-</td><td>95.9</td><td>-</td><td>81.9</td><td>87.5</td></tr><tr><td>ColBERT (Khattab and Zaharia, 2020)*</td><td>36.0</td><td>82.9</td><td>96.8</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DPR-PAQ (Oguz et al., 2022)*</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>- BERTbase</td><td>31.4</td><td>-</td><td>-</td><td>74.5</td><td>83.7</td><td>88.6</td></tr><tr><td>- RoBERTa(base)</td><td>32.3</td><td>-</td><td>-</td><td>74.2</td><td>84.0</td><td>89.2</td></tr><tr><td>Condenser (Gao and Callan, 2021)*</td><td>36.6</td><td>-</td><td>97.4</td><td>-</td><td>83.2</td><td>88.4</td></tr><tr><td>coCondenser (Gao and Callan, 2022)*</td><td>38.2</td><td>-</td><td>98.4</td><td>75.8</td><td>84.3</td><td>89.0</td></tr><tr><td>Our coCondenser (Teacher) (Gao and Callan, 2022)†</td><td>37.8</td><td>85.5</td><td>98.2</td><td>75.9</td><td>85.0</td><td>89.4</td></tr><tr><td>Student coCondenser</td><td>39.4</td><td>86.7</td><td>98.5</td><td>77.0</td><td>85.5</td><td>89.5</td></tr></table>
+
+![](images/435276f9c36238ea120a7a1e5ec9fe53c983ddd312d4a106263f1ff155b9ca25.jpg)
+
+![](images/18c5824ab0eaea25b401bd159cadbb6d6e175cfb06dcf2f623d6d020613d9524.jpg)  
+Figure 2: Results on MS-MARCO dev and SCIDOCS test sets when using different pre-trained models.
+
+Table 2: Results on MS-MARCO dev set and Natural Questions test set. * indicates results directly copied from Gao and Callan (2022) and Khattab and Zaharia (2020). † indicates our implementation. The best results are marked bold and unavailable results are left blank.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">MS-MARCO</td><td colspan="2">Natural Questions</td></tr><tr><td>MRR@10</td><td>R@1</td><td>R@5</td><td>R@20</td></tr><tr><td>Teacher Reranker</td><td>43.0</td><td>61.9</td><td>80.0</td><td>86.4</td></tr><tr><td>Student Reranker</td><td>43.7</td><td>62.7</td><td>80.3</td><td>86.8</td></tr></table>
+
+ing BERT (Devlin et al., 2019), Contriever (Izacard et al., 2021), Condenser (Gao and Callan, 2021), and RetroMAE (Xiao et al., 2022). Note that all models share the same architecture and only differ in their initial weights. Figure 2 shows the results on MS-MARCO. It is observed that the student model achieves consistent improvements on all pretrained models. Moreover, for worse teachers, the benefits gained by adopting the proposed method are generally more significant (e.g., 2.3 MRR@10 on BERT vs 1.3 MRR@10 on RetroMAE).
+
+Table 3: Compared results between the teacher and student reranker. The top-50 and top-100 predictions from student coCondenser are re-ranked on MS-MARCO and Natural Questions, respectively.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">MS-MARCO</td><td colspan="2">NQ</td></tr><tr><td>MRR@10</td><td>R@50</td><td>R@5</td><td>R@20</td></tr><tr><td>Student coCondenser</td><td>39.4</td><td>86.7</td><td>77.0</td><td>85.5</td></tr><tr><td>w/o noises</td><td>38.8</td><td>86.6</td><td>76.7</td><td>85.5</td></tr><tr><td>w/o pseudo labels</td><td>38.8</td><td>84.5</td><td>75.9</td><td>84.2</td></tr><tr><td>consistency filtering</td><td>38.8</td><td>86.2</td><td>76.5</td><td>85.3</td></tr><tr><td>joint training</td><td>38.4</td><td>86.1</td><td>76.8</td><td>85.3</td></tr></table>
+
+Table 4: Results of different variants compared to the student model on MS-MARCO and Natural Questions.
+
+We also adapt the proposed framework to see if it can be applied to improve an expensive reranker. Similarly, a teacher reranker that was fully trained on labelled data $^2$  is used to generate soft labels on synthetic data, and an identical student reranker is pre-trained. A cross-encoder model is used as the backbone to jointly encode both query and passage. As shown in Table 3, we see that by reranking the student coCondenser's top- $k$  predictions, the student reranker can outperform the teacher, boosting the performance with another  $0.7\%$  and  $0.8\%$  gains on MRR@10 and R@1, respectively.
+
+# 4.4 Ablation Studies
+
+We conduct ablation studies to further understand our methods, with results reported in Table 4.
+
+Noise Injection We remove noise injection during pre-training, and the pre-training objective changes from Eq. 4 to Eq. 3. It is observed that the noisy injection strategies show positive impacts in helping the model rank correct answers higher.
+
+Pseudo Labels We remove line 4 in Algorithm 1 and directly use synthetic labels  $\{(q', p'^+)\}$  together with hard negatives for pre-training in line 5. The results show that directly pre-training on synthetic query-passage pairs leads to inferior performance on both datasets, resulting in significant performance degradation on all metrics, especially on Natural Questions where the R@20 score is even worse than the coCondenser baseline. This indicates that synthetic query-passage pairs contain a large number of noises, and adopting our proposed method can effectively alleviate the negative impacts of noises, which validates our hypothesis.
+
+Consistency Filtering We take the teacher as a consistency filter (Alberti et al., 2019) to remove noises contained in synthetic data. More specifically, for a given synthetic query-passage pair  $(q^{\prime},p^{\prime +})$ , if  $p^{\prime +}$  can be retrieved by the teacher in the top-1 position, this pair is kept; otherwise, it will be discarded. Although this strategy can effectively improve the quality of synthetic data and also yields competitive performance, taking the teacher as a pseudo-label generator leads to better results.
+
+Joint Training We jointly train the student model on both synthetic and labelled data, where the loss becomes  $\mathcal{L} = \mathcal{L}_{\mathrm{ST}} + \mathcal{L}_{\mathrm{CL}}$ . Different batch sizes are also used for synthetic and labelled data to ensure the overall update steps are roughly the same as in pre-train + finetune. We observe that joint training leads to significantly worse results on MS-MARCO but comparable performance on Natural Questions.
+
+# 4.5 Analysis
+
+Data Efficiency We further analyse how the size of labelled query-passage pairs used for training affects the retrieval performance of our method. Smaller datasets are randomly sampled from the full MS-MARCO training data with different sizes, ranging from  $1\%$  to  $70\%$ . Note that in each data size setting, all involved models are restricted to that specific amount of labelled data samples. Figure 3 compares the performance of the teacher and student model trained on labelled data with different sizes. The results reveal that the student surpasses the teacher under all data sizes, and the performance gap becomes larger as the number of
+
+![](images/7640f5b1767d32a25868846f57e3f1e050d7266389282a7fb6914bce78e6c7aa.jpg)  
+Figure 3: Impacts of labelled training data size on MS-MARCO dev set. Teacher refers to coCondenser.
+
+![](images/ba38a14199520b7a11b37c0c0b7001dbd6c8cc78db8cb1f9304584dfe586ee93.jpg)  
+Figure 4: Impacts of randomly shuffling different proportions of tokens on MS-MARCO dev queries at test time. Teacher refers to coCondenser.
+
+labelled data samples decreases. Moreover, the student is superior in terms of data efficiency, achieving performance comparable to the teacher model using full data samples (i.e., our coCondenser in Table 2) with as little as  $30\%$  of labelled data, and the performance continues to improve when more data is available.
+
+Robustness to Shuffled Query Figure 4 compares the performance when different proportions of tokens in each test query are randomly shuffled. We observe that the student model is significantly more robust to queries with tokens in random orders. By removing noise injection during training, the performance drops become increasingly larger with more tokens shuffled, indicating that the noise injection strategy not only leads to performance gains in normal settings but also improves the robustness towards perturbed inputs. This also signifies our model learns lexical matching in capturing salient phrases to some extent, a property typically held in sparse retrievers (e.g., BM25).
+
+# 4.6 Out-of-Domain Experiments
+
+We further test the domain adaptation ability of our proposed method by reporting the performance on the BEIR benchmark (Thakur et al., 2021), which
+
+<table><tr><td rowspan="2">Model (→)</td><td colspan="3">Lexical &amp; Sparse</td><td colspan="6">Dense Retriever</td><td colspan="3">Retriever + Reranker</td></tr><tr><td>BM25*</td><td>DeepCT*</td><td>docT5- query*§</td><td>Teacher</td><td>GenQ§</td><td>GPL§</td><td>GPL-S</td><td>PTR§</td><td>Student Retriever§</td><td>PTR+§</td><td>Teacher Reranker</td><td>Student Reranker§</td></tr><tr><td>Model Size</td><td>-</td><td>-</td><td>-</td><td>110M</td><td>110M×18</td><td>66M×18</td><td>110M</td><td>110M×11</td><td>110M</td><td>110M×11</td><td>110M</td><td>110M</td></tr><tr><td>QGen. Size</td><td>-</td><td>-</td><td>220M</td><td>-</td><td>220M</td><td>220M</td><td>220M</td><td>137B</td><td>220M</td><td>137B</td><td>-</td><td>220M</td></tr><tr><td>TREC-COVID</td><td>65.6</td><td>40.6</td><td>71.3</td><td>74.1</td><td>68.4</td><td>70.0</td><td>73.6</td><td>72.7</td><td>76.7</td><td>76.0</td><td>77.6</td><td>76.5</td></tr><tr><td>BioASQ</td><td>46.5</td><td>40.7</td><td>43.1</td><td>29.5</td><td>31.0</td><td>44.2</td><td>44.7</td><td>-</td><td>35.7</td><td>-</td><td>46.8</td><td>46.8</td></tr><tr><td>NFCorpus</td><td>32.5</td><td>28.3</td><td>32.8</td><td>33.1</td><td>34.6</td><td>34.5</td><td>34.9</td><td>33.4</td><td>34.9</td><td>36.0</td><td>31.0</td><td>32.6</td></tr><tr><td>NQ</td><td>32.9</td><td>18.8</td><td>39.9</td><td>48.4</td><td>44.4</td><td>48.3</td><td>49.0</td><td>-</td><td>50.3</td><td>-</td><td>58.5</td><td>59.1</td></tr><tr><td>HotpotQA</td><td>60.3</td><td>50.3</td><td>58.0</td><td>56.9</td><td>56.5</td><td>58.2</td><td>56.3</td><td>60.4</td><td>58.4</td><td>71.2</td><td>66.6</td><td>68.4</td></tr><tr><td>FiQA-2018</td><td>23.6</td><td>19.1</td><td>29.1</td><td>30.1</td><td>33.0</td><td>34.4</td><td>34.6</td><td>40.4</td><td>33.3</td><td>45.9</td><td>39.1</td><td>39.6</td></tr><tr><td>Signal-1M (RT)</td><td>33.0</td><td>26.9</td><td>30.7</td><td>27.8</td><td>27.0</td><td>27.6</td><td>27.0</td><td>-</td><td>29.5</td><td>-</td><td>26.5</td><td>28.1</td></tr><tr><td>TREC-NEWS</td><td>39.8</td><td>22.0</td><td>42.0</td><td>40.1</td><td>40.7</td><td>42.1</td><td>42.8</td><td>-</td><td>43.8</td><td>-</td><td>43.9</td><td>44.2</td></tr><tr><td>Robust04</td><td>40.8</td><td>28.7</td><td>43.7</td><td>43.6</td><td>40.6</td><td>43.7</td><td>46.9</td><td>-</td><td>46.8</td><td>-</td><td>46.2</td><td>48.6</td></tr><tr><td>ArguAna</td><td>31.5</td><td>30.9</td><td>34.9</td><td>37.3</td><td>46.3</td><td>55.7</td><td>53.4</td><td>53.8</td><td>50.1</td><td>52.1</td><td>55.1</td><td>57.8</td></tr><tr><td>Touché-2020</td><td>36.7</td><td>15.6</td><td>34.7</td><td>30.5</td><td>17.9</td><td>25.5</td><td>24.4</td><td>26.6</td><td>30.4</td><td>27.8</td><td>33.3</td><td>35.0</td></tr><tr><td>CQADupStack</td><td>29.9</td><td>26.8</td><td>32.5</td><td>32.6</td><td>35.5</td><td>35.7</td><td>36.2</td><td>-</td><td>35.8</td><td>-</td><td>35.0</td><td>35.5</td></tr><tr><td>Quora</td><td>78.9</td><td>69.1</td><td>80.2</td><td>85.6</td><td>86.1</td><td>83.6</td><td>83.6</td><td>-</td><td>86.8</td><td>-</td><td>80.0</td><td>82.5</td></tr><tr><td>DBPedia</td><td>31.3</td><td>17.1</td><td>33.1</td><td>37.5</td><td>36.3</td><td>38.4</td><td>38.1</td><td>36.4</td><td>37.2</td><td>41.3</td><td>41.5</td><td>43.3</td></tr><tr><td>SCIDOCS</td><td>15.8</td><td>12.4</td><td>16.2</td><td>14.3</td><td>15.2</td><td>16.9</td><td>16.7</td><td>16.3</td><td>16.1</td><td>19.1</td><td>15.6</td><td>16.4</td></tr><tr><td>FEVER</td><td>75.3</td><td>35.3</td><td>71.4</td><td>71.8</td><td>69.2</td><td>75.9</td><td>75.9</td><td>76.2</td><td>72.2</td><td>83.8</td><td>82.1</td><td>82.1</td></tr><tr><td>Climate-FEVER</td><td>21.3</td><td>6.6</td><td>20.1</td><td>17.6</td><td>21.8</td><td>23.5</td><td>23.4</td><td>21.4</td><td>22.2</td><td>22.6</td><td>23.3</td><td>23.3</td></tr><tr><td>SciFact</td><td>66.5</td><td>63.0</td><td>67.5</td><td>61.0</td><td>66.2</td><td>67.4</td><td>67.6</td><td>62.3</td><td>65.1</td><td>73.2</td><td>70.8</td><td>72.1</td></tr><tr><td colspan="13">Avg. Performance</td></tr><tr><td>PTR-11 Subsets</td><td>41.8</td><td>29.0</td><td>42.6</td><td>42.2</td><td>42.3</td><td>45.5</td><td>45.4</td><td>45.5</td><td>45.2</td><td>49.9</td><td>48.7</td><td>49.7</td></tr><tr><td>All</td><td>42.3</td><td>30.7</td><td>43.4</td><td>42.9</td><td>42.8</td><td>45.9</td><td>46.1</td><td>-</td><td>45.9</td><td>-</td><td>48.5</td><td>49.6</td></tr></table>
+
+Table 5: Results on BEIR benchmark (nDCG@10). Best results are marked bold and second best results are underlined. * indicates results copied from Thakur et al. (2021). † indicates our implementation. § means methods using synthetic queries. ¶ means methods learning from cross-encoder rerankers, thus are not directly comparable to ours. ×n means that these methods train specialised models for each datasets.
+
+contains 18 datasets from retrieval tasks of diverse formats and domains. The average nDCG@10 score over all datasets is used for evaluation.
+
+Implementation Details We use the coCondenser model that was fine-tuned on MS-MARCO as the teacher model in our algorithm, which we call Teacher henceforth. For synthetic queries, we directly use the publicly available ones released by Wang et al. (2022). Since labelled data is not available, the fine-tuning step is eliminated. Unlike previous methods that train specialised retrievers for each task, we train a single universal retriever on the union of synthetic queries from all tasks. We directly use the last checkpoint for evaluation on the test set of each task. To ensure fair comparison, we also re-implement GPL by following the same configurations as in Wang et al. (2022) except that
+
+we employ the Teacher retriever for initialisation and train a single model, denoted as GPL-S.
+
+We also investigate if the proposed method can improve the reranker's out-of-domain performance. Similarly, the reranker trained on MS-MARCO serves as the teacher: Teacher-Reranker, guided by which an indetical Student Reranker is trained. At inference, both rerankers are applied to rerank the top-100 predictions of the Student Retriever. The official evaluation script is used for experiments.6 More details are in Appendix A.2.
+
+Main Results We compare our proposed model with a variety of models, including BM25 (Robertson and Zaragoza, 2009), DocT5query (Nogueira and Lin, 2019), DeepCT (Dai and Callan, 2020), GenQ (Thakur et al., 2021), GPL (Wang et al.,
+
+2022) and PTR (Dai et al., 2023).<sup>7</sup> Table 5 shows the experimental results. We first notice that the Teacher model already achieves strong results compared to other baseline models. GenQ which further finetunes Teacher on synthetic data does not show positive effects, indicating that synthetic queries are extremely noisy and directly using them for training is not beneficial. By contrast, the student retriever trained using our noisy self-training framework significantly improves over Teacher, increasing the averaged nDCG@10 by  $3.0\%$ . Even when being compared to GPL and GPL-S, the models that also use query generation for training data augmentation but take a prohibitively expensive reranker as a pseudo-label generator, and PTR that prompts a 137B instruction-tuned large language model for query generation, our model can still achieve competitive averaged performance and beats these task-specialised retrievers on 6 out of 18 tasks while being comparable on the rest. Note that our model does not rely on any external models, and it is improved in a self-evolution manner, managing to exhibit a high degree of efficiency. More specifically, when being compared to GPL, our method significantly diminishes training time by roughly a factor of three  $(204\mathrm{h} \rightarrow 74\mathrm{h})$  and it achieves a  $25\times$  speed-up in relevance labelling  $(20\mathrm{ms/q} \rightarrow 0.8\mathrm{ms/q})$ . Figure 2 shows that our method achieves consistent improvements and beats GenQ across different pre-trained models on SCIDOCS, a dataset that is significantly different from MS-MARCO in retrieval needs and domain. See Table 10 for full details.
+
+When adopting the same framework to rerankers, the student rereanker can further boost the performance with another 1.1 average points over the teacher reranker. The results again confirm that our method is general and can be extended to improve a powerful reranker in out-of-domain settings.
+
+Discussion The adaptation of our proposed framework in out-of-domain tasks can be interpreted as a specific type of unsupervised domain adaptation (UDA) algorithm (Wang and Deng, 2018), where a model aims to maximise its performance on target domains when only labelled data from in-domain sources and unlabelled data from target domains are available. In our case, the query generator trained on in-domain data (i.e., MS-MARCO) is employed to generate synthetic queries on unlabelled target-domain data (i.e., cor
+
+pus in BEIR datasets). Meanwhile, the retriever trained on in-domain data (i.e. Teacher) generates soft labels for these synthetic queries. As a result, the query generator fabricates distributions of potential queries that could be possibly asked in the target domain; while the teacher retriever captures prevalent patterns of correspondence between queries and passages within the target domain, by leveraging the knowledge it has acquired from the labelled in-domain data. When exposing it to such silver target-domain data during training, a new student retriever is enforced to acquire relevant knowledge that is required to complete retrieval tasks in target domains. Consequently, the exposure to pseudo target-domain data mandates the improvement of domain-specific aptitude within the student retriever, allowing it to effectively retrieve information within target domains.
+
+# 4.7 Self-Training Iterations
+
+Table 6 compares the models trained with varying numbers of iterations. We observe that employing a single iteration yields optimal results, while performance diminishes with more iterations. We conjecture that errors produced in relevance labelling may accumulate over successive iterations, potentially reinforcing the model's bias towards particular error types as the training process continues.
+
+# 4.8 Quality of Pseudo Labels
+
+To assess the quality of the generated soft labels, we randomly sampled 100 synthetic queries and verified the relatedness of the originating passage to each synthetic query through manual examination. We observed that 49 of these queries were indeed related to their source passages. Among this subset of 49 queries, we noted that,  $93\%$  of the time, the soft labels effectively identified alternative passages capable of accurately responding to the query. This was indicated by the assignment of a high probability mass to these alternative passages. In contrast, for the remaining 51 query-passage pairs, where the association was found to be incorrect, soft labels can identify better-matched passages for 42 of them. This observation provides strong evidence for the effectiveness of our proposed self-training methodology.
+
+# 5 Related Work
+
+Neural Dense Retriever Neural retrievers adopt neural networks to encode texts into low
+
+<table><tr><td>Iteration</td><td>MM</td><td>TC</td><td>BA</td><td>NF</td><td>NQ</td><td>HP</td><td>FQ</td><td>SG</td><td>TN</td><td>RB</td><td>AA</td><td>T2</td><td>CQ</td><td>QU</td><td>DB</td><td>SD</td><td>FE</td><td>CF</td><td>SF</td><td>Avg</td></tr><tr><td>0</td><td>37.3</td><td>74.1</td><td>29.5</td><td>33.1</td><td>48.4</td><td>56.9</td><td>30.1</td><td>27.8</td><td>40.1</td><td>43.6</td><td>37.3</td><td>30.5</td><td>32.6</td><td>85.6</td><td>37.5</td><td>14.3</td><td>71.8</td><td>17.6</td><td>61.0</td><td>42.9</td></tr><tr><td>1</td><td>39.4</td><td>76.7</td><td>35.7</td><td>34.9</td><td>50.3</td><td>58.4</td><td>33.3</td><td>29.5</td><td>43.8</td><td>46.8</td><td>50.1</td><td>30.4</td><td>35.8</td><td>86.8</td><td>37.2</td><td>16.1</td><td>72.2</td><td>22.2</td><td>65.1</td><td>45.9</td></tr><tr><td>2</td><td>39.1</td><td>75.7</td><td>36.9</td><td>34.9</td><td>49.2</td><td>57.5</td><td>33.7</td><td>28.7</td><td>42.8</td><td>46.4</td><td>50.6</td><td>29.2</td><td>36.5</td><td>86.2</td><td>35.8</td><td>16.5</td><td>70.9</td><td>22.2</td><td>66.7</td><td>45.6</td></tr></table>
+
+Table 6: Ablations on the number of self-training iterations. Tasks are ordered as in Table 5. MM indicates MS-MARCO and is excluded when computing the average.
+
+dimensional embeddings and show superiority over lexical retrievers when being trained on sufficient data. Karpukhin et al. (2020) adopts a dual-encoder structure to use two independent neural encoders to encode queries and passages into fix-sized vectors separately with their dot product as the relevance score. The model is trained to discriminate positive passages from randomly-sampled irrelevant ones or more informative negatives (Xiong et al., 2021). Other works adopt the poly-encoder architecture (Humeau et al., 2020), where query and documents are represented as multiple vectors to allow token-level interactions (Khattab and Zaharia, 2020). Although effective, the benefits come at the cost of increased index size and a more complex scoring function. In this work, we adopt the dual-encoder structure for its simplicity. However, our proposed method is orthogonal to model architectures, and we believe its combination with poly-encoder retrievers is worth further exploration.
+
+Synthetic Queries for Information Retrieval Synthetic queries have been widely used in information retrieval. Early work expands passages with synthetic queries for lexical retrievers (Nogueira and Lin, 2019). Recent neural models take synthetic queries and their originating passages as positive pairs for model pre-training, resulting in boosted performance (Lu et al., 2021), better domain adaptation ability (Ma et al., 2021; Gangi Reddy et al., 2022) and promising zero-shot results (Dai et al., 2023). In order to reduce the noise in synthetic data, Wang et al. (2022) exploits a cross-encoder reranker to generate pseudo labels. It incurs expensive costs in generating soft labels and training task-customised retrievers, resulting in slow adaptation. Our work follows this direction but significantly differs in that we do not rely on any external model for synthetic data labelling and create dataset-specialised retrievers. By contrast, we empirically show that by using the retriever itself as a more efficient pseudo-label generator, it can be improved in a self-evolution manner with our introduced noisy self-training framework. Moreover, we show this framework is general and can
+
+be extended to boost reranking performance.
+
+Self-Training Self-training (Scudder, 1965) refers to a class of approaches that learn from unlabelled data with pseudo labels. A good teacher model is first trained on labelled data, which is later used to label unlabelled data. Another student model is then pre-trained on unlabelled data first and further finetuned on labelled data. More advanced methods train multiple teachers using features of disjoint partitions on labelled data and a student is learned from their ensembles (Blum and Mitchell, 1998). Recently, the effectiveness of self-training has been verified in a wide range of tasks, including machine translation with back translation (Wu et al., 2019) and language generation (He et al., 2020). These approaches are termed noisy self-training, as the inputs of the student are perturbed. In this work, we follow this direction to show that, for the first time, noisy self-training can be adopted to make better use of synthetic queries to improve neural retrievers in both general-domain and out-of-domain settings.
+
+# 6 Conclusion
+
+In this paper, we present a novel noisy self-training framework for neural retrieval models. It shows that when combined with automatically generated queries, neural retrievers can be improved in a self-evolution manner without relying on any external models. Empirical results on both general-domain and out-of-domain benchmarks confirm the superiority of our proposed method, significantly outperforming a wide range of competitive existing models. We further adapt our method to show it can be applied to improve an expensive reranker.
+
+# Limitations
+
+Although our proposed method does not change model architectures and the resulting models can perform as efficiently as previous models, the introduced training framework does incur the additional training cost, including query generation and the associated hard negative mining. Despite these ex
+
+tra costs, the training of our methods has a fairly modest footprint by modern standards, taking about 3 days of a server with  $4 \times \mathrm{A}100$  GPUs and 450G CPU RAM.
+
+We mainly experiment with dual-encoder-based neural dense retrieval models and additionally adapt the method to reranker training in this work. Alternative retriever methods such as ColBERT (Khattab and Zaharia, 2020) and SPLADE (Formal et al., 2021) are compatible with our approach, and their incorporation should lead to further gains. However, they introduce extra complexity compared to the dual-encoder, e.g., requiring an index over tokens rather than complete passages. We leave their efficient integration as future work.
+
+For out-of-domain experiments, the current method still relies on using general-domain datasets to obtain the teacher model and the query generator. How to achieve this in an unsupervised setting needs further exploration. Moreover, on tasks that have different retrieval needs from general passage retrieval, synthetic queries are normally quite different from the gold standard. For instance, some gold queries in DBPedia are general and may have multiple matching passages (e.g., Give me all people that were born in Vienna and died in Berlin.), but synthetic queries are usually only related to their originating passages (e.g., Who was Johannes Mayer). How to generate synthetic queries with high quality and similar properties to gold standard queries is the key to better retrievers.
+
+As this work uses pre-trained language models for generating synthetic queries, it is possible that undesirable biases (e.g., gender and cultural) from the language models (Wei et al., 2022) is propagated to downstream models. This is in additional to existing biases in training and evaluation datasets (Bigdeli et al., 2021). Evaluating the extent to which biases affect the synthetic data and the resulting model is an inherently complex problem, and remains an open question for future work.
+
+# Acknowledgements
+
+We thank the anonymous reviewers for their helpful feedback and suggestions. The first author is supported by the Graduate Research Scholarships funded by the University of Melbourne. This work was funded by the Australian Research Council, Discovery grant DP230102775. This research was undertaken using the LIEF HPCGPGPU Facil
+
+ity hosted at the University of Melbourne, which was established with the assistance of LIEF Grant LE170100200.
+
+# References
+
+Chris Alberti, Daniel Andor, Emily Pitler, Jacob Devlin, and Michael Collins. 2019. Synthetic QA corpora generation with roundtrip consistency. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6168-6173, Florence, Italy. Association for Computational Linguistics.  
+Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Tong Wang. 2016. Ms marco: A human generated machine reading comprehension dataset.  
+Amin Bigdeli, Negar Arabzadeh, Morteza Zihayat, and Ebrahim Bagheri. 2021. Exploring gender biases in information retrieval relevance judgement datasets. In Advances in Information Retrieval: 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28 - April 1, 2021, Proceedings, Part II, page 216-224, Berlin, Heidelberg. Springer-Verlag.  
+Avrim Blum and Tom Mitchell. 1998. Combining labeled and unlabeled data with co-training. In Proceedings of the Eleventh Annual Conference on Computational Learning Theory, COLT'98, page 92-100, New York, NY, USA. Association for Computing Machinery.  
+Alexander Bondarenko, Maik Frobe, Meriem Beloucif, Lukas Gienapp, Yamen Ajjour, Alexander Panchenko, Chris Biemann, Benno Stein, Henning Wachsmuth, Martin Potthast, and Matthias Hagen. 2020. Overview of touche 2020: Argument retrieval. In Experimental IR Meets Multilinguality, Multimodality, and Interaction, pages 384-395, Cham. Springer International Publishing.  
+Vera Boteva, Demian Gholipour, Artem Sokolov, and Stefan Riezler. 2016. A full-text learning to rank dataset for medical information retrieval. In Advances in Information Retrieval, pages 716-722, Cham. Springer International Publishing.  
+Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1597-1607. PMLR.  
+Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, and Daniel Weld. 2020. SPECTER: Document-level representation learning using citation-informed transformers. In Proceedings of the 58th Annual Meeting of the Association
+
+for Computational Linguistics, pages 2270-2282, Online. Association for Computational Linguistics.  
+Zhuyun Dai and Jamie Callan. 2020. Context-aware term weighting for first stage passage retrieval. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '20, page 1533-1536, New York, NY, USA. Association for Computing Machinery.  
+Zhuyun Dai, Vincent Y Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith Hall, and Ming-Wei Chang. 2023. Promptagator: Few-shot dense retrieval from 8 examples. In The Eleventh International Conference on Learning Representations.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Thibault Formal, Benjamin Piwowarski, and Stéphane Clinchant. 2021. Splade: Sparse lexical and expansion model for first stage ranking. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '21, page 2288-2292, New York, NY, USA. Association for Computing Machinery.  
+Revanth Gangi Reddy, Vikas Yadav, Md Arafat Sultan, Martin Franz, Vittorio Castelli, Heng Ji, and Avirup Sil. 2022. Towards robust neural retrieval with source domain synthetic pre-finetuning. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1065-1070, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Luyu Gao and Jamie Callan. 2021. Condenser: a pretraining architecture for dense retrieval. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 981-993, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Luyu Gao and Jamie Callan. 2022. Unsupervised corpus aware language model pre-training for dense passage retrieval. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2843-2853, Dublin, Ireland. Association for Computational Linguistics.  
+Faegheh Hasibi, Fedor Nikolaev, Chenyan Xiong, Krisztian Balog, Svein Erik Bratsberg, Alexander Kotov, and Jamie Callan. 2017. Dbpedia-entity v2: A test collection for entity search. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval,
+
+SIGIR '17, page 1265-1268, New York, NY, USA. Association for Computing Machinery.  
+Junxian He, Jiatao Gu, Jiajun Shen, and Marc'Aurelio Ranzato. 2020. Revisiting self-training for neural sequence generation. In International Conference on Learning Representations.  
+Doris Hoogeveen, Karin M. Verspoor, and Timothy Baldwin. 2015. Cqadupstack: A benchmark data set for community question-answering research. In Proceedings of the 20th Australasian Document Computing Symposium, ADCS '15, New York, NY, USA. Association for Computing Machinery.  
+Samuel Humeau, Kurt Shuster, Marie-Anne Lachaux, and Jason Weston. 2020. Poly-encoders: Architectures and pre-training strategies for fast and accurate multi-sentence scoring. In International Conference on Learning Representations.  
+Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning.  
+Jeff Johnson, Matthijs Douze, and Herve Jegou. 2021. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547.  
+Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online. Association for Computational Linguistics.  
+Omar Khattab and Matei Zaharia. 2020. *Colbert: Efficient and effective passage search via contextualized late interaction over bert.* In *Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval*, SIGIR '20, page 39-48, New York, NY, USA. Association for Computing Machinery.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466.  
+Markus Leippold and Thomas Diggelmann. 2020. Climate-fever: A dataset for verification of real-world climate claims. In NeurIPS 2020 Workshop on Tackling Climate Change with Machine Learning.  
+Jimmy Lin, Xueguang Ma, Sheng-Chieh Lin, Jheng-Hong Yang, Ronak Pradeep, and Rodrigo Nogueira. 2021. Pyserini: A Python toolkit for reproducible information retrieval research with sparse and dense representations. In Proceedings of the 44th Annual
+
+International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2021), pages 2356-2362.  
+Jing Lu, Gustavo Hernandez Abrego, Ji Ma, Jianmo Ni, and Yinfei Yang. 2021. Multi-stage training with improved negative contrast for neural passage retrieval. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6091-6103, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Ji Ma, Ivan Korotkov, Yinfei Yang, Keith Hall, and Ryan McDonald. 2021. Zero-shot neural passage retrieval via domain-targeted synthetic question generation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1075-1088, Online. Association for Computational Linguistics.  
+Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. 2018. Ww'18 open challenge: Financial opinion mining and question answering. In Companion Proceedings of the The Web Conference 2018, WWW '18, page 1941-1942, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee.  
+Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schütze. 2008. Introduction to Information Retrieval. Cambridge University Press, Cambridge, UK.  
+Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Generation-augmented retrieval for open-domain question answering. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4089-4100, Online. Association for Computational Linguistics.  
+Rodrigo Nogueira and Jimmy Lin. 2019. From doc2query to doctttttquery.  
+Barlas Oguz, Kushal Lakhotia, Anchit Gupta, Patrick Lewis, Vladimir Karpukhin, Aleksandra Piktus, Xilun Chen, Sebastian Riedel, Scott Yih, Sonal Gupta, and Yashar Mehdad. 2022. Domain-matched pre-training tasks for dense retrieval. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 1524-1534, Seattle, United States. Association for Computational Linguistics.  
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning Library. Curran Associates Inc., Red Hook, NY, USA.
+
+Stephen Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: Bm25 and beyond. Found. Trends Inf. Retr., 3(4):333-389.  
+H. Scudder. 1965. Probability of error of some adaptive pattern-recognition machines. IEEE Transactions on Information Theory, 11(3):363-371.  
+Ian Soboroff, Shudong Huang, and Donna Harman. 2019. Trec 2019 news track overview. In TREC.  
+Axel Suarez, Dyaa Albakour, David Corney, Miguel Martinez, and José Esquivel. 2018. A data collection for evaluating the retrieval of related tweets to news articles. In Advances in Information Retrieval, pages 780-786, Cham. Springer International Publishing.  
+Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).  
+James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics.  
+George Tsatsaronis, Georgios Balikas, Prodromos Malakasiotis, Ioannis Partalas, Matthias Zschunke, Michael R Alvers, Dirk Weissenborn, Anastasia Krithara, Sergios Petridis, Dimitris Polychronopoulos, Yannis Almirantis, John Pavlopoulos, Nicolas Baskiotis, Patrick Gallinari, Thierry Artieres, Axel Ngonga, Norman Heino, Eric Gaussier, Liliana Barrio-Alvers, Michael Schroeder, Ion Androutsopoulos, and Georgios Paliouras. 2015. An overview of the bioasq large-scale biomedical semantic indexing and question answering competition. BMC Bioinformatics, 16:138.  
+Ellen Voorhees. 2005. Overview of the trec 2004 robust retrieval track.  
+Ellen Voorhees, Tasmeer Alam, Steven Bedrick, Dina Demner-Fushman, William R. Hersh, Kyle Lo, Kirk Roberts, Ian Soboroff, and Lucy Lu Wang. 2021. Trec-covid: Constructing a pandemic information retrieval test collection. SIGIR Forum, 54(1).  
+Henning Wachsmuth, Shahbaz Syed, and Benno Stein. 2018. Retrieval of the best counterargument without prior topic knowledge. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 241-251, Melbourne, Australia. Association for Computational Linguistics.
+
+David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. 2020. Fact or fiction: Verifying scientific claims. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7534-7550, Online. Association for Computational Linguistics.  
+Kexin Wang, Nandan Thakur, Nils Reimers, and Iryna Gurevych. 2022. GPL: Generative pseudo labeling for unsupervised domain adaptation of dense retrieval. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2345-2360, Seattle, United States. Association for Computational Linguistics.  
+Mei Wang and Weihong Deng. 2018. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153.  
+Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V Le. 2022. Finetuned language models are zero-shot learners. In International Conference on Learning Representations.  
+Lijun Wu, Yiren Wang, Yingce Xia, Tao Qin, Jianhuang Lai, and Tie-Yan Liu. 2019. Exploiting monolingual data at scale for neural machine translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4207-4216, Hong Kong, China. Association for Computational Linguistics.  
+Shitao Xiao, Zheng Liu, Yingxia Shao, and Zhao Cao. 2022. Retromae: Pre-training retrieval-oriented language models via masked auto-encoder.  
+Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. 2020. Self-training with noisy student improvesImagenet classification. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).  
+Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics.  
+David Yarowsky. 1995. Unsupervised word sense disambiguation rivaling supervised methods. In Proceedings of the 33rd Annual Meeting on Association for Computational Linguistics, ACL '95, page
+
+189-196, USA. Association for Computational Linguistics.
+
+# A Implementation Details
+
+Statistics about the datasets used in our experiments are reported in Table 7. Tables 8 and 9 show the hyperparameters used in our experiments.
+
+# A.1 General Domain Datasets
+
+co-condenser-marco $^{8}$  is used for model initialisation on MS-MARCO and co-condenser-wiki $^{9}$  on Natural Questions. For retrievers, the weights of dual encoders are tied. The teacher is trained using the two-stage training following the hyperparameters used in (Gao and Callan, 2022) and (Karpukhin et al., 2020). The student is pre-trained for 1 epoch for all datasets, with a learning rate  $1 \times 10^{-5}$  and batch size 32. The top-100 passages retrieved by the teacher are used as hard negatives, from which 7 negatives are sampled for each query in a minibatch. We empirically find conducting the training with one iteration is enough for achieving the best result. The training takes about 48 hours on both datasets, with up to 4 A100 GPUs.
+
+For generating synthetic queries, we use the publicly available model $^{10}$  to generate queries on MS-MARCO with top-k sampling. For Natural Questions, we first finetune a T5 model on labelled query-passage pairs for 200 epochs with learning rate  $2 \times 10^{-5}$ , which takes about 10 hours on 1 A100 GPU. One query per passage is generated for all datasets, which requires approximately 7 hours on MS-MARCO and 16 hours on Natural Questions using 2 A100 GPUs.
+
+As for the reranker described in §3, ERNIE-base $^{11}$  is used for model initialisation. The teacher reranker is trained with the teacher retriever hard negatives (Figure 1). It is trained for 2 epochs on MS-MARCO, with learning rate set to  $1 \times 10^{-5}$ , batch size 12, and weight decay 0.1. Each query is paired with 40 sampled negatives; while for Natural Questions, we train it for 10 epochs with 15 sampled negatives. For pre-training the student rereanker, the batch size is set to 48 and 128 on MS-MARCO and Natural Questions, respectively, and other hyperparameters remain the same. For the finetuning stage, the settings used for training
+
+the teacher reranker are adopted. Completing the whole training pipeline requires approximately 80 hours on MS-MARCO and 67 hours on Natural Questions with up to 4 A100 GPUs.
+
+# A.2 Out-of-Domain Datasets
+
+On BEIR benchmark, the Teacher retriever is used to mine hard negatives for synthetic queries on each dataset. The top-100 passages retrieved by the Teacher retriever are regarded as the hard negative pool. Student Retriever is initialised from the co-condenser-marco checkpoint. On each dataset, we train a Student Retriever for 10 epochs with learning rate  $1\times 10^{-5}$ . The batch size is set to 32, and each query is paired with 7 sampled negatives. Training a single model on each dataset takes about 8 hours using 2 A100 GPUs and two iterations are used to achieve the best average performance. For GenQ, we follow Thakur et al. (2021) to further finetune the Teacher retriever on synthetic queries of each dataset for 1 epoch with batch size 64 and only in-batch negatives.
+
+The reranker is initialised from ERNIE-base and finetuned for 5 epochs with learning rate  $1 \times 10^{-5}$  and batch size 12. Each query is paired with 15 negatives sampled from the same hard negative pool as above. Training a single model takes about 18 hours using 2 A100 GPUs.
+
+# B Baselines on BEIR
+
+BM25 (Robertson and Zaragoza, 2009) is a lexical retriever based on token matching. DocT5query (Nogueira and Lin, 2019) uses a query generator to synthesise queries and append them to passages as expansion. DeepCT (Dai and Callan, 2020) uses BERT model trained on MS-MARCO to compute the weight of each term in the vocabulary and each passage is represented with keywords multiplied by the term weights. GenQ (Thakur et al., 2021) first trains a dense retrieval model on MS-MARCO and continues to finetune it on synthetic queries with in-batch negatives through contrastive learning. GPL (Wang et al., 2022) uses knowledge distillation to train a dense retriever by learning from a reranker trained on MS-MARCO. PTR (Dai et al., 2023) generates a large number of synthetic queries by prompting an instruction-tuned large language model and trains task-specialised retrievers.
+
+<table><tr><td>Task</td><td>Domain</td><td>Dataset</td><td>Train</td><td>dev</td><td>test</td><td>#Passages</td></tr><tr><td>Passage Retrieval</td><td>Misc.</td><td>MS-MARCO</td><td>502,939</td><td>6,980</td><td>-</td><td>8,841,823</td></tr><tr><td>Open Domain QA</td><td>Wikipedia</td><td>Natural Questions</td><td>58,880</td><td>8,757</td><td>3,610</td><td>21,015,324</td></tr><tr><td colspan="7">BEIR Benchmark</td></tr><tr><td>Biomedical</td><td>Biomedical</td><td>TREC-COVID</td><td>-</td><td>-</td><td>50</td><td>171,332</td></tr><tr><td>Information</td><td>Biomedical</td><td>BioASQ</td><td>-</td><td>-</td><td>500</td><td>14,914,602</td></tr><tr><td>Retrieval</td><td>Biomedical</td><td>NFCorpus</td><td>-</td><td>-</td><td>323</td><td>3,633</td></tr><tr><td rowspan="3">Question Answering</td><td>Wikipedia</td><td>NQ</td><td>-</td><td>-</td><td>3,452</td><td>2,681,468</td></tr><tr><td>Wikipedia</td><td>HotpotQA</td><td>-</td><td>-</td><td>7,405</td><td>5,233,329</td></tr><tr><td>Financial</td><td>FiQA-2018</td><td>-</td><td>-</td><td>648</td><td>57,638</td></tr><tr><td>Twitter Retrieval</td><td>Twitter</td><td>Signal-1M (RT)</td><td>-</td><td>-</td><td>97</td><td>2,866,316</td></tr><tr><td rowspan="2">News Retrieval</td><td>News</td><td>TREC-NEWS</td><td>-</td><td>-</td><td>57</td><td>594,977</td></tr><tr><td>News</td><td>Robust04</td><td>-</td><td>-</td><td>249</td><td>528,155</td></tr><tr><td rowspan="2">Argument Retrieval</td><td>Misc.</td><td>ArguAna</td><td>-</td><td>-</td><td>1,406</td><td>8,674</td></tr><tr><td>Misc.</td><td>Touché-2020</td><td>-</td><td>-</td><td>49</td><td>382,545</td></tr><tr><td rowspan="2">Duplicate Question Retrieval</td><td>StackEchange</td><td>CQADupStack</td><td>-</td><td>-</td><td>13,145</td><td>457,199</td></tr><tr><td>Quora</td><td>Quora</td><td>-</td><td>-</td><td>10,000</td><td>522,931</td></tr><tr><td>Entity Retrieval</td><td>Wikipedia</td><td>DBPedia</td><td>-</td><td>-</td><td>400</td><td>4,635,922</td></tr><tr><td>Citation Prediction</td><td>Scientific</td><td>SCIDOCS</td><td>-</td><td>-</td><td>1,000</td><td>25,657</td></tr><tr><td rowspan="3">Fact Checking</td><td>Wikipedia</td><td>FEVER</td><td>-</td><td>-</td><td>6,666</td><td>5,416,568</td></tr><tr><td>Wikipedia</td><td>Climate-FEVER</td><td>-</td><td>-</td><td>1,535</td><td>5,416,593</td></tr><tr><td>Scientific</td><td>SciFact</td><td>-</td><td>-</td><td>300</td><td>5,183</td></tr></table>
+
+Table 7: Number of query examples in Train/Dev/Test sets and passage number in each target corpus. - means unavailable data that was not used in our experiments. Datasets in BEIR benchmark are grouped according to Thakur et al. (2021), please refer to their paper for full details. TREC-COVID (Voorhees et al., 2021), BioASQ (Tsatsaronis et al., 2015), NFCorpus (Boteva et al., 2016), NQ (Kwiatkowski et al., 2019), HotpotQA (Yang et al., 2018), FiQA-2018 (Maia et al., 2018), Signal-1M (Suarez et al., 2018), TREC-NEWS (Soboroff et al., 2019), Robust04 (Voorhees, 2005), ArguAna (Wachsmuth et al., 2018), Touché (Bondarenko et al., 2020), CQADupStack (Hoogeveen et al., 2015), DBPedia (Hasibi et al., 2017), SCIDOCS (Cohan et al., 2020), FEVER (Thorne et al., 2018), ClimateFEVER (Leippold and Diggelmann, 2020), SciFact (Wadden et al., 2020).
+
+# C Adopting Different Pre-trained Models on BEIR
+
+We test our method on the BEIR benchmark with different pre-trained models. As Table 10 shows, the student model trained using our method achieves consistent improvements on all kinds of pre-trained models, confirming the proposed algorithm is general enough to be adapted to diverse models on out-of-domain settings. Moreover, the patterns for improvements are similar to those in MS-MARCO, with more obvious gains on less-performing teachers (e.g.,  $+4.0\%$  on Contriever vs  $+2.6\%$  on RetroMAE). GenQ shows positive effects on less-performing pre-trained models (i.e., BERT, Condenser, and Contriever). However, it struggles with beating the teacher retriever on more advanced pre-trained models and even degrades average performance. Besides, our method can consistently surpass GenQ across all settings, again confirming its superiority.
+
+# D License
+
+1. MS-MARCO is licensed under the Creative Commons Attribution 4.0 International.  
+2. Natural Questions is licensed under the Apache License 2.0.  
+3. The BEIR benchmark is licensed under the Apache License 2.0.
+
+<table><tr><td></td><td>Hyperparameters</td><td>MS-MARCO</td><td>Natural Questions</td><td>BEIR</td></tr><tr><td rowspan="7">Shared</td><td>Max query length</td><td>32</td><td>32</td><td>32</td></tr><tr><td>Max Passage length</td><td>128</td><td>128</td><td>128</td></tr><tr><td>Optimizer</td><td>AdamW</td><td>AdamW</td><td>AdamW</td></tr><tr><td>Scheduler</td><td>Linear warmup</td><td>Linear warmup</td><td>Linear Warmup</td></tr><tr><td>Warmup proportion</td><td>0.1</td><td>0.1</td><td>0.1</td></tr><tr><td>Weight Tying</td><td>Yes</td><td>Yes</td><td>Yes</td></tr><tr><td>#Params</td><td>110M</td><td>110M</td><td>110M</td></tr><tr><td colspan="5">Teacher Preparation</td></tr><tr><td rowspan="5">Teacher Stage 1</td><td>Learning rate</td><td>5 × 10-6</td><td>1 × 10-5</td><td>-</td></tr><tr><td>Batch size</td><td>8</td><td>128</td><td>-</td></tr><tr><td>#hard negatives</td><td>7</td><td>1</td><td>-</td></tr><tr><td>Hard negative source</td><td>BM25</td><td>BM25</td><td>-</td></tr><tr><td>#Epochs</td><td>3</td><td>40</td><td>-</td></tr><tr><td rowspan="5">Teacher Stage 2</td><td>Learning rate</td><td>5 × 10-6</td><td>1 × 10-5</td><td></td></tr><tr><td>Batch size</td><td>8</td><td>128</td><td>-</td></tr><tr><td>#hard negatives</td><td>7</td><td>1</td><td>-</td></tr><tr><td>Hard negative source</td><td>stage 1 teacher</td><td>stage 1 teacher</td><td>-</td></tr><tr><td>#Epochs</td><td>2</td><td>40</td><td>-</td></tr><tr><td colspan="5">Self Training</td></tr><tr><td rowspan="5">Pre-train</td><td>Learning rate</td><td>1 × 10-5</td><td>1 × 10-5</td><td>1 × 10-5</td></tr><tr><td>Batch size</td><td>32</td><td>32</td><td>32</td></tr><tr><td>#hard negatives</td><td>7</td><td>7</td><td>7</td></tr><tr><td>Hard negative source</td><td>stage 2 teacher</td><td>stage 2 teacher</td><td>stage 2 teacher</td></tr><tr><td>#Epochs</td><td>1</td><td>1</td><td>10</td></tr><tr><td rowspan="5">Finetune</td><td>Learning rate</td><td>5 × 10-6</td><td>1 × 10-5</td><td>-</td></tr><tr><td>Batch size</td><td>8</td><td>128</td><td>-</td></tr><tr><td>#hard negatives</td><td>7</td><td>1</td><td>-</td></tr><tr><td>Hard negative source</td><td>stage 2 teacher</td><td>stage 2 teacher</td><td>-</td></tr><tr><td>#Epochs</td><td>2</td><td>40</td><td>-</td></tr></table>
+
+Table 8: Hyperparameter settings for retriever training.  
+
+<table><tr><td></td><td>Hyperparameters</td><td>MS-MARCO</td><td>Natural Questions</td><td>BEIR</td></tr><tr><td rowspan="6">Shared</td><td>Max length</td><td>156</td><td>156</td><td>351</td></tr><tr><td>Optimizer</td><td>AdamW</td><td>AdamW</td><td>AdamW</td></tr><tr><td>Scheduler</td><td>Linear warmup</td><td>Linear warmup</td><td>Linear Warmup</td></tr><tr><td>Warmup proportion</td><td>0.1</td><td>0.1</td><td>0.1</td></tr><tr><td>Weight decay</td><td>0.1</td><td>0.1</td><td>0.1</td></tr><tr><td>#Params</td><td>110M</td><td>110M</td><td>110M</td></tr><tr><td rowspan="5">Teacher Reranker</td><td>Learning rate</td><td>\( 1 \times 10^{-5} \)</td><td>\( 1 \times 10^{-5} \)</td><td>-</td></tr><tr><td>Batch size</td><td>12</td><td>12</td><td>-</td></tr><tr><td>#hard negatives</td><td>40</td><td>15</td><td>-</td></tr><tr><td>Hard negative source</td><td>stage 2 teacher retriever</td><td>stage 2 teacher retriever</td><td>-</td></tr><tr><td>#Epochs</td><td>2</td><td>10</td><td>-</td></tr><tr><td colspan="5">Self Training</td></tr><tr><td rowspan="5">Pre-train</td><td>Learning rate</td><td>\( 1 \times 10^{-5} \)</td><td>\( 1 \times 10^{-5} \)</td><td>\( 1 \times 10^{-5} \)</td></tr><tr><td>Batch size</td><td>48</td><td>128</td><td>32</td></tr><tr><td>#hard negatives</td><td>40</td><td>15</td><td>7</td></tr><tr><td>Hard negative source</td><td>stage 2 teacher retriever</td><td>stage 2 teacher retriever</td><td>stage 2 teacher retriever</td></tr><tr><td>#Epochs</td><td>1</td><td>1</td><td>5</td></tr><tr><td rowspan="5">Finetune</td><td>Learning rate</td><td>\( 1 \times 10^{-5} \)</td><td>\( 1 \times 10^{-5} \)</td><td>-</td></tr><tr><td>Batch size</td><td>12</td><td>12</td><td>-</td></tr><tr><td>#hard negatives</td><td>40</td><td>15</td><td>-</td></tr><tr><td>Hard negative source</td><td>stage 2 teacher retriever</td><td>stage 2 teacher retriever</td><td>-</td></tr><tr><td>#Epochs</td><td>2</td><td>10</td><td>- -</td></tr></table>
+
+Table 9: Hyperparameter settings for reranker training.
+
+<table><tr><td>Model (→)</td><td colspan="3">BERT</td><td colspan="3">Condenser</td><td colspan="3">Contrlver</td><td colspan="3">coCondenser</td><td colspan="3">RetroMAE</td></tr><tr><td>Dataset</td><td>GenQ</td><td>Tea-cher</td><td>Stu-dent</td><td>GenQ</td><td>Tea-cher</td><td>Stu-dent</td><td>GenQ</td><td>Tea-cher</td><td>Stu-dent</td><td>GenQ</td><td>Tea-cher</td><td>Stu-dent</td><td>GenQ</td><td>Tea-cher</td><td>Stu-dent</td></tr><tr><td>TREC-COVID</td><td>66.3</td><td>60.4</td><td>67.9</td><td>64.2</td><td>65.3</td><td>67.6</td><td>58.0</td><td>46.2</td><td>62.1</td><td>68.4</td><td>74.1</td><td>76.7</td><td>70.0</td><td>70.0</td><td>74.1</td></tr><tr><td>BioASQ</td><td>33.3</td><td>25.7</td><td>30.1</td><td>29.7</td><td>24.6</td><td>32.0</td><td>34.9</td><td>30.6</td><td>35.2</td><td>31.0</td><td>29.5</td><td>35.7</td><td>40.6</td><td>39.0</td><td>43.0</td></tr><tr><td>NFCorpus</td><td>28.8</td><td>26.0</td><td>30.2</td><td>31.3</td><td>26.0</td><td>30.9</td><td>33.8</td><td>32.5</td><td>34.4</td><td>34.6</td><td>33.1</td><td>34.9</td><td>34.8</td><td>34.3</td><td>35.4</td></tr><tr><td>NQ</td><td>40.8</td><td>45.1</td><td>45.9</td><td>43.2</td><td>44.6</td><td>46.9</td><td>38.6</td><td>47.1</td><td>48.4</td><td>44.4</td><td>48.4</td><td>50.3</td><td>44.2</td><td>50.3</td><td>50.5</td></tr><tr><td>HotpotQA</td><td>52.5</td><td>53.0</td><td>55.1</td><td>53.3</td><td>52.5</td><td>56.2</td><td>58.7</td><td>62.2</td><td>62.8</td><td>56.5</td><td>56.9</td><td>58.4</td><td>59.0</td><td>62.3</td><td>62.1</td></tr><tr><td>FiQA-2018</td><td>28.1</td><td>24.4</td><td>28.2</td><td>27.7</td><td>23.1</td><td>29.1</td><td>32.9</td><td>29.4</td><td>33.7</td><td>33.0</td><td>30.1</td><td>33.3</td><td>34.9</td><td>32.1</td><td>34.5</td></tr><tr><td>Signal-1M (RT)</td><td>26.2</td><td>24.8</td><td>27.0</td><td>28.3</td><td>27.3</td><td>27.3</td><td>27.3</td><td>26.7</td><td>28.4</td><td>27.0</td><td>27.8</td><td>29.5</td><td>27.8</td><td>29.0</td><td>30.5</td></tr><tr><td>TREC-NEWS</td><td>35.8</td><td>33.7</td><td>38.3</td><td>37.3</td><td>32.3</td><td>40.5</td><td>40.0</td><td>39.4</td><td>42.9</td><td>40.7</td><td>40.1</td><td>43.8</td><td>39.4</td><td>37.3</td><td>40.7</td></tr><tr><td>Robust04</td><td>35.4</td><td>35.1</td><td>38.6</td><td>37.9</td><td>35.7</td><td>40.9</td><td>36.8</td><td>41.5</td><td>44.6</td><td>40.6</td><td>43.6</td><td>46.8</td><td>40.0</td><td>42.9</td><td>44.0</td></tr><tr><td>ArguAna</td><td>43.1</td><td>32.9</td><td>42.8</td><td>44.1</td><td>31.7</td><td>45.2</td><td>42.3</td><td>31.2</td><td>49.4</td><td>46.3</td><td>37.3</td><td>50.1</td><td>40.3</td><td>32.9</td><td>43.0</td></tr><tr><td>Touché-2020</td><td>16.8</td><td>28.7</td><td>28.5</td><td>17.4</td><td>29.1</td><td>30.1</td><td>16.9</td><td>20.4</td><td>19.4</td><td>17.9</td><td>30.5</td><td>30.4</td><td>16.0</td><td>26.9</td><td>26.1</td></tr><tr><td>CQADupStack</td><td>31.9</td><td>26.0</td><td>32.3</td><td>32.7</td><td>27.7</td><td>32.3</td><td>33.2</td><td>32.4</td><td>36.1</td><td>35.5</td><td>32.6</td><td>35.8</td><td>35.5</td><td>33.8</td><td>37.2</td></tr><tr><td>Quora</td><td>83.8</td><td>82.5</td><td>84.6</td><td>85.6</td><td>84.9</td><td>85.8</td><td>84.5</td><td>84.0</td><td>86.1</td><td>86.1</td><td>85.6</td><td>86.8</td><td>85.3</td><td>85.2</td><td>86.2</td></tr><tr><td>DBPedia</td><td>32.8</td><td>32.5</td><td>32.0</td><td>34.7</td><td>34.0</td><td>33.9</td><td>37.4</td><td>40.4</td><td>38.7</td><td>36.3</td><td>37.5</td><td>37.2</td><td>35.4</td><td>37.7</td><td>35.7</td></tr><tr><td>SCIDOCS</td><td>14.1</td><td>12.4</td><td>14.6</td><td>14.3</td><td>12.6</td><td>15.5</td><td>15.8</td><td>14.8</td><td>17.0</td><td>15.2</td><td>14.3</td><td>16.1</td><td>16.1</td><td>14.7</td><td>16.9</td></tr><tr><td>FEVER</td><td>66.9</td><td>66.9</td><td>72.0</td><td>69.4</td><td>68.5</td><td>72.2</td><td>74.8</td><td>73.2</td><td>75.6</td><td>69.2</td><td>71.8</td><td>72.2</td><td>73.0</td><td>72.4</td><td>76.5</td></tr><tr><td>Climate-FEVER</td><td>22.2</td><td>18.4</td><td>22.3</td><td>20.2</td><td>14.7</td><td>22.3</td><td>23.4</td><td>17.9</td><td>23.6</td><td>21.8</td><td>17.6</td><td>22.2</td><td>19.7</td><td>17.2</td><td>23.8</td></tr><tr><td>SciFact</td><td>65.4</td><td>57.0</td><td>63.4</td><td>64.4</td><td>57.1</td><td>63.2</td><td>68.6</td><td>66.3</td><td>69.8</td><td>66.2</td><td>61.0</td><td>65.1</td><td>67.8</td><td>64.7</td><td>68.7</td></tr><tr><td>Avg. Performance</td><td>40.2</td><td>38.1</td><td>41.9</td><td>40.9</td><td>38.4</td><td>42.9</td><td>42.1</td><td>40.9</td><td>44.9</td><td>42.8</td><td>42.9</td><td>45.9</td><td>43.3</td><td>43.5</td><td>46.1</td></tr></table>
+
+Table 10: Results on BEIR benchmark (nDCG@10) when adopting different pre-trained models. Teacher refers to the model finetuned on MS-MARCO.
+
+![](images/cb62c1ec3f995c317ef44ef3a3aaad31fd5fc4d9f7af0b92d30e630408e52860.jpg)  
+Figure 5: The overview of adapting the self-training framework to reranker.
+
+<table><tr><td colspan="4">Motivation</td></tr><tr><td>Practical</td><td>Cognitive</td><td>Intrinsic</td><td>Fairness</td></tr><tr><td colspan="4">Generalisation type</td></tr><tr><td rowspan="2">Compositional</td><td rowspan="2">Structural</td><td>Cross Task</td><td>Cross Language Cross Domain Robustness</td></tr><tr><td colspan="2">Shift type</td></tr><tr><td>Covariate</td><td>Label</td><td>Full</td><td>Assumed</td></tr><tr><td colspan="4">Shift source</td></tr><tr><td>Naturally occurring</td><td>Partitioned natural</td><td colspan="2">Generated shift Fully generated</td></tr><tr><td colspan="4">Shift locus</td></tr><tr><td>Train-test</td><td>Finetune train-test</td><td>Pretrain-train</td><td>Pretrain-test</td></tr></table>

2023/Noisy Self-Training with Synthetic Queries for Dense Retrieval/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6153d1ffc57841be277238e0109fd1a20499e1c40de403a5f13a3300c9471257
+size 1179888

2023/Non-Autoregressive Document-Level Machine Translation/6f06de2c-1532-40db-88ad-1021607cb22a_content_list.json

@@ -0,0 +1,2057 @@
+[
+    {
+        "type": "text",
+        "text": "Non-Autoregressive Document-Level Machine Translation",
+        "text_level": 1,
+        "bbox": [
+            196,
+            89,
+            801,
+            109
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Guangsheng Bao $^{1,2,*}$ , Zhiyang Teng $^{3,*}$ , Hao Zhou $^{4}$ , Jianhao Yan $^{1,2}$ , and Yue Zhang $^{2,5,\\dagger}$",
+        "bbox": [
+            147,
+            124,
+            863,
+            142
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$  Zhejiang University  $^{2}$  School of Engineering, Westlake University",
+        "bbox": [
+            216,
+            143,
+            784,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$  Nanyang Technological University  $^{4}$  Tsinghua University",
+        "bbox": [
+            250,
+            159,
+            749,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{5}$  Institute of Advanced Technology, Westlake Institute for Advanced Study",
+        "bbox": [
+            196,
+            175,
+            805,
+            191
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2{baoguangsheng, yanjianhao, zhangyue}@westlake.edu.cn",
+        "bbox": [
+            225,
+            192,
+            774,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "3 zhiyang.teng@ntu.edu.sg 4 zhouhao@air.tsinghua.edu.cn",
+        "bbox": [
+            221,
+            209,
+            779,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Non-autoregressive translation (NAT) models achieve comparable performance and superior speed compared to auto-regressive translation (AT) models in the context of sentence-level machine translation (MT). However, their abilities are unexplored in document-level MT, hindering their usage in real scenarios. In this paper, we conduct a comprehensive examination of typical NAT models in the context of document-level MT and further propose a simple but effective design of sentence alignment between source and target. Experiments show that NAT models achieve high acceleration on documents, and sentence alignment significantly enhances their performance.",
+        "bbox": [
+            141,
+            282,
+            460,
+            497
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, current NAT models still have a significant performance gap compared to their AT counterparts. Further investigation reveals that NAT models suffer more from the multi-modality and misalignment issues in the context of document-level MT, and current NAT models struggle with exploiting document context and handling discourse phenomena. We delve into these challenges and provide our code at https://github.com/ baoguangsheng/nat-on-doc.",
+        "bbox": [
+            141,
+            502,
+            460,
+            659
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            676,
+            258,
+            690
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Non-autoregressive neural machine translation (NAT) models achieve significant acceleration on the inference, with even better translation performance, when compared to auto-regressive translation (AT) models on sentence-level MT (Gu et al., 2018, 2019; Stern et al., 2019; Ma et al., 2019; Lee et al., 2020; Huang et al., 2022; Shao and Feng, 2022). However, real applications (e.g., Google Translation and ChatGPT) typically need to understand and respond in discourse, which requires a quick process of document-level content. Different",
+        "bbox": [
+            112,
+            703,
+            490,
+            881
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "from sentence-level MT, document-level MT considers a much larger inter-sentential context and models discourse dependence, such as anaphora, ellipsis, and lexical cohesion (Voita et al., 2019). The sentence-by-sentence translation done by NAT models cannot possibly produce contextual coherent translations, hindering the usage of NAT models in real scenarios.",
+        "bbox": [
+            507,
+            252,
+            884,
+            380
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Document-level context enhances the translation quality of AT models significantly (Werlen et al., 2018; Maruf et al., 2019; Liu et al., 2020; Bao et al., 2021; Sun et al., 2022; Bao et al., 2023). However, no precedent research explores the possibility of applying NAT models to document-level MT, which leaves open research questions including: 1) Can NAT models leverage document context to improve translation quality? 2) Can NAT models generate translations with cross-sentential cohesion and coherence? and 3) Can NAT models achieve the same performance as AT models? To address these questions, we investigate the challenges and opportunities of NAT models on document-level MT.",
+        "bbox": [
+            507,
+            385,
+            884,
+            608
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "NAT models primarily face two challenges: multi-modality (Gu et al., 2018), which causes failure when one source has multiple possible translations, and misalignment (Saharia et al., 2020; Shao and Feng, 2022), which causes repetition and disorder in translations. Previous studies reduce modalities using knowledge distillation (Zhou et al., 2019) and improve alignment using new loss (Saharia et al., 2020; Shao and Feng, 2022; Huang et al., 2022). These methods have proven effective on the sentence level, while their efficacy is uncertain on the document level. First, document-level MT necessitates the modeling of cross-sentential coherence and discourse structure. Second, the extended length of input and output sequences intensifies the potential modalities. Last, the alignment between the source and target becomes harder because the enlarged input and output sequences increase the alignment space exponentially.",
+        "bbox": [
+            507,
+            613,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Contributed equally.",
+        "bbox": [
+            134,
+            891,
+            268,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author.",
+        "bbox": [
+            134,
+            904,
+            280,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we first assess recent NAT models on document-level MT tasks to understand their abilities and limitations, where we observe model failures and unstable performance. According to these observations, we further introduce sentence alignment to NAT models, which equips the encoder and decoder of the NAT models with the group-attention (Bao et al., 2021), restricting the target-to-source attention inside each target and source sentence pairs. Since each target sentence is aligned with a source sentence, we do not need to consider the possibility of aligning the target sentence with other source sentences. Therefore, the alignment space between the input and output sequences is significantly reduced.",
+        "bbox": [
+            115,
+            84,
+            485,
+            324
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Experiments on three benchmark datasets TED, News, and Europarl show that NAT models achieve a translation speed of more than 30 times faster than their AT counterparts on document-level MT, but still lag behind on the translation performance. Sentence alignment significantly enhances the translation performance for NAT models, closing the gap from 2.59 to 1.46 d-BLEU (document-level BLEU score) compared to the best AT results. To the best of our knowledge, we are the first to discuss the challenges and opportunities of NAT models in the context of document-level MT. We release our code and data to stimulate future research.",
+        "bbox": [
+            115,
+            326,
+            485,
+            533
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            115,
+            551,
+            267,
+            565
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Non-Autoregressive Machine Translation. Existing NAT models include three types: fully NAT models (Gu et al., 2018; Libovicky and Helcl, 2018; Qian et al., 2021; Huang et al., 2022), iterative NAT models (Ghazvininejad et al., 2019; Stern et al., 2019; Gu et al., 2019; Chan et al., 2020), and semi-autoregressive model (Ran et al., 2020). All these models generate tokens in parallel at some level. The fully NAT models generate tokens in a single forward pass, while the other two types generate tokens in multiple iterations or steps. In this paper, we take the fully NAT models to investigate document-level MT.",
+        "bbox": [
+            115,
+            580,
+            485,
+            787
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "NAT models face two fundamental challenges: the multi-modality issue and the misalignment issue. The multi-modality issue happens when one source has multiple possible translations, breaking the conditional independence assumption. For example, \"thank you .\" in English can be translated into \"Vielen Dank.\" or \"Danke.\" in German. The second output token could either be \"Dank\" or \". . .",
+        "bbox": [
+            115,
+            790,
+            485,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "which cannot be determined without a condition on the first token. The multi-modality issue causes vanilla NAT models to fail on complex datasets. Previous research leverages knowledge distillation to reduce modalities of a dataset, so that the conditional independence assumption is more likely satisfied (Gu et al., 2018). They generate a translation for each training sample using an AT model and then train NAT models on this generated training set. In this paper, we compare NAT models trained on both raw and knowledge-distilled data.",
+        "bbox": [
+            512,
+            84,
+            880,
+            260
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The misalignment issue elicits various NAT techniques. Vanilla NAT (Gu et al., 2018) uses an implicit token-alignment approach, enforcing a monotonic alignment between source and target by copy encoder outputs to decoder inputs. NAT+CTC (Libovický and Helcl, 2018) introduces connectionist temporal classification (CTC) (Graves et al., 2006) to model the monotonic alignment explicitly. Others introduce non-monotonic alignment between source and target, such as n-gram matching (Shao et al., 2020; Shao and Feng, 2022), aligned cross-entropy (AXE) (Ghazvininejad et al., 2020), and order-agnostic cross-entropy (OAXE) (Du et al., 2021). In this paper, we investigate NAT models using the first two techniques.",
+        "bbox": [
+            512,
+            266,
+            880,
+            506
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Document-Level Machine Translation. Previous methods on document-level MT are dominated by auto-regressive models, which can be categorized into two approaches. The first approach splits a document into sentences and translates each sentence using its context as additional inputs (Zhang et al., 2018; Maruf et al., 2019; Zheng et al., 2021). The second approach takes a document as a whole translation unit and translates the document in one beam search (Liu et al., 2020; Bao et al., 2021; Sun et al., 2022; Bao et al., 2023). In this paper, we follow the second approach for our investigation of NAT models.",
+        "bbox": [
+            512,
+            512,
+            880,
+            719
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Efficient Model for Long Sequence. Recent advances in efficient models such as Longformer (Beltagy et al., 2020), Reformer (Kitaev et al., 2020), Linformer (Wang et al., 2020), and FlashAttention (Dao et al., 2022) improve the time and space complexity of the attention mechanism, which may affect the inference speed of both the auto-regressive and non-autoregressive MT models. In this paper, we focus on the relative inference acceleration of NAT compared to AT models with standard multi-head attention, leaving advanced attention mechanisms for the future.",
+        "bbox": [
+            512,
+            726,
+            880,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Methods",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            225,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 AT Baselines for Document-Level MT",
+        "text_level": 1,
+        "bbox": [
+            112,
+            108,
+            457,
+            124
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Document-level MT can be formulated as a seq2seq problem, where the input document  $x$  and output document  $y$  are represented in token sequence. The auto-regressive factorization is",
+        "bbox": [
+            112,
+            129,
+            487,
+            193
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\theta} (y | x) = \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(y _ {t} \\mid y _ {<   t}, x\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            200,
+            487,
+            244
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $T$  denotes the number of target tokens and  $y_{t}$  conditions on all previous tokens  $y_{< t}$ .",
+        "bbox": [
+            112,
+            250,
+            485,
+            282
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We choose two AT baselines.",
+        "bbox": [
+            132,
+            284,
+            352,
+            297
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Transformer (Vaswani et al., 2017) is the standard encoder-decoder model, which is widely used as a baseline. We represent each document as a sequence of tokens and train the model to do the seq2seq mapping.",
+        "bbox": [
+            112,
+            299,
+            487,
+            380
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "G-Transformer (Bao et al., 2021) extends Transformer by separating the self-attention and cross-attention into sentence-level group-attention and document-level global-attention, as Figure 1 illustrates. The group-attention enables sentence alignment between input and output sequences.",
+        "bbox": [
+            112,
+            380,
+            487,
+            475
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Knowledge Distillation (KD). In addition to the raw data, we also experiment with KD data for training NAT models (Zhou et al., 2019), following previous NAT studies. We use G-Transformer finetuned on sentence Transformer as the teacher to generate distilled document translations.",
+        "bbox": [
+            112,
+            476,
+            487,
+            571
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Existing NAT Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            583,
+            327,
+            599
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Due to the large amount of NAT models (a nonexhaustive search shows more than 60 recent NAT models proposed between 2020 and 2023), we follow the common practice (Huang et al., 2022; Qian et al., 2021) to choose representative NAT models for our investigation, leaving more complex iterative methods, semi-autoregressive methods, and pre-training settings for future. Specifically, we select representative fully NAT models for document-level experiments, where all the models are transformer-based and implemented in Fairseq (Ott et al., 2019).",
+        "bbox": [
+            112,
+            604,
+            487,
+            796
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Vanilla NAT (Gu et al., 2018) is the earliest NAT model, which factorizes translation into two parts",
+        "bbox": [
+            112,
+            797,
+            487,
+            829
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\theta} (y | x) = p _ {\\theta} (T | x) \\cdot \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(y _ {t} | x\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            171,
+            837,
+            487,
+            879
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where the first part predicts the length  $T$  of target  $y$ , the second part predicts target tokens  $y_{t}$  given",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the length, and  $x$  denotes the source input. For simplicity, we take the model as a whole and use the  $\\theta$  to denote all the parameters, which includes the parameters of the length model and the translation model.",
+        "bbox": [
+            507,
+            84,
+            880,
+            162
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "An implicit token alignment is used between the source and target by copying the outputs of the encoder to the inputs of the decoder, where if the source and target have different lengths, they interpolate the decoder inputs uniformly. During training, it initializes the decoder with  $T$  token positions, where  $T$  denotes the real number of target tokens. During inference, it first predicts a number  $T$  using  $p_{\\theta}(T|x)$ , and then initializes the decoder with  $T$  token positions to predict target tokens.",
+        "bbox": [
+            507,
+            165,
+            882,
+            325
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "GLAT (Qian et al., 2021) improves vanilla NAT with a glancing mechanism, which adopts an adaptive glancing sampling strategy, exposing some target fragments to the decoder inputs so that the decoder can fit the training target easier. The glancing mechanism reduces the difficulty of independent training of target tokens and improves the performance on sentence-level MT significantly.",
+        "bbox": [
+            507,
+            325,
+            882,
+            454
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Latent-GLAT (Bao et al., 2022) further improves GLAT using latent variables, where the latent variables are supposed to alleviate the multimodality problem. It represents the modality and alignment in the latent variables, dividing the learning of the target tokens into the modeling of the latent variables and the reconstruction of the target sequence.",
+        "bbox": [
+            507,
+            455,
+            882,
+            582
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "$\\mathbf{NAT} + \\mathbf{CTC}$  (Libovicky and Helcl, 2018) applies connectionist temporal classification (CTC) (Graves et al., 2006) alignment to decoder outputs",
+        "bbox": [
+            507,
+            583,
+            882,
+            632
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\theta} (y | x) = \\sum_ {a \\in \\beta (y)} \\prod_ {i = 1} ^ {M} p _ {\\theta} \\left(a _ {i} | x\\right), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            579,
+            640,
+            882,
+            686
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $a$  denotes an alignment between  $y$  and the reserved  $M$  token positions in the decoder while  $\\beta(y)$  denotes all possible alignments. The alignment  $a$  is assumed conditional independent given the source input  $x$ . The reserved token positions are usually set as times of the length of the source input.",
+        "bbox": [
+            507,
+            694,
+            882,
+            804
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "CTC alignment in Eq. 3 requires a summation over all possible alignments, which is generally intractable. However, with conditional independence assumption for  $a$ , the summation can be achieved using dynamic programming.",
+        "bbox": [
+            507,
+            806,
+            882,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "GLAT+CTC (Qian et al., 2021) combines GLAT with CTC alignment.",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/dbe285da9a0fdf5ef41c154a1f88b05721183927629a4d30490e0c9dda53f7b1.jpg",
+        "image_caption": [
+            "Figure 1: G-Transformer with sentence alignment between source and target documents, where the low layers use the group-attention and only the top 2 layers use the global-attention."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            136,
+            84,
+            463,
+            225
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "DA-Transformer (Huang et al., 2022) represents the output alignment in a directed acyclic graph (DAG), which contains vertices and edges. A path in DAG represents a possible alignment. It models translation as",
+        "bbox": [
+            112,
+            319,
+            489,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\theta} (y | x) = \\sum_ {a \\in \\beta (y)} p _ {\\theta} (a | x) p _ {\\theta} (y | a, x), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            164,
+            409,
+            487,
+            445
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $a$  denotes a path represented in a sequence of vertex indexes and  $\\beta (y)$  here denotes all paths having the same length as the target  $y$ . The right expression in Eq. 4 can further be expanded as",
+        "bbox": [
+            112,
+            456,
+            487,
+            521
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\theta} (a | x) = \\prod_ {i = 1} ^ {M - 1} p _ {\\theta} \\left(a _ {i + 1} \\mid a _ {i}, x\\right), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            531,
+            487,
+            574
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\theta} (y | a, x) = \\prod_ {i = 1} ^ {M} p _ {\\theta} \\left(y _ {i} \\mid a _ {i}, x\\right), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            174,
+            576,
+            487,
+            619
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $a_{i}$  are dependent on each other in a linear chain and  $y_{i}$  are independent with each other given  $a_{i}$  and  $x$ .",
+        "bbox": [
+            112,
+            630,
+            487,
+            677
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Similar to CTC alignment, DA-Transformer also adopts dynamic programming to marginalize all possible paths. The model provides three decoding strategies: greedy, lookahead, and beam search. In this paper, we choose lookahead decoding for balanced performance and inference speed.",
+        "bbox": [
+            112,
+            678,
+            487,
+            776
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 NAT Models with Sentence Alignment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            785,
+            460,
+            802
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We propose novel NAT models with sentence alignment, as shown in Figure 2. The sentence alignment is a key feature in the G-Transformer (Bao et al., 2021), which is implemented through a group-attention module. The group-attention stabilizes self-attention and cross-attention in the model over long sequences. Inspired by the success, we",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/e460e4143e5080f776b4ff06e80cdd4a5ea08084c3ee4dee0229e8725d1c4238.jpg",
+        "image_caption": [
+            "Figure 2: NAT model with sentence alignment, where we replace NAT encoder and decoder layers with G-Transformer layers and we remove the causal mask from G-Transformer decoder layers."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            84,
+            860,
+            225
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "adopt it in our NAT models to reduce the alignment space.",
+        "bbox": [
+            507,
+            319,
+            880,
+            350
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we adopt the G-Transformer encoder and decoder layers by removing the causal mask from self-attention in the decoder layers. We replace the encoder and decoder layers in NAT models with G-Transformer layers and redesign the initial output to include the special token of begin-of-sentence “<s>” and end-of-sentence “</s>” for each target sentence.",
+        "bbox": [
+            507,
+            351,
+            882,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Substituting the causal masking layer in G-Transformer is only a necessary part at the attention level to facilitate sentence alignment. Its effectiveness hinges on both model design and training loss. Specifically, we introduce new length prediction per sentence in G-Trans+GLAT, improve CTC loss for sentence alignment in G-Trans+GLAT+CTC, and improve DAG design to restrict the transition in each sentence in G-Trans+DAT as follows.",
+        "bbox": [
+            507,
+            480,
+            882,
+            623
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "G-Trans+GLAT predicts the target length per sentence instead of the whole sequence. It factorizes the translation as",
+        "bbox": [
+            507,
+            625,
+            882,
+            671
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\theta} (y | x) = \\prod_ {j = 1} ^ {K} \\left[ p _ {\\theta} \\left(T _ {j} \\mid x _ {j}\\right) \\cdot \\prod_ {t = 1} ^ {T _ {j}} p _ {\\theta} \\left(y _ {j, t} \\mid x _ {j}, x\\right) \\right], \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            539,
+            677,
+            882,
+            719
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $K$  denotes the number of sentences. We predict the length  $T_{j}$  of the  $j$ -th target sentence and generate tokens  $y_{j,t}$  accordingly. In addition to Eq. 2, where the corresponding source sentence of  $y_{t}$  is unknown,  $y_{j,t}$  also conditions on the source sentence  $x_{j}$ .",
+        "bbox": [
+            507,
+            726,
+            882,
+            821
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each source sentence  $x_{j}$ , we predict the length  $T_{j}$  of the target sentence using a linear classifier based on the mean pool of output features of the source sentence tokens. Consequently, we calculate the length  $T$  of the target sequence by aggregating sentence lengths  $T = \\sum_{j=1}^{K} T_{j}$ .",
+        "bbox": [
+            507,
+            822,
+            882,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/4c63ecfc31fb0ca25562e43d855fb184178afac08c75a28218c748d7a66140c1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">TED (d-BLEU)</td><td colspan=\"2\">News (d-BLEU)</td><td colspan=\"2\">Europarl (d-BLEU)</td><td colspan=\"2\">Average (d-BLEU)</td><td rowspan=\"2\">Speed -up</td></tr><tr><td>Raw</td><td>KD</td><td>Raw</td><td>KD</td><td>Raw</td><td>KD</td><td>Raw</td><td>KD</td></tr><tr><td colspan=\"10\">AT Baselines</td></tr><tr><td>G-Transformer (Bao et al., 2021)</td><td>27.23</td><td>26.42</td><td>27.22</td><td>26.38</td><td>34.09</td><td>32.87</td><td>29.51</td><td>28.56</td><td>-</td></tr><tr><td>Transformer (Vaswani et al., 2017)</td><td>0.69</td><td>26.04</td><td>0.23</td><td>0.43</td><td>33.41</td><td>26.43</td><td>11.44</td><td>17.63</td><td>1.0x</td></tr><tr><td colspan=\"10\">Existing NAT Models</td></tr><tr><td>Vanilla NAT (Gu et al., 2018)</td><td>0.45</td><td>0.34</td><td>0.12</td><td>0.61</td><td>1.30</td><td>2.43</td><td>0.62</td><td>1.13</td><td>41.2x</td></tr><tr><td>GLAT (Qian et al., 2021)</td><td>1.77</td><td>0.03</td><td>0.01</td><td>2.56</td><td>2.13</td><td>8.17</td><td>1.30</td><td>3.59</td><td>40.0x</td></tr><tr><td>Latent-GLAT (Bao et al., 2022)</td><td>2.10</td><td>1.87</td><td>0.30</td><td>2.30</td><td>5.26</td><td>4.25</td><td>2.55</td><td>2.81</td><td>29.2x</td></tr><tr><td>NAT+CTC (Libovický and Helcl, 2018)</td><td>21.54</td><td>24.98</td><td>15.95</td><td>24.03</td><td>0.00</td><td>31.58</td><td>12.50</td><td>26.86</td><td>27.7x</td></tr><tr><td>GLAT+CTC (Qian et al., 2021)</td><td>18.49</td><td>25.31</td><td>10.23</td><td>20.01</td><td>0.00</td><td>29.47</td><td>9.57</td><td>24.93</td><td>26.3x</td></tr><tr><td>DA-Transformer (Huang et al., 2022)</td><td>20.47</td><td>25.02</td><td>13.99</td><td>23.37</td><td>30.34</td><td>32.14</td><td>21.60</td><td>26.84</td><td>30.8x</td></tr><tr><td colspan=\"10\">NAT Models with Sentence Alignment</td></tr><tr><td>G-Trans+GLAT (ours)</td><td>19.96*</td><td>24.23*</td><td>15.14*</td><td>23.04*</td><td>25.67*</td><td>31.45*</td><td>20.26</td><td>26.24</td><td>30.0x</td></tr><tr><td>G-Trans+GLAT+CTC (ours)</td><td>24.09*</td><td>26.31*</td><td>21.68*</td><td>25.61*</td><td>30.35*</td><td>32.24*</td><td>25.37</td><td>28.05</td><td>20.0x</td></tr><tr><td>G-Trans+DA-Trans (ours)</td><td>23.45*</td><td>25.73*</td><td>21.43*</td><td>24.70*</td><td>30.36</td><td>32.29</td><td>25.08</td><td>27.57</td><td>25.1x</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            884,
+            291
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Main results on raw data (Raw) and knowledge distilled data (KD), where the failures are marked and investigated in section 5.1. We use the official code from the respective papers of NAT models, except NAT+CTC, which is implemented by ourselves based on the GLAT code. “*” denotes a statistical significance at the level of  $p < 0.01$  using a t-test, compared to the corresponding baseline NAT model without sentence alignment.",
+        "bbox": [
+            110,
+            299,
+            882,
+            357
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "G-Trans+GLAT+CTC integrates the sentence alignment with the CTC alignment. The default CTC algorithm aggregates all possible latent alignments across the entire sequence, which may align a source sentence to a wrong target sentence (e.g., the first source sentence to the second target sentence). Such global alignment not only slows down the training process but also causes unstable model performance. Different from the global alignment in Eq. 3, we apply CTC alignment on each sentence",
+        "bbox": [
+            112,
+            382,
+            487,
+            542
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {\\theta} (y | x) = \\prod_ {j = 1} ^ {K} \\left[ \\sum_ {a _ {j} \\in \\beta \\left(y _ {j}\\right)} \\prod_ {i = 1} ^ {M _ {j}} p _ {\\theta} \\left(a _ {j, i} \\mid x _ {j}, x\\right) \\right], \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            154,
+            556,
+            487,
+            599
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $M_{j}$  denotes the reserved token positions for the  $j$ -th target sentence. Since CTC alignment is restricted inside each target sentence, the alignment space  $\\beta(y_{j})$  is enormously reduced compared to the  $\\beta(y)$  in Eq. 3.",
+        "bbox": [
+            112,
+            611,
+            487,
+            690
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "G-Trans+DA-Trans introduces the sentence alignment into the output DAG of the DA-Transformer. The default DAG models the whole sequence, which enables the transition from a vertex in one sentence to a vertex in another sentence, making the transition space linearly increase with the enlargement of the sequence length.",
+        "bbox": [
+            112,
+            693,
+            487,
+            804
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To address the issue, we enforce a constraint to isolate the vertex transitions of each sentence, forcing the path  $a$  on each sentence starting with a special vertex  $\" <   s > \"$  and ending with a special vertex  $\" <   / s > \"$ . The transition between sentences only happens from the vertex  $\" <   / s > \"$  of a previous sentence to the vertex  $\" <   s > \"$  of the current sentence.",
+        "bbox": [
+            112,
+            806,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Formally, we use the same factorization as Eq. 4 but with a different collection  $\\beta(y)$  of paths. At the implementation level, we simply mask the transition matrix to disable transitions from one sentence to another sentence, so that the dynamic programming algorithm keeps unchanged.",
+        "bbox": [
+            507,
+            382,
+            882,
+            479
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            489,
+            653,
+            506
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            514,
+            729,
+            530
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the models using document-level MT benchmark (Maruf et al., 2019), which includes three datasets TED, News, and Europarl, representing three domains and various data scales for English-German translation. More details about each dataset and the preprocessing are in Appendix A.1. We follow Liu et al. (2020), evaluating the model performance in sentence-level BLEU score  $(s\\text{-BLEU})$  and document-level BLEU score  $(d\\text{-BLEU})$ , which are explained in detail as Appendix A.2. We experiment on Base model (where Big model does not provide a stronger baseline) and evaluate the speedup using 1 GPU and 4 CPUs of a Tesla V100 environment as Appendix A.3.",
+        "bbox": [
+            507,
+            533,
+            882,
+            760
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Overall Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            769,
+            678,
+            784
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 1, NAT models achieve high speedup but still suffer from a significant performance gap with their AT counterparts.",
+        "bbox": [
+            507,
+            790,
+            882,
+            838
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Speedup. The inference accelerations of NAT models on the document level are between 25x and 41x, which surpasses the accelerations on the sentence level (between 2x to 15x) by a big margin. Specifically, NAT and GLAT provide the biggest",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "acceleration by around  $40\\mathrm{x}$ . More complex models, such as Latent-GLAT, DA-Transformer, and G-Trans+GLAT, accelerate inference by approximately  $30\\mathrm{x}$ . CTC-based models have lower accelerations between  $20\\mathrm{x}$  and  $27\\mathrm{x}$ . On average, the acceleration on the document level is about  $30\\mathrm{x}$ , which means that document-level MT systems could potentially save  $96\\%$  computational resources and energy consumption by using NAT models.",
+        "bbox": [
+            112,
+            84,
+            489,
+            228
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Performance. Though NAT models underperform G-Transformer, some outpace the Transformer. The lengthy text sequence challenges both the AT models and the NAT models, resulting in exceptionally low d-BLEU scores (<10) in some settings. We treat these low scores as model failures and investigate them in section 5.1.",
+        "bbox": [
+            112,
+            230,
+            489,
+            343
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "With the help of sentence alignment, the performance gap between NAT models and AT baselines is largely reduced, especially when trained on KD data. For example, G-Trans+GLAT+CTC achieves an average d-BLEU of 28.05 on KD data, which is only 0.51 points lower than the d-BLEU of 28.56 of G-Transformer. However, its performance on Raw data is 4.14 points lower than G-Transformer, suggesting that NAT models experience severe challenges with raw data. These results demonstrate that even though knowledge distillation and sentence alignment enhance NAT models largely, there is still a gap between the NAT models and the strongest AT baseline.",
+        "bbox": [
+            115,
+            344,
+            489,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Breakdown Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            587,
+            314,
+            602
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Sentence-level alignment. Sentence alignment substantially enhances the performance of NAT models. Specifically, for GLAT, sentence alignment enhances the performance on KD data from an average of 3.59 to 26.24 d-BLEU of G-Trans+GLAT. For GLAT+CTC, sentence alignment elevates the performance on KD data from 24.93 to 28.05 d-BLEU on average. This score is comparable to G-Transformer on KD data, leaving a minor gap of 0.51 d-BLEU on average. The better results of G-Trans+GLAT+CTC than G-Trans+GLAT suggest that sentence alignment complements output alignment techniques such as CTC.",
+        "bbox": [
+            112,
+            611,
+            489,
+            820
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The sentence alignment brings more benefits to models trained on raw data. As the average scores in Table 1 show, G-Trans+GLAT+CTC outperforms GLAT+CTC by 15.80 points on raw data and 3.12 points on KD data. G-Trans+DA-Trans outperforms DA-Transformer by 3.48 points on",
+        "bbox": [
+            112,
+            822,
+            489,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "raw data and 0.73 points on KD data. These results suggest that sentence alignment mitigates the challenge brought by more modalities in raw data.",
+        "bbox": [
+            507,
+            84,
+            884,
+            131
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Token-level alignment. NAT models with implicit alignment such as vanilla NAT, GLAT, and Latent-GLAT fail on document-level MT, resulting in messy translations (repetitions, wrong words, and meaningless phrases). Conversely, NAT models with explicit alignments, such as NAT+CTC, GLAT+CTC, and DA-Transformer, produce superior document-level performance. DA-Transformer delivers the best overall performance among existing NAT models, even outperforming Transformer on Europarl (KD). These results suggest that token-level alignment plays an important role in NAT models to achieve good performance.",
+        "bbox": [
+            507,
+            135,
+            885,
+            344
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "However, we also observe that training with CTC loss on documents is not always stable, leading to occasional failure and a high variance in their performance (e.g., +/-2.66 d-BLEU on raw data and +/-0.84 on KD data for GLAT+CTC). The failure happens more frequently on raw data than on KD data. We speculate that this instability is caused by the increased alignment space on long sequences, which can be mitigated by sentence alignment. Experiments on G-Trans+GLAT+CTC show stable training and lower variance in performance (+/- 0.19 d-BLEU). These results suggest that token-level alignment alone is not enough to achieve stable training and good performance.",
+        "bbox": [
+            507,
+            347,
+            885,
+            571
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Knowledge distillation. Intuitively, documents have more modalities because they are much longer than sentences. If it is the case, knowledge distillation will become more critical for training NAT models on document-level MT than on sentence-level MT. We compare NAT models trained with and without knowledge distillation.",
+        "bbox": [
+            507,
+            575,
+            884,
+            688
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As Table 1 shows, NAT models trained on KD data generally outperform the same model trained on raw data. The improvement is especially significant for NAT models with explicit alignment. For example, DA-Transformer obtains an average of 26.84 d-BLEU on KD data, which is 5.24 points higher than 21.60 d-BLEU on raw data. In contrast, DA-Transformer on sentence-level MT achieves similar results on raw data and on KD data (Huang et al., 2022). The results suggest that compared to sentences, documents have more severe multimodality issues.",
+        "bbox": [
+            507,
+            690,
+            885,
+            883
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Although knowledge distillation enhances the performance of NAT models, it also sets a ceiling",
+        "bbox": [
+            507,
+            887,
+            884,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5ee81d0b79c72258487b0ab4486f91c348b87b205008822352882329a3e312bb.jpg",
+        "image_caption": [
+            "Figure 3: Speedup rises as long as the sequence length grows, evaluated on TED and News."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            171,
+            82,
+            421,
+            195
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "to the performance on document-level MT. Comparing the performance of G-Transformer on KD data to raw data, we see that KD data downgrades the performance by about 1 d-BLEU on average (from 29.51 to 28.56). These results differ from previous results on sentence-level MT (Huang et al., 2022), where knowledge distillation maintains or even enhances the performance of AT models. The discrepancy calls for further research on techniques to reduce the modalities on the document level.",
+        "bbox": [
+            112,
+            266,
+            489,
+            426
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Speedup on different sequence lengths. We evaluate GLAT and GLAT+CTC on various sequence lengths, including the single sentence and segments with a maximum length of 64, 128, 256, and 512. As Figure 3 shows, the speedup displays consistent trends among different datasets and NAT models. The GLAT generally has a higher speedup than GLAT-CTC, especially when the segment length goes above 256. The speedups on TED and News are almost identical, which is expected because the time costs are supposed to be irrelevant to the data domains. The trends suggest that we benefit more from the inference acceleration on longer documents, where document-level MT tasks provide NAT models with the best scenario.",
+        "bbox": [
+            115,
+            430,
+            489,
+            671
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Speedup on different batch sizes. The previous study (Helcl et al., 2022) reports that NAT models have limited acceleration on the sentence-level MT in parallel inference settings. We evaluate the models in the settings with different batch sizes, including 1, 2, 4, and 8 instances per batch. Given that document-level MT sequences are much longer than sentence-level MT sequences, we do not evaluate using a batch size larger than 8.",
+        "bbox": [
+            112,
+            675,
+            489,
+            819
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As Figure 4 shows, when we increase the batch size from 1 to 8, the overall speedup of GLAT decreases from  $40\\mathrm{x}$  to  $9\\mathrm{x}$ . However, if we consider a more strict calculation of the speedup, excluding the time for initializing the models (which takes almost constant time) from the total evaluation time,",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf8650d2403274b90126466ea4a7406a816f9a9c72e32a7c5552cfde358763da.jpg",
+        "image_caption": [
+            "Figure 4: Speedup with various batch sizes evaluated on News, where the “/ex” conducts a strict calculation of speedup by excluding the time cost for model initialization from both the AT baseline and the NAT models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            566,
+            82,
+            818,
+            195
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the inference speedup of GLAT is  $125\\mathrm{x}$  and  $44\\mathrm{x}$  for the batch size of 1 and 8, respectively. These results suggest that although the speedup ratio decreases for bigger batch sizes, NAT models on document-level MT still show significant acceleration in the parallel inference settings.",
+        "bbox": [
+            507,
+            287,
+            884,
+            385
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 The Challenge of Long Sequence",
+        "text_level": 1,
+        "bbox": [
+            507,
+            395,
+            826,
+            413
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The long input and output in document-level MT bring unexplored challenges to NAT models. In section 4, we learn that NAT models have a significant performance gap with their AT counterparts, and various NAT models even fail in some settings. We investigate these challenges in this section, leaving other discussions in Appendix D.",
+        "bbox": [
+            507,
+            420,
+            884,
+            533
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 The Failure of NAT Models",
+        "text_level": 1,
+        "bbox": [
+            507,
+            544,
+            769,
+            558
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Previous study (Bao et al., 2021) suggests that Transformer fails on long sequences because of local minima of loss values during the training process. We first investigate this possibility for the cause of the failure on NAT models. We evaluate GLAT, GLAT+CTC, Transformer, and GTransformer on different lengths of sequences, as shown in Figure 5a. Overall, the four models produce different patterns of trends. The d-BLEU score of the Transformer rises as long as the sequence length increases until the failure happens at the length of 512 tokens. In contrast, the d-BLEU scores of the GLAT and GLAT+CTC descend progressively as long as the sequence length increases, which suggests a performance decline instead of a training failure on NAT models.",
+        "bbox": [
+            507,
+            565,
+            884,
+            821
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Further evidence on the bigger dataset Europarl confirms that the low performance of NAT models is different from the failure of Transformer. Bao et al. (2021) suggests that the local minima encountered by Transformer can be escaped by using a bigger dataset, resulting in normal scores on Europarl.",
+        "bbox": [
+            507,
+            822,
+            885,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d58843f75fcbc9506a34733eaa2356f44d310e8a4feeb40cdc6c84c76a17a0b4.jpg",
+        "image_caption": [
+            "(a) NAT and AT performance on different sequence length."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            80,
+            349,
+            187
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9f141243d8aa4c436811c51bbc230d887438983f828d307b2059a4b74eed3174.jpg",
+        "image_caption": [
+            "(b) GLAT detailed BLEU scores on different sequence lengths.",
+            "Figure 5: NAT vs. AT models on trends evaluated on News, where Transformer fails abruptly at the length of 512 while NAT models decrease progressively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            381,
+            80,
+            613,
+            186
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9e84efc128d6b4550bae17b17f5d0bea87828d989f05db40bbf2fc78271511b6.jpg",
+        "image_caption": [
+            "(c)  $GLAT + CTC$  detailed BLEU scores on different sequence lengths."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            648,
+            80,
+            880,
+            186
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We evaluate Transformer, GLAT, and GLAT+CTC on Europarl, obtaining d-BLEU scores of 33.41, 2.13, and 0.00, respectively. We could see that GLAT and GLAT+CTC still give low scores on the bigger dataset, suggesting a different cause from the local minima.",
+        "bbox": [
+            112,
+            282,
+            487,
+            378
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We look into the d-BLEU score details, where BLEU- $n$  ( $n \\in [1,2,3,4]$ ) denotes the precision on  $n$ -gram. As the d-BLEU scores on GLAT in Figure 5b show, the BLEU-3/4 decreases rapidly, reaching almost 0.0 at the lengths of 256 and 512. The BLEU-1 remains relatively high at 27.3 on the length of 256, resulting in messy translations, where the generated tokens are related to the content but are repeated, disordered, or in the wrong collocation as shown in Table 2. In comparison, the d-BLEU scores on GLAT+CTC in Figure 5c decrease slowly, where the BLEU-1 even increases on the length of 512. We speculate the rapid decrease in BLEU-3/4 on GLAT is caused by the multi-modality and misalignment issues, which can be mitigated by explicit alignments, such as CTC.",
+        "bbox": [
+            112,
+            382,
+            489,
+            640
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The exceptional zero scores. Table 1 reveals some unexpected results. Both NAT+CTC and GLAT+CTC score a 0.00 d-BLEU on the raw Europarl dataset, a performance notably inferior to that of NAT and GLAT. This unprecedented anomaly may stem from the challenges of applying CTC loss on long sequences. The table further indicates that both combinations, NAT/GLAT+CTC and NAT/GLAT, experience training failures on the raw Europarl dataset. These failures are likely due to multi-modality and misalignment issues. These results empirically demonstrate that, while CTC loss can be effective, its application to document-level training is not consistently stable. We hypothesize that this instability arises from the expanded alignment space between lengthy input and output sequences, as detailed in the token-level alignment",
+        "bbox": [
+            112,
+            645,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Source: companies with a high percentage of floating rate debt stand to lose the most, Goldman said. outside pure stock plays, consumers stand to benefit as well through the rising dollar.",
+        "bbox": [
+            517,
+            282,
+            873,
+            332
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Target: Unternehmen mit einem hohen Anteil an flexiblen Zinsen werden am meisten verlieren, sagte Goldman. außerhalb der reinen Aktienspiele werden Verbraucher ebenfalls durch den steigenden Dollar profitieren.",
+        "bbox": [
+            517,
+            337,
+            875,
+            386
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "GLAT: Unternehmen mit einem hohen, freiem freiem freien erverlieren die die meisten meisten, . . . te te dazu kräften profitieren profitieren profitieren profitieren durch durch steigenden steigenden profitieren profitieren.",
+        "bbox": [
+            517,
+            391,
+            873,
+            451
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "GLAT+CTC: Unternehmen mit hohen der Schulden, die verlieren. außerhalb spielten Verbrauchener den profitieren.",
+        "bbox": [
+            517,
+            456,
+            873,
+            482
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 2: GLAT and GLAT+CTC produce good translation at the beginning but downgrade later with repetitions and missing translations.",
+        "bbox": [
+            507,
+            494,
+            884,
+            538
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "discussion in Section 4.3.",
+        "bbox": [
+            507,
+            564,
+            700,
+            579
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2 Document Context",
+        "text_level": 1,
+        "bbox": [
+            507,
+            592,
+            702,
+            606
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We compare AT and NAT models with document context to the Transformer baseline without document context (trained on sentence) in Appendix B. The results suggest that document context enhances AT model (G-Transformer) by 0.70 s-BLEU. However, NAT models with document context still underperform the Transformer baseline, indicating that current NAT models have a limited ability to utilize document context.",
+        "bbox": [
+            505,
+            612,
+            884,
+            756
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We further apply an ablation study on the document context to quantify its contribution. As Table 3 illustrates, the performance of G-Trans+GLAT+CTC without document context drops by approximately 0.28 s-BLEU on average over the three benchmarks. Specifically, the target-side context contributes merely 0.01, while the source-side context contributes 0.27. The context contributions in G-Trans+GLAT+CTC are less than that in G-Transformer (0.23 and 0.70 s-BLEU on",
+        "bbox": [
+            507,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/6ec6a5ab82f0c1eea9bef9325f6994ddb6ccdf731bf829c9011d670ce2535cb5.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 3: Impact of document context, evaluated in sBLEU.  $\\diamond$  - the scores are from the paper report, where the model is trained on the raw datasets. The NAT models are trained on the KD datasets."
+        ],
+        "table_body": "<table><tr><td>Method (s-BLEU)</td><td>TED</td><td>News</td><td>Europarl</td><td>Drop</td></tr><tr><td>G-Transformer◇</td><td>25.12</td><td>25.52</td><td>32.39</td><td>-</td></tr><tr><td>- target-side context</td><td>25.05</td><td>25.41</td><td>32.16</td><td>-0.14</td></tr><tr><td>- source-side context</td><td>24.56</td><td>24.58</td><td>31.39</td><td>-0.70</td></tr><tr><td>G-Trans+GLAT+CTC</td><td>24.16</td><td>24.09</td><td>30.57</td><td>-</td></tr><tr><td>- target-side context</td><td>24.31</td><td>24.00</td><td>30.47</td><td>-0.01</td></tr><tr><td>- source-side context</td><td>23.96</td><td>24.00</td><td>30.02</td><td>-0.27</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            487,
+            168
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "target and source contexts, respectively). The minor contribution from the target-side context is expected, given that NAT models predict target sentences independently. The relatively low contribution of source-side context indicates that NAT models do not fully exploit the source-side contextual information.",
+        "bbox": [
+            112,
+            249,
+            487,
+            360
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Leak of source context information on adjacent sentences? We observe serious repetitions in the translations generated by NAT models without sentence alignment, raising the concern that the information in a source sentence may leak into its adjacent sentences during translation. We measure the repetitions on NAT models with sentence alignment, where the sentence boundaries are assured by the model design. We find that these models do not generate obvious cross-sentential repetitions. For example, on the TED test set, G-Trans+GLAT generates translations with repetition ratios of 0.14 and 0.02 for 1 and 2 grams, respectively, which are almost identical to the ratios of the reference.",
+        "bbox": [
+            112,
+            363,
+            487,
+            588
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.3 Discourse Phenomena",
+        "text_level": 1,
+        "bbox": [
+            112,
+            604,
+            332,
+            618
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We assess the discourse ability of NAT models using a human-annotated test suite in English-Russian MT (Voita et al., 2019). We train both the AT baselines and NAT models using the 1.5M document (only 4 sentences) pairs. In contrast to previous work (Bao et al., 2021), we do not use the additional 6M sentence pairs for training for the purpose of highlighting their discourse capabilities.",
+        "bbox": [
+            112,
+            627,
+            487,
+            756
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "As Table 4 shows, the NAT models significantly underperform G-Transformer across the four discourse phenomena. The performance of G-Trans+GLAT+CTC matches the Transformer baseline (without document context) on deixis and lexical cohesion (lexcoh), but excels on the ellipsis of inflection (el.infl.) and ellipsis of verb phrase (el.VP). G-Trans+DA-Trans achieve relatively higher deixis than G-Trans+GLAT+CTC because its DAG link models the target-side depen",
+        "bbox": [
+            112,
+            758,
+            487,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/c4eadaa525f72c5dbf867643b01cc27a10da718e88465d6874f229f0ad9c4c9b.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 4: Discourse phenomena. el.infl. - ellipsis of inflection. el.VP - ellipsis of verb phrase. lexcoh - lexical cohesion.  $\\diamondsuit$  - Transformer baseline trained on sentences.  $\\diamondsuit$  - G-Transformer baseline trained on documents."
+        ],
+        "table_body": "<table><tr><td>Method</td><td>deixis</td><td>el.infl.</td><td>el.VP</td><td>lexcoh</td></tr><tr><td>Transformer (sent)♥</td><td>50.0</td><td>53.0</td><td>28.4</td><td>45.9</td></tr><tr><td>G-Transformer◇</td><td>87.1</td><td>82.4</td><td>79.8</td><td>58.6</td></tr><tr><td>G-Trans+GLAT+CTC</td><td>50.0</td><td>55.2</td><td>46.6</td><td>45.9</td></tr><tr><td>G-Trans+DA-Trans</td><td>56.7</td><td>33.0</td><td>21.0</td><td>45.2</td></tr></table>",
+        "bbox": [
+            509,
+            80,
+            882,
+            145
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "dence somehow. These results suggest that current NAT models have some discourse abilities but still struggle with handling discourse phenomena.",
+        "bbox": [
+            507,
+            239,
+            882,
+            287
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            299,
+            796,
+            313
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We investigated NAT models on document-level MT, revealing more severe affections of multimodality and misalignment issues on documents than on sentences. We proposed NAT models with sentence alignment, reducing the possible alignment space, and achieving the best results across three benchmarks. Our experiments show that NAT models significantly accelerate text generation in documents while their performance still lags behind their AT counterparts. Further analysis shows that fully NAT models underutilize document context, leading to loose discourse relations.",
+        "bbox": [
+            507,
+            325,
+            882,
+            517
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "As the first research of NAT models on document-level MT, we hope this work could stimulate future research on reducing the modalities, exploiting document contexts, and modeling discourse dependence.",
+        "bbox": [
+            507,
+            519,
+            882,
+            599
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            612,
+            613,
+            627
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We do not enumerate all recent NAT models. For the purpose of our investigation, we only evaluate the pure (fully) NAT models, leaving other NAT models, such as semi-autoregressive and iterative NAT models, out of our scope.",
+        "bbox": [
+            507,
+            638,
+            880,
+            718
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            507,
+            732,
+            680,
+            747
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We would like to thank the anonymous reviewers for their valuable feedback. This work is funded by the China Strategic Scientific and Technological Innovation Cooperation Project Grant No. 2022YFE0204900 (The Macao counterpart project Grant No. FDCT/0070/2022/AMJ) and the National Natural Science Foundation of China (grant NSFC No. 62161160339). Zhiyang Teng is partially supported by CAAI-Huawei MindSpore Open Fund (CAAIXSJLJJ-2021-046A).",
+        "bbox": [
+            507,
+            758,
+            882,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            83,
+            213,
+            98
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Guangsheng Bao, Zhiyang Teng, and Yue Zhang. 2023. Target-side augmentation for document-level machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10725-10742.",
+            "Guangsheng Bao, Yue Zhang, Zhiyang Teng, Boxing Chen, and Weihua Luo. 2021. G-transformer for document-level machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3442-3455.",
+            "Yu Bao, Hao Zhou, Shujian Huang, Dongqi Wang, Lihua Qian, Xinyu Dai, Jiajun Chen, and Lei Li. 2022. latent-glat: Glancing at latent variables for parallel text generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8398-8409.",
+            "Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150.",
+            "William Chan, Chitwan Sahara, Geoffrey Hinton, Mohammad Norouzi, and Navdeep Jaitly. 2020. Imputer: Sequence modelling via imputation and dynamic programming. In International Conference on Machine Learning, pages 1403-1413. PMLR.",
+            "Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359.",
+            "Cunxiao Du, Zhaopeng Tu, and Jing Jiang. 2021. Order-agnostic cross entropy for non-autoregressive machine translation. In International Conference on Machine Learning, pages 2849-2859. PMLR.",
+            "Marjan Ghazvininejad, Vladimir Karpukhin, Luke Zettlemoyer, and Omer Levy. 2020. Aligned cross entropy for non-autoregressive machine translation. In International Conference on Machine Learning, pages 3515-3523. PMLR.",
+            "Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask-predict: Parallel decoding of conditional masked language models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6112-6121.",
+            "Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376."
+        ],
+        "bbox": [
+            115,
+            107,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jiatao Gu, James Bradbury, Caiming Xiong, Victor OK Li, and Richard Socher. 2018. Non-autoregressive neural machine translation. In International Conference on Learning Representations (ICLR).",
+            "Jiatao Gu, Changhan Wang, and Junbo Zhao. 2019. Levenshtein transformer. Advances in Neural Information Processing Systems, 32.",
+            "Jindrich Helcl, Barry Haddow, and Alexandra Birch. 2022. Non-autoregressive machine translation: It's not as fast as it seems. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1780-1790.",
+            "Fei Huang, Hao Zhou, Yang Liu, Hang Li, and Minlie Huang. 2022. Directed acyclic transformer for nonautoregressive machine translation. In International Conference on Machine Learning, pages 9410-9428. PMLR.",
+            "Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer. arXiv preprint arXiv:2001.04451.",
+            "Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th annual meeting of the association for computational linguistics companion volume proceedings of the demo and poster sessions, pages 177-180.",
+            "Jason Lee, Elman Mansimov, and Kyunghyun Cho. 2020. Deterministic non-autoregressive neural sequence modeling by iterative refinement. In 2018 Conference on Empirical Methods in Natural Language Processing, EMNLP 2018, pages 1173-1182. Association for Computational Linguistics.",
+            "Jindrich Libovický and Jindrich Helcl. 2018. End-to-end non-autoregressive neural machine translation with connectionist temporal classification. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3016-3021.",
+            "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742.",
+            "Xuezhe Ma, Chunting Zhou, Xian Li, Graham Neubig, and Eduard Hovy. 2019. Flowseq: Non-autoregressive conditional sequence generation with generative flow. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4282-4292."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Sameen Maruf, André FT Martins, and Gholamreza Haffari. 2019. Selective attention for context-aware neural machine translation. In Proceedings of NAACL-HLT, pages 3092-3102.",
+            "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. fairseq: A fast, extensible toolkit for sequence modeling. arXiv preprint arXiv:1904.01038.",
+            "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.",
+            "Lihua Qian, Hao Zhou, Yu Bao, Mingxuan Wang, Lin Qiu, Weinan Zhang, Yong Yu, and Lei Li. 2021. Glancing transformer for non-autoregressive neural machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1993-2003.",
+            "Qiu Ran, Yankai Lin, Peng Li, and Jie Zhou. 2020. Learning to recover from multi-modality errors for non-autoregressive neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3059-3069.",
+            "Chitwan Sahara, William Chan, Saurabh Saxena, and Mohammad Norouzi. 2020. Non-autoregressive machine translation with latent alignments. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1098-1108.",
+            "Chenze Shao and Yang Feng. 2022. Non-monotonic latent alignments for ctc-based non-autoregressive machine translation. arXiv preprint arXiv:2210.03953.",
+            "Chenze Shao, Jinchao Zhang, Yang Feng, Fandong Meng, and Jie Zhou. 2020. Minimizing the bag-of-ngrams difference for non-autoregressive neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 198-205.",
+            "Mitchell Stern, William Chan, Jamie Kiros, and Jakob Uszkoreit. 2019. Insertion transformer: Flexible sequence generation via insertion operations. In International Conference on Machine Learning, pages 5976-5985. PMLR.",
+            "Zewei Sun, Mingxuan Wang, Hao Zhou, Chengqi Zhao, Shujian Huang, Jiajun Chen, and Lei Li. 2022. Rethinking document-level neural machine translation. In *Findings of the Association for Computational Linguistics: ACL 2022*, pages 3537-3548.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Elena Voita, Rico Sennrich, and Ivan Titov. 2019. When a good translation is wrong in context: Context-aware machine translation improves on deixis, ellipsis, and lexical cohesion. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1198-1212.",
+            "Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. 2020. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768.",
+            "Lesly Miculicich Werlen, Dhananjay Ram, Nikolaos Pappas, and James Henderson. 2018. Document-level neural machine translation with hierarchical attention networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2947-2954.",
+            "Jiacheng Zhang, Huanbo Luan, Maosong Sun, Feifei Zhai, Jingfang Xu, Min Zhang, and Yang Liu. 2018. Improving the transformer translation model with document-level context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 533-542.",
+            "Zaixiang Zheng, Xiang Yue, Shujian Huang, Jiajun Chen, and Alexandra Birch. 2021. Towards making the most of context in neural machine translation. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 3983-3989.",
+            "Chunting Zhou, Graham Neubig, and Jiatao Gu. 2019. Understanding knowledge distillation in non-autoregressive machine translation. arXiv preprint arXiv:1911.02727."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            542
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/02913b16de3919f404a40be94932820be6209e28ada085fcbc74a381c3d8bac1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Sentences train / dev / test</td><td>Documents train / dev / test</td><td>Segments train / dev / test</td><td>Sents per Segment train / dev / test</td><td>Tokens per Segment train / dev / test</td></tr><tr><td>TED</td><td>0.21M / 9K / 2.3K</td><td>1.7K / 92 / 22</td><td>11K / 483 / 123</td><td>18.3 / 18.5 / 18.3</td><td>436 / 428 / 429</td></tr><tr><td>News</td><td>0.24M / 2K / 3K</td><td>6K / 80 / 154</td><td>18.5K / 172 / 263</td><td>12.8 / 12.6 / 11.3</td><td>380 / 355 / 321</td></tr><tr><td>Europarl</td><td>1.67M / 3.6K / 5.1K</td><td>118K / 239 / 359</td><td>162K / 346 / 498</td><td>10.3 / 10.4 / 10.3</td><td>320 / 326 / 323</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            877,
+            145
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/55aad8d8f7dbd45568b404a4364489cd7de380564fa52640b885f92bc0cf0f93.jpg",
+        "table_caption": [
+            "Table 5: English-German datasets for evaluation, where we split each document into segments with a maximum of 512 tokens."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method (s-BLEU)</td><td colspan=\"2\">TED</td><td colspan=\"2\">News</td><td colspan=\"2\">Europarl</td><td colspan=\"2\">Average</td></tr><tr><td>Raw</td><td>KD</td><td>Raw</td><td>KD</td><td>Raw</td><td>KD</td><td>Raw</td><td>KD</td></tr><tr><td colspan=\"9\">AT Baseline</td></tr><tr><td>Transformer (sent baseline)</td><td>24.63</td><td>-</td><td>24.97</td><td>-</td><td>31.34</td><td>-</td><td>26.98</td><td>-</td></tr><tr><td>G-Transformer (doc baseline)</td><td>25.12</td><td>-</td><td>25.52</td><td>-</td><td>32.39</td><td>-</td><td>27.68</td><td>-</td></tr><tr><td colspan=\"9\">NAT Models</td></tr><tr><td>G-Trans+GLAT</td><td>17.05</td><td>21.81</td><td>13.37</td><td>21.16</td><td>23.49</td><td>29.66</td><td>17.97</td><td>24.21</td></tr><tr><td>G-Trans+GLAT+CTC</td><td>21.87</td><td>24.16</td><td>20.18</td><td>24.09</td><td>28.80</td><td>30.57</td><td>23.62</td><td>26.27</td></tr><tr><td>G-Trans+DA-Trans</td><td>21.02</td><td>23.48</td><td>19.79</td><td>23.10</td><td>28.64</td><td>30.55</td><td>23.15</td><td>25.71</td></tr></table>",
+        "bbox": [
+            181,
+            195,
+            818,
+            309
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Sentence-level performance with and without document context.",
+        "bbox": [
+            248,
+            318,
+            744,
+            332
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            114,
+            357,
+            344,
+            374
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            384,
+            231,
+            398
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the models using document-level MT benchmark (Maruf et al., 2019), which includes three datasets covering three domains and various data scales for English-German translation.",
+        "bbox": [
+            112,
+            404,
+            485,
+            467
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "TED is from IWSLT17, which is transcribed from TED talks that each talk forms a document. tst2016-2017 is used to test the model, and the rest for development.",
+        "bbox": [
+            112,
+            468,
+            485,
+            532
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "News is from News Commentary v11 for training set. For testing and development sets, it uses newtest2016 and newtest2015, respectively.",
+        "bbox": [
+            112,
+            533,
+            487,
+            580
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Europarl is from Europarl v7, where the training, development, and testing sets are randomly split.",
+        "bbox": [
+            112,
+            581,
+            487,
+            627
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 5 shows the detailed statistics. We preprocess the documents by tokenizing and truecasing using MOSES (Koehn et al., 2007) and applying BPE with 30,000 merge operations. We follow Bao et al. (2021) to split each document into segments with a maximum length of 512 tokens.",
+        "bbox": [
+            112,
+            629,
+            487,
+            726
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.2 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            114,
+            737,
+            315,
+            751
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We follow Liu et al. (2020), evaluating the model performance in  $s$ -BLEU and  $d$ -BLEU.",
+        "bbox": [
+            112,
+            757,
+            485,
+            789
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "s-BLEU is calculated on each pair of sentences, which are obtained using the sentence alignment between source and target documents.",
+        "bbox": [
+            112,
+            790,
+            485,
+            837
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "d-BLEU is calculated on each pair of segments, taking the whole segment as a translation unit to compute the BLEU score.",
+        "bbox": [
+            112,
+            838,
+            485,
+            885
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We calculate the BLEU score (Papineni et al., 2002) using sacreBLEU on the detokenized case",
+        "bbox": [
+            112,
+            887,
+            485,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "words.",
+        "bbox": [
+            509,
+            359,
+            561,
+            372
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.3 Model Configuration",
+        "text_level": 1,
+        "bbox": [
+            509,
+            387,
+            724,
+            403
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "All the experiments are run on the Base model, which has 6 layers, 8 heads, 512 embedding dimensions, and 2048 hidden dimensions. We train the models on 4 Tesla V100/A100 GPUs for both AT and NAT. By default, we use the Tesla V100, but in case out-of-memory happens, we switch to Tesla A100 to re-train the model. We do not change the code of existing NAT models, and we obtain the default training and testing arguments from their official code. We update the arguments max-source-positions and max-target-positions to fit the enlarged input and output sequences. We run all main experiments three times and report the median.",
+        "bbox": [
+            507,
+            409,
+            882,
+            632
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We assess the speedup on the test set using a batch size of 1 within a virtual environment equipped with 1 GPU and 4 CPUs of a Tesla V100.",
+        "bbox": [
+            507,
+            636,
+            882,
+            684
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Contribution of Document Context",
+        "text_level": 1,
+        "bbox": [
+            507,
+            697,
+            850,
+            713
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Previous studies demonstrate that document context can significantly enhance MT performance (Zhang et al., 2018; Zheng et al., 2021; Bao et al., 2021). We evaluate its contribution as shown in Table 6. Comparing Transformer on sentence-level MT and G-Transformer on document-level MT, we can see that document context enhances S-BLEU by 0.49, 0.55, and 1.05 on TED, News, and Europarl, respectively. However, the best NAT results produced by G-Trans+GLAT+CTC on KD data are still lower than the Transformer baseline by 0.47, 0.88, and 0.77 on TED, News, and Europarl,",
+        "bbox": [
+            507,
+            726,
+            884,
+            919
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Source: companies with a high percentage of floating rate debt stand to lose the most, Goldman said. outside pure stock plays, consumers stand to benefit as well through the rising dollar. savers could see gains as well through higher yields at the, though experts differ on how quickly that will take hold.",
+        "bbox": [
+            122,
+            85,
+            870,
+            123
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Target: Unternehmen mit einem hohen Anteil an flexiblen Zinsen werden am meisten verlieren, sagte Goldman. außerhalb der reinen Aktienspiele werden Verbraucher ebenfalls durch den steigenden Dollar profitieren. Sparer konnten Gewinne durch höhere Erträge sehen, auch wenn Experten unterschiedlicher Meinung sind, wiechnell das staatfinden wird.",
+        "bbox": [
+            122,
+            127,
+            868,
+            165
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "GLAT: Unternehmen mit einem hohen, freien freien freien erverlieren die die meisten meisten, . . . te te dazu kräften profitieren profitieren profitieren profitieren durch durch steigenden steigenden profitieren. die konnten konden Gewinne Erträge höhere hochere erzielen erzielen erzielen, obwohl sich sich sich sich , , schnell schnell diese diese Fuß wird wird . werden werden . . . . . der die die die , , , , , , , erzielen erzielen erzielen , , sich sich sich sich , , , sich sich sich sich , , , , die die die , , haltenephmen .",
+        "bbox": [
+            122,
+            168,
+            868,
+            230
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "G-Trans+GLAT: Unternehmen mit einem hohen Prozentsatz freier Zinsen werden am am meisten verlieren, sagte Goldman Goldman. außerhalb reinreinen Aktien wurden die Verbraucher ebenso durch durch den steigenden Dollar profitieren. die Sparer konnten durch höhere Erträge Erträge Gewinne erzielen, auch sich die Experten darüber unterscheiden, wiechnel das das einnehmen wird.",
+        "bbox": [
+            122,
+            234,
+            868,
+            282
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "GLAT+CTC: Unternehmen mit hohen der Schulden, die verlieren. außerhalb spiel Verbraucher den profitieren. konnten Gewinne höhere Renditen, sich Experten sich, schnell sich. Verbraucher Aktien werden wird Aktien.",
+        "bbox": [
+            122,
+            288,
+            868,
+            313
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "G-Trans+GLAT+CTC: Unternehmen mit einem hohen Prozentsatz freier Zinsen wurden am meisten verlieren, sagte Goldman. reine Aktienspielen wurden die Verbraucher durch den steigenden Dollar. Sparer konnten Gewinne etwa durch höhere Renditen sehen, auch sich die Experten darüber sind, wie schnell dies haben wird.",
+        "bbox": [
+            122,
+            317,
+            868,
+            355
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Case study. GLAT and GLAT-CTC show good translation quality at the beginning but poor quality at the end of the document. G-Trans+GLAT and G-Trans+GLAT+CTC show more consistent translation quality for sentences at different positions.",
+        "bbox": [
+            112,
+            368,
+            882,
+            411
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "respectively, not to mention the G-Transformer baseline. These results indicate that NAT models have a limited ability to utilize document context.",
+        "bbox": [
+            112,
+            437,
+            485,
+            483
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Case Study.",
+        "text_level": 1,
+        "bbox": [
+            112,
+            498,
+            253,
+            514
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "As a case in Table 7 shows, even though the source is not a lengthy document, sentence alignment enhances the translation quality significantly. Specifically, on the one hand, G-Trans+GLAT and G-Trans+GLAT+CTC estimate the target length more accurately than GLAT and GLAT+CTC because of their fine-grained length prediction on each target sentence. On the other, G-Trans+GLAT and G-Trans+GLAT+CTC show consistent translation quality between the beginning and the end of the document, while GLAT and GLAT+CTC show good quality at the beginning and poor quality at the end. This case illustrates the challenge of long text sequences to NAT models that the translation quality degrades as long as the distance to the beginning of the document increases, and demonstrates the advantage of sentence alignment to stop the degradation.",
+        "bbox": [
+            112,
+            525,
+            487,
+            815
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D More Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            828,
+            294,
+            843
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Can a larger model handle longer input sequences? A larger model does not necessarily solve the longer input sequence issue given the same amount of training data. For pre-trained",
+        "bbox": [
+            112,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "language models, a larger model typically shows a stronger ability to handle long sequences because the increased parameters enable the model to capture more complex long-context dependencies. However, for a non-pretraining setting, a larger model does not necessarily show better performance on the current document-level MT benchmarks. The longer input makes the model more likely to overfit and a larger model will make it worse given the limited training corpus. Take the AT baseline G-Transformer as an example. When we increase the model size from Base to Big, the d-BLEU scores decline by 0.36, 1.41, and 0.10 on TED, News, and Europarl, respectively. When we further increase the model size to Large, the d-BLEU scores further decline by 16.53, 8.49, and 0.56 on TED, News, and Europarl, respectively.",
+        "bbox": [
+            507,
+            437,
+            882,
+            709
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Can the model handle more complex sentence alignments, such as one source sentence divided into multiple target simple sentences? Actually, the current model can handle the case. The sentence-level alignment between the source and target is achieved by the same group tag assigned to the source tokens and the target tokens. We can treat the multiple simple sentences as a whole translation unit, assigning them the same group tag to map them to a single complex sentence.",
+        "bbox": [
+            507,
+            711,
+            882,
+            871
+        ],
+        "page_idx": 12
+    }
+]