Add Batch b86eac8f-e502-4988-8448-2b61c847d979 data

.gitattributes

@@ -6361,3 +6361,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/RoMQA_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Robust,[[:space:]]Multi-evidence,[[:space:]]Multi-answer[[:space:]]Question[[:space:]]Answering/196afee1-17ad-4853-89f8-175284b5619b_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/RobustEmbed_[[:space:]]Robust[[:space:]]Sentence[[:space:]]Embeddings[[:space:]]Using[[:space:]]Self-Supervised[[:space:]]Contrastive[[:space:]]Pre-Training/9ca14927-64ae-4a85-b127-81acac6ddba5_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Robustness[[:space:]]Tests[[:space:]]for[[:space:]]Automatic[[:space:]]Machine[[:space:]]Translation[[:space:]]Metrics[[:space:]]with[[:space:]]Adversarial[[:space:]]Attacks/741055da-30ed-408c-92f2-b92276fe890a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Robustness[[:space:]]of[[:space:]]Named-Entity[[:space:]]Replacements[[:space:]]for[[:space:]]In-Context[[:space:]]Learning/92b3e48a-ba9b-4031-b53b-6f1de232b262_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Role[[:space:]]of[[:space:]]Context[[:space:]]in[[:space:]]Unsupervised[[:space:]]Sentence[[:space:]]Representation[[:space:]]Learning_[[:space:]]the[[:space:]]Case[[:space:]]of[[:space:]]Dialog[[:space:]]Act[[:space:]]Modeling/7f70510d-2f4b-4269-8c9d-3b663839a931_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Roles[[:space:]]of[[:space:]]Scaling[[:space:]]and[[:space:]]Instruction[[:space:]]Tuning[[:space:]]in[[:space:]]Language[[:space:]]Perception_[[:space:]]Model[[:space:]]vs.[[:space:]]Human[[:space:]]Attention/4d81e6a0-662e-4a1d-8420-5bca7d9f2692_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Romanization-based[[:space:]]Large-scale[[:space:]]Adaptation[[:space:]]of[[:space:]]Multilingual[[:space:]]Language[[:space:]]Models/1a02cba7-8ebc-4f20-8832-cb43efde101d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SAC3_[[:space:]]Reliable[[:space:]]Hallucination[[:space:]]Detection[[:space:]]in[[:space:]]Black-Box[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Semantic-aware[[:space:]]Cross-check[[:space:]]Consistency/4a5107be-dc68-42ed-bca9-6c454b388710_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SDOH-NLI_[[:space:]]a[[:space:]]Dataset[[:space:]]for[[:space:]]Inferring[[:space:]]Social[[:space:]]Determinants[[:space:]]of[[:space:]]Health[[:space:]]from[[:space:]]Clinical[[:space:]]Notes/a4f22374-5104-49b6-a9f1-43a434ee9541_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SELFOOD_[[:space:]]Self-Supervised[[:space:]]Out-Of-Distribution[[:space:]]Detection[[:space:]]via[[:space:]]Learning[[:space:]]to[[:space:]]Rank/f2ac6bc4-d902-4552-a8de-3700ee905922_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SGP-TOD_[[:space:]]Building[[:space:]]Task[[:space:]]Bots[[:space:]]Effortlessly[[:space:]]via[[:space:]]Schema-Guided[[:space:]]LLM[[:space:]]Prompting/586b129b-aeb9-4225-8c30-4a6d80f38793_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SHARCS_[[:space:]]Efficient[[:space:]]Transformers[[:space:]]Through[[:space:]]Routing[[:space:]]with[[:space:]]Dynamic[[:space:]]Width[[:space:]]Sub-networks/29c91769-471e-4e25-9798-c851bcedbfe6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SIR-ABSC_[[:space:]]Incorporating[[:space:]]Syntax[[:space:]]into[[:space:]]RoBERTa-based[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]Models[[:space:]]with[[:space:]]a[[:space:]]Special[[:space:]]Aggregator[[:space:]]Token/2a704cf1-766c-4f6e-90e0-05c66a3ec5d6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SQLPrompt_[[:space:]]In-Context[[:space:]]Text-to-SQL[[:space:]]with[[:space:]]Minimal[[:space:]]Labeled[[:space:]]Data/d03d4059-cd6d-446d-ae0d-8da6124a479a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/STEER_[[:space:]]Unified[[:space:]]Style[[:space:]]Transfer[[:space:]]with[[:space:]]Expert[[:space:]]Reinforcement/3bbeeef2-3085-4c7b-94de-46f2df9fd813_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SWEET[[:space:]]-[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Person[[:space:]]Name[[:space:]]Extraction[[:space:]]for[[:space:]]Fighting[[:space:]]Human[[:space:]]Trafficking/51a490a2-5e7e-4b10-904d-5fd5062a1f3a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SYMPTOMIFY_[[:space:]]Transforming[[:space:]]Symptom[[:space:]]Annotations[[:space:]]with[[:space:]]Language[[:space:]]Model[[:space:]]Knowledge[[:space:]]Harvesting/34e7002f-74f2-4990-ab6c-c9c196334fbd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Salespeople[[:space:]]vs[[:space:]]SalesBot_[[:space:]]Exploring[[:space:]]the[[:space:]]Role[[:space:]]of[[:space:]]Educational[[:space:]]Value[[:space:]]in[[:space:]]Conversational[[:space:]]Recommender[[:space:]]Systems/4d0cb677-47cd-4086-a853-d17aea24efc4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Scaling[[:space:]]Law[[:space:]]for[[:space:]]Document[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/d1d71bc0-53c3-42f5-96a3-b8faa324e888_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Scaling[[:space:]]Laws[[:space:]]vs[[:space:]]Model[[:space:]]Architectures_[[:space:]]How[[:space:]]does[[:space:]]Inductive[[:space:]]Bias[[:space:]]Influence[[:space:]]Scaling_/21fba291-c310-416b-86a7-b4a40df8a17b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Scaling[[:space:]]Vision-Language[[:space:]]Models[[:space:]]with[[:space:]]Sparse[[:space:]]Mixture[[:space:]]of[[:space:]]Experts/477a5679-8917-4064-acb5-fe7ac9e31291_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Scene[[:space:]]Graph[[:space:]]Enhanced[[:space:]]Pseudo-Labeling[[:space:]]for[[:space:]]Referring[[:space:]]Expression[[:space:]]Comprehension/f59292b1-1736-4310-bbdc-cf9b3092cfb5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Schema-adaptable[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Construction/7b905e37-b58b-4bb5-a541-f568a0a5526f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Search[[:space:]]Augmented[[:space:]]Instruction[[:space:]]Learning/dbc8d3b6-3908-4a2f-bfbe-5f06e762e736_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SegAugment_[[:space:]]Maximizing[[:space:]]the[[:space:]]Utility[[:space:]]of[[:space:]]Speech[[:space:]]Translation[[:space:]]Data[[:space:]]with[[:space:]]Segmentation-based[[:space:]]Augmentations/346e3771-765b-48d1-a35a-0db773f44c6a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Segmented[[:space:]]Recurrent[[:space:]]Transformer_[[:space:]]An[[:space:]]Efficient[[:space:]]Sequence-to-Sequence[[:space:]]Model/843bad54-cb4c-424a-baeb-0e5a7a070610_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SelectNoise_[[:space:]]Unsupervised[[:space:]]Noise[[:space:]]Injection[[:space:]]to[[:space:]]Enable[[:space:]]Zero-Shot[[:space:]]Machine[[:space:]]Translation[[:space:]]for[[:space:]]Extremely[[:space:]]Low-resource[[:space:]]Languages/29f3935c-fb5b-4050-a8ad-521bc9fe7631_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Selecting[[:space:]]Key[[:space:]]Views[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Entity[[:space:]]Linking/06f7bda4-973b-4bc1-aff2-28c32ca54c97_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Selective[[:space:]]Demonstrations[[:space:]]for[[:space:]]Cross-domain[[:space:]]Text-to-SQL/e2d63869-ca04-4735-be27-319213b7db7c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-Knowledge[[:space:]]Guided[[:space:]]Retrieval[[:space:]]Augmentation[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/998f7293-c03b-45ed-bf56-cdb482e089ec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-Polish_[[:space:]]Enhance[[:space:]]Reasoning[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Problem[[:space:]]Refinement/c7f78779-0505-4b54-9d8c-1100f1efcbf4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-Supervised[[:space:]]Behavior[[:space:]]Cloned[[:space:]]Transformers[[:space:]]are[[:space:]]Path[[:space:]]Crawlers[[:space:]]for[[:space:]]Text[[:space:]]Games/20196f57-25bc-4147-a092-f2da72cb61a9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-Supervised[[:space:]]Rule[[:space:]]Learning[[:space:]]to[[:space:]]Link[[:space:]]Text[[:space:]]Segments[[:space:]]to[[:space:]]Relational[[:space:]]Elements[[:space:]]of[[:space:]]Structured[[:space:]]Knowledge/414cb7d8-f28a-4147-86db-8782a97e36bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-distilled[[:space:]]Transitive[[:space:]]Instance[[:space:]]Weighting[[:space:]]for[[:space:]]Denoised[[:space:]]Distantly[[:space:]]Supervised[[:space:]]Relation[[:space:]]Extraction/d91426af-5287-4140-a5cb-6a259295cf81_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-prompted[[:space:]]Chain-of-Thought[[:space:]]on[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Open-domain[[:space:]]Multi-hop[[:space:]]Reasoning/a17aeaa2-dcfd-44eb-b9fa-cbe1fbd08f03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-supervised[[:space:]]Meta-Prompt[[:space:]]Learning[[:space:]]with[[:space:]]Meta-Gradient[[:space:]]Regularization[[:space:]]for[[:space:]]Few-shot[[:space:]]Generalization/cba196e6-3604-4774-8b70-1747605c92a9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-supervised[[:space:]]Post-processing[[:space:]]Method[[:space:]]to[[:space:]]Enrich[[:space:]]Pretrained[[:space:]]Word[[:space:]]Vectors/8e5cce6f-e78d-452c-85ca-e7d3cbafc61d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantic[[:space:]]Decomposition[[:space:]]of[[:space:]]Question[[:space:]]and[[:space:]]SQL[[:space:]]for[[:space:]]Text-to-SQL[[:space:]]Parsing/24c1b45e-a8f4-46f9-9983-ebff85cb144b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantic[[:space:]]Parsing[[:space:]]by[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Intricate[[:space:]]Updating[[:space:]]Strategies[[:space:]]of[[:space:]]Zero-Shot[[:space:]]Dialogue[[:space:]]State[[:space:]]Tracking/743bb9e3-7bb3-4a87-aacd-21ec129787ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantic[[:space:]]Similarity[[:space:]]Covariance[[:space:]]Matrix[[:space:]]Shrinkage/61a0db41-237c-4393-b759-b8a6269d1f6e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semi-Structured[[:space:]]Object[[:space:]]Sequence[[:space:]]Encoders/6e327045-ce68-4822-a6c3-3e26d2925114_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sharing,[[:space:]]Teaching[[:space:]]and[[:space:]]Aligning_[[:space:]]Knowledgeable[[:space:]]Transfer[[:space:]]Learning[[:space:]]for[[:space:]]Cross-Lingual[[:space:]]Machine[[:space:]]Reading[[:space:]]Comprehension/94860b0e-b6b8-46ac-9884-3c17f12798f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Show,[[:space:]]Write,[[:space:]]and[[:space:]]Retrieve_[[:space:]]Entity-aware[[:space:]]Article[[:space:]]Generation[[:space:]]and[[:space:]]Retrieval/c2ee6168-3c12-4507-a324-57969290b9e7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SiMFy_[[:space:]]A[[:space:]]Simple[[:space:]]Yet[[:space:]]Effective[[:space:]]Approach[[:space:]]for[[:space:]]Temporal[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Reasoning/4cb09716-7156-43ce-8c2e-0b4763cf3744_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SimCKP_[[:space:]]Simple[[:space:]]Contrastive[[:space:]]Learning[[:space:]]of[[:space:]]Keyphrase[[:space:]]Representations/f5d22dff-d4ee-41ab-8281-25b7ba3cad7b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Simple[[:space:]]Hardware-Efficient[[:space:]]PCFGs[[:space:]]with[[:space:]]Independent[[:space:]]Left[[:space:]]and[[:space:]]Right[[:space:]]Productions/06f8a66f-f3b6-444e-bf2e-472e017b8bbd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Simpler[[:space:]]neural[[:space:]]networks[[:space:]]prefer[[:space:]]subregular[[:space:]]languages/66af9635-ca8c-4d95-821b-351292c187e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Simultaneous[[:space:]]Machine[[:space:]]Translation[[:space:]]with[[:space:]]Tailored[[:space:]]Reference/71acfca9-9272-4b42-91c8-9299d69b3815_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Smart[[:space:]]“Chef”_[[:space:]]Verifying[[:space:]]the[[:space:]]Effect[[:space:]]of[[:space:]]Role-based[[:space:]]Paraphrasing[[:space:]]for[[:space:]]Aspect[[:space:]]Term[[:space:]]Extraction/76d67012-9195-430d-8599-8f48f8506969_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SmartSpanNER_[[:space:]]Making[[:space:]]SpanNER[[:space:]]Robust[[:space:]]in[[:space:]]Low[[:space:]]Resource[[:space:]]Scenarios/b80e326d-e69c-4af3-9d89-4f2b37ce32cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Social[[:space:]]Commonsense-Guided[[:space:]]Search[[:space:]]Query[[:space:]]Generation[[:space:]]for[[:space:]]Open-Domain[[:space:]]Knowledge-Powered[[:space:]]Conversations/73da73d7-0016-4ed0-abaf-c5ed9fda597f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Solving[[:space:]]the[[:space:]]Right[[:space:]]Problem[[:space:]]is[[:space:]]Key[[:space:]]for[[:space:]]Translational[[:space:]]NLP_[[:space:]]A[[:space:]]Case[[:space:]]Study[[:space:]]in[[:space:]]UMLS[[:space:]]Vocabulary[[:space:]]Insertion/15d5736c-6b1d-42c8-8d4a-58ba27f2ad52_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SoulChat_[[:space:]]Improving[[:space:]]LLMs’[[:space:]]Empathy,[[:space:]]Listening,[[:space:]]and[[:space:]]Comfort[[:space:]]Abilities[[:space:]]through[[:space:]]Fine-tuning[[:space:]]with[[:space:]]Multi-turn[[:space:]]Empathy[[:space:]]Conversations/74f3215a-0faa-4c0d-ab2e-a847453e3539_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sound[[:space:]]of[[:space:]]Story_[[:space:]]Multi-modal[[:space:]]Storytelling[[:space:]]with[[:space:]]Audio/99521d4b-eefb-4d04-87ae-40101ca7f8f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sources[[:space:]]of[[:space:]]Hallucination[[:space:]]by[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]on[[:space:]]Inference[[:space:]]Tasks/cf9798f7-c301-41e3-9b65-cde2050b49d8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sparse[[:space:]]Black-Box[[:space:]]Multimodal[[:space:]]Attack[[:space:]]for[[:space:]]Vision-Language[[:space:]]Adversary[[:space:]]Generation/d54be74f-477a-407c-a5e8-5c4936d96dff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sparse[[:space:]]Frame[[:space:]]Grouping[[:space:]]Network[[:space:]]with[[:space:]]Action[[:space:]]Centered[[:space:]]for[[:space:]]Untrimmed[[:space:]]Video[[:space:]]Paragraph[[:space:]]Captioning/b6e9cd8e-141b-4100-a1f2-76a7ec8ca1a4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Speaking[[:space:]]Style[[:space:]]Conversion[[:space:]]in[[:space:]]the[[:space:]]Waveform[[:space:]]Domain[[:space:]]Using[[:space:]]Discrete[[:space:]]Self-Supervised[[:space:]]Units/90c3c132-92bd-4d68-aac0-d2e1088568f8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Speculative[[:space:]]Decoding_[[:space:]]Exploiting[[:space:]]Speculative[[:space:]]Execution[[:space:]]for[[:space:]]Accelerating[[:space:]]Seq2seq[[:space:]]Generation/afabe768-91ed-4314-a1b3-7c94661ccb97_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SpeechGPT_[[:space:]]Empowering[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Intrinsic[[:space:]]Cross-Modal[[:space:]]Conversational[[:space:]]Abilities/85d328a2-4439-4f42-90fc-54c5bc9b7daf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Statistically[[:space:]]Profiling[[:space:]]Biases[[:space:]]in[[:space:]]Natural[[:space:]]Language[[:space:]]Reasoning[[:space:]]Datasets[[:space:]]and[[:space:]]Models/5c107149-777b-43a2-a3a4-33d7d122d935_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SteerLM_[[:space:]]Attribute[[:space:]]Conditioned[[:space:]]SFT[[:space:]]as[[:space:]]an[[:space:]](User-Steerable)[[:space:]]Alternative[[:space:]]to[[:space:]]RLHF/b74fd9b2-3087-4740-b345-13f453f014de_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Steering[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Machine[[:space:]]Translation[[:space:]]with[[:space:]]Finetuning[[:space:]]and[[:space:]]In-Context[[:space:]]Learning/3356a2d7-0dd3-44c6-85e6-7157e3682b8c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Strong[[:space:]]and[[:space:]]Efficient[[:space:]]Baselines[[:space:]]for[[:space:]]Open[[:space:]]Domain[[:space:]]Conversational[[:space:]]Question[[:space:]]Answering/e4d4b958-cdad-450e-bcb0-13eba56e0063_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Structure[[:space:]]and[[:space:]]Label[[:space:]]Constrained[[:space:]]Data[[:space:]]Augmentation[[:space:]]for[[:space:]]Cross-domain[[:space:]]Few-shot[[:space:]]NER/5e408da6-5360-4cd0-8091-cb592e563355_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Style-Aware[[:space:]]Radiology[[:space:]]Report[[:space:]]Generation[[:space:]]with[[:space:]]RadGraph[[:space:]]and[[:space:]]Few-Shot[[:space:]]Prompting/e948d766-c9e9-46c5-a899-aabf24f48fe1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/StyleBART_[[:space:]]Decorate[[:space:]]Pretrained[[:space:]]Model[[:space:]]with[[:space:]]Style[[:space:]]Adapters[[:space:]]for[[:space:]]Unsupervised[[:space:]]Stylistic[[:space:]]Headline[[:space:]]Generation/06bfaf10-740c-4fa6-bba5-20457cd25380_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Robustness of Named-Entity Replacements for In-Context Learning/92b3e48a-ba9b-4031-b53b-6f1de232b262_content_list.json

@@ -0,0 +1,2098 @@
+[
+    {
+        "type": "text",
+        "text": "Robustness of Named Entity Replacements for In-Context Learning",
+        "text_level": 1,
+        "bbox": [
+            147,
+            79,
+            850,
+            101
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Saeed Goodarzi† Nikhil Kagita† Dennis Minn† Shufan Wang†",
+        "bbox": [
+            206,
+            105,
+            794,
+            124
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Roberto Dessi $^{\\infty, \\pi}$  Shubham Toshniwal $^{\\nabla}$  Adina Williams $^{\\infty}$",
+        "bbox": [
+            228,
+            131,
+            766,
+            149
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jack Lanchantin\\*,Koustuv Sinha\\*,",
+        "bbox": [
+            321,
+            156,
+            673,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "†University of Massachusetts Amherst; †Universitat Pompeu Fabra  \n∇NVIDIA; †FAIR, Meta",
+        "bbox": [
+            228,
+            181,
+            771,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "sgoodarzitae@umass.edu, koustuvs@meta.com",
+        "bbox": [
+            258,
+            217,
+            742,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            341,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A key feature of modern large language models (LLMs) is their ability to perform in-context learning, a prompting technique where query-answer demonstrations are shown before the final query. This allows for generalization to novel distributions at inference time where the LLM can learn new rules without parameter updates. However, the choice of demonstrations and their relationship to a particular query can have a profound impact on model accuracy, raising concerns about the true in-context generalization capabilities (Zhao et al., 2021). In this work, we explore the robustness of the in-context learning paradigm by focusing on entities. In particular, we seek to understand the robustness of LLM in-context learning with respect to named entity replacements. We discover a significant variance in downstream performance based on the choice of the named entities, across three popular reasoning tasks and two popular LLMs. Specifically, model accuracy on the test sets can fluctuate between -2.7 to +8.0 points depending on the choice of named entity replacements. Our analysis exposes the sensitivity of LLM in-context learning with respect to named entities, and offers a simple recipe to improve test performance by hyper-parameter tuning the named entities for a given dataset. Code and datasets for reproducing the results are publicly available. $^{1}$",
+        "bbox": [
+            141,
+            280,
+            460,
+            708
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            732,
+            258,
+            747
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Modern large language models (LLMs) routinely demonstrate in-context learning, where the model can solve a wide range of tasks from a few demonstration examples (or exemplars) (Brown et al., 2020). In-context learning, or few-shot prompting, is a powerful learning paradigm as it allows for fast adaptation to new tasks without the need to train",
+        "bbox": [
+            112,
+            758,
+            489,
+            870
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "https://github.com/DennisMinn/ entities-in-context",
+        "bbox": [
+            112,
+            879,
+            416,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "*Equal Leadership.",
+        "bbox": [
+            136,
+            904,
+            255,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Original Demonstration",
+        "text_level": 1,
+        "bbox": [
+            638,
+            254,
+            747,
+            263
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "[...] context: Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops.",
+        "bbox": [
+            519,
+            263,
+            855,
+            288
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "question: How many lollipops did Jason give to Denny?  \nanswer: Jason started with 20 lollipops. Then he had 12 after giving some to Denny. So he gave Denny 20 - 12 = 8.",
+        "bbox": [
+            519,
+            288,
+            873,
+            315
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "[...] context: A bakery produces 60 loaves of bread each day. Two-thirds of the loaves are sold in the morning and half of what is left is sold equally in the afternoon and evening. question: How many loaves of bread are sold in the afternoon? answer:  $60 \\times 2 / 3 = 40$  loaves of bread are sold in the morning.  $60 - 40 = 20$  loaves of bread are left.  $20 \\times 2 = 40$  loaves of bread are sold in the afternoon. ###",
+        "bbox": [
+            519,
+            316,
+            868,
+            363
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Named Entity Replaced Demonstration",
+        "text_level": 1,
+        "bbox": [
+            606,
+            373,
+            781,
+            382
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "[...] context: Douglas had 20 lollipops. He gave Denny some lollipops. Now Douglas has 12 lollipops.",
+        "bbox": [
+            519,
+            382,
+            873,
+            407
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "question: How many lollipops did Douglas give to Denny?  \nanswer: Douglas started with 20 lollipops. Then he had 12 after giving some to Denny.  \nSo he gave Denny 20 - 12 = 8.",
+        "bbox": [
+            519,
+            407,
+            868,
+            434
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "[...] context: A bakery produces 60 loaves of bread each day. Two-thirds of the loaves are sold in the morning and half of what is left is sold equally in the afternoon and evening. question: How many loaves of bread are sold in the afternoon? answer:  $60 \\times 2 / 3 = 40$  loaves of bread are sold in the morning.  $60 - 40 = 20$  loaves of bread are left.  $20 \\times 1 / 2 = 10$  loaves of bread are sold in the afternoon. ### 10",
+        "bbox": [
+            519,
+            434,
+            868,
+            483
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: Named entity replacements for In-Context Learning. Example from FLAN-T5 on GSM8k (Cobbe et al., 2021). Here we replace all male entities  $(e^{o})$  in demonstrations with Douglas, and observe a change in the query answer prediction (from the incorrect answer 40 to the correct answer 10). Only 1 out of  $k = 5$  demonstrations  $(D)$  is depicted to conserve space, with the remaining demonstrations denoted by “[...]”.",
+        "bbox": [
+            507,
+            497,
+            882,
+            613
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "model parameters. However, model performance in the in-context setting can vary significantly with the choice of demonstrations. Previous work has uncovered performance instability when altering the demonstrations, particularly when the demonstrations change semantically (Schick and Schütze, 2020; Zhao et al., 2021; Perez et al., 2021; Gonen et al., 2022a). Naturally, a question arises whether LLMs are robust to semantically preserving, minimal alterations to the demonstrations.",
+        "bbox": [
+            507,
+            640,
+            884,
+            800
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we seek to address this question using simple, semantically preserving alterations: named entity replacements. Because changing named entities should not, in principle, change the correct response to the query, the model is expected to remain invariant to such modifications. For example, given the demonstration \"Alice has",
+        "bbox": [
+            507,
+            806,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10914",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10914-10931",
+        "bbox": [
+            208,
+            945,
+            786,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            277,
+            959,
+            719,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the ball. Alice went to the kitchen. Where is the ball? kitchen.\", the model should be invariant to the choice of the named entity \"Alice\"-i.e., a model should behave the same if we swap all instances of \"Alice\" with another named entity, e.g. \"Eve\". However, we observe that the model behavior is not always robust to such alterations (Figure 1)—prompting us to critically evaluate named entity invariance within in-context learning.",
+        "bbox": [
+            112,
+            84,
+            489,
+            228
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments to investigate the robustness of LLM in-context learning with respect to named entity replacements in three popular named entity heavy reasoning tasks, namely, bAbI, Weston et al. 2015; CLUTRR, Sinha et al. 2019; GSM8k, Cobbe et al. 2021, and two popular models, Flant5-XL (Chung et al., 2022a) and GPT-J (Wang and Komatsuzaki, 2021). Our key findings are:",
+        "bbox": [
+            112,
+            230,
+            489,
+            357
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Neither Flan-T5-XL nor GPT-J are robust to entity replacements for in-context learning. On bAbI tasks, the best replacement entity yields a +8.0 point performance increase (for Task 16), and the worst replacement entity yields a -2.7 point performance drop (Task 12). Similarly, for CLUTRR, we observe a 1.6 point improvement. For GSM8k, we only find <1 point increase/decrease with entity replacements.",
+            "- Performance gains from best performing entities are not random. To verify, we ran a bootstrap analysis and observe consistent gains with low variance.",
+            "- Best/worst performing entities are task-specific.",
+            "- We also analyze frequency, token length and perplexity of replaced entities, and find none of these alone explain our results, introducing an open problem for future research."
+        ],
+        "bbox": [
+            122,
+            362,
+            489,
+            668
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            682,
+            321,
+            700
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The goal of this study is to investigate the robustness of LLMs' in-context learning performance by changing entities in the demonstrations and queries. We consider a fixed set of prompt demonstrations aside from the named entities. By replacing the named entities and observing any accuracy changes, we can better understand the effect of certain entities on a model's accuracy. While many prior works have studied the robustness of LLMs with respect to perturbations (Ribeiro et al., 2020; Jin et al., 2020; Balasubramanian et al., 2020), we focus our analysis on in-context learning. A more in-depth analysis of related work is provided in",
+        "bbox": [
+            112,
+            709,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Appendix subsection A.1.",
+        "bbox": [
+            509,
+            84,
+            705,
+            99
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 In-Context Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            112,
+            719,
+            128
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "With in-context learning, a test sample is comprised of a sequence of  $k$  demonstrations  $D = d_{1},d_{2},\\ldots ,d_{k}$ . Each  $d_{i}$  consists of a context, question and answer string. The final query  $q$ , consisting only of a context and question string, follows  $D$ . The goal of in-context learning is to predict the answer to  $q$  given a small set of demonstrations at inference time. In other words, learning is done by eliciting activations and not changing model parameters.",
+        "bbox": [
+            507,
+            133,
+            884,
+            294
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Modifying In-Context Learning with Named Entity Replacement",
+        "text_level": 1,
+        "bbox": [
+            507,
+            306,
+            843,
+            337
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We seek to measure the change in accuracy when we replace named entities in the demonstrations  $(D)$  and/or query  $(q)$ . We perform entity replacement in the following way.",
+        "bbox": [
+            507,
+            343,
+            882,
+            406
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We first extract the named entities from each demonstration and query using an off-the-shelf NER model, and flag the subject entity,  $e^o$ , for replacement. We then select a replacement entity  $e^r$  from a large set of entities from Tzioumis (2018), which is composed of 4,250 unique first names. Finally, we perform named entity replacement in one of the following three settings:  $(D)$  replace all  $e^o$  entities in the demonstrations,  $e_{d_1}^o, e_{d_2}^o, \\dots, e_{d_k}^o$ ,  $(q)$  replace all  $e^o$  entities in the query,  $e_q^o$ , or  $(D$  and  $q)$  replace all  $e^o$  entities in both the query and demonstrations,  $e_{d_1}^o, e_{d_2}^o, \\dots, e_{d_k}^o, e_q^o$ . Note that each demonstration  $d_i$  can contain multiple instances of the entity  $e^o$ . These settings allow us to isolate and evaluate the effect of named entities in different parts of an in-context learning sample on accuracy.",
+        "bbox": [
+            507,
+            407,
+            884,
+            665
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We run multiple experiments using bootstrap sampling (with replacement) on three sets of names, half labeled as 'male' and half as 'female'. For each name  $e^r$  in a particular bootstrap set, we replace the original entities  $e^o$  with  $e^r$  in the test samples according to one of the three previously mentioned settings. We then evaluate the performance on the entity-replaced dataset. This procedure is repeated for each bootstrap set.",
+        "bbox": [
+            507,
+            665,
+            882,
+            810
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://huggingface.co/dslim/bert-base-NER, see Appendix Section A.3 for details on the selection of  $e^{o}$ .",
+        "bbox": [
+            507,
+            819,
+            882,
+            854
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "3 Appendix Section A.4 describes each setting in detail.",
+        "bbox": [
+            532,
+            856,
+            867,
+            869
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "4While our data source provided a binary gender annotation on names based on self-report, we note that this is a proxy and moreover, that using names as a proxy for gender is indicative but can be noisy.",
+        "bbox": [
+            507,
+            869,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10915",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/17fab8028c5b111fa1080156a13637f492bdd2e2c8845e213fbafebb9fc96144.jpg",
+        "image_caption": [
+            "Figure 2: Bootstrap results of Flan-T5-XL on validation sets for 3 distinct sets of 200 names. Error bars represent the variation across bootstrap sets of names."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            87,
+            480,
+            422
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Baseline results are reported using the original prompt demonstrations and query. However, our replacement strategy introduces a frequency bias, as  $e^r$  is used to replace multiple demonstrations in a given input, resulting in artificially increasing the usage of a given entity in an input. Thus, we also report the median accuracy of replacement, which serves as a proxy baseline for replacing all demonstrations and/or query with a random entity.",
+        "bbox": [
+            112,
+            508,
+            489,
+            653
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Models and Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            665,
+            324,
+            682
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our entity replacement strategy on both an encoder-decoder LLM (Flan-T5-XL, 3B; Chung et al. 2022b) as well as a decoder-only LLM (GPT-J, 6B; Wang and Komatsuzaki 2021). We test these models across the following datasets: bAbI (Tasks 1, 12, 14, 16; Weston et al. 2015), CLUTRR (Sinha et al., 2019), GSM8k (Cobbe et al., 2021). We chose these datasets primarily because they include a significant presence of named entities.",
+        "bbox": [
+            112,
+            688,
+            487,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For each dataset, we select five samples from the training set to serve as demonstrations. To maintain grammatical correctness and clarity, we",
+        "bbox": [
+            112,
+            833,
+            487,
+            882
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "only do replacements that preserve gender when a pronoun is contained in the sentence (see Appendix A.5). This step specifically addresses any potential uncontrolled effects on the performance from entity replacement, such as perplexity spikes that may results from swaps where the genders of  $e^{o}$  and  $e^{r}$  differ, and the model is expected to accommodate unexpected pronominal coreference with  $e^{r}$ .",
+        "bbox": [
+            507,
+            84,
+            882,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            230,
+            608,
+            244
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Performance Across Bootstrap Sets. We find a noticeable change in accuracy with the best and worst performing entity compared to the baseline or median performing entity. Figure 2 shows Flant5-XL results on all validation datasets (analogous GPT-J results are in Figure 5 in the Appendix, and follow a similar pattern). This indicates that specific entities can noticeably alter in-context learning performance. Furthermore, the standard deviations for the median and best performing entity across bootstraps are low, suggesting that the selection of different name sets will likely yield similar median and best performance. However, the variation in the worst performing entities is large.",
+        "bbox": [
+            507,
+            260,
+            884,
+            485
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Performance Within a Single Bootstrap Set. To analyze the results in more detail, we measure the performance for Flan-T5-XL and GPT-J on a single bootstrap set of 200 examples. We show the results for bAbI task 14 in Figure 3. When we replace names only in the demonstrations  $(D)$ , the variance in model performance is lower compared to both when we replace names only in the query  $(q)$  and in both query and demonstration  $(D$  and  $q)$ . The  $(D$  and  $q)$  setup exhibits the highest variance, as this scenario involves the maximum change in the tokens of the original prompt. Nevertheless, we see a consistent variance from the baseline by simply replacing named entities in any of the three replacement settings. Other datasets exhibit a similar behavior; see Appendix Figures 6 and 7.",
+        "bbox": [
+            507,
+            492,
+            882,
+            750
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Best and Worst Entity Generalization. Thus far, we've used the valid set to demonstrate accuracy variation. To evaluate the best performing entities on the test set, we implement a majority voting method as follows. We first find the 10 best performing entities on the valid set. For each sample in the test set, we take the majority answer prediction from the 10 different best names, denoted \"Maj@10\". Table 1 shows results of entity replacements across the validation and test sets by replac",
+        "bbox": [
+            507,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>5</sup>Except GSM8k, where we select 5 of the 8 prompts from Wei et al. (2022).",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10916",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ing entities in the demonstration and query (both  $D$  and  $q$ ). We observe a maximum 8.0 point accuracy increase, seen on bAbI task 16. $^{6}$  The best and worst performing entities vary by dataset, but overall entities generalize across the validation and test sets, with a uniform increase in accuracy from the best entity. A list of these entities for each dataset and further evaluations are in Appendix Table 2.",
+        "bbox": [
+            112,
+            84,
+            487,
+            212
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For all datasets, there are entities with appreciable deviations from the baseline, and these entities generalize across the validation and test splits. This finding indicates that the observed accuracy shifts are not coincidental and highlights the influence of entity swaps, including those that preserve overall example semantics, such as for names, on model performance.",
+        "bbox": [
+            112,
+            212,
+            487,
+            342
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            355,
+            221,
+            370
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Why are the models not invariant to entity replacements? We conducted several analyses in attempts to find the root cause of this result, but they were largely inconclusive.",
+        "bbox": [
+            112,
+            381,
+            487,
+            445
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Does the frequency of names affect how much they shift model performance? The lack of robustness could well be explained by the frequency effects of named entities in the training corpus of the pre-trained model (Wei et al., 2021). We use the entity frequency data provided by Tzioumis (2018), and assessed performance changes across names with varying frequencies. We do not find any correlation between name frequency and the model performance post entity replacement (see Appendix Section B.3).",
+        "bbox": [
+            112,
+            451,
+            487,
+            629
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Does the token length of names affect performance? Another hypothesis of the invariance could be due to the inconsistent tokenization of named entities. Rare names can be tokenized into longer tokens based on the learned subwords of the model. We analyzed token length post tokenization and compared it with model performance, but we are unable to find any meaningful relationship between the two (see Appendix Section B.4).",
+        "bbox": [
+            112,
+            634,
+            487,
+            778
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Does improved performance result from a decrease in prompt perplexity due to entity replacement? Finally, the surprisal of the LLM to certain entities could be a promising hypothesis to explain the change in model performance (Gonen et al., 2022b). We compare prompt perplexity af",
+        "bbox": [
+            112,
+            785,
+            487,
+            883
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/806ba71760a2083e986f1d9e3c8486ceedd8450c702759db96c9852bcfc8b229.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            90,
+            858,
+            261
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/a884b3da2e0d17034a9e33ec7f497e383486e897e60a3367f94c029904ee1c03.jpg",
+        "image_caption": [
+            "bAbI Task 14 (Baseline Accuracy=70)",
+            "Figure 3: Accuracy for Flan-T5-XL and GPT-J-6B on bAbI task 14 for a single bootstrap set of 200 samples."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            277,
+            858,
+            437
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ter entity replacement to LLM performance, and yet again we are unable to find a statistical correlation (see Appendix Section B.5). While Gonen et al. (2022b) suggests improvement of model performance with decrease in perplexity, we observe certain entities displaying the opposite behavior.",
+        "bbox": [
+            507,
+            512,
+            882,
+            609
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            624,
+            640,
+            638
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this study, we find that LLMs exhibit a consistent lack of robustness with respect to named entity replacement for the in-context learning setting, fur",
+        "bbox": [
+            507,
+            650,
+            882,
+            699
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/99f3f7f108c2d1d531a75f9dc65d14aecbc1534f3cb3314b2d990ab3697ac9ae.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"2\">Baseline</td><td colspan=\"2\">Best Entity</td><td colspan=\"2\">Worst Entity</td></tr><tr><td>Valid</td><td>Test</td><td>Valid</td><td>Test (Maj@10)</td><td>Valid</td><td>Test</td></tr><tr><td>bAbI 12</td><td>95.1</td><td>95.5</td><td>98.5 [+3.4]</td><td>99.0 [+3.5]</td><td>92.4 [-2.7]</td><td>93.0 [-2.5]</td></tr><tr><td>bAbI 14</td><td>76.1</td><td>77.5</td><td>78.4 [+2.3]</td><td>79.5 [+2.0]</td><td>75.2 [-0.9]</td><td>76.6 [-0.9]</td></tr><tr><td>bAbI 16</td><td>53.5</td><td>53.3</td><td>61.7 [+8.2]</td><td>61.3 [+8.0]</td><td>55.5 [+2.0]</td><td>55.9 [+2.6]</td></tr><tr><td>GSM8k</td><td>14.3</td><td>10.5</td><td>14.6 [+0.3]</td><td>11.3 [+0.8]</td><td>12.7 [-1.6]</td><td>10.1 [-0.4]</td></tr><tr><td>CLUTRR</td><td>59.0</td><td>53.0</td><td>61.8 [+2.8]</td><td>54.6 [+1.6]</td><td>51.4 [-7.6]</td><td>53.1 [+0.1]</td></tr></table>",
+        "bbox": [
+            512,
+            721,
+            882,
+            804
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Performance of FLAN-T5-XL, with accuracy shift from baseline provided in [square brackets]. Entity replacements are done in the  $D$  and  $q$  setting. bAbI task 1 reaches perfect accuracy and is thus omitted from this table. Performance between the validation and test datasets is consistent, and we observe a generalizable increase in accuracy across all datasets.",
+        "bbox": [
+            507,
+            814,
+            882,
+            914
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>6</sup>We did not include bAbI Task 1 because the baseline accuracy is already  $100\\%$ , leaving no space for improvement.",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10917",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ther questioning the process of reasoning employed by these models (Min et al., 2022). We demonstrate this across three axes: the complexity of the task, the model architecture, and which named entities are to be replaced (in the demonstrations, query, or both). The cause for this effect remains elusive, which opens up exciting research directions. One intriguing result is that the entities that can boost performance are dataset specific. This points to potential avenues for enhancing LLM downstream performance by performing hyperparameter search over a range of entities.",
+        "bbox": [
+            112,
+            84,
+            492,
+            277
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            290,
+            250,
+            305
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We acknowledge that the selection of names in our study may not provide an unbiased representation of the complete set of possible names. The data provided by Tzioumis (2018) was derived from mortgage applications which may introduce a bias against popular names within marginalized communities. Moreover, this issue is compounded when we sub-sample the set of names due to computational limitations. Additionally, we used a NER model available through Hugging Face to identify  $e^o$ . It was BERT-based (Devlin et al., 2018) and trained on a CONLL-2003 corpus (Tjong Kim Sang and De Meulder, 2003) (Wolf et al., 2019), and we anecdotally observed that sometimes it struggles to identify rare names as named entities, for example, it incorrectly identified \"TreQuan\" as \"TreQ.\" We spot checked for this issue and removed names that posed problems to the best of our abilities.",
+        "bbox": [
+            112,
+            316,
+            489,
+            605
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "7 Ethics and Broader Impact",
+        "text_level": 1,
+        "bbox": [
+            112,
+            619,
+            384,
+            636
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this work, we explore the effect of name swaps in an in-context learning setup and find that particular names can increase accuracy. In our efforts to align names with their respective pronouns, we assumed binary gender, but acknowledge that other genders exist. We made this assumption because our evaluated datasets used binary pronouns, and our names dataset was also only binarily annotated. While our paper assumes name-gender associations, we do not want to discredit negative evidence. A name that is not explicitly associated with a gender doesn't imply evidence of absence of association (imagine that only one person named \"Alex\" was sampled, and she was a woman—that doesn't mean that there aren't also men or nonbinary people called \"Alex\"). At the moment, there aren't many annotated name datasets that consists",
+        "bbox": [
+            112,
+            645,
+            490,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "of gender-agnostic, gender-ambiguous, or gender-neutral names. Hopefully, additional work such as (Dev et al., 2021; Lauscher et al., 2022; Ovalle et al., 2023) will expand such resources, but regardless of these complications, our conclusions are still useful for trying to understand why some names can shift model performance under some settings.",
+        "bbox": [
+            507,
+            84,
+            885,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            509,
+            240,
+            608,
+            255
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "2023. Genderize.io. https://genderize.io/. Accessed: June 15, 2023.",
+            "Oshin Agarwal, Yinfei Yang, Byron C Wallace, and Ani Nenkova. 2020. Entity-switched datasets: An approach to auditing the in-domain robustness of named entity recognition models. arXiv preprint arXiv:2004.04123.",
+            "Sriram Balasubramanian, Naman Jain, Gaurav Jindal, Abhijeet Awasthi, and Sunita Sarawagi. 2020. What's in a name? are BERT named entity representations just as good for any other name? In Proceedings of the 5th Workshop on Representation Learning for NLP, pages 205-214, Online. Association for Computational Linguistics.",
+            "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.",
+            "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022a. Scaling instruction-finetuned language models. arXiv:2210.11416.",
+            "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022b. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416.",
+            "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.",
+            "Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff Phillips, and Kai-Wei Chang."
+        ],
+        "bbox": [
+            509,
+            263,
+            885,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10918",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "2021. Harms of gender exclusivity and challenges in non-binary representation in language technologies. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1968-1994, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805.",
+            "Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. 2017. Hotflip: White-box adversarial examples for text classification. arXiv preprint arXiv:1712.06751.",
+            "Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. 2018. Black-box generation of adversarial text sequences to evade deep learning classifiers. In 2018 IEEE Security and Privacy Workshops (SPW), pages 50-56. IEEE.",
+            "Hila Gonen, Srini Iyer, Terra Blevins, Noah A Smith, and Luke Zettlemoyer. 2022a. Demystifying prompts in language models via perplexity estimation. arXiv preprint arXiv:2212.04037.",
+            "Hila Gonen, Shauli Ravfogel, and Yoav Goldberg. 2022b. Analyzing gender representation in multilingual models. In Proceedings of the 7th Workshop on Representation Learning for NLP, pages 67-77, Dublin, Ireland. Association for Computational Linguistics.",
+            "Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8018-8025.",
+            "Anne Lauscher, Archie Crowley, and Dirk Hovy. 2022. Welcome to the modern world of pronouns: Identity-inclusive natural language processing beyond gender. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1221-1232, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.",
+            "Bill Yuchen Lin, Wenyang Gao, Jun Yan, Ryan Moreno, and Xiang Ren. 2021. RockNER: A simple method to create adversarial examples for evaluating the robustness of named entity recognition models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3728-3737, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2021. What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Robert L Logan IV, Ivana Balažević, Eric Wallace, Fabio Petroni, Sameer Singh, and Sebastian Riedel. 2021. Cutting down on prompts and parameters: Simple few-shot learning with language models. arXiv preprint arXiv:2106.13353.",
+            "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837.",
+            "Anaelia Ovalle, Palash Goyal, Jwala Dhamala, Zachary Jaggers, Kai-Wei Chang, Aram Galstyan, Richard Zemel, and Rahul Gupta. 2023. \"i'm fully who i am\": Towards centering transgender and non-binary voices to measure biases in open language generation. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1246-1266.",
+            "Nicolas Papernot, Patrick McDaniel, Ananthram Swami, and Richard Harang. 2016. Crafting adversarial input sequences for recurrent neural networks. In MILCOM 2016-2016 IEEE Military Communications Conference, pages 49-54. IEEE.",
+            "Ethan Perez, Douwe Kiela, and Kyunghyun Cho. 2021. True few-shot learning with language models. Advances in neural information processing systems, 34:11054-11070.",
+            "Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902-4912, Online. Association for Computational Linguistics.",
+            "Timo Schick and Hinrich Schütze. 2020. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676.",
+            "Timo Schick and Hinrich Schütze. 2022. True few-shot learning with prompts—a real-world perspective. Transactions of the Association for Computational Linguistics, 10:716-731.",
+            "Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. 2020. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980.",
+            "Koustuv Sinha, Shagun Sodhani, Jin Dong, Joelle Pineau, and William L Hamilton. 2019. Clutrr: A diagnostic benchmark for inductive reasoning from text. arXiv preprint arXiv:1908.06177.",
+            "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            916
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10919",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Miles Turpin, Julian Michael, Ethan Perez, and Samuel R Bowman. 2023. Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting. arXiv preprint arXiv:2305.04388.",
+            "Konstantinos Tzioumis. 2018. Demographic aspects of first names. Scientific data, 5(1):1-9.",
+            "Ben Wang and Aran Komatsuzaki. 2021. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model. https://github.com/kingoflolz/mesh-transformer-jax.",
+            "Xiao Wang, Qin Liu, Tao Gui, Qi Zhang, Yicheng Zou, Xin Zhou, Jiacheng Ye, Yongxin Zhang, Rui Zheng, Zexiong Pang, Qinzhuo Wu, Zhengyan Li, Chong Zhang, Ruotian Ma, Zichu Fei, Ruijian Cai, Jun Zhao, Xingwu Hu, Zhiheng Yan, Yiding Tan, Yuan Hu, Qiyuan Bian, Zhihua Liu, Shan Qin, Bolin Zhu, Xiaoyu Xing, Jinlan Fu, Yue Zhang, Minlong Peng, Xiaqing Zheng, Yaqian Zhou, Zhongyu Wei, Xipeng Qiu, and Xuanjing Huang. 2021. TextFlint: Unified multilingual robustness evaluation toolkit for natural language processing. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations, pages 347-355, Online. Association for Computational Linguistics.",
+            "Jason Wei, Dan Garrette, Tal Linzen, and Ellie Pavlick. 2021. Frequency effects on syntactic rule learning in transformers.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903.",
+            "Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart Van Merrienboer, Armand Joulin, and Tomas Mikolov. 2015. Towards ai-complete question answering: A set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698.",
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.",
+            "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            822
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10920",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            238,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A.1 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            109,
+            273,
+            124
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Several works investigated the robustness of language models with respect to input rewrites (Papernot et al., 2016; Ebrahimi et al., 2017; Gao et al., 2018; Jin et al., 2020). In particular, Balasubramanian et al. (2020) found that while fine-tuned BERT models were brittle in certain tasks, they demonstrated robustness against entity replacement for question answering. Other works have shown that many transformer-based models are insufficiently robust to name swaps (Ribeiro et al. 2020; Agarwal et al. 2020; Lin et al. 2021; Wang et al. 2021, i.a.), Our work differs from these in that we focus on the in-context learning setting.",
+        "bbox": [
+            112,
+            130,
+            489,
+            338
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The ability to do in-context learning first emerged from Brown et al. (2020). Subsequently, many works sought to improve and understand the abilities of LLMs with respect to in-context learning (Liu et al., 2021; Zhao et al., 2021). Perez et al. (2021); Logan IV et al. (2021); Schick and Schütze (2022); Turpin et al. (2023) show an improvement over the original paper in the the \"true few-shot learning\" setup where there is no access to a large test set to tune in-context learning hyperparameters. Despite the impressive results of this new paradigm, little has been understood regarding the generalization of certain in-context demonstrations.",
+        "bbox": [
+            115,
+            340,
+            489,
+            546
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Gonen et al. (2022a) bring to attention the importance of the specific prompts in model performance. They show that the lower the perplexity of the prompt is, the better the prompt is able to perform the task. Min et al. (2022) show that the actual labels in the prompt are not very important. However, the labels still have a small impact in performance.",
+        "bbox": [
+            112,
+            549,
+            489,
+            676
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In recent years, several studies have explored the significance of prompt quality in few-shot learning, particularly in the context of question-answering tasks. One such study was conducted by Shin et al. (2020) who proposed AutoPrompt, an automated approach for prompt generation in few-shot learning scenarios that has been shown to achieve the highest accuracy. The AutoPrompt model leverages gradient-based prompt search techniques to find the optimal prompt for a given task, thereby reducing the need for human intervention. Additionally, AutoPrompt utilizes Masked Language Models (MLMs) to perform sentiment analysis and natural language inference tasks without the need for fine-tuning, highlighting the potential of MLMs",
+        "bbox": [
+            112,
+            678,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "in few-shot learning scenarios. In another study, Gonen et al. (2022a) introduced a novel technique for prompt generation that involves using a small set of manually created prompts and then applying paraphrasing and backtranslation to generate additional prompts automatically. Specifically, they utilized GPT-3 to perform the paraphrasing step, which involves rephrasing the original prompts while preserving their underlying meaning. The backtranslation step then involves translating the paraphrased prompts into another language and then back into the original language to generate new prompts that are semantically similar to the original prompts but with different phrasing and wording. This approach enables the generation of a large number of diverse prompts without requiring extensive manual effort, which can be particularly useful in few-shot learning scenarios where data is limited. Gonen et al. (2022a) also demonstrated an inverse relationship between the perplexity of the prompts and the model's performance across various tasks.",
+        "bbox": [
+            507,
+            84,
+            884,
+            437
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our work is related to previous studies as we focus on the true few-shot learning setup for the task of question answering and explore the robustness of NLP models against entity replacement. There are three main ways in which our work differs from previous studies. First, our investigation focuses on analyzing the robustness of language models against entity replacement specifically for the task of question answering in the ICL setup, which has not been explored in previous studies. Second, we utilize left-to-right language models, which are more widely used and powerful than masked language models. Finally, our approach involves prompt-learning instead of fine-tuning, whereby the parameters of the language model are modified to improve its performance on specific tasks. By leveraging the true few-shot learning setup and prompt-learning techniques, we aim to develop a more robust and effective approach for few-shot question-answering tasks that can handle entity replacement and other challenges more effectively.",
+        "bbox": [
+            507,
+            441,
+            884,
+            778
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A.2 Dataset Details",
+        "text_level": 1,
+        "bbox": [
+            509,
+            797,
+            680,
+            810
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "bAbI (Weston et al., 2015): The bAbI dataset consists of 20 tasks that evaluate various aspects of language understanding by NLP models. In our experiments, we select four tasks representing varying degrees of difficulty from the bAbI dataset: task 1, task 12, task 14, and task 16.",
+        "bbox": [
+            507,
+            822,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10921",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [],
+        "code_body": "context: Mary and  $[e_{d1}]$  travelled to the   \nbathroom. John and  $[e_{d1}]$  travelled to the office.   \nquestion: Where is  $[e_{d1}]$ ?   \nanswer: office   \n????   \ncontext:  $[e_{d2}]$  and Mary went back to the office.   \nDaniel and Sandra went to the bedroom.   \nquestion: Where is  $[e_{d2}]$ ?   \nanswer: bedroom   \n????   \ncontext: Daniel and John moved to the hallway.   \n $[e_q]$  and Daniel journeyed to the bathroom.   \nquestion: Where is  $[e_q]$ ?   \nanswer:",
+        "bbox": [
+            144,
+            87,
+            460,
+            281
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: Named entity replacement overview for in context learning. Here we show an in-context learning example with  $k = 2$  demonstrations before the final query. We consider 3 scenarios: replacing only the demonstration entities  $(e_{d1}, e_{d2})$ , only the query entity  $(e_q)$ , or both  $(e_{d1}, e_{d2}, e_q)$ . All replacements are done with the same replacement entity  $e^r$ .",
+        "bbox": [
+            112,
+            296,
+            489,
+            397
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "CLUTRR (Sinha et al., 2019): This dataset is similar to a subset of the bAbI dataset, but contains more difficult multi-step questions. This allows us to determine whether named entity replacement robustness is affected by task difficulty. Given a multi-sentence context describing kinship relations between certain entities and a question, the task is to predict the kinship relation between the entities that can be inferred from the context but is not made explicit. We convert this into a multiple-choice question where we expect the model to pick the answer from the four choices provided.",
+        "bbox": [
+            112,
+            426,
+            487,
+            620
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "GSM8k (Cobbe et al., 2021): This dataset is a collection of elementary multi-step mathematical reasoning questions. We specifically selected this dataset due to the challenging nature of arithmetic operations, where existing models often struggle to achieve accurate results. We subsampled GSM8k to include only examples which contained identifiable named entities.",
+        "bbox": [
+            112,
+            623,
+            489,
+            751
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The Tzioumis (2018) dataset, contains of 4,250 unique first names, associated with their unigram frequency, and self-reported demographic information.",
+        "bbox": [
+            112,
+            755,
+            489,
+            818
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "For each sample in the dataset, which is comprised of a context, question, and answer triplet, we transform it into a formatted string. This transformation involves concatenating the field name, a colon, space, the respective value corresponding to that field, and newline character. We ap",
+        "bbox": [
+            112,
+            822,
+            490,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "ply this transformation to  $k$  demonstrations:  $D = d_{1},d_{2},\\ldots ,d_{k}$  and the final query  $q$  , where the answer to  $q$  is omitted. We concatenate the formatted  $k$  demonstrations and query  $q$  , delimiting each string with a newline character followed by  $\"\\# \\# \\# \"$  and a newline character. Figure 4 shows an incontext learning example.",
+        "bbox": [
+            507,
+            84,
+            885,
+            197
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "A.3 Entity Selection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            210,
+            685,
+            225
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In general, we determine  $e^o$  by selecting the first encountered entity within the question, as it often represents the subject of the question. If the question contains interrogative pronouns, such as \"who,\" no entities will be present within the question. In this case, we consider  $e^o$  as the first entity encountered within the answer since it typically represents the subject being referred to. If the subject of the question and answer is referred to using pronouns then neither the question nor answer will contain any entities, we assume that noun ambiguity is not an issue and select the first encountered entity within the context as  $e^o$ .",
+        "bbox": [
+            507,
+            233,
+            885,
+            441
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "A.4 Entity-replacement Scenarios",
+        "text_level": 1,
+        "bbox": [
+            507,
+            456,
+            793,
+            472
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In the demonstration-only replacement scenario, we perform entity replacement exclusively on the demonstrations  $e_d$  while keeping the query unmodified 4. By selectively replacing entities within the context, question, and answer of the demonstrations, we aim to examine the extent to which this replacement strategy influences the model's performance.",
+        "bbox": [
+            507,
+            479,
+            882,
+            607
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In the query-only replacement setting, we solely augment the entities within the query  $e_{q}$  while leaving the demonstrations unaltered Fig 4. This allows us to investigate the impact of enhancing the query's entity representation on the model's ability to generate accurate responses.",
+        "bbox": [
+            507,
+            608,
+            882,
+            703
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Applying both demonstration and query replacement, we replace  $e_d$  and  $e_q$  in both the demonstrations and query simultaneously. By enriching the entity information in both the context and the question, we explore the combined effect of entity replacement on the overall performance of the model.",
+        "bbox": [
+            507,
+            706,
+            882,
+            816
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "A.5 Ensuring Grammatical Correctness",
+        "text_level": 1,
+        "bbox": [
+            507,
+            832,
+            840,
+            847
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "To begin, we verify the pronouns associated with both  $e^o$  and  $e^r$  using (gen, 2023). If the gender associated with  $e^o$  corresponds with that of  $e^r$ , we proceed to substitute  $e^o$  with  $e^r$ . However, if there",
+        "bbox": [
+            507,
+            854,
+            880,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10922",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "is no alignment in gender, we refrain from altering  $e^o$  and maintain its original form.",
+        "bbox": [
+            112,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B Analysis",
+        "text_level": 1,
+        "bbox": [
+            114,
+            130,
+            226,
+            145
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B.1 Analyzing Best/Worst performing entities",
+        "text_level": 1,
+        "bbox": [
+            112,
+            156,
+            485,
+            172
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "In order to examine the generalizability of the best and worst performing entities in each dataset, we extracted named entities from the validation dataset and tested them on the test dataset. More specifically, we replaced the named entities in the prompts of the validation dataset with a set of names that included 100 male and 100 female names. We then calculated the accuracy for each name, selecting the 10 names that exhibited the highest performance and the 10 names that showed the lowest performance. Next, we replaced the named entities in the test dataset prompts with these 20 selected names and calculated the corresponding accuracy for the test dataset. Table 2 presents the best and worst performing entities for each dataset along with the absolute percentage of accuracy changes. The ensemble model functions by replacing the named entity in the prompt with the highest performing entity and then generating an answer at each step. After cycling through the 10 best performing entities, it selects the answer with the most votes as the final response. This approach is referred to as majority@10. It's important to note that the bAbI task 1 doesn't have a 'best performing entity' because the baseline accuracy is already  $100\\%$ , leaving no room for improvement.",
+        "bbox": [
+            112,
+            178,
+            489,
+            595
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Figure 8 provides an example of an generated answer for the GSM8k test dataset. The original prompt failed to produce the correct answer, while the transformed prompt was able to successfully answer the question. The absolute improvement of  $0.8\\%$  in the accuracy of GSM8k through the majority@10 is remarkable, given the following facts: 1) The baseline accuracy for this challenging dataset is  $10.5\\%$ , making even a small improvement significant. 2) The number of named entities in GSM8k is fewer than in other datasets, meaning only a few words are replaced in the transformed prompt, as shown in Figure 8. The ratio of replaced words to the total number of words in the prompt is 0.032. 3) Of the 1319 examples in the GSM8k test dataset, 663 contain a male name, 377 contain a female name, and 279 don't contain any named entity in the query. We could have achieved greater changes in percentage if we had only considered examples where the gender of the best perform",
+        "bbox": [
+            112,
+            596,
+            489,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "ing entity matched the gender of the name in the query (as replacements wouldn't occur otherwise). However, we chose not to adjust the test dataset to ensure that our accuracy improvement could be compared with other studies.",
+        "bbox": [
+            507,
+            84,
+            882,
+            164
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "The consistency in performance between the validation and test datasets in CLUTRR is lower than that in the other datasets. This could be attributed to the structure of CLUTRR, where four options are provided in the prompt, from which the model chooses an answer. This setup introduces a degree of randomness to the generated answer, potentially leading to less consistency between the performance of the validation and test datasets.",
+        "bbox": [
+            507,
+            165,
+            884,
+            309
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B.2 LLM Architecture",
+        "text_level": 1,
+        "bbox": [
+            507,
+            319,
+            705,
+            334
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "The architecture of LLMs plays a crucial role in determining their behavior and performance, including their vulnerability to entity replacement. The design choices and architectural components of LLMs can influence how they process and generate text, affecting their sensitivity to changes in named input entities. By comparing the Flant5-XL, which employs an encoder-decoder architecture, with the GPT-J-6B, which represents a decoder-only architecture, we aimed to investigate the impact of LLM architecture on their brittleness towards entity replacement. The high variance in the performance of GPT-J (as shown in Fig 7) underscores that the brittleness towards entity replacement is not exclusive to the encoder-decoder architecture but is also a feature of decoder-only architecture models.",
+        "bbox": [
+            507,
+            340,
+            884,
+            613
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B.3 Effect of entity frequency",
+        "text_level": 1,
+        "bbox": [
+            507,
+            625,
+            756,
+            640
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Frequently recurring named entities in the web, which subsequently appear in the training data, can impact both the model's ppl and performance. Entities appearing frequently in the training data are more likely to be well-represented in the model's understanding, contributing to lower ppl values. This occurs because the frequent presence of these entities reduces the unpredictability of the next word in a sequence, making it easier for the LLM to anticipate. The hypothesis is that names appearing more frequently would contribute to better model performance. To examine this hypothesis, the frequency of names, extracted from (Tzioumis, 2018), were categorized into five groups of equal numbers, labeled G1 to G5, ranging from the most to the least frequent names. Fig 9 presents the performance of Flan-T5-XL across different datasets.",
+        "bbox": [
+            507,
+            646,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10923",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "image",
+        "img_path": "images/d2fa12070fd32ac897205bd220ddb0c766c2a9efe46e048bc8ea4736ad8eec2d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            192,
+            490,
+            473
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/074580c419c423622844c55c98b779242616403502e4dfd01ca02db84eff0410.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            505,
+            192,
+            870,
+            473
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/149a1c5187a9b12426b10c9b18c1810803a37bb3390a2b5e830472973d64304c.jpg",
+        "image_caption": [
+            "Figure 5: Bootstrap results of GPT-J-6B for 3 distinct sets of names, each comprising 200 names. Error bars represent the variation across different name sets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            482,
+            490,
+            762
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/b595f85513bb0796b82cd377b64cef7f712d753e68c94765d5913129c854c29d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            507,
+            482,
+            868,
+            762
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "10924",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/84c5071efe92c658740b73b43fb7ab9a82d2445158c42c57023db64eadb65699.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            219,
+            489,
+            388
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/9af7ee133001d94dd6145685fdd994c10f2f19ca3508f836765a296b8d4237c7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            219,
+            870,
+            388
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/a3d0c964c46fa72e5dcfc45acd7e059152ae75558ae08d160fe211bd7efe33e5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            395,
+            489,
+            561
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/a8be2c992d5164ed570c4c8fb0a62549d1c9061d5f2323a0edcd583e7e83e934.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            395,
+            870,
+            561
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/fa2270a4e0ab7c99a1855f2a278ec33e6211bd361f83ee9d0cafdcc0942deaa0.jpg",
+        "image_caption": [
+            "Figure 6: Deviation in accuracy from the baseline (without data augmentation) for Flan-T5-XL across different datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            568,
+            489,
+            736
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/6adbcbf66a45bfc8c4b68034db00705847cb60e07ccc10e61116be7cf63afd4a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            568,
+            870,
+            736
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "10925",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/fc6aa65b5f61908d8679283485b4f204d69bc634e25afac91d976a623963c33e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            137,
+            104,
+            494,
+            277
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/ddbd6b645183aa081d4e287cc0f5f48cce9b986b4fe698cc6812c380a7380779.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            103,
+            873,
+            277
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/dcec4185a5fef4c3dd60e4caa8629a2f2c8aead8bf54a79c699154c42db4b6dc.jpg",
+        "image_caption": [
+            "Figure 7: Deviation in accuracy of GPT-J from the baseline across various bAbI tasks. Like the encoder-decoder architecture, the decoder-only architecture also exhibits brittleness towards entity replacement."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            282,
+            492,
+            454
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd0e5c65da0264ed6197fea3b66ecd58100960523f9ca9ebb2bff5d611931eb4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            282,
+            873,
+            454
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/7125dc817e6dab1e6722d82340e62eee3454ac6ec98b07e07996c0c5ee4e27ef.jpg",
+        "image_caption": [
+            "Figure 8: An example of a generated answer from the GSM8k test dataset. In this instance, the original prompt fails to generate the correct answer, whereas the transformed prompt — with the best-performing entity replaced — successfully produces the correct response. This generated text is reproducible using a greedy decoding."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            546,
+            882,
+            840
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "10926",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/f221659095d2f3d58305b8b5881a62c7f2a1dda85f5423d06a99a52b9d2c80ce.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td rowspan=\"2\">Index</td><td colspan=\"3\">Best performing entity</td><td colspan=\"3\">Worst performing entity</td></tr><tr><td>entity</td><td>Acc change (valid)</td><td>Acc change (test)</td><td>entity</td><td>Acc change (valid)</td><td>Acc change (test)</td></tr><tr><td rowspan=\"10\">bAbI task 12 valid baseline Acc = 95.1 test baseline Acc = 95.5</td><td>1</td><td>Penelope</td><td>3.6</td><td>3.0</td><td>Mihai</td><td>-4.0</td><td>-2.3</td></tr><tr><td>2</td><td>Loreen</td><td>3.6</td><td>3.0</td><td>Raisa</td><td>-4.0</td><td>-3.6</td></tr><tr><td>3</td><td>Dorcas</td><td>3.6</td><td>3.2</td><td>Ethel</td><td>-3.7</td><td>-3.5</td></tr><tr><td>4</td><td>Toby</td><td>3.4</td><td>3.4</td><td>Migdalia</td><td>-3.0</td><td>-4.0</td></tr><tr><td>5</td><td>Flora</td><td>3.3</td><td>2.9</td><td>Azucena</td><td>-2.6</td><td>-3.2</td></tr><tr><td>6</td><td>Nataliya</td><td>3.3</td><td>3.4</td><td>Bhavana</td><td>-2.6</td><td>-3.6</td></tr><tr><td>7</td><td>Nathaniel</td><td>3.3</td><td>3.4</td><td>Surinder</td><td>-2.1</td><td>-0.9</td></tr><tr><td>8</td><td>Lucretia</td><td>3.3</td><td>3.3</td><td>Sanjeev</td><td>-1.9</td><td>-1.7</td></tr><tr><td>9</td><td>Keiko</td><td>3.2</td><td>3.0</td><td>Armida</td><td>-1.8</td><td>-1.1</td></tr><tr><td>10</td><td>Bernardo</td><td>3.2</td><td>3.1</td><td>Jianwei</td><td>-1.7</td><td>-1.4</td></tr><tr><td colspan=\"4\">Majority@10</td><td>3.5</td><td></td><td></td><td></td></tr><tr><td rowspan=\"10\">bAbI task 14 valid baseline Acc = 76.1 test baseline Acc = 77.5</td><td>1</td><td>Erich</td><td>2.9</td><td>2.0</td><td>Frederick</td><td>-1.4</td><td>-1.2</td></tr><tr><td>2</td><td>Jacqueline</td><td>2.6</td><td>1.8</td><td>Raisa</td><td>-1.3</td><td>-1.3</td></tr><tr><td>3</td><td>Zoltan</td><td>2.4</td><td>2.3</td><td>Maryellen</td><td>-0.5</td><td>-0.6</td></tr><tr><td>4</td><td>Chryystal</td><td>2.3</td><td>1.5</td><td>Meera</td><td>-0.5</td><td>-0.5</td></tr><tr><td>5</td><td>Petra</td><td>2.2</td><td>1.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>6</td><td>Pravin</td><td>2.2</td><td>1.8</td><td>—</td><td>—</td><td>—</td></tr><tr><td>7</td><td>Claudio</td><td>2.2</td><td>1.0</td><td>—</td><td>—</td><td>—</td></tr><tr><td>8</td><td>Sanjeev</td><td>2.2</td><td>1.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>9</td><td>Soledad</td><td>2.1</td><td>2.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>10</td><td>Prashant</td><td>2.1</td><td>2.3</td><td>—</td><td>—</td><td>—</td></tr><tr><td colspan=\"4\">Majority@10</td><td>2.0</td><td></td><td></td><td></td></tr><tr><td rowspan=\"10\">bAbI task 16 valid baseline Acc = 53.5 test baseline Acc = 53.3</td><td>1</td><td>Loyd</td><td>8.9</td><td>8.4</td><td>—</td><td>—</td><td>—</td></tr><tr><td>2</td><td>Armida</td><td>8.7</td><td>8.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>3</td><td>Brannon</td><td>8.7</td><td>9.5</td><td>—</td><td>—</td><td>—</td></tr><tr><td>4</td><td>Pravin</td><td>8.4</td><td>8.8</td><td>—</td><td>—</td><td>—</td></tr><tr><td>5</td><td>Bozena</td><td>8.1</td><td>8.2</td><td>—</td><td>—</td><td>—</td></tr><tr><td>6</td><td>Surinder</td><td>8.0</td><td>7.2</td><td>—</td><td>—</td><td>—</td></tr><tr><td>7</td><td>Bonita</td><td>8.0</td><td>7.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>8</td><td>Trena</td><td>8.0</td><td>6.9</td><td>—</td><td>—</td><td>—</td></tr><tr><td>9</td><td>Rosalia</td><td>7.9</td><td>6.2</td><td>—</td><td>—</td><td>—</td></tr><tr><td>10</td><td>Leopoldo</td><td>7.7</td><td>9.3</td><td>—</td><td>—</td><td>—</td></tr><tr><td colspan=\"4\">Majority@10</td><td>8.0</td><td></td><td></td><td></td></tr><tr><td rowspan=\"10\">GSM8k valid baseline Acc = 14.3 test baseline Acc = 10.5</td><td>1</td><td>Douglas</td><td>0.5</td><td>0.9</td><td>Leopoldo</td><td>-2.5</td><td>-0.8</td></tr><tr><td>2</td><td>Juliann</td><td>0.4</td><td>0.2</td><td>Youssef</td><td>-1.9</td><td>-0.3</td></tr><tr><td>3</td><td>Kathleen</td><td>0.3</td><td>0.4</td><td>Epifanioo</td><td>-1.7</td><td>0.5</td></tr><tr><td>4</td><td>Ernst</td><td>0.3</td><td>0.4</td><td>Sanjeev</td><td>-1.6</td><td>-0.8</td></tr><tr><td>5</td><td>Tonja</td><td>0.3</td><td>0.0</td><td>Zoltan</td><td>-1.5</td><td>-0.3</td></tr><tr><td>6</td><td>Angela</td><td>0.2</td><td>0.2</td><td>Jianwei</td><td>-1.4</td><td>-0.4</td></tr><tr><td>7</td><td>Hina</td><td>0.2</td><td>-0.3</td><td>Abelardo</td><td>-1.3</td><td>-0.7</td></tr><tr><td>8</td><td>Diego</td><td>0.2</td><td>0.4</td><td>Alphonse</td><td>-1.3</td><td>-0.7</td></tr><tr><td>9</td><td>Bradley</td><td>0.1</td><td>0.9</td><td>Dagoberto</td><td>-1.3</td><td>-0.5</td></tr><tr><td>10</td><td>Jonathan</td><td>0.1</td><td>0.8</td><td>Nathaniel</td><td>-1.2</td><td>-0.1</td></tr><tr><td colspan=\"4\">Majority@10</td><td>0.8</td><td></td><td></td><td></td></tr><tr><td rowspan=\"10\">CLUTRR valid baseline Acc = 59.0 test baseline Acc = 53.0</td><td>1</td><td>Ashlee</td><td>4.4</td><td>1.2</td><td>Rob</td><td>-8.2</td><td>0.6</td></tr><tr><td>2</td><td>Penelope</td><td>3.4</td><td>3.4</td><td>Ahmad</td><td>-8.0</td><td>-0.2</td></tr><tr><td>3</td><td>Allyson</td><td>3.2</td><td>1.0</td><td>Mihai</td><td>-7.6</td><td>0.6</td></tr><tr><td>4</td><td>Chryystal</td><td>2.8</td><td>2.6</td><td>Frederick</td><td>-7.6</td><td>-0.4</td></tr><tr><td>5</td><td>Keiko</td><td>2.8</td><td>-4.4</td><td>Martin</td><td>-7.4</td><td>0.6</td></tr><tr><td>6</td><td>Alisha</td><td>2.6</td><td>3.0</td><td>Jake</td><td>-7.4</td><td>-0.4</td></tr><tr><td>7</td><td>Rosalba</td><td>2.4</td><td>1.8</td><td>Edward</td><td>-7.4</td><td>0.2</td></tr><tr><td>8</td><td>Jacqueline</td><td>2.2</td><td>2.0</td><td>Erik</td><td>-7.4</td><td>0.2</td></tr><tr><td>9</td><td>Zoltan</td><td>2.2</td><td>1.2</td><td>Giovanni</td><td>-7.4</td><td>-0.4</td></tr><tr><td>10</td><td>Dillon</td><td>2.0</td><td>1.4</td><td>Douglas</td><td>-7.2</td><td>-0.2</td></tr><tr><td colspan=\"4\">Majority@10</td><td>1.6</td><td></td><td></td><td></td></tr></table>",
+        "bbox": [
+            114,
+            153,
+            884,
+            760
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Performance of the best and worst performing entities in the validation and test datasets. The green color signifies consistent performance between the validation and test datasets, while the orange color indicates inconsistency. The entities, selected based on their performance in the validation dataset, show consistent behavior in the test dataset, suggesting that they generalize in each dataset. We have noted less than 10 entities in some columns because the remaining entities did not reduce the accuracy compared to the baseline.",
+        "bbox": [
+            112,
+            766,
+            884,
+            843
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "10927",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Contrary to our hypothesis and expectations, there appears to be no correlation between the frequency of names and the performance of LLMs in entity replacement.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "B.4 Effect of entity token length",
+        "text_level": 1,
+        "bbox": [
+            112,
+            159,
+            381,
+            174
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "The token length of a prompt can directly influence the performance of a QA task, particularly as it relates to the computational demands on the model. A longer prompt, represented by a larger number of tokens, increases the amount of data the model must process, potentially straining the model's capabilities and affecting the overall efficiency and accuracy of the task. Changes in the token length of a replaced name, which in turn modify the prompt's overall token length due to multiple instances of the name within the prompt, provide an interesting variable for analysis. We delved into this potential effect by examining the model's performance in relation to the token length of names, as illustrated in Fig 10. Contrary to expectations, our study found no observable correlation between the model's performance and the token length of names. This suggests that, within the ranges examined, the token length of names does not significantly affect the model's performance in QA tasks.",
+        "bbox": [
+            112,
+            180,
+            489,
+            501
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "B.5 Perplexity Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            512,
+            315,
+            527
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Our expectation was to observe a strong inverse correlation between the perplexity and accuracy of the model across all datasets, according to the findings reported in (Gonen et al., 2022a). However, we noted this inverse correlation only in some datasets, such as the bAbI task 14 (as shown in Figure Fig 11), but not in all of them. A plausible explanation for this inconsistency could be significant differences in our methodology compared to the approach adopted in (Gonen et al., 2022a). In our study, we limited our alterations to substituting named entities within the prompts, as opposed to changing the entire prompt. Therefore, this discrepancy in data augmentation techniques could potentially account for the lack of a significant correlation in some datasets.",
+        "bbox": [
+            112,
+            533,
+            489,
+            788
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Fig 12 shows the deviation of ppl from the baseline across various tasks. Regardless of the task, a more reduction in prompt ppl was observed when entities were replaced both in demonstrations and queries, compared to the other two data augmentation settings. This can likely be attributed to the repeated usage of the same replaced named entity within the prompt. The frequent repetition",
+        "bbox": [
+            112,
+            791,
+            489,
+            919
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "of the same named entity can contribute to lower ppl due to the decreased uncertainty and increased predictability of the next word in the sequence.",
+        "bbox": [
+            507,
+            84,
+            882,
+            133
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "10928",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/58e32b9a1d8be3aca18b8d21e84a5e900f70751678a1380cd248c6f16e43b277.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            181,
+            494,
+            368
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "image",
+        "img_path": "images/0cc51f357f2b90ed53ae8166969af204f584d8fafa2185109bbc91b4a56ce1de.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            181,
+            870,
+            368
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "image",
+        "img_path": "images/37ac92f5d8397bfd8fa8e32f952baaca36d88b46979b43e1f016f85873041a4a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            141,
+            378,
+            492,
+            565
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "image",
+        "img_path": "images/1d19d6b25aead79dfcb0ff6c926cc9fd0417248d8d82c84e24c3aea76fb21d11.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            376,
+            870,
+            565
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "image",
+        "img_path": "images/e25d5ca2003cba21a913c6ef1a965a09b79d7e307a69ec6dc759e92aa9bc3f36.jpg",
+        "image_caption": [
+            "Figure 9: The performance of Flan-T5-XL in relation to the frequency of names derived from a demographic dataset. Contrary to expectations, no correlation appears to exist between the frequency of names and the performance of Flan-T5-XL."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            141,
+            574,
+            494,
+            760
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "image",
+        "img_path": "images/6a5d7a6d2b38215a0d6e2ad3e8f5a130ea23aae41f6b5fa36ac2822f535aa51a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            574,
+            870,
+            760
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "page_number",
+        "text": "10929",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb4910993580662683da331b1c3466154c0a8e1cdebb80164a4dc0471c38e127.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            181,
+            494,
+            368
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "image",
+        "img_path": "images/c7f334e2cef818d5180ef97130361a86b3d660760b22cb7e08f9a0d0fb42075f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            181,
+            870,
+            369
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "image",
+        "img_path": "images/ee690188a4278ea50a5461c0a43e0aaf3f0ff4d7fff300a5ba0fa3646094d636.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            141,
+            378,
+            494,
+            565
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "image",
+        "img_path": "images/45a432677bb3ac51bada3ed1361b45190d4ef9e46315b0587fa17bf3fff65c92.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            376,
+            870,
+            565
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "image",
+        "img_path": "images/e631edc615afeea6f5496310d1725428481ff0d60956520a0768f6f8d03ae95f.jpg",
+        "image_caption": [
+            "Figure 10: The performance of various names in relation to their token length when using Flan-T5-XL. There is no apparent correlation between the performance and the token length in names. The quantity of names corresponding to each group of token lengths is illustrated on the plot."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            142,
+            574,
+            494,
+            760
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a8ab99e5b23ee2e077f59376c0bde6c68b9ac10248c552ba1a7fba4f3196ec0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            574,
+            870,
+            760
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "page_number",
+        "text": "10930",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "image",
+        "img_path": "images/55ef7e12e623edcf93b4f8edc5372fd7fc5edd4d38bf0d9b0f3204b7b0536358.jpg",
+        "image_caption": [
+            "Figure 11: Correlation between prompt perplexity (ppl) and LLM performance for task 14 and GPT-J-6B across various data augmentation settings. The accuracy and perplexity of prompt for the baseline are shown by red lines. Although a statistically significant inverse correlation between accuracy and ppl is observed in this dataset, this trend is not consistent across all datasets. The absence of this trend, contrary to expectation, could be attributed to the nature of our data augmentation approach, which only involved entity replacement instead of entire prompt modification."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            126,
+            118,
+            384,
+            292
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/308344b9ef32d4e7ad0cee1a5995f4c16dc916fdad0a39aaf452168c7b965240.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            386,
+            118,
+            625,
+            292
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/ab2d347feb646d3499a5a59a68a7b05e460ec4ac40d0a38f083b45d11ed2fbde.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            625,
+            118,
+            870,
+            292
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/f5c65aec909b08d9053bcfa5d800e94e9356bb5254188cd6537a95db895722bf.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            470,
+            495,
+            646
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/350b9857cc3b55f107480b196ec3f22045615f87988c8087faec994472bf696b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            499,
+            470,
+            873,
+            646
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/df0cfa71990b174c091e6d0f0b0f68a287f359214d474e3a5b816523961a6f4e.jpg",
+        "image_caption": [
+            "Figure 12: Deviation of prompt ppl for GPT-J-6B from the baseline across different tasks. A greater reduction in prompt ppl is observed when entities are replaced in both demonstrations and queries, a trend likely driven by the recurrent usage of the same replaced named entity within the prompt."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            649,
+            494,
+            822
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "image",
+        "img_path": "images/b762292e2ec4bcf50af5382be903b24fe5be2077e24f0663b2a059c110810f73.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            499,
+            649,
+            873,
+            822
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "page_number",
+        "text": "10931",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 17
+    }
+]

2023/Robustness of Named-Entity Replacements for In-Context Learning/92b3e48a-ba9b-4031-b53b-6f1de232b262_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e0704ead5e413959fb7c0d4559ebc2b00fa2d7c299d6b78463234f6440f5272
+size 1144672

2023/Robustness of Named-Entity Replacements for In-Context Learning/full.md

@@ -0,0 +1,362 @@
+# Robustness of Named Entity Replacements for In-Context Learning
+
+Saeed Goodarzi† Nikhil Kagita† Dennis Minn† Shufan Wang†
+
+Roberto Dessi $^{\infty, \pi}$  Shubham Toshniwal $^{\nabla}$  Adina Williams $^{\infty}$
+
+Jack Lanchantin\*,Koustuv Sinha\*,
+
+†University of Massachusetts Amherst; †Universitat Pompeu Fabra  
+∇NVIDIA; †FAIR, Meta
+
+sgoodarzitae@umass.edu, koustuvs@meta.com
+
+# Abstract
+
+A key feature of modern large language models (LLMs) is their ability to perform in-context learning, a prompting technique where query-answer demonstrations are shown before the final query. This allows for generalization to novel distributions at inference time where the LLM can learn new rules without parameter updates. However, the choice of demonstrations and their relationship to a particular query can have a profound impact on model accuracy, raising concerns about the true in-context generalization capabilities (Zhao et al., 2021). In this work, we explore the robustness of the in-context learning paradigm by focusing on entities. In particular, we seek to understand the robustness of LLM in-context learning with respect to named entity replacements. We discover a significant variance in downstream performance based on the choice of the named entities, across three popular reasoning tasks and two popular LLMs. Specifically, model accuracy on the test sets can fluctuate between -2.7 to +8.0 points depending on the choice of named entity replacements. Our analysis exposes the sensitivity of LLM in-context learning with respect to named entities, and offers a simple recipe to improve test performance by hyper-parameter tuning the named entities for a given dataset. Code and datasets for reproducing the results are publicly available. $^{1}$
+
+# 1 Introduction
+
+Modern large language models (LLMs) routinely demonstrate in-context learning, where the model can solve a wide range of tasks from a few demonstration examples (or exemplars) (Brown et al., 2020). In-context learning, or few-shot prompting, is a powerful learning paradigm as it allows for fast adaptation to new tasks without the need to train
+
+https://github.com/DennisMinn/ entities-in-context
+
+*Equal Leadership.
+
+# Original Demonstration
+
+[...] context: Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops.
+
+question: How many lollipops did Jason give to Denny?  
+answer: Jason started with 20 lollipops. Then he had 12 after giving some to Denny. So he gave Denny 20 - 12 = 8.
+
+[...] context: A bakery produces 60 loaves of bread each day. Two-thirds of the loaves are sold in the morning and half of what is left is sold equally in the afternoon and evening. question: How many loaves of bread are sold in the afternoon? answer:  $60 \times 2 / 3 = 40$  loaves of bread are sold in the morning.  $60 - 40 = 20$  loaves of bread are left.  $20 \times 2 = 40$  loaves of bread are sold in the afternoon. ###
+
+# Named Entity Replaced Demonstration
+
+[...] context: Douglas had 20 lollipops. He gave Denny some lollipops. Now Douglas has 12 lollipops.
+
+question: How many lollipops did Douglas give to Denny?  
+answer: Douglas started with 20 lollipops. Then he had 12 after giving some to Denny.  
+So he gave Denny 20 - 12 = 8.
+
+[...] context: A bakery produces 60 loaves of bread each day. Two-thirds of the loaves are sold in the morning and half of what is left is sold equally in the afternoon and evening. question: How many loaves of bread are sold in the afternoon? answer:  $60 \times 2 / 3 = 40$  loaves of bread are sold in the morning.  $60 - 40 = 20$  loaves of bread are left.  $20 \times 1 / 2 = 10$  loaves of bread are sold in the afternoon. ### 10
+
+Figure 1: Named entity replacements for In-Context Learning. Example from FLAN-T5 on GSM8k (Cobbe et al., 2021). Here we replace all male entities  $(e^{o})$  in demonstrations with Douglas, and observe a change in the query answer prediction (from the incorrect answer 40 to the correct answer 10). Only 1 out of  $k = 5$  demonstrations  $(D)$  is depicted to conserve space, with the remaining demonstrations denoted by “[...]”.
+
+model parameters. However, model performance in the in-context setting can vary significantly with the choice of demonstrations. Previous work has uncovered performance instability when altering the demonstrations, particularly when the demonstrations change semantically (Schick and Schütze, 2020; Zhao et al., 2021; Perez et al., 2021; Gonen et al., 2022a). Naturally, a question arises whether LLMs are robust to semantically preserving, minimal alterations to the demonstrations.
+
+In this paper, we seek to address this question using simple, semantically preserving alterations: named entity replacements. Because changing named entities should not, in principle, change the correct response to the query, the model is expected to remain invariant to such modifications. For example, given the demonstration "Alice has
+
+the ball. Alice went to the kitchen. Where is the ball? kitchen.", the model should be invariant to the choice of the named entity "Alice"-i.e., a model should behave the same if we swap all instances of "Alice" with another named entity, e.g. "Eve". However, we observe that the model behavior is not always robust to such alterations (Figure 1)—prompting us to critically evaluate named entity invariance within in-context learning.
+
+We conduct experiments to investigate the robustness of LLM in-context learning with respect to named entity replacements in three popular named entity heavy reasoning tasks, namely, bAbI, Weston et al. 2015; CLUTRR, Sinha et al. 2019; GSM8k, Cobbe et al. 2021, and two popular models, Flant5-XL (Chung et al., 2022a) and GPT-J (Wang and Komatsuzaki, 2021). Our key findings are:
+
+- Neither Flan-T5-XL nor GPT-J are robust to entity replacements for in-context learning. On bAbI tasks, the best replacement entity yields a +8.0 point performance increase (for Task 16), and the worst replacement entity yields a -2.7 point performance drop (Task 12). Similarly, for CLUTRR, we observe a 1.6 point improvement. For GSM8k, we only find <1 point increase/decrease with entity replacements.  
+- Performance gains from best performing entities are not random. To verify, we ran a bootstrap analysis and observe consistent gains with low variance.  
+- Best/worst performing entities are task-specific.  
+- We also analyze frequency, token length and perplexity of replaced entities, and find none of these alone explain our results, introducing an open problem for future research.
+
+# 2 Experimental Setup
+
+The goal of this study is to investigate the robustness of LLMs' in-context learning performance by changing entities in the demonstrations and queries. We consider a fixed set of prompt demonstrations aside from the named entities. By replacing the named entities and observing any accuracy changes, we can better understand the effect of certain entities on a model's accuracy. While many prior works have studied the robustness of LLMs with respect to perturbations (Ribeiro et al., 2020; Jin et al., 2020; Balasubramanian et al., 2020), we focus our analysis on in-context learning. A more in-depth analysis of related work is provided in
+
+Appendix subsection A.1.
+
+# 2.1 In-Context Learning
+
+With in-context learning, a test sample is comprised of a sequence of  $k$  demonstrations  $D = d_{1},d_{2},\ldots ,d_{k}$ . Each  $d_{i}$  consists of a context, question and answer string. The final query  $q$ , consisting only of a context and question string, follows  $D$ . The goal of in-context learning is to predict the answer to  $q$  given a small set of demonstrations at inference time. In other words, learning is done by eliciting activations and not changing model parameters.
+
+# 2.2 Modifying In-Context Learning with Named Entity Replacement
+
+We seek to measure the change in accuracy when we replace named entities in the demonstrations  $(D)$  and/or query  $(q)$ . We perform entity replacement in the following way.
+
+We first extract the named entities from each demonstration and query using an off-the-shelf NER model, and flag the subject entity,  $e^o$ , for replacement. We then select a replacement entity  $e^r$  from a large set of entities from Tzioumis (2018), which is composed of 4,250 unique first names. Finally, we perform named entity replacement in one of the following three settings:  $(D)$  replace all  $e^o$  entities in the demonstrations,  $e_{d_1}^o, e_{d_2}^o, \dots, e_{d_k}^o$ ,  $(q)$  replace all  $e^o$  entities in the query,  $e_q^o$ , or  $(D$  and  $q)$  replace all  $e^o$  entities in both the query and demonstrations,  $e_{d_1}^o, e_{d_2}^o, \dots, e_{d_k}^o, e_q^o$ . Note that each demonstration  $d_i$  can contain multiple instances of the entity  $e^o$ . These settings allow us to isolate and evaluate the effect of named entities in different parts of an in-context learning sample on accuracy.
+
+We run multiple experiments using bootstrap sampling (with replacement) on three sets of names, half labeled as 'male' and half as 'female'. For each name  $e^r$  in a particular bootstrap set, we replace the original entities  $e^o$  with  $e^r$  in the test samples according to one of the three previously mentioned settings. We then evaluate the performance on the entity-replaced dataset. This procedure is repeated for each bootstrap set.
+
+![](images/17fab8028c5b111fa1080156a13637f492bdd2e2c8845e213fbafebb9fc96144.jpg)  
+Figure 2: Bootstrap results of Flan-T5-XL on validation sets for 3 distinct sets of 200 names. Error bars represent the variation across bootstrap sets of names.
+
+Baseline results are reported using the original prompt demonstrations and query. However, our replacement strategy introduces a frequency bias, as  $e^r$  is used to replace multiple demonstrations in a given input, resulting in artificially increasing the usage of a given entity in an input. Thus, we also report the median accuracy of replacement, which serves as a proxy baseline for replacing all demonstrations and/or query with a random entity.
+
+# 2.3 Models and Datasets
+
+We evaluate our entity replacement strategy on both an encoder-decoder LLM (Flan-T5-XL, 3B; Chung et al. 2022b) as well as a decoder-only LLM (GPT-J, 6B; Wang and Komatsuzaki 2021). We test these models across the following datasets: bAbI (Tasks 1, 12, 14, 16; Weston et al. 2015), CLUTRR (Sinha et al., 2019), GSM8k (Cobbe et al., 2021). We chose these datasets primarily because they include a significant presence of named entities.
+
+For each dataset, we select five samples from the training set to serve as demonstrations. To maintain grammatical correctness and clarity, we
+
+only do replacements that preserve gender when a pronoun is contained in the sentence (see Appendix A.5). This step specifically addresses any potential uncontrolled effects on the performance from entity replacement, such as perplexity spikes that may results from swaps where the genders of  $e^{o}$  and  $e^{r}$  differ, and the model is expected to accommodate unexpected pronominal coreference with  $e^{r}$ .
+
+# 3 Results
+
+Performance Across Bootstrap Sets. We find a noticeable change in accuracy with the best and worst performing entity compared to the baseline or median performing entity. Figure 2 shows Flant5-XL results on all validation datasets (analogous GPT-J results are in Figure 5 in the Appendix, and follow a similar pattern). This indicates that specific entities can noticeably alter in-context learning performance. Furthermore, the standard deviations for the median and best performing entity across bootstraps are low, suggesting that the selection of different name sets will likely yield similar median and best performance. However, the variation in the worst performing entities is large.
+
+Performance Within a Single Bootstrap Set. To analyze the results in more detail, we measure the performance for Flan-T5-XL and GPT-J on a single bootstrap set of 200 examples. We show the results for bAbI task 14 in Figure 3. When we replace names only in the demonstrations  $(D)$ , the variance in model performance is lower compared to both when we replace names only in the query  $(q)$  and in both query and demonstration  $(D$  and  $q)$ . The  $(D$  and  $q)$  setup exhibits the highest variance, as this scenario involves the maximum change in the tokens of the original prompt. Nevertheless, we see a consistent variance from the baseline by simply replacing named entities in any of the three replacement settings. Other datasets exhibit a similar behavior; see Appendix Figures 6 and 7.
+
+Best and Worst Entity Generalization. Thus far, we've used the valid set to demonstrate accuracy variation. To evaluate the best performing entities on the test set, we implement a majority voting method as follows. We first find the 10 best performing entities on the valid set. For each sample in the test set, we take the majority answer prediction from the 10 different best names, denoted "Maj@10". Table 1 shows results of entity replacements across the validation and test sets by replac
+
+ing entities in the demonstration and query (both  $D$  and  $q$ ). We observe a maximum 8.0 point accuracy increase, seen on bAbI task 16. $^{6}$  The best and worst performing entities vary by dataset, but overall entities generalize across the validation and test sets, with a uniform increase in accuracy from the best entity. A list of these entities for each dataset and further evaluations are in Appendix Table 2.
+
+For all datasets, there are entities with appreciable deviations from the baseline, and these entities generalize across the validation and test splits. This finding indicates that the observed accuracy shifts are not coincidental and highlights the influence of entity swaps, including those that preserve overall example semantics, such as for names, on model performance.
+
+# 4 Analysis
+
+Why are the models not invariant to entity replacements? We conducted several analyses in attempts to find the root cause of this result, but they were largely inconclusive.
+
+Does the frequency of names affect how much they shift model performance? The lack of robustness could well be explained by the frequency effects of named entities in the training corpus of the pre-trained model (Wei et al., 2021). We use the entity frequency data provided by Tzioumis (2018), and assessed performance changes across names with varying frequencies. We do not find any correlation between name frequency and the model performance post entity replacement (see Appendix Section B.3).
+
+Does the token length of names affect performance? Another hypothesis of the invariance could be due to the inconsistent tokenization of named entities. Rare names can be tokenized into longer tokens based on the learned subwords of the model. We analyzed token length post tokenization and compared it with model performance, but we are unable to find any meaningful relationship between the two (see Appendix Section B.4).
+
+Does improved performance result from a decrease in prompt perplexity due to entity replacement? Finally, the surprisal of the LLM to certain entities could be a promising hypothesis to explain the change in model performance (Gonen et al., 2022b). We compare prompt perplexity af
+
+![](images/806ba71760a2083e986f1d9e3c8486ceedd8450c702759db96c9852bcfc8b229.jpg)
+
+![](images/a884b3da2e0d17034a9e33ec7f497e383486e897e60a3367f94c029904ee1c03.jpg)  
+bAbI Task 14 (Baseline Accuracy=70)  
+Figure 3: Accuracy for Flan-T5-XL and GPT-J-6B on bAbI task 14 for a single bootstrap set of 200 samples.
+
+ter entity replacement to LLM performance, and yet again we are unable to find a statistical correlation (see Appendix Section B.5). While Gonen et al. (2022b) suggests improvement of model performance with decrease in perplexity, we observe certain entities displaying the opposite behavior.
+
+# 5 Conclusion
+
+In this study, we find that LLMs exhibit a consistent lack of robustness with respect to named entity replacement for the in-context learning setting, fur
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="2">Baseline</td><td colspan="2">Best Entity</td><td colspan="2">Worst Entity</td></tr><tr><td>Valid</td><td>Test</td><td>Valid</td><td>Test (Maj@10)</td><td>Valid</td><td>Test</td></tr><tr><td>bAbI 12</td><td>95.1</td><td>95.5</td><td>98.5 [+3.4]</td><td>99.0 [+3.5]</td><td>92.4 [-2.7]</td><td>93.0 [-2.5]</td></tr><tr><td>bAbI 14</td><td>76.1</td><td>77.5</td><td>78.4 [+2.3]</td><td>79.5 [+2.0]</td><td>75.2 [-0.9]</td><td>76.6 [-0.9]</td></tr><tr><td>bAbI 16</td><td>53.5</td><td>53.3</td><td>61.7 [+8.2]</td><td>61.3 [+8.0]</td><td>55.5 [+2.0]</td><td>55.9 [+2.6]</td></tr><tr><td>GSM8k</td><td>14.3</td><td>10.5</td><td>14.6 [+0.3]</td><td>11.3 [+0.8]</td><td>12.7 [-1.6]</td><td>10.1 [-0.4]</td></tr><tr><td>CLUTRR</td><td>59.0</td><td>53.0</td><td>61.8 [+2.8]</td><td>54.6 [+1.6]</td><td>51.4 [-7.6]</td><td>53.1 [+0.1]</td></tr></table>
+
+Table 1: Performance of FLAN-T5-XL, with accuracy shift from baseline provided in [square brackets]. Entity replacements are done in the  $D$  and  $q$  setting. bAbI task 1 reaches perfect accuracy and is thus omitted from this table. Performance between the validation and test datasets is consistent, and we observe a generalizable increase in accuracy across all datasets.
+
+ther questioning the process of reasoning employed by these models (Min et al., 2022). We demonstrate this across three axes: the complexity of the task, the model architecture, and which named entities are to be replaced (in the demonstrations, query, or both). The cause for this effect remains elusive, which opens up exciting research directions. One intriguing result is that the entities that can boost performance are dataset specific. This points to potential avenues for enhancing LLM downstream performance by performing hyperparameter search over a range of entities.
+
+# 6 Limitations
+
+We acknowledge that the selection of names in our study may not provide an unbiased representation of the complete set of possible names. The data provided by Tzioumis (2018) was derived from mortgage applications which may introduce a bias against popular names within marginalized communities. Moreover, this issue is compounded when we sub-sample the set of names due to computational limitations. Additionally, we used a NER model available through Hugging Face to identify  $e^o$ . It was BERT-based (Devlin et al., 2018) and trained on a CONLL-2003 corpus (Tjong Kim Sang and De Meulder, 2003) (Wolf et al., 2019), and we anecdotally observed that sometimes it struggles to identify rare names as named entities, for example, it incorrectly identified "TreQuan" as "TreQ." We spot checked for this issue and removed names that posed problems to the best of our abilities.
+
+# 7 Ethics and Broader Impact
+
+In this work, we explore the effect of name swaps in an in-context learning setup and find that particular names can increase accuracy. In our efforts to align names with their respective pronouns, we assumed binary gender, but acknowledge that other genders exist. We made this assumption because our evaluated datasets used binary pronouns, and our names dataset was also only binarily annotated. While our paper assumes name-gender associations, we do not want to discredit negative evidence. A name that is not explicitly associated with a gender doesn't imply evidence of absence of association (imagine that only one person named "Alex" was sampled, and she was a woman—that doesn't mean that there aren't also men or nonbinary people called "Alex"). At the moment, there aren't many annotated name datasets that consists
+
+of gender-agnostic, gender-ambiguous, or gender-neutral names. Hopefully, additional work such as (Dev et al., 2021; Lauscher et al., 2022; Ovalle et al., 2023) will expand such resources, but regardless of these complications, our conclusions are still useful for trying to understand why some names can shift model performance under some settings.
+
+# References
+
+2023. Genderize.io. https://genderize.io/. Accessed: June 15, 2023.  
+Oshin Agarwal, Yinfei Yang, Byron C Wallace, and Ani Nenkova. 2020. Entity-switched datasets: An approach to auditing the in-domain robustness of named entity recognition models. arXiv preprint arXiv:2004.04123.  
+Sriram Balasubramanian, Naman Jain, Gaurav Jindal, Abhijeet Awasthi, and Sunita Sarawagi. 2020. What's in a name? are BERT named entity representations just as good for any other name? In Proceedings of the 5th Workshop on Representation Learning for NLP, pages 205-214, Online. Association for Computational Linguistics.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022a. Scaling instruction-finetuned language models. arXiv:2210.11416.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022b. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.  
+Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff Phillips, and Kai-Wei Chang.
+
+2021. Harms of gender exclusivity and challenges in non-binary representation in language technologies. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1968-1994, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805.  
+Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. 2017. Hotflip: White-box adversarial examples for text classification. arXiv preprint arXiv:1712.06751.  
+Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. 2018. Black-box generation of adversarial text sequences to evade deep learning classifiers. In 2018 IEEE Security and Privacy Workshops (SPW), pages 50-56. IEEE.  
+Hila Gonen, Srini Iyer, Terra Blevins, Noah A Smith, and Luke Zettlemoyer. 2022a. Demystifying prompts in language models via perplexity estimation. arXiv preprint arXiv:2212.04037.  
+Hila Gonen, Shauli Ravfogel, and Yoav Goldberg. 2022b. Analyzing gender representation in multilingual models. In Proceedings of the 7th Workshop on Representation Learning for NLP, pages 67-77, Dublin, Ireland. Association for Computational Linguistics.  
+Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8018-8025.  
+Anne Lauscher, Archie Crowley, and Dirk Hovy. 2022. Welcome to the modern world of pronouns: Identity-inclusive natural language processing beyond gender. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1221-1232, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Bill Yuchen Lin, Wenyang Gao, Jun Yan, Ryan Moreno, and Xiang Ren. 2021. RockNER: A simple method to create adversarial examples for evaluating the robustness of named entity recognition models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3728-3737, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2021. What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804.
+
+Robert L Logan IV, Ivana Balažević, Eric Wallace, Fabio Petroni, Sameer Singh, and Sebastian Riedel. 2021. Cutting down on prompts and parameters: Simple few-shot learning with language models. arXiv preprint arXiv:2106.13353.  
+Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837.  
+Anaelia Ovalle, Palash Goyal, Jwala Dhamala, Zachary Jaggers, Kai-Wei Chang, Aram Galstyan, Richard Zemel, and Rahul Gupta. 2023. "i'm fully who i am": Towards centering transgender and non-binary voices to measure biases in open language generation. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1246-1266.  
+Nicolas Papernot, Patrick McDaniel, Ananthram Swami, and Richard Harang. 2016. Crafting adversarial input sequences for recurrent neural networks. In MILCOM 2016-2016 IEEE Military Communications Conference, pages 49-54. IEEE.  
+Ethan Perez, Douwe Kiela, and Kyunghyun Cho. 2021. True few-shot learning with language models. Advances in neural information processing systems, 34:11054-11070.  
+Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902-4912, Online. Association for Computational Linguistics.  
+Timo Schick and Hinrich Schütze. 2020. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676.  
+Timo Schick and Hinrich Schütze. 2022. True few-shot learning with prompts—a real-world perspective. Transactions of the Association for Computational Linguistics, 10:716-731.  
+Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. 2020. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980.  
+Koustuv Sinha, Shagun Sodhani, Jin Dong, Joelle Pineau, and William L Hamilton. 2019. Clutrr: A diagnostic benchmark for inductive reasoning from text. arXiv preprint arXiv:1908.06177.  
+Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147.
+
+Miles Turpin, Julian Michael, Ethan Perez, and Samuel R Bowman. 2023. Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting. arXiv preprint arXiv:2305.04388.  
+Konstantinos Tzioumis. 2018. Demographic aspects of first names. Scientific data, 5(1):1-9.  
+Ben Wang and Aran Komatsuzaki. 2021. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model. https://github.com/kingoflolz/mesh-transformer-jax.  
+Xiao Wang, Qin Liu, Tao Gui, Qi Zhang, Yicheng Zou, Xin Zhou, Jiacheng Ye, Yongxin Zhang, Rui Zheng, Zexiong Pang, Qinzhuo Wu, Zhengyan Li, Chong Zhang, Ruotian Ma, Zichu Fei, Ruijian Cai, Jun Zhao, Xingwu Hu, Zhiheng Yan, Yiding Tan, Yuan Hu, Qiyuan Bian, Zhihua Liu, Shan Qin, Bolin Zhu, Xiaoyu Xing, Jinlan Fu, Yue Zhang, Minlong Peng, Xiaqing Zheng, Yaqian Zhou, Zhongyu Wei, Xipeng Qiu, and Xuanjing Huang. 2021. TextFlint: Unified multilingual robustness evaluation toolkit for natural language processing. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations, pages 347-355, Online. Association for Computational Linguistics.  
+Jason Wei, Dan Garrette, Tal Linzen, and Ellie Pavlick. 2021. Frequency effects on syntactic rule learning in transformers.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903.  
+Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart Van Merrienboer, Armand Joulin, and Tomas Mikolov. 2015. Towards ai-complete question answering: A set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.  
+Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR.
+
+# A Appendix
+
+# A.1 Related Work
+
+Several works investigated the robustness of language models with respect to input rewrites (Papernot et al., 2016; Ebrahimi et al., 2017; Gao et al., 2018; Jin et al., 2020). In particular, Balasubramanian et al. (2020) found that while fine-tuned BERT models were brittle in certain tasks, they demonstrated robustness against entity replacement for question answering. Other works have shown that many transformer-based models are insufficiently robust to name swaps (Ribeiro et al. 2020; Agarwal et al. 2020; Lin et al. 2021; Wang et al. 2021, i.a.), Our work differs from these in that we focus on the in-context learning setting.
+
+The ability to do in-context learning first emerged from Brown et al. (2020). Subsequently, many works sought to improve and understand the abilities of LLMs with respect to in-context learning (Liu et al., 2021; Zhao et al., 2021). Perez et al. (2021); Logan IV et al. (2021); Schick and Schütze (2022); Turpin et al. (2023) show an improvement over the original paper in the the "true few-shot learning" setup where there is no access to a large test set to tune in-context learning hyperparameters. Despite the impressive results of this new paradigm, little has been understood regarding the generalization of certain in-context demonstrations.
+
+Gonen et al. (2022a) bring to attention the importance of the specific prompts in model performance. They show that the lower the perplexity of the prompt is, the better the prompt is able to perform the task. Min et al. (2022) show that the actual labels in the prompt are not very important. However, the labels still have a small impact in performance.
+
+In recent years, several studies have explored the significance of prompt quality in few-shot learning, particularly in the context of question-answering tasks. One such study was conducted by Shin et al. (2020) who proposed AutoPrompt, an automated approach for prompt generation in few-shot learning scenarios that has been shown to achieve the highest accuracy. The AutoPrompt model leverages gradient-based prompt search techniques to find the optimal prompt for a given task, thereby reducing the need for human intervention. Additionally, AutoPrompt utilizes Masked Language Models (MLMs) to perform sentiment analysis and natural language inference tasks without the need for fine-tuning, highlighting the potential of MLMs
+
+in few-shot learning scenarios. In another study, Gonen et al. (2022a) introduced a novel technique for prompt generation that involves using a small set of manually created prompts and then applying paraphrasing and backtranslation to generate additional prompts automatically. Specifically, they utilized GPT-3 to perform the paraphrasing step, which involves rephrasing the original prompts while preserving their underlying meaning. The backtranslation step then involves translating the paraphrased prompts into another language and then back into the original language to generate new prompts that are semantically similar to the original prompts but with different phrasing and wording. This approach enables the generation of a large number of diverse prompts without requiring extensive manual effort, which can be particularly useful in few-shot learning scenarios where data is limited. Gonen et al. (2022a) also demonstrated an inverse relationship between the perplexity of the prompts and the model's performance across various tasks.
+
+Our work is related to previous studies as we focus on the true few-shot learning setup for the task of question answering and explore the robustness of NLP models against entity replacement. There are three main ways in which our work differs from previous studies. First, our investigation focuses on analyzing the robustness of language models against entity replacement specifically for the task of question answering in the ICL setup, which has not been explored in previous studies. Second, we utilize left-to-right language models, which are more widely used and powerful than masked language models. Finally, our approach involves prompt-learning instead of fine-tuning, whereby the parameters of the language model are modified to improve its performance on specific tasks. By leveraging the true few-shot learning setup and prompt-learning techniques, we aim to develop a more robust and effective approach for few-shot question-answering tasks that can handle entity replacement and other challenges more effectively.
+
+# A.2 Dataset Details
+
+bAbI (Weston et al., 2015): The bAbI dataset consists of 20 tasks that evaluate various aspects of language understanding by NLP models. In our experiments, we select four tasks representing varying degrees of difficulty from the bAbI dataset: task 1, task 12, task 14, and task 16.
+
+context: Mary and  $[e_{d1}]$  travelled to the   
+bathroom. John and  $[e_{d1}]$  travelled to the office.   
+question: Where is  $[e_{d1}]$ ?   
+answer: office   
+????   
+context:  $[e_{d2}]$  and Mary went back to the office.   
+Daniel and Sandra went to the bedroom.   
+question: Where is  $[e_{d2}]$ ?   
+answer: bedroom   
+????   
+context: Daniel and John moved to the hallway.   
+ $[e_q]$  and Daniel journeyed to the bathroom.   
+question: Where is  $[e_q]$ ?   
+answer:
+
+Figure 4: Named entity replacement overview for in context learning. Here we show an in-context learning example with  $k = 2$  demonstrations before the final query. We consider 3 scenarios: replacing only the demonstration entities  $(e_{d1}, e_{d2})$ , only the query entity  $(e_q)$ , or both  $(e_{d1}, e_{d2}, e_q)$ . All replacements are done with the same replacement entity  $e^r$ .
+
+CLUTRR (Sinha et al., 2019): This dataset is similar to a subset of the bAbI dataset, but contains more difficult multi-step questions. This allows us to determine whether named entity replacement robustness is affected by task difficulty. Given a multi-sentence context describing kinship relations between certain entities and a question, the task is to predict the kinship relation between the entities that can be inferred from the context but is not made explicit. We convert this into a multiple-choice question where we expect the model to pick the answer from the four choices provided.
+
+GSM8k (Cobbe et al., 2021): This dataset is a collection of elementary multi-step mathematical reasoning questions. We specifically selected this dataset due to the challenging nature of arithmetic operations, where existing models often struggle to achieve accurate results. We subsampled GSM8k to include only examples which contained identifiable named entities.
+
+The Tzioumis (2018) dataset, contains of 4,250 unique first names, associated with their unigram frequency, and self-reported demographic information.
+
+For each sample in the dataset, which is comprised of a context, question, and answer triplet, we transform it into a formatted string. This transformation involves concatenating the field name, a colon, space, the respective value corresponding to that field, and newline character. We ap
+
+ply this transformation to  $k$  demonstrations:  $D = d_{1},d_{2},\ldots ,d_{k}$  and the final query  $q$  , where the answer to  $q$  is omitted. We concatenate the formatted  $k$  demonstrations and query  $q$  , delimiting each string with a newline character followed by  $"\# \# \# "$  and a newline character. Figure 4 shows an incontext learning example.
+
+# A.3 Entity Selection
+
+In general, we determine  $e^o$  by selecting the first encountered entity within the question, as it often represents the subject of the question. If the question contains interrogative pronouns, such as "who," no entities will be present within the question. In this case, we consider  $e^o$  as the first entity encountered within the answer since it typically represents the subject being referred to. If the subject of the question and answer is referred to using pronouns then neither the question nor answer will contain any entities, we assume that noun ambiguity is not an issue and select the first encountered entity within the context as  $e^o$ .
+
+# A.4 Entity-replacement Scenarios
+
+In the demonstration-only replacement scenario, we perform entity replacement exclusively on the demonstrations  $e_d$  while keeping the query unmodified 4. By selectively replacing entities within the context, question, and answer of the demonstrations, we aim to examine the extent to which this replacement strategy influences the model's performance.
+
+In the query-only replacement setting, we solely augment the entities within the query  $e_{q}$  while leaving the demonstrations unaltered Fig 4. This allows us to investigate the impact of enhancing the query's entity representation on the model's ability to generate accurate responses.
+
+Applying both demonstration and query replacement, we replace  $e_d$  and  $e_q$  in both the demonstrations and query simultaneously. By enriching the entity information in both the context and the question, we explore the combined effect of entity replacement on the overall performance of the model.
+
+# A.5 Ensuring Grammatical Correctness
+
+To begin, we verify the pronouns associated with both  $e^o$  and  $e^r$  using (gen, 2023). If the gender associated with  $e^o$  corresponds with that of  $e^r$ , we proceed to substitute  $e^o$  with  $e^r$ . However, if there
+
+is no alignment in gender, we refrain from altering  $e^o$  and maintain its original form.
+
+# B Analysis
+
+# B.1 Analyzing Best/Worst performing entities
+
+In order to examine the generalizability of the best and worst performing entities in each dataset, we extracted named entities from the validation dataset and tested them on the test dataset. More specifically, we replaced the named entities in the prompts of the validation dataset with a set of names that included 100 male and 100 female names. We then calculated the accuracy for each name, selecting the 10 names that exhibited the highest performance and the 10 names that showed the lowest performance. Next, we replaced the named entities in the test dataset prompts with these 20 selected names and calculated the corresponding accuracy for the test dataset. Table 2 presents the best and worst performing entities for each dataset along with the absolute percentage of accuracy changes. The ensemble model functions by replacing the named entity in the prompt with the highest performing entity and then generating an answer at each step. After cycling through the 10 best performing entities, it selects the answer with the most votes as the final response. This approach is referred to as majority@10. It's important to note that the bAbI task 1 doesn't have a 'best performing entity' because the baseline accuracy is already  $100\%$ , leaving no room for improvement.
+
+Figure 8 provides an example of an generated answer for the GSM8k test dataset. The original prompt failed to produce the correct answer, while the transformed prompt was able to successfully answer the question. The absolute improvement of  $0.8\%$  in the accuracy of GSM8k through the majority@10 is remarkable, given the following facts: 1) The baseline accuracy for this challenging dataset is  $10.5\%$ , making even a small improvement significant. 2) The number of named entities in GSM8k is fewer than in other datasets, meaning only a few words are replaced in the transformed prompt, as shown in Figure 8. The ratio of replaced words to the total number of words in the prompt is 0.032. 3) Of the 1319 examples in the GSM8k test dataset, 663 contain a male name, 377 contain a female name, and 279 don't contain any named entity in the query. We could have achieved greater changes in percentage if we had only considered examples where the gender of the best perform
+
+ing entity matched the gender of the name in the query (as replacements wouldn't occur otherwise). However, we chose not to adjust the test dataset to ensure that our accuracy improvement could be compared with other studies.
+
+The consistency in performance between the validation and test datasets in CLUTRR is lower than that in the other datasets. This could be attributed to the structure of CLUTRR, where four options are provided in the prompt, from which the model chooses an answer. This setup introduces a degree of randomness to the generated answer, potentially leading to less consistency between the performance of the validation and test datasets.
+
+# B.2 LLM Architecture
+
+The architecture of LLMs plays a crucial role in determining their behavior and performance, including their vulnerability to entity replacement. The design choices and architectural components of LLMs can influence how they process and generate text, affecting their sensitivity to changes in named input entities. By comparing the Flant5-XL, which employs an encoder-decoder architecture, with the GPT-J-6B, which represents a decoder-only architecture, we aimed to investigate the impact of LLM architecture on their brittleness towards entity replacement. The high variance in the performance of GPT-J (as shown in Fig 7) underscores that the brittleness towards entity replacement is not exclusive to the encoder-decoder architecture but is also a feature of decoder-only architecture models.
+
+# B.3 Effect of entity frequency
+
+Frequently recurring named entities in the web, which subsequently appear in the training data, can impact both the model's ppl and performance. Entities appearing frequently in the training data are more likely to be well-represented in the model's understanding, contributing to lower ppl values. This occurs because the frequent presence of these entities reduces the unpredictability of the next word in a sequence, making it easier for the LLM to anticipate. The hypothesis is that names appearing more frequently would contribute to better model performance. To examine this hypothesis, the frequency of names, extracted from (Tzioumis, 2018), were categorized into five groups of equal numbers, labeled G1 to G5, ranging from the most to the least frequent names. Fig 9 presents the performance of Flan-T5-XL across different datasets.
+
+![](images/d2fa12070fd32ac897205bd220ddb0c766c2a9efe46e048bc8ea4736ad8eec2d.jpg)
+
+![](images/074580c419c423622844c55c98b779242616403502e4dfd01ca02db84eff0410.jpg)
+
+![](images/149a1c5187a9b12426b10c9b18c1810803a37bb3390a2b5e830472973d64304c.jpg)  
+Figure 5: Bootstrap results of GPT-J-6B for 3 distinct sets of names, each comprising 200 names. Error bars represent the variation across different name sets.
+
+![](images/b595f85513bb0796b82cd377b64cef7f712d753e68c94765d5913129c854c29d.jpg)
+
+![](images/84c5071efe92c658740b73b43fb7ab9a82d2445158c42c57023db64eadb65699.jpg)
+
+![](images/9af7ee133001d94dd6145685fdd994c10f2f19ca3508f836765a296b8d4237c7.jpg)
+
+![](images/a3d0c964c46fa72e5dcfc45acd7e059152ae75558ae08d160fe211bd7efe33e5.jpg)
+
+![](images/a8be2c992d5164ed570c4c8fb0a62549d1c9061d5f2323a0edcd583e7e83e934.jpg)
+
+![](images/fa2270a4e0ab7c99a1855f2a278ec33e6211bd361f83ee9d0cafdcc0942deaa0.jpg)  
+Figure 6: Deviation in accuracy from the baseline (without data augmentation) for Flan-T5-XL across different datasets.
+
+![](images/6adbcbf66a45bfc8c4b68034db00705847cb60e07ccc10e61116be7cf63afd4a.jpg)
+
+![](images/fc6aa65b5f61908d8679283485b4f204d69bc634e25afac91d976a623963c33e.jpg)
+
+![](images/ddbd6b645183aa081d4e287cc0f5f48cce9b986b4fe698cc6812c380a7380779.jpg)
+
+![](images/dcec4185a5fef4c3dd60e4caa8629a2f2c8aead8bf54a79c699154c42db4b6dc.jpg)  
+Figure 7: Deviation in accuracy of GPT-J from the baseline across various bAbI tasks. Like the encoder-decoder architecture, the decoder-only architecture also exhibits brittleness towards entity replacement.
+
+![](images/dd0e5c65da0264ed6197fea3b66ecd58100960523f9ca9ebb2bff5d611931eb4.jpg)
+
+![](images/7125dc817e6dab1e6722d82340e62eee3454ac6ec98b07e07996c0c5ee4e27ef.jpg)  
+Figure 8: An example of a generated answer from the GSM8k test dataset. In this instance, the original prompt fails to generate the correct answer, whereas the transformed prompt — with the best-performing entity replaced — successfully produces the correct response. This generated text is reproducible using a greedy decoding.
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Index</td><td colspan="3">Best performing entity</td><td colspan="3">Worst performing entity</td></tr><tr><td>entity</td><td>Acc change (valid)</td><td>Acc change (test)</td><td>entity</td><td>Acc change (valid)</td><td>Acc change (test)</td></tr><tr><td rowspan="10">bAbI task 12 valid baseline Acc = 95.1 test baseline Acc = 95.5</td><td>1</td><td>Penelope</td><td>3.6</td><td>3.0</td><td>Mihai</td><td>-4.0</td><td>-2.3</td></tr><tr><td>2</td><td>Loreen</td><td>3.6</td><td>3.0</td><td>Raisa</td><td>-4.0</td><td>-3.6</td></tr><tr><td>3</td><td>Dorcas</td><td>3.6</td><td>3.2</td><td>Ethel</td><td>-3.7</td><td>-3.5</td></tr><tr><td>4</td><td>Toby</td><td>3.4</td><td>3.4</td><td>Migdalia</td><td>-3.0</td><td>-4.0</td></tr><tr><td>5</td><td>Flora</td><td>3.3</td><td>2.9</td><td>Azucena</td><td>-2.6</td><td>-3.2</td></tr><tr><td>6</td><td>Nataliya</td><td>3.3</td><td>3.4</td><td>Bhavana</td><td>-2.6</td><td>-3.6</td></tr><tr><td>7</td><td>Nathaniel</td><td>3.3</td><td>3.4</td><td>Surinder</td><td>-2.1</td><td>-0.9</td></tr><tr><td>8</td><td>Lucretia</td><td>3.3</td><td>3.3</td><td>Sanjeev</td><td>-1.9</td><td>-1.7</td></tr><tr><td>9</td><td>Keiko</td><td>3.2</td><td>3.0</td><td>Armida</td><td>-1.8</td><td>-1.1</td></tr><tr><td>10</td><td>Bernardo</td><td>3.2</td><td>3.1</td><td>Jianwei</td><td>-1.7</td><td>-1.4</td></tr><tr><td colspan="4">Majority@10</td><td>3.5</td><td></td><td></td><td></td></tr><tr><td rowspan="10">bAbI task 14 valid baseline Acc = 76.1 test baseline Acc = 77.5</td><td>1</td><td>Erich</td><td>2.9</td><td>2.0</td><td>Frederick</td><td>-1.4</td><td>-1.2</td></tr><tr><td>2</td><td>Jacqueline</td><td>2.6</td><td>1.8</td><td>Raisa</td><td>-1.3</td><td>-1.3</td></tr><tr><td>3</td><td>Zoltan</td><td>2.4</td><td>2.3</td><td>Maryellen</td><td>-0.5</td><td>-0.6</td></tr><tr><td>4</td><td>Chryystal</td><td>2.3</td><td>1.5</td><td>Meera</td><td>-0.5</td><td>-0.5</td></tr><tr><td>5</td><td>Petra</td><td>2.2</td><td>1.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>6</td><td>Pravin</td><td>2.2</td><td>1.8</td><td>—</td><td>—</td><td>—</td></tr><tr><td>7</td><td>Claudio</td><td>2.2</td><td>1.0</td><td>—</td><td>—</td><td>—</td></tr><tr><td>8</td><td>Sanjeev</td><td>2.2</td><td>1.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>9</td><td>Soledad</td><td>2.1</td><td>2.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>10</td><td>Prashant</td><td>2.1</td><td>2.3</td><td>—</td><td>—</td><td>—</td></tr><tr><td colspan="4">Majority@10</td><td>2.0</td><td></td><td></td><td></td></tr><tr><td rowspan="10">bAbI task 16 valid baseline Acc = 53.5 test baseline Acc = 53.3</td><td>1</td><td>Loyd</td><td>8.9</td><td>8.4</td><td>—</td><td>—</td><td>—</td></tr><tr><td>2</td><td>Armida</td><td>8.7</td><td>8.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>3</td><td>Brannon</td><td>8.7</td><td>9.5</td><td>—</td><td>—</td><td>—</td></tr><tr><td>4</td><td>Pravin</td><td>8.4</td><td>8.8</td><td>—</td><td>—</td><td>—</td></tr><tr><td>5</td><td>Bozena</td><td>8.1</td><td>8.2</td><td>—</td><td>—</td><td>—</td></tr><tr><td>6</td><td>Surinder</td><td>8.0</td><td>7.2</td><td>—</td><td>—</td><td>—</td></tr><tr><td>7</td><td>Bonita</td><td>8.0</td><td>7.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>8</td><td>Trena</td><td>8.0</td><td>6.9</td><td>—</td><td>—</td><td>—</td></tr><tr><td>9</td><td>Rosalia</td><td>7.9</td><td>6.2</td><td>—</td><td>—</td><td>—</td></tr><tr><td>10</td><td>Leopoldo</td><td>7.7</td><td>9.3</td><td>—</td><td>—</td><td>—</td></tr><tr><td colspan="4">Majority@10</td><td>8.0</td><td></td><td></td><td></td></tr><tr><td rowspan="10">GSM8k valid baseline Acc = 14.3 test baseline Acc = 10.5</td><td>1</td><td>Douglas</td><td>0.5</td><td>0.9</td><td>Leopoldo</td><td>-2.5</td><td>-0.8</td></tr><tr><td>2</td><td>Juliann</td><td>0.4</td><td>0.2</td><td>Youssef</td><td>-1.9</td><td>-0.3</td></tr><tr><td>3</td><td>Kathleen</td><td>0.3</td><td>0.4</td><td>Epifanioo</td><td>-1.7</td><td>0.5</td></tr><tr><td>4</td><td>Ernst</td><td>0.3</td><td>0.4</td><td>Sanjeev</td><td>-1.6</td><td>-0.8</td></tr><tr><td>5</td><td>Tonja</td><td>0.3</td><td>0.0</td><td>Zoltan</td><td>-1.5</td><td>-0.3</td></tr><tr><td>6</td><td>Angela</td><td>0.2</td><td>0.2</td><td>Jianwei</td><td>-1.4</td><td>-0.4</td></tr><tr><td>7</td><td>Hina</td><td>0.2</td><td>-0.3</td><td>Abelardo</td><td>-1.3</td><td>-0.7</td></tr><tr><td>8</td><td>Diego</td><td>0.2</td><td>0.4</td><td>Alphonse</td><td>-1.3</td><td>-0.7</td></tr><tr><td>9</td><td>Bradley</td><td>0.1</td><td>0.9</td><td>Dagoberto</td><td>-1.3</td><td>-0.5</td></tr><tr><td>10</td><td>Jonathan</td><td>0.1</td><td>0.8</td><td>Nathaniel</td><td>-1.2</td><td>-0.1</td></tr><tr><td colspan="4">Majority@10</td><td>0.8</td><td></td><td></td><td></td></tr><tr><td rowspan="10">CLUTRR valid baseline Acc = 59.0 test baseline Acc = 53.0</td><td>1</td><td>Ashlee</td><td>4.4</td><td>1.2</td><td>Rob</td><td>-8.2</td><td>0.6</td></tr><tr><td>2</td><td>Penelope</td><td>3.4</td><td>3.4</td><td>Ahmad</td><td>-8.0</td><td>-0.2</td></tr><tr><td>3</td><td>Allyson</td><td>3.2</td><td>1.0</td><td>Mihai</td><td>-7.6</td><td>0.6</td></tr><tr><td>4</td><td>Chryystal</td><td>2.8</td><td>2.6</td><td>Frederick</td><td>-7.6</td><td>-0.4</td></tr><tr><td>5</td><td>Keiko</td><td>2.8</td><td>-4.4</td><td>Martin</td><td>-7.4</td><td>0.6</td></tr><tr><td>6</td><td>Alisha</td><td>2.6</td><td>3.0</td><td>Jake</td><td>-7.4</td><td>-0.4</td></tr><tr><td>7</td><td>Rosalba</td><td>2.4</td><td>1.8</td><td>Edward</td><td>-7.4</td><td>0.2</td></tr><tr><td>8</td><td>Jacqueline</td><td>2.2</td><td>2.0</td><td>Erik</td><td>-7.4</td><td>0.2</td></tr><tr><td>9</td><td>Zoltan</td><td>2.2</td><td>1.2</td><td>Giovanni</td><td>-7.4</td><td>-0.4</td></tr><tr><td>10</td><td>Dillon</td><td>2.0</td><td>1.4</td><td>Douglas</td><td>-7.2</td><td>-0.2</td></tr><tr><td colspan="4">Majority@10</td><td>1.6</td><td></td><td></td><td></td></tr></table>
+
+Table 2: Performance of the best and worst performing entities in the validation and test datasets. The green color signifies consistent performance between the validation and test datasets, while the orange color indicates inconsistency. The entities, selected based on their performance in the validation dataset, show consistent behavior in the test dataset, suggesting that they generalize in each dataset. We have noted less than 10 entities in some columns because the remaining entities did not reduce the accuracy compared to the baseline.
+
+Contrary to our hypothesis and expectations, there appears to be no correlation between the frequency of names and the performance of LLMs in entity replacement.
+
+# B.4 Effect of entity token length
+
+The token length of a prompt can directly influence the performance of a QA task, particularly as it relates to the computational demands on the model. A longer prompt, represented by a larger number of tokens, increases the amount of data the model must process, potentially straining the model's capabilities and affecting the overall efficiency and accuracy of the task. Changes in the token length of a replaced name, which in turn modify the prompt's overall token length due to multiple instances of the name within the prompt, provide an interesting variable for analysis. We delved into this potential effect by examining the model's performance in relation to the token length of names, as illustrated in Fig 10. Contrary to expectations, our study found no observable correlation between the model's performance and the token length of names. This suggests that, within the ranges examined, the token length of names does not significantly affect the model's performance in QA tasks.
+
+# B.5 Perplexity Analysis
+
+Our expectation was to observe a strong inverse correlation between the perplexity and accuracy of the model across all datasets, according to the findings reported in (Gonen et al., 2022a). However, we noted this inverse correlation only in some datasets, such as the bAbI task 14 (as shown in Figure Fig 11), but not in all of them. A plausible explanation for this inconsistency could be significant differences in our methodology compared to the approach adopted in (Gonen et al., 2022a). In our study, we limited our alterations to substituting named entities within the prompts, as opposed to changing the entire prompt. Therefore, this discrepancy in data augmentation techniques could potentially account for the lack of a significant correlation in some datasets.
+
+Fig 12 shows the deviation of ppl from the baseline across various tasks. Regardless of the task, a more reduction in prompt ppl was observed when entities were replaced both in demonstrations and queries, compared to the other two data augmentation settings. This can likely be attributed to the repeated usage of the same replaced named entity within the prompt. The frequent repetition
+
+of the same named entity can contribute to lower ppl due to the decreased uncertainty and increased predictability of the next word in the sequence.
+
+![](images/58e32b9a1d8be3aca18b8d21e84a5e900f70751678a1380cd248c6f16e43b277.jpg)
+
+![](images/0cc51f357f2b90ed53ae8166969af204f584d8fafa2185109bbc91b4a56ce1de.jpg)
+
+![](images/37ac92f5d8397bfd8fa8e32f952baaca36d88b46979b43e1f016f85873041a4a.jpg)
+
+![](images/1d19d6b25aead79dfcb0ff6c926cc9fd0417248d8d82c84e24c3aea76fb21d11.jpg)
+
+![](images/e25d5ca2003cba21a913c6ef1a965a09b79d7e307a69ec6dc759e92aa9bc3f36.jpg)  
+Figure 9: The performance of Flan-T5-XL in relation to the frequency of names derived from a demographic dataset. Contrary to expectations, no correlation appears to exist between the frequency of names and the performance of Flan-T5-XL.
+
+![](images/6a5d7a6d2b38215a0d6e2ad3e8f5a130ea23aae41f6b5fa36ac2822f535aa51a.jpg)
+
+![](images/eb4910993580662683da331b1c3466154c0a8e1cdebb80164a4dc0471c38e127.jpg)
+
+![](images/c7f334e2cef818d5180ef97130361a86b3d660760b22cb7e08f9a0d0fb42075f.jpg)
+
+![](images/ee690188a4278ea50a5461c0a43e0aaf3f0ff4d7fff300a5ba0fa3646094d636.jpg)
+
+![](images/45a432677bb3ac51bada3ed1361b45190d4ef9e46315b0587fa17bf3fff65c92.jpg)
+
+![](images/e631edc615afeea6f5496310d1725428481ff0d60956520a0768f6f8d03ae95f.jpg)  
+Figure 10: The performance of various names in relation to their token length when using Flan-T5-XL. There is no apparent correlation between the performance and the token length in names. The quantity of names corresponding to each group of token lengths is illustrated on the plot.
+
+![](images/5a8ab99e5b23ee2e077f59376c0bde6c68b9ac10248c552ba1a7fba4f3196ec0.jpg)
+
+![](images/55ef7e12e623edcf93b4f8edc5372fd7fc5edd4d38bf0d9b0f3204b7b0536358.jpg)  
+Figure 11: Correlation between prompt perplexity (ppl) and LLM performance for task 14 and GPT-J-6B across various data augmentation settings. The accuracy and perplexity of prompt for the baseline are shown by red lines. Although a statistically significant inverse correlation between accuracy and ppl is observed in this dataset, this trend is not consistent across all datasets. The absence of this trend, contrary to expectation, could be attributed to the nature of our data augmentation approach, which only involved entity replacement instead of entire prompt modification.
+
+![](images/308344b9ef32d4e7ad0cee1a5995f4c16dc916fdad0a39aaf452168c7b965240.jpg)
+
+![](images/ab2d347feb646d3499a5a59a68a7b05e460ec4ac40d0a38f083b45d11ed2fbde.jpg)
+
+![](images/f5c65aec909b08d9053bcfa5d800e94e9356bb5254188cd6537a95db895722bf.jpg)
+
+![](images/350b9857cc3b55f107480b196ec3f22045615f87988c8087faec994472bf696b.jpg)
+
+![](images/df0cfa71990b174c091e6d0f0b0f68a287f359214d474e3a5b816523961a6f4e.jpg)  
+Figure 12: Deviation of prompt ppl for GPT-J-6B from the baseline across different tasks. A greater reduction in prompt ppl is observed when entities are replaced in both demonstrations and queries, a trend likely driven by the recurrent usage of the same replaced named entity within the prompt.
+
+![](images/b762292e2ec4bcf50af5382be903b24fe5be2077e24f0663b2a059c110810f73.jpg)

2023/Robustness of Named-Entity Replacements for In-Context Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62769932b79c8cebcbe58ff0a8823473fb4843a7d3d2367060cf39e16cfa7fd8
+size 1140430

2023/Role of Context in Unsupervised Sentence Representation Learning_ the Case of Dialog Act Modeling/7f70510d-2f4b-4269-8c9d-3b663839a931_content_list.json

@@ -0,0 +1,1699 @@
+[
+    {
+        "type": "text",
+        "text": "Role of Context in Unsupervised Sentence Representation Learning: the Case of Dialog Act Tagging",
+        "text_level": 1,
+        "bbox": [
+            127,
+            87,
+            870,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Rastislav Hronsky",
+        "bbox": [
+            253,
+            143,
+            413,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jheronimus Academy of Data Science",
+        "bbox": [
+            176,
+            160,
+            489,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sint Janssingel 92, 5211 DA 's-Hertogenbosch",
+        "bbox": [
+            144,
+            175,
+            522,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Netherlands",
+        "bbox": [
+            282,
+            193,
+            383,
+            206
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "r.hronsky@tue.nl",
+        "bbox": [
+            252,
+            210,
+            415,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Emmanuel Keuleers",
+        "bbox": [
+            576,
+            143,
+            754,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Tilburg University",
+        "bbox": [
+            589,
+            160,
+            739,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Warandelaan 2, 5037 AB Tilburg",
+        "bbox": [
+            531,
+            177,
+            801,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Netherlands",
+        "bbox": [
+            615,
+            193,
+            715,
+            206
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "E.A.Keuleers@tilburguniversity.edu",
+        "bbox": [
+            492,
+            210,
+            836,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised learning of word representations involves capturing the contextual information surrounding word occurrences, which can be grounded in the observation that word form is largely disconnected from word meaning. While there are fewer reasons to believe that the same holds for sentences, learning through context has been carried over to learning representations of word sequences. However, this work pays minimal to no attention to the role of context in inferring sentence representations. In this article, we present a dialog act tag probing task designed to explicitly compare content- and context-oriented sentence representations inferred on utterances of telephone conversations (SwDA). Our results suggest that there is no clear benefit of context-based sentence representations over content-based sentence representations. However, there is a very clear benefit of increasing the dimensionality of the sentence vectors in nearly all approaches.",
+        "bbox": [
+            141,
+            278,
+            460,
+            577
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            588,
+            258,
+            602
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised methods for constructing word representations vary widely in their implementation (Landauer and Dumais, 1997; Mikolov et al., 2013a; Pennington et al., 2014), but share the core property that they try to capture the contextual information surrounding word occurrences. This can be grounded in the observation that word form is largely disconnected from word meaning (De Saussure, 1916) and that, therefore, trying to extract meaning from form directly would be pointless.",
+        "bbox": [
+            112,
+            613,
+            489,
+            772
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, there are fewer reasons to believe that representing the meaning of a fully composed sentence requires the description of its context in the way that a word does. Still, much effort has been devoted to developing and evaluating context-oriented methods for sentence representation (Kiros et al., 2015; Logeswaran and Lee, 2018; Cer et al., 2018). Interestingly, the theoretical justification for the approach –roughly paraphrased– is",
+        "bbox": [
+            112,
+            774,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "that it works well for words (Kiros et al., 2015).",
+        "bbox": [
+            507,
+            253,
+            865,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In the present article, we examine whether a contextual signal is of any significant benefit to general-purpose sentence representation.",
+        "bbox": [
+            507,
+            269,
+            880,
+            318
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To accommodate this goal, we infer two sets of utterance representations from telephone conversation transcripts, one by a model trained to encode the transcripts' content, and one by a model trained to encode the transcripts context in a Skip-Thought-like fashion. We evaluate the learned representations with a linear probe classifying the dialog act tags. Below, we discuss these choices more extensively.",
+        "bbox": [
+            507,
+            319,
+            884,
+            463
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The reason for the linear probing technique is to avoid any further representation learning in the downstream task training, such that the resulting accuracy numbers solely reflect the unsupervised representations. This technique is commonly used to assess to what extent sentence representations exhibit certain linguistic properties, such as part-of-speech or syntax (Conneau et al., 2018; Chrupała and Alishahi, 2019; Perone et al., 2018; Belinkov, 2022a).",
+        "bbox": [
+            507,
+            467,
+            882,
+            626
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We chose the Switchboard Dialog Act Corpus (SwDA) (Jurafsky, 1997) corpus for the following properties: (1) its size allows for inferring representations with both traditional, statistical learning methods and modern neural network based methods; (2) at the same time, the corpus is fairly narrow-domain, enabling small models to converge to meaningful representations<sup>1</sup>; (3) it contains utterance labels that we consider suitable for a probing task with respect to contextuality: while a dialog act tag unquestionably depends on the utterance contents, the surrounding utterances likely provide a good signal about its label too. See example utterances from the corpus in Table 1.",
+        "bbox": [
+            507,
+            629,
+            882,
+            854
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>As opposed to some other well-known sentence-level tasks, such as NLI (Bowman et al., 2015), SICK (Marelli et al., 2014), SST-2 (Socher et al., 2013), that require \"universal\" sentence representations.",
+        "bbox": [
+            507,
+            869,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8784",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8784-8792",
+        "bbox": [
+            216,
+            945,
+            779,
+            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": "table",
+        "img_path": "images/189a4430dcd0169a034f6f3dda776a9304f7ed13e516d4f2b82cfc1788c69a66.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>SD^E</td><td>But I, my grandparents were looking into it before.</td></tr><tr><td>SD</td><td>So I know what they’ve said.</td></tr><tr><td>B</td><td>Uh-huh.</td></tr><tr><td>^H</td><td>Well, I’m trying to think.</td></tr></table>",
+        "bbox": [
+            152,
+            80,
+            448,
+            166
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Example dialog act tags and utterances from the SwDA corpus. (SD^E = Statement expanding y/n answer, SD = Statement-non-opinion, B = Backchannel, ^H = Hold before answer/agreement).",
+        "bbox": [
+            112,
+            177,
+            489,
+            233
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Using context to enhance dialog act predictions is not a novel idea (Raheja and Tetreault, 2019; Li and Wu, 2019; Kalchbrenner and Blunsom, 2013). To ascertain the usefulness of context in this task, we checked for the reduction in entropy of a tag in a \"unigram\" setting vs. conditioned on a previous tag, a \"bigram\" setting. The results were  $H(\\text{currenttag}) = 2.1$  bit and  $H(\\text{currenttag|previousat}) = 1.86$  bit, meaning that such a context-enhanced baseline would, at least, increase the chance-level accuracy.",
+        "bbox": [
+            112,
+            261,
+            489,
+            438
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For the content-oriented representation variants, the representations of utterances were inferred via classic topic modeling methods, word vector averaging, or an LSTM-based auto-encoder (see Section 3 for details). The resulting sentence vectors were then used as input for a Skip-Thought-like training regime, where the goal was to learn their context-oriented version by projecting them such that they are predictive of surrounding utterances. Additionally, we trained a version where the utterance encoding and projection are trained jointly, with an LSTM-based encoder of identical proportions with the one used for auto-encoding.",
+        "bbox": [
+            112,
+            439,
+            489,
+            650
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            664,
+            268,
+            680
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In the literature review, we summarize some of the most important work in sentence representation and relate it to the issues outlined in previous paragraphs.",
+        "bbox": [
+            112,
+            692,
+            487,
+            756
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Before the deep learning era of NLP, text-processing methods relied mainly on lexical features of sentences and were invariant to word order. Using word frequency transformations, e.g., tfidf (Ramos et al., 2003), has proven effective to base the representations on more semantically laden words. Techniques such as singular value decomposition (SVD) (Landauer and Dumais, 1997; Landauer et al., 1998), or topic modeling (Blei et al., 2003; Blei, 2012; Hofmann, 2001) were used",
+        "bbox": [
+            112,
+            758,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "to transform the sparse word-count matrices into dense vectors.",
+        "bbox": [
+            507,
+            84,
+            880,
+            115
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "With the subsequent success of simple neural networks such as word2vec (Mikolov et al., 2013a,b; Goldberg and Levy, 2014) or GloVe (Pennington et al., 2014) at representing words and phrases, a seemingly naive method of averaging these word vectors was shown to produce useful sentence/paragraph/document representations (Arora et al., 2017).",
+        "bbox": [
+            507,
+            120,
+            884,
+            249
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "With further success of neural networks in text processing, the research on the topic grew markedly. Among the unsupervised methods, recursive auto-encoders (RAEs) (Socher et al., 2011a,b; Pollack, 1990) were trained to reconstruct sentences via recursively encoding nodes in a dependency parse tree and performed well at a paraphrasing task (Dolan et al., 2004). Tree-LSTMs (Tai et al., 2015) later succeeded the RAEs in the same recursive set up but using an LSTM encoder (Hochreiter and Schmidhuber, 1997), and evaluated on predicting semantic relatedness between two sentences (Marelli et al., 2014) and classifying sentiment (Socher et al., 2013). The current state-of-the-art unsupervised, general-purpose sentence representation systems are based on pre-trained Transformer language models (Vaswani et al., 2017) and fine-tuned via contrastive learning (Gutmann and Hyvarinen, 2012; Chen et al., 2020), i.e., pulling representations of alternative sentence formulations closer together and those of unrelated, random sentence pairs apart (Gao et al., 2021; Chuang et al., 2022).",
+        "bbox": [
+            507,
+            255,
+            884,
+            624
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The success of skip-gram-based word representations inspired researchers to transfer the idea to sentence representation. Skip-Thought (Kiros et al., 2015) is a large, LSTM-based (Hochreiter and Schmidhuber, 1997), sentence encoder that was trained to produce representations predictive of surrounding sentences. Quick-Thought (Logeswaran and Lee, 2018) is a similar system that differs in that it casts the problem as contrastive learning (Gutmann and Hyvarinen, 2012). The Universal Sentence Encoder (USE) is a Transformer-based (Vaswani et al., 2017) system introduced by Cer et al. (2018) that uses the Skip-Thought as one of the tasks in a multi-task learning set-up (including some supervised tasks). The task of deciding whether two sentences are consecutive or not (NSP task) was also used in training of BERT (Devlin et al., 2019). The latest dialog processing sys",
+        "bbox": [
+            507,
+            629,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8785",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tems, built on top of Transformer language models (Vaswani et al., 2017), typically learn to predict contextual utterances within the contrastive learning framework and achieve superior performance on dialog tasks compared to universal representations (Wu et al., 2020; Zhou et al., 2022; Zeng et al., 2023; Oh et al., 2023).",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "While sentence contextuality is used in all of the above-mentioned studies, there is little to no focus on assessing the necessity for using context to derive sentence representations. An ablation of the NSP task is presented by Devlin et al. (2019), suggesting a slight performance benefit over a version where the task is omitted. However, Devlin et al. (2019) only evaluate the fine-tuning pipeline, while the raw BERT representations were shown to perform poorly overall (Reimers and Gurevych, 2019). Further, results by Liu et al. (2019) are contradictory and do not credit a performance improvement to the NSP task. Lastly, Zhou et al. (2022) additionally compare the representations resulting from dropout augmentations (Gao et al., 2021) to those of consecutive utterances and found a significant benefit in favor of the latter augmentations. Their results are perplexing, however, since the dropout version they implemented performed even worse than that of SimCSE (Gao et al., 2021), which also uses dropout augmentation, and should in fact perform worse, not being tailored for dialog.",
+        "bbox": [
+            115,
+            199,
+            489,
+            551
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Using the natural language inference (NLI) task has been effective in several general sentence representation learning systems (Conneau et al., 2017; Reimers and Gurevych, 2019; Zhou et al., 2022). However, along with machine translation models, these systems fall outside the category of unsupervised sentence representations.",
+        "bbox": [
+            112,
+            554,
+            489,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            682,
+            218,
+            697
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We trained several variants of content-based representations, their context-based counterparts, and an end-to-end, Skip-Thought like system, trained to encode and predict sentences.",
+        "bbox": [
+            112,
+            712,
+            487,
+            775
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We evaluated the following variants of content-based sentence representation<sup>2</sup>:",
+        "bbox": [
+            112,
+            778,
+            489,
+            810
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Latent Semantic Analysis (LSA)  $\\star$",
+            "2. Latent Dirichlet Allocation (LDA)\\* (Blei et al., 2003)"
+        ],
+        "bbox": [
+            127,
+            826,
+            485,
+            889
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "3. Mean of GloVe word vectors  $\\star$  (Pennington et al., 2014)",
+            "4. LSTM-based auto-encoder (Hochreiter and Schmidhuber, 1997)",
+            "5. Mean of arbitrarily assigned word vectors"
+        ],
+        "bbox": [
+            522,
+            84,
+            880,
+            187
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We also varied the final dimension of the sentence vector, namely  $D \\in \\{6, 12, 25\\}$ , either by setting the number_of_topics parameter, setting the auto-encoder bottleneck size, or truncating GloVe vectors. The auto-encoder was based on a 2-layer, bidirectional Long Short-Term Memory (LSTM) neural network with hidden size of  $D_{hidden} = 64$ , dropout of  $p = 0.1$ . The variant 5 is one where we randomly assigned a feature vector from an 8-dimensional raster to each vocabulary entry (a procedure by which we want to emulate the usage of 1-hot vector encoding, but with dense, real-valued vectors), and created the final sentence representation by averaging corresponding word vectors.",
+        "bbox": [
+            507,
+            200,
+            884,
+            426
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To create the context-based representations, we devised a learning regimen for 2 linear layers for each variant. One layer was used to project the sentence content vector to the context space (encode), and the other layer was used to predict the content vectors of surrounding sentences from the context vector (decode).",
+        "bbox": [
+            507,
+            426,
+            882,
+            537
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Finally, we trained a Skip-Thought-like model, where each sentence was represented by a sequence of word vectors (obtained via the same procedure as in variant 5), optionally truncated and padded to 20 tokens, processed via an LSTM (identical with the auto-encoder in variant 4), and projected such that, conditioned on the resulting vector, a decoder-LSTM reconstructs the surrounding sentences well via teacher forcing.",
+        "bbox": [
+            507,
+            539,
+            882,
+            684
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In all of the context-based variants, we varied the context size between 1 and 2 sentences, on each side.",
+        "bbox": [
+            507,
+            684,
+            880,
+            731
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Dataset and Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            745,
+            741,
+            759
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To infer and evaluate the representations, we used the Switchboard Dialog Act Corpus (SwDA) (Jurafsky, 1997), with exception of the GloVe vectors, where we used the pre-trained glove-twitter-25 vectors. The SwDA contains transcripts of casual telephone conversations in which the two parties were instructed to discuss a certain topic, e.g. \"care of the elderly\". Every utterance is assigned a label",
+        "bbox": [
+            507,
+            766,
+            884,
+            894
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "Entries marked with  $\\star$  use the GenSim implementation.",
+        "bbox": [
+            134,
+            903,
+            480,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "3We used the pre-trained glove-twitter-25 version.",
+        "bbox": [
+            529,
+            904,
+            835,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8786",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/4ded5239487167f661787aa1277fe7588a4a13ad65d90ac36d36c036b1cd23b8.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 2: Summary of the SwDA dataset. In the upper section of the table, we summarize the corpus statistics. In the lower section we list the percentages of dialog act labels per corpus split (SD=\"Statement-non-opinion\", B=\"Backchannel\", SV=\"Statement-opinion\", AA=\"Agree/Accept\")."
+        ],
+        "table_body": "<table><tr><td></td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td>Utterances</td><td>212,145</td><td>26,518</td><td>26,519</td></tr><tr><td>Tokens</td><td>1,833M</td><td>227K</td><td>227K</td></tr><tr><td>Tokens per att.</td><td>8.64</td><td>8.56</td><td>8.32</td></tr><tr><td>SD</td><td>35.47%</td><td>33.52%</td><td>35.00%</td></tr><tr><td>%/Other</td><td>27.45%</td><td>25.77%</td><td>25.60%</td></tr><tr><td>B</td><td>18.52%</td><td>21.54%</td><td>23.02%</td></tr><tr><td>SV</td><td>13.76%</td><td>14.60%</td><td>13.73%</td></tr><tr><td>AA</td><td>4.81%</td><td>4.57%</td><td>2.65%</td></tr></table>",
+        "bbox": [
+            115,
+            82,
+            480,
+            230
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "describing the \"role\" of the act, in a similar way that part-of-speech tags describe roles of words in sentences 4.",
+        "bbox": [
+            110,
+            349,
+            487,
+            397
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To pre-process the texts, (1) we tokenized the utterances with the reference Morfessor 2.0 (Virpioja et al., 2013) implementation, (2) long utterances were broken down into multiple shorter utterances with a sliding-window of 20 words and a step-size of 18 words, and (3) we split the corpus into a train, dev, and test split with respective proportions of  $80\\%$ ,  $10\\%$ , and  $10\\%$  of the utterances. For an overview of the final corpus statistics, including the dialog act tags, see Table 2.",
+        "bbox": [
+            110,
+            398,
+            487,
+            558
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each representation type, we trained a linear ridge-regression $^{5}$ -based classifier with 10-fold cross-validation (on the training set) to predict the dialog act labels. Altogether, there were 42 distinct dialog acts. To simplify the probe, we only selected the 10 most frequent labels to classify, while grouping the remaining labels under a single category.",
+        "bbox": [
+            110,
+            560,
+            489,
+            671
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            683,
+            213,
+            697
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The classification accuracy results are presented in Table 3. We established baseline performance as the frequency of the majority class. The only variant performing below the majority baseline was the Random variant ( $D_{hidden} = 6$ ), which used means of arbitrarily assigned word vectors. The other variants achieved scores in the range of up to  $15 - 21\\%$  above the majority baseline. Except",
+        "bbox": [
+            110,
+            708,
+            487,
+            837
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/8a17389fd4e21eaa37b2d880780239217e537085a1e11c87bbfc31ec1d281a18.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 3: Results. The three right-most columns contain the accuracy percentages for the content-only, context of distance 1, and context of distance 2 variants."
+        ],
+        "table_body": "<table><tr><td></td><td></td><td>Context</td><td>1 att.</td><td>2 att.</td></tr><tr><td>Algo.</td><td>Dim.</td><td colspan=\"3\">Accuracy [%]</td></tr><tr><td>Maj. class</td><td></td><td>35.00</td><td></td><td></td></tr><tr><td rowspan=\"3\">GLOVE</td><td>6</td><td>51.62</td><td>50.37</td><td>50.14</td></tr><tr><td>12</td><td>52.82</td><td>52.60</td><td>52.55</td></tr><tr><td>25</td><td>57.29</td><td>57.24</td><td>56.52</td></tr><tr><td rowspan=\"3\">LDA</td><td>6</td><td>51.10</td><td>51.16</td><td>51.05</td></tr><tr><td>12</td><td>53.32</td><td>52.99</td><td>52.73</td></tr><tr><td>25</td><td>56.82</td><td>56.36</td><td>56.13</td></tr><tr><td rowspan=\"3\">LSA</td><td>6</td><td>51.07</td><td>50.62</td><td>50.58</td></tr><tr><td>12</td><td>51.94</td><td>51.44</td><td>51.04</td></tr><tr><td>25</td><td>52.02</td><td>52.11</td><td>52.12</td></tr><tr><td rowspan=\"3\">Random</td><td>6</td><td>36.52</td><td>41.55</td><td>41.83</td></tr><tr><td>12</td><td>46.99</td><td>47.17</td><td>46.47</td></tr><tr><td>25</td><td>48.46</td><td>47.69</td><td>47.59</td></tr><tr><td rowspan=\"3\">Auto-enc.</td><td>6</td><td>53.88</td><td>52.27</td><td>51.63</td></tr><tr><td>12</td><td>54.46</td><td>53.89</td><td>52.34</td></tr><tr><td>25</td><td>52.43</td><td>53.18</td><td>52.57</td></tr><tr><td rowspan=\"3\">Skip-Th.</td><td>6</td><td></td><td>51.13</td><td>39.66</td></tr><tr><td>12</td><td></td><td>53.33</td><td>52.38</td></tr><tr><td>25</td><td></td><td>55.21</td><td>57.54</td></tr></table>",
+        "bbox": [
+            509,
+            83,
+            870,
+            442
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "for LSA, each algorithm achieved the highest score with  $D_{hidden} = 25$ .",
+        "bbox": [
+            507,
+            521,
+            882,
+            552
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "There were no big differences and no discerning patterns in the performance of the contextually projected variants.",
+        "bbox": [
+            507,
+            555,
+            882,
+            602
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The only case where the training regime of predicting two neighboring sentences on each side scored the best was with the end-to-end, SkipThought like variant, which also achieved the overall highest score.",
+        "bbox": [
+            507,
+            604,
+            884,
+            684
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Discussion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            699,
+            636,
+            713
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "All tested models performed well over the majority baseline, including the averaged random word vectors which exceeded the majority baseline by a large margin when coupled with high dimensionality. One reason for this might be that, when averaging word vectors of higher dimension, there is a higher chance that some important lexical features will remain separable from the noise (also applicable to GloVe). Only in case of LSA and AutoEncoder, the effect of dimensionality was small. For the latter, using a variational or contrastive learning set-up might allow for better scaling of",
+        "bbox": [
+            505,
+            726,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "4See the Coder's manual (Jurafsky, 1997) for more information.",
+        "bbox": [
+            112,
+            844,
+            487,
+            868
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>5</sup>We opted for ridge regression, instead of logistic regression, because of high colinearity (except for LDA and LSA) across feature vectors. See correlation matrices in the appendix Section A. We used the Scikit-Learn implementation.",
+        "bbox": [
+            112,
+            869,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8787",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "performance with capacity. Based on the patterns observed here, increasing the dimensionality increases performance on the probe, with or without context.",
+        "bbox": [
+            112,
+            84,
+            487,
+            146
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We did not find either of these aspects discussed in the mainstream related work (Conneau et al., 2018; Belinkov, 2022b). The question of how far increasing the dimensionality helps performance should be explored in future work.",
+        "bbox": [
+            112,
+            149,
+            487,
+            228
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The presented results on the dialog act tag probing task are in line with related literature in that sequence/document representations with word embeddings and classical topic modeling approaches are a tough-to-beat baseline for many neural networks (Conneau et al., 2018; Reimers and Gurevych, 2019). Our results suggest that transforming existing, content-oriented representations to context-oriented ones provides minimal to no meaningful improvement in representation. An improvement of small, but perhaps meaningful magnitude is one from the best performing AutoEncoder to Skip-Thought variant  $(3\\%)$ , suggesting that some benefit in this training scheme may be found for sentences, too.",
+        "bbox": [
+            115,
+            231,
+            489,
+            470
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "However, our experimental design limits us in drawing conclusions about what the Skip-Thought set-up benefited from specifically: the more general sentence representations may simply be an artifact of the more difficult learning task (predicting neighboring sentences), as opposed to being a result of form disambiguation via context, the way it works with words. Future work should employ methodology that separates the contribution of the competing sources of signal more clearly.",
+        "bbox": [
+            112,
+            473,
+            487,
+            633
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "When compared to the obvious benefits of context on learning word representations, it is striking that the benefits for learning sentence representations appear to be very small at best.",
+        "bbox": [
+            112,
+            634,
+            487,
+            697
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This observation is suggestive of a more serious blind-spot in NLP: because context-based representations work well for words, many researchers assume that the same holds for any level of analysis. This is akin to claiming arbitrariness of the sign at any level of language processing.",
+        "bbox": [
+            112,
+            700,
+            487,
+            796
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            810,
+            245,
+            826
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We presented evidence that explicitly examines the differences between content-, and context-oriented unsupervised sentence representations. While results suggest that dimensionality, rather than context, is the determining factor in building good",
+        "bbox": [
+            112,
+            839,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "sentence representations, end-to-end approaches based on reconstructing surrounding sentences are a promising way forward.",
+        "bbox": [
+            507,
+            84,
+            880,
+            131
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            143,
+            613,
+            159
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Since this work only presents evidence from one sentence task, the conclusions drawn are task-specific. Although we believe the chosen task to be fairly general, one can easily think of tasks that would likely challenge the drawn conclusions, e.g., recognition of humor or sarcasm. However, these are rather rare cases and it is questionable, where the best position of the boundary between unsupervised and task-specific lies. Additionally, the selected range of dimensions is limited to a small amount, and at larger scales the patterns of results might change.",
+        "bbox": [
+            507,
+            168,
+            882,
+            363
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8788",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Sanjeev Arora, Yingyu Liang, and Tengyu Ma. 2017. A simple but tough-to-beat baseline for sentence embeddings. In International conference on learning representations.",
+            "Yonatan Belinkov. 2022a. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219.",
+            "Yonatan Belinkov. 2022b. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219.",
+            "David M Blei. 2012. Probabilistic topic models. Communications of the ACM, 55(4):77-84.",
+            "David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993-1022.",
+            "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics.",
+            "Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, et al. 2018. Universal sentence encoder for english. In Proceedings of the 2018 conference on empirical methods in natural language processing: system demonstrations, pages 169-174.",
+            "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR.",
+            "Grzegorz Chrupa and Afra Alishahi. 2019. Correlating neural and symbolic representations of language. arXiv preprint arXiv:1905.06401.",
+            "Yung-Sung Chuang, Rumen Dangovski, Hongyin Luo, Yang Zhang, Shiyu Chang, Marin Soljacic, Shang-Wen Li, Scott Yih, Yoon Kim, and James Glass. 2022. DiffCSE: Difference-based contrastive learning for sentence embeddings. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4207-4218, Seattle, United States. Association for Computational Linguistics.",
+            "Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. arXiv preprint arXiv:1705.02364.",
+            "Alexis Conneau, German Kruszewski, Guillaume Lample, Loic Barrault, and Marco Baroni. 2018. What"
+        ],
+        "bbox": [
+            115,
+            107,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "you can cram into a single vector: Probing sentence embeddings for linguistic properties. arXiv preprint arXiv:1805.01070.",
+            "Ferdinand De Saussure. 1916. Nature of the linguistic sign. Course in general linguistics, 1:65-70.",
+            "Jacob Devlin, Ming-Wei Chang, and Kenton Lee. 2019. Google, kt, language, ai: Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186.",
+            "William Dolan, Chris Quirk, Chris Brockett, and Bill Dolan. 2004. Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources.",
+            "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple contrastive learning of sentence embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6894-6910, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Yoav Goldberg and Omer Levy. 2014. word2vec explained: deriving mikolov et al.'s negative-sampling word-embedding method. arXiv preprint arXiv:1402.3722.",
+            "Michael U Gutmann and Aapo Hyvarinen. 2012. Noisecontrastive estimation of unnormalized statistical models, with applications to natural image statistics. Journal of machine learning research, 13(2).",
+            "Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural computation*, 9(8):1735-1780.",
+            "Thomas Hofmann. 2001. Unsupervised learning by probabilistic latent semantic analysis. Machine learning, 42(1-2):177.",
+            "Dan Jurafsky. 1997. Switchboard swbd-damsl shallow-discourse-function annotation coders manual. Technical report.",
+            "Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent convolutional neural networks for discourse compositionality. arXiv preprint arXiv:1306.3584.",
+            "Ryan Kiros, Yukun Zhu, Russ R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Skip-thought vectors. Advances in neural information processing systems, 28.",
+            "Thomas K Landauer and Susan T Dumais. 1997. A solution to Plato's problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological review, 104(2):211.",
+            "Thomas K Landauer, Peter W Foltz, and Darrell Laham. 1998. An introduction to latent semantic analysis. Discourse processes, 25(2-3):259-284."
+        ],
+        "bbox": [
+            510,
+            86,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8789",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Wei Li and Yunfang Wu. 2019. Multi-level gated recurrent neural network for dialog act classification. arXiv preprint arXiv:1910.01822.",
+            "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.",
+            "Lajanugen Logeswaran and Honglak Lee. 2018. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893.",
+            "Marco Marelli, Luisa Bentivogli, Marco Baroni, Raffaella Bernardi, Stefano Menini, and Roberto Zamparelli. 2014. Semeval-2014 task 1: Evaluation of compositional distributional semantic models on full sentences through semantic relatedness and textual entailment. In Proceedings of the 8th international workshop on semantic evaluation (SemEval 2014), pages 1-8.",
+            "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781.",
+            "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26.",
+            "Minsik Oh, Jiwei Li, and Guoyin Wang. 2023. Tadse: Template-aware dialogue sentence embeddings.",
+            "Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543.",
+            "Christian S Perone, Roberto Silveira, and Thomas S Paula. 2018. Evaluation of sentence embeddings in downstream and linguistic probing tasks. arXiv preprint arXiv:1806.06259.",
+            "Jordan B Pollack. 1990. Recursive distributed representations. Artificial Intelligence, 46(1-2):77-105.",
+            "Vipul Raheja and Joel Tetreault. 2019. Dialogue act classification with context-aware self-attention. arXiv preprint arXiv:1904.02594.",
+            "Juan Ramos et al. 2003. Using tfidf to determine word relevance in document queries. In Proceedings of the first instructional conference on machine learning, volume 242, pages 29-48. Citeseer.",
+            "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Richard Socher, Eric Huang, Jeffrey Pennin, Christopher D Manning, and Andrew Ng. 2011a. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. Advances in neural information processing systems, 24.",
+            "Richard Socher, Jeffrey Pennington, Eric H Huang, Andrew Y Ng, and Christopher D Manning. 2011b. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the 2011 conference on empirical methods in natural language processing, pages 151-161.",
+            "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642.",
+            "Kai Sheng Tai, Richard Socher, and Christopher D Manning. 2015. Improved semantic representations from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075.",
+            "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.",
+            "Sami Virpioja, Peter Smit, Stig-Arne Gronroos, Mikko Kurimo, et al. 2013. Morfessor 2.0: Python implementation and extensions for morfessor baseline. Technical report.",
+            "Chien-Sheng Wu, Steven C.H. Hoi, Richard Socher, and Caiming Xiong. 2020. TOD-BERT: Pre-trained natural language understanding for task-oriented dialogue. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 917-929, Online. Association for Computational Linguistics.",
+            "Weihao Zeng, Keqing He, Yejie Wang, Chen Zeng, Jinggang Wang, Yunsen Xian, and Weiran Xu. 2023. Futuretod: Teaching future knowledge to pre-trained language model for task-oriented dialogue.",
+            "Zhihan Zhou, Dejiao Zhang, Wei Xiao, Nicholas Dingwall, Xiaofei Ma, Andrew Arnold, and Bing Xiang. 2022. Learning dialogue representations from consecutive utterances. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 754-768, Seattle, United States. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            822
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8790",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e3bd7dab7f09a78492363b6a1387f06f781a32a9a2d1431da995acd3a7d59e8a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            147,
+            257,
+            248
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b2b7228fad90e69a59d321d64ac43cb69620915ca6ae02157f52ff20191a9668.jpg",
+        "image_caption": [
+            "Pearson correlation between features (dim=6)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            148,
+            413,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f4dc16a9bfd9f146d249cfd0a918b088a34fefa7f600a26f3c8189f1ab1e1e62.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            431,
+            148,
+            549,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b4aa368cbcbdcb09b1eb7f8f9d572459db95e964c62dec886e638d0c5861ca07.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            578,
+            148,
+            697,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a1b7a46a28ae57dbbbe6c0e2e4376fe4dc32594ad8fba9383dfaf5a21b1a4af9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            726,
+            148,
+            843,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b8318ad01d3dc6d703407573d070942519a0b3e2105f51dc1970dbe829d44af9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            250,
+            255,
+            348
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/406714a29b6928cbbd3cf99a5c1cf477acea5e0d3f634d05d710097bcadf6b8c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            250,
+            413,
+            348
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ab49c52879c9e63344ebb0b11d288945b4623c0921773c8bc1e507da73951071.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            431,
+            250,
+            549,
+            348
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/18b2be039e660f5a21f72a97b37fb0712ffcabb9d570f6289638adbe73da9de6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            578,
+            250,
+            695,
+            348
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/bbae9db3c2c61be51038065c6271133afbf794e7d81fb6f6dd2c9407dcc9e875.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            726,
+            250,
+            843,
+            348
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/eaaff27a405d9fe37685a0491a51dae754afaba045402119ea5d7937a27e3471.jpg",
+        "image_caption": [
+            "Figure 1: Correlation matrices of feature vectors across representation variants with dimension of 6."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            351,
+            255,
+            451
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3f510d7fa00621c384e8521f1580fa5bdcf8eafda028390758d19d87d1eb36f8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            351,
+            413,
+            451
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3fb5ed6ae184b0b7cee9a431f65693bd100a4d0695b429e6f3bb0fd8fb051bd3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            431,
+            351,
+            549,
+            451
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9f5cf7fbe539f5feb4e41b80a7565a289b306aa39c619b848002750a27de7930.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            578,
+            351,
+            697,
+            451
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cdef12ac6c191c98ddd2445564c513d1bed3178d3106f1330f56cc272e3c1568.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            351,
+            873,
+            451
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ba962b57c0c5246fc90ef58b76a0c12cca8d16889acbe6c156a5f6a472c20ce7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            552,
+            263,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a366fdfbea52d788a92af08818565a327e98c3df86a050e41e372e00b9c010f6.jpg",
+        "image_caption": [
+            "Pearson correlation between features (dim=12)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            552,
+            418,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/04ac375c04e26358ccd6ea6389ef32d851582187466f15170f15e29747ddb1c6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            430,
+            552,
+            551,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/464693b6622a338abc6c1f8268461dc4b3dc6d49ede288df8fa4cc1885724f2b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            552,
+            700,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f3ba839187d7c35414596f45ae59a132d0142f60358af367f1c0a80a6f8e52eb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            724,
+            552,
+            847,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ebe9284b92ab7aa72dc45aead00eec0f87619fb570715e1ca389850afabbb711.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            653,
+            262,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/72fab356a9886508898e03f65065d7c6843406057da271191f22b7c4db25ed9b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            653,
+            418,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1b8161abc276285198cd56cd2876819edc5098c492e311e7ed0a6ce3071f0cf0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            431,
+            653,
+            551,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fd1b29c5fdd1b56235c96ab54b644854c2e4899736b1437390b8ebb0da8a0739.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            653,
+            700,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7559ab50de6be62847564c41822757c71d9f1024d785f2024d9f2d2b23101e08.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            724,
+            653,
+            847,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/69bdb5a9c1f7be3ac3ecea2e05d8fa810b85df666a6f76d0cabf96354bb29a86.jpg",
+        "image_caption": [
+            "Figure 2: Correlation matrices of feature vectors across representation variants with dimension of 12."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            756,
+            262,
+            854
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1841577c8eea6af3ff3767280aed87b54821d4228cccdf4386de82f0e4736ec4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            756,
+            418,
+            854
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2b0b460584ecf83a5f860bce74ea515a663989ca952fbfeccb6b2a3e242e649f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            430,
+            756,
+            551,
+            854
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1057ade4d2ac1b65a8b24ba0436c30938bab50136818954bad3380fa91e22a9b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            756,
+            700,
+            854
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cd79acc52bec3506d06e3e8e662ff62d9650b35e9e7e0f21552102c3e9b4882f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            756,
+            873,
+            854
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "header",
+        "text": "A Correlation Matrices",
+        "bbox": [
+            114,
+            84,
+            334,
+            98
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8791",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Pearson correlation between features (dim=25)",
+        "text_level": 1,
+        "bbox": [
+            337,
+            322,
+            653,
+            334
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/4150446ab607544a5eb7c0f00ce24ce092873d5b85f8b9b72527d72fe720221a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            346,
+            263,
+            445
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/c8cce625139a6879322e78635b20ef8a0bc1b8f4cbb939704211c3e9582da226.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            449,
+            263,
+            545
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/8930cdf5fe5c77ffa84035a1d0d737b490b30ceb6625ca733e5b5019c9527df0.jpg",
+        "image_caption": [
+            "Figure 3: Correlation matrices of feature vectors across representation variants with dimension of 25."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            548,
+            262,
+            646
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/e47ecb0b45c1943cf4122e45fa8a6cf9b64a9cb4763ce1b7a2a826caeb3b7d5e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            346,
+            420,
+            445
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/cf9627ac9b4e98c9ebee65949b37c162d6ef97cc4f18665daaa5047d90c22ba9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            448,
+            420,
+            545
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/d82564006893445346d5e0d61c487f3ef32857a20dd2f3262f19758aa4ba2d86.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            548,
+            420,
+            646
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/7b20f1378f18117882eeae91bcff18b5198d07eb49123bf3a6e61543d743e9f7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            428,
+            346,
+            554,
+            445
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/b2d98184a46d3e4682108937827856479f67325ff4e3edd3677090a21c527ad5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            428,
+            448,
+            552,
+            545
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/763d9c57c03a73a650ca8d36c638af7a1a82d0472a7d623ae8c9a9f28aa33ff4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            428,
+            548,
+            552,
+            646
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/1cfd29e61a7083a3c9fd13f24a4282967ff1005d386da95ca8944188ed647fc3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            574,
+            346,
+            700,
+            445
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/247cf832c26e0ffc9c8c1c35201d0cdd38890345da7d16fa13e7991d145877fd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            448,
+            700,
+            545
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/39a24e18f9c9208d3b5d828bb77ee788f8111bd5534b1873320dbc4a4adbd332.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            548,
+            700,
+            646
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/028d2dd7a7a963c8fe04ee18e5dbdf982408ca9235cf9a731442d6334b6aeb33.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            722,
+            346,
+            847,
+            445
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/d2f0df741541b95d244eb4b1fc61eec1602867cddbe37ed93fb16e1da9e21f0d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            722,
+            448,
+            847,
+            545
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/21a5f9311fc83cebd8d0f22c68e1bd69a752bfd38b6c4e584aee0d552567b8e7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            548,
+            873,
+            648
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8792",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 8
+    }
+]

2023/Role of Context in Unsupervised Sentence Representation Learning_ the Case of Dialog Act Modeling/7f70510d-2f4b-4269-8c9d-3b663839a931_model.json

@@ -0,0 +1,2198 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.128,
+                0.089,
+                0.872,
+                0.13
+            ],
+            "angle": 0,
+            "content": "Role of Context in Unsupervised Sentence Representation Learning: the Case of Dialog Act Tagging"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.254,
+                0.144,
+                0.414,
+                0.16
+            ],
+            "angle": 0,
+            "content": "Rastislav Hronsky"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.178,
+                0.161,
+                0.49,
+                0.176
+            ],
+            "angle": 0,
+            "content": "Jheronimus Academy of Data Science"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.146,
+                0.177,
+                0.523,
+                0.193
+            ],
+            "angle": 0,
+            "content": "Sint Janssingel 92, 5211 DA 's-Hertogenbosch"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.284,
+                0.194,
+                0.384,
+                0.208
+            ],
+            "angle": 0,
+            "content": "Netherlands"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.253,
+                0.211,
+                0.416,
+                0.226
+            ],
+            "angle": 0,
+            "content": "r.hronsky@tue.nl"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.577,
+                0.144,
+                0.756,
+                0.159
+            ],
+            "angle": 0,
+            "content": "Emmanuel Keuleers"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.591,
+                0.161,
+                0.741,
+                0.177
+            ],
+            "angle": 0,
+            "content": "Tilburg University"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.178,
+                0.802,
+                0.193
+            ],
+            "angle": 0,
+            "content": "Warandelaan 2, 5037 AB Tilburg"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.616,
+                0.194,
+                0.716,
+                0.208
+            ],
+            "angle": 0,
+            "content": "Netherlands"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.494,
+                0.211,
+                0.838,
+                0.226
+            ],
+            "angle": 0,
+            "content": "E.A.Keuleers@tilburguniversity.edu"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.268
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.279,
+                0.461,
+                0.578
+            ],
+            "angle": 0,
+            "content": "Unsupervised learning of word representations involves capturing the contextual information surrounding word occurrences, which can be grounded in the observation that word form is largely disconnected from word meaning. While there are fewer reasons to believe that the same holds for sentences, learning through context has been carried over to learning representations of word sequences. However, this work pays minimal to no attention to the role of context in inferring sentence representations. In this article, we present a dialog act tag probing task designed to explicitly compare content- and context-oriented sentence representations inferred on utterances of telephone conversations (SwDA). Our results suggest that there is no clear benefit of context-based sentence representations over content-based sentence representations. However, there is a very clear benefit of increasing the dimensionality of the sentence vectors in nearly all approaches."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.589,
+                0.26,
+                0.604
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.614,
+                0.49,
+                0.774
+            ],
+            "angle": 0,
+            "content": "Unsupervised methods for constructing word representations vary widely in their implementation (Landauer and Dumais, 1997; Mikolov et al., 2013a; Pennington et al., 2014), but share the core property that they try to capture the contextual information surrounding word occurrences. This can be grounded in the observation that word form is largely disconnected from word meaning (De Saussure, 1916) and that, therefore, trying to extract meaning from form directly would be pointless."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.775,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "However, there are fewer reasons to believe that representing the meaning of a fully composed sentence requires the description of its context in the way that a word does. Still, much effort has been devoted to developing and evaluating context-oriented methods for sentence representation (Kiros et al., 2015; Logeswaran and Lee, 2018; Cer et al., 2018). Interestingly, the theoretical justification for the approach –roughly paraphrased– is"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.254,
+                0.866,
+                0.269
+            ],
+            "angle": 0,
+            "content": "that it works well for words (Kiros et al., 2015)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.271,
+                0.882,
+                0.319
+            ],
+            "angle": 0,
+            "content": "In the present article, we examine whether a contextual signal is of any significant benefit to general-purpose sentence representation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.321,
+                0.885,
+                0.464
+            ],
+            "angle": 0,
+            "content": "To accommodate this goal, we infer two sets of utterance representations from telephone conversation transcripts, one by a model trained to encode the transcripts' content, and one by a model trained to encode the transcripts context in a Skip-Thought-like fashion. We evaluate the learned representations with a linear probe classifying the dialog act tags. Below, we discuss these choices more extensively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.468,
+                0.884,
+                0.627
+            ],
+            "angle": 0,
+            "content": "The reason for the linear probing technique is to avoid any further representation learning in the downstream task training, such that the resulting accuracy numbers solely reflect the unsupervised representations. This technique is commonly used to assess to what extent sentence representations exhibit certain linguistic properties, such as part-of-speech or syntax (Conneau et al., 2018; Chrupała and Alishahi, 2019; Perone et al., 2018; Belinkov, 2022a)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.63,
+                0.884,
+                0.855
+            ],
+            "angle": 0,
+            "content": "We chose the Switchboard Dialog Act Corpus (SwDA) (Jurafsky, 1997) corpus for the following properties: (1) its size allows for inferring representations with both traditional, statistical learning methods and modern neural network based methods; (2) at the same time, the corpus is fairly narrow-domain, enabling small models to converge to meaningful representations<sup>1</sup>; (3) it contains utterance labels that we consider suitable for a probing task with respect to contextuality: while a dialog act tag unquestionably depends on the utterance contents, the surrounding utterances likely provide a good signal about its label too. See example utterances from the corpus in Table 1."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.87,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>As opposed to some other well-known sentence-level tasks, such as NLI (Bowman et al., 2015), SICK (Marelli et al., 2014), SST-2 (Socher et al., 2013), that require \"universal\" sentence representations."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8784"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.78,
+                0.959
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8784-8792"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.278,
+                0.959,
+                0.72,
+                0.973
+            ],
+            "angle": 0,
+            "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.154,
+                0.082,
+                0.449,
+                0.167
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>SD^E</td><td>But I, my grandparents were looking into it before.</td></tr><tr><td>SD</td><td>So I know what they’ve said.</td></tr><tr><td>B</td><td>Uh-huh.</td></tr><tr><td>^H</td><td>Well, I’m trying to think.</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.178,
+                0.49,
+                0.234
+            ],
+            "angle": 0,
+            "content": "Table 1: Example dialog act tags and utterances from the SwDA corpus. (SD^E = Statement expanding y/n answer, SD = Statement-non-opinion, B = Backchannel, ^H = Hold before answer/agreement)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.262,
+                0.49,
+                0.439
+            ],
+            "angle": 0,
+            "content": "Using context to enhance dialog act predictions is not a novel idea (Raheja and Tetreault, 2019; Li and Wu, 2019; Kalchbrenner and Blunsom, 2013). To ascertain the usefulness of context in this task, we checked for the reduction in entropy of a tag in a \"unigram\" setting vs. conditioned on a previous tag, a \"bigram\" setting. The results were \\( H(\\text{currenttag}) = 2.1 \\) bit and \\( H(\\text{currenttag|previousat}) = 1.86 \\) bit, meaning that such a context-enhanced baseline would, at least, increase the chance-level accuracy."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.441,
+                0.49,
+                0.651
+            ],
+            "angle": 0,
+            "content": "For the content-oriented representation variants, the representations of utterances were inferred via classic topic modeling methods, word vector averaging, or an LSTM-based auto-encoder (see Section 3 for details). The resulting sentence vectors were then used as input for a Skip-Thought-like training regime, where the goal was to learn their context-oriented version by projecting them such that they are predictive of surrounding utterances. Additionally, we trained a version where the utterance encoding and projection are trained jointly, with an LSTM-based encoder of identical proportions with the one used for auto-encoding."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.665,
+                0.27,
+                0.681
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.693,
+                0.489,
+                0.757
+            ],
+            "angle": 0,
+            "content": "In the literature review, we summarize some of the most important work in sentence representation and relate it to the issues outlined in previous paragraphs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Before the deep learning era of NLP, text-processing methods relied mainly on lexical features of sentences and were invariant to word order. Using word frequency transformations, e.g., tfidf (Ramos et al., 2003), has proven effective to base the representations on more semantically laden words. Techniques such as singular value decomposition (SVD) (Landauer and Dumais, 1997; Landauer et al., 1998), or topic modeling (Blei et al., 2003; Blei, 2012; Hofmann, 2001) were used"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.882,
+                0.116
+            ],
+            "angle": 0,
+            "content": "to transform the sparse word-count matrices into dense vectors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.121,
+                0.885,
+                0.25
+            ],
+            "angle": 0,
+            "content": "With the subsequent success of simple neural networks such as word2vec (Mikolov et al., 2013a,b; Goldberg and Levy, 2014) or GloVe (Pennington et al., 2014) at representing words and phrases, a seemingly naive method of averaging these word vectors was shown to produce useful sentence/paragraph/document representations (Arora et al., 2017)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.256,
+                0.885,
+                0.625
+            ],
+            "angle": 0,
+            "content": "With further success of neural networks in text processing, the research on the topic grew markedly. Among the unsupervised methods, recursive auto-encoders (RAEs) (Socher et al., 2011a,b; Pollack, 1990) were trained to reconstruct sentences via recursively encoding nodes in a dependency parse tree and performed well at a paraphrasing task (Dolan et al., 2004). Tree-LSTMs (Tai et al., 2015) later succeeded the RAEs in the same recursive set up but using an LSTM encoder (Hochreiter and Schmidhuber, 1997), and evaluated on predicting semantic relatedness between two sentences (Marelli et al., 2014) and classifying sentiment (Socher et al., 2013). The current state-of-the-art unsupervised, general-purpose sentence representation systems are based on pre-trained Transformer language models (Vaswani et al., 2017) and fine-tuned via contrastive learning (Gutmann and Hyvarinen, 2012; Chen et al., 2020), i.e., pulling representations of alternative sentence formulations closer together and those of unrelated, random sentence pairs apart (Gao et al., 2021; Chuang et al., 2022)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.63,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The success of skip-gram-based word representations inspired researchers to transfer the idea to sentence representation. Skip-Thought (Kiros et al., 2015) is a large, LSTM-based (Hochreiter and Schmidhuber, 1997), sentence encoder that was trained to produce representations predictive of surrounding sentences. Quick-Thought (Logeswaran and Lee, 2018) is a similar system that differs in that it casts the problem as contrastive learning (Gutmann and Hyvarinen, 2012). The Universal Sentence Encoder (USE) is a Transformer-based (Vaswani et al., 2017) system introduced by Cer et al. (2018) that uses the Skip-Thought as one of the tasks in a multi-task learning set-up (including some supervised tasks). The task of deciding whether two sentences are consecutive or not (NSP task) was also used in training of BERT (Devlin et al., 2019). The latest dialog processing sys"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8785"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.198
+            ],
+            "angle": 0,
+            "content": "tems, built on top of Transformer language models (Vaswani et al., 2017), typically learn to predict contextual utterances within the contrastive learning framework and achieve superior performance on dialog tasks compared to universal representations (Wu et al., 2020; Zhou et al., 2022; Zeng et al., 2023; Oh et al., 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.2,
+                0.49,
+                0.552
+            ],
+            "angle": 0,
+            "content": "While sentence contextuality is used in all of the above-mentioned studies, there is little to no focus on assessing the necessity for using context to derive sentence representations. An ablation of the NSP task is presented by Devlin et al. (2019), suggesting a slight performance benefit over a version where the task is omitted. However, Devlin et al. (2019) only evaluate the fine-tuning pipeline, while the raw BERT representations were shown to perform poorly overall (Reimers and Gurevych, 2019). Further, results by Liu et al. (2019) are contradictory and do not credit a performance improvement to the NSP task. Lastly, Zhou et al. (2022) additionally compare the representations resulting from dropout augmentations (Gao et al., 2021) to those of consecutive utterances and found a significant benefit in favor of the latter augmentations. Their results are perplexing, however, since the dropout version they implemented performed even worse than that of SimCSE (Gao et al., 2021), which also uses dropout augmentation, and should in fact perform worse, not being tailored for dialog."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.555,
+                0.49,
+                0.668
+            ],
+            "angle": 0,
+            "content": "Using the natural language inference (NLI) task has been effective in several general sentence representation learning systems (Conneau et al., 2017; Reimers and Gurevych, 2019; Zhou et al., 2022). However, along with machine translation models, these systems fall outside the category of unsupervised sentence representations."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.683,
+                0.219,
+                0.699
+            ],
+            "angle": 0,
+            "content": "3 Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.713,
+                0.489,
+                0.776
+            ],
+            "angle": 0,
+            "content": "We trained several variants of content-based representations, their context-based counterparts, and an end-to-end, Skip-Thought like system, trained to encode and predict sentences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.779,
+                0.49,
+                0.811
+            ],
+            "angle": 0,
+            "content": "We evaluated the following variants of content-based sentence representation<sup>2</sup>:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.827,
+                0.409,
+                0.843
+            ],
+            "angle": 0,
+            "content": "1. Latent Semantic Analysis (LSA) \\(\\star\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.859,
+                0.486,
+                0.89
+            ],
+            "angle": 0,
+            "content": "2. Latent Dirichlet Allocation (LDA)\\* (Blei et al., 2003)"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.129,
+                0.827,
+                0.486,
+                0.89
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.524,
+                0.085,
+                0.882,
+                0.116
+            ],
+            "angle": 0,
+            "content": "3. Mean of GloVe word vectors \\(\\star\\) (Pennington et al., 2014)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.524,
+                0.129,
+                0.882,
+                0.16
+            ],
+            "angle": 0,
+            "content": "4. LSTM-based auto-encoder (Hochreiter and Schmidhuber, 1997)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.524,
+                0.173,
+                0.858,
+                0.189
+            ],
+            "angle": 0,
+            "content": "5. Mean of arbitrarily assigned word vectors"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.524,
+                0.085,
+                0.882,
+                0.189
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.202,
+                0.885,
+                0.427
+            ],
+            "angle": 0,
+            "content": "We also varied the final dimension of the sentence vector, namely \\( D \\in \\{6, 12, 25\\} \\), either by setting the number_of_topics parameter, setting the auto-encoder bottleneck size, or truncating GloVe vectors. The auto-encoder was based on a 2-layer, bidirectional Long Short-Term Memory (LSTM) neural network with hidden size of \\( D_{hidden} = 64 \\), dropout of \\( p = 0.1 \\). The variant 5 is one where we randomly assigned a feature vector from an 8-dimensional raster to each vocabulary entry (a procedure by which we want to emulate the usage of 1-hot vector encoding, but with dense, real-valued vectors), and created the final sentence representation by averaging corresponding word vectors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.428,
+                0.884,
+                0.538
+            ],
+            "angle": 0,
+            "content": "To create the context-based representations, we devised a learning regimen for 2 linear layers for each variant. One layer was used to project the sentence content vector to the context space (encode), and the other layer was used to predict the content vectors of surrounding sentences from the context vector (decode)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.54,
+                0.884,
+                0.685
+            ],
+            "angle": 0,
+            "content": "Finally, we trained a Skip-Thought-like model, where each sentence was represented by a sequence of word vectors (obtained via the same procedure as in variant 5), optionally truncated and padded to 20 tokens, processed via an LSTM (identical with the auto-encoder in variant 4), and projected such that, conditioned on the resulting vector, a decoder-LSTM reconstructs the surrounding sentences well via teacher forcing."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.686,
+                0.882,
+                0.732
+            ],
+            "angle": 0,
+            "content": "In all of the context-based variants, we varied the context size between 1 and 2 sentences, on each side."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.746,
+                0.742,
+                0.76
+            ],
+            "angle": 0,
+            "content": "3.1 Dataset and Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.767,
+                0.885,
+                0.895
+            ],
+            "angle": 0,
+            "content": "To infer and evaluate the representations, we used the Switchboard Dialog Act Corpus (SwDA) (Jurafsky, 1997), with exception of the GloVe vectors, where we used the pre-trained glove-twitter-25 vectors. The SwDA contains transcripts of casual telephone conversations in which the two parties were instructed to discuss a certain topic, e.g. \"care of the elderly\". Every utterance is assigned a label"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.135,
+                0.904,
+                0.482,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Entries marked with \\(\\star\\) use the GenSim implementation."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.531,
+                0.905,
+                0.836,
+                0.919
+            ],
+            "angle": 0,
+            "content": "3We used the pre-trained glove-twitter-25 version."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8786"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.083,
+                0.482,
+                0.231
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td>Utterances</td><td>212,145</td><td>26,518</td><td>26,519</td></tr><tr><td>Tokens</td><td>1,833M</td><td>227K</td><td>227K</td></tr><tr><td>Tokens per att.</td><td>8.64</td><td>8.56</td><td>8.32</td></tr><tr><td>SD</td><td>35.47%</td><td>33.52%</td><td>35.00%</td></tr><tr><td>%/Other</td><td>27.45%</td><td>25.77%</td><td>25.60%</td></tr><tr><td>B</td><td>18.52%</td><td>21.54%</td><td>23.02%</td></tr><tr><td>SV</td><td>13.76%</td><td>14.60%</td><td>13.73%</td></tr><tr><td>AA</td><td>4.81%</td><td>4.57%</td><td>2.65%</td></tr></table>"
+        },
+        {
+            "type": "table_footnote",
+            "bbox": [
+                0.112,
+                0.24,
+                0.49,
+                0.327
+            ],
+            "angle": 0,
+            "content": "Table 2: Summary of the SwDA dataset. In the upper section of the table, we summarize the corpus statistics. In the lower section we list the percentages of dialog act labels per corpus split (SD=\"Statement-non-opinion\", B=\"Backchannel\", SV=\"Statement-opinion\", AA=\"Agree/Accept\")."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.35,
+                0.488,
+                0.398
+            ],
+            "angle": 0,
+            "content": "describing the \"role\" of the act, in a similar way that part-of-speech tags describe roles of words in sentences 4."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.399,
+                0.489,
+                0.56
+            ],
+            "angle": 0,
+            "content": "To pre-process the texts, (1) we tokenized the utterances with the reference Morfessor 2.0 (Virpioja et al., 2013) implementation, (2) long utterances were broken down into multiple shorter utterances with a sliding-window of 20 words and a step-size of 18 words, and (3) we split the corpus into a train, dev, and test split with respective proportions of \\(80\\%\\), \\(10\\%\\), and \\(10\\%\\) of the utterances. For an overview of the final corpus statistics, including the dialog act tags, see Table 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.561,
+                0.49,
+                0.673
+            ],
+            "angle": 0,
+            "content": "For each representation type, we trained a linear ridge-regression\\(^{5}\\)-based classifier with 10-fold cross-validation (on the training set) to predict the dialog act labels. Altogether, there were 42 distinct dialog acts. To simplify the probe, we only selected the 10 most frequent labels to classify, while grouping the remaining labels under a single category."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.684,
+                0.214,
+                0.699
+            ],
+            "angle": 0,
+            "content": "4 Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.709,
+                0.489,
+                0.838
+            ],
+            "angle": 0,
+            "content": "The classification accuracy results are presented in Table 3. We established baseline performance as the frequency of the majority class. The only variant performing below the majority baseline was the Random variant (\\(D_{hidden} = 6\\)), which used means of arbitrarily assigned word vectors. The other variants achieved scores in the range of up to \\(15 - 21\\%\\) above the majority baseline. Except"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.845,
+                0.489,
+                0.869
+            ],
+            "angle": 0,
+            "content": "4See the Coder's manual (Jurafsky, 1997) for more information."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.87,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>5</sup>We opted for ridge regression, instead of logistic regression, because of high colinearity (except for LDA and LSA) across feature vectors. See correlation matrices in the appendix Section A. We used the Scikit-Learn implementation."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.113,
+                0.845,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.51,
+                0.084,
+                0.871,
+                0.443
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td></td><td>Context</td><td>1 att.</td><td>2 att.</td></tr><tr><td>Algo.</td><td>Dim.</td><td colspan=\"3\">Accuracy [%]</td></tr><tr><td>Maj. class</td><td></td><td>35.00</td><td></td><td></td></tr><tr><td rowspan=\"3\">GLOVE</td><td>6</td><td>51.62</td><td>50.37</td><td>50.14</td></tr><tr><td>12</td><td>52.82</td><td>52.60</td><td>52.55</td></tr><tr><td>25</td><td>57.29</td><td>57.24</td><td>56.52</td></tr><tr><td rowspan=\"3\">LDA</td><td>6</td><td>51.10</td><td>51.16</td><td>51.05</td></tr><tr><td>12</td><td>53.32</td><td>52.99</td><td>52.73</td></tr><tr><td>25</td><td>56.82</td><td>56.36</td><td>56.13</td></tr><tr><td rowspan=\"3\">LSA</td><td>6</td><td>51.07</td><td>50.62</td><td>50.58</td></tr><tr><td>12</td><td>51.94</td><td>51.44</td><td>51.04</td></tr><tr><td>25</td><td>52.02</td><td>52.11</td><td>52.12</td></tr><tr><td rowspan=\"3\">Random</td><td>6</td><td>36.52</td><td>41.55</td><td>41.83</td></tr><tr><td>12</td><td>46.99</td><td>47.17</td><td>46.47</td></tr><tr><td>25</td><td>48.46</td><td>47.69</td><td>47.59</td></tr><tr><td rowspan=\"3\">Auto-enc.</td><td>6</td><td>53.88</td><td>52.27</td><td>51.63</td></tr><tr><td>12</td><td>54.46</td><td>53.89</td><td>52.34</td></tr><tr><td>25</td><td>52.43</td><td>53.18</td><td>52.57</td></tr><tr><td rowspan=\"3\">Skip-Th.</td><td>6</td><td></td><td>51.13</td><td>39.66</td></tr><tr><td>12</td><td></td><td>53.33</td><td>52.38</td></tr><tr><td>25</td><td></td><td>55.21</td><td>57.54</td></tr></table>"
+        },
+        {
+            "type": "table_footnote",
+            "bbox": [
+                0.508,
+                0.453,
+                0.883,
+                0.496
+            ],
+            "angle": 0,
+            "content": "Table 3: Results. The three right-most columns contain the accuracy percentages for the content-only, context of distance 1, and context of distance 2 variants."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.523,
+                0.883,
+                0.554
+            ],
+            "angle": 0,
+            "content": "for LSA, each algorithm achieved the highest score with \\( D_{hidden} = 25 \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.556,
+                0.884,
+                0.603
+            ],
+            "angle": 0,
+            "content": "There were no big differences and no discerning patterns in the performance of the contextually projected variants."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.605,
+                0.885,
+                0.685
+            ],
+            "angle": 0,
+            "content": "The only case where the training regime of predicting two neighboring sentences on each side scored the best was with the end-to-end, SkipThought like variant, which also achieved the overall highest score."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.7,
+                0.637,
+                0.714
+            ],
+            "angle": 0,
+            "content": "5 Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.727,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "All tested models performed well over the majority baseline, including the averaged random word vectors which exceeded the majority baseline by a large margin when coupled with high dimensionality. One reason for this might be that, when averaging word vectors of higher dimension, there is a higher chance that some important lexical features will remain separable from the noise (also applicable to GloVe). Only in case of LSA and AutoEncoder, the effect of dimensionality was small. For the latter, using a variational or contrastive learning set-up might allow for better scaling of"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8787"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.489,
+                0.147
+            ],
+            "angle": 0,
+            "content": "performance with capacity. Based on the patterns observed here, increasing the dimensionality increases performance on the probe, with or without context."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.15,
+                0.489,
+                0.229
+            ],
+            "angle": 0,
+            "content": "We did not find either of these aspects discussed in the mainstream related work (Conneau et al., 2018; Belinkov, 2022b). The question of how far increasing the dimensionality helps performance should be explored in future work."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.232,
+                0.49,
+                0.472
+            ],
+            "angle": 0,
+            "content": "The presented results on the dialog act tag probing task are in line with related literature in that sequence/document representations with word embeddings and classical topic modeling approaches are a tough-to-beat baseline for many neural networks (Conneau et al., 2018; Reimers and Gurevych, 2019). Our results suggest that transforming existing, content-oriented representations to context-oriented ones provides minimal to no meaningful improvement in representation. An improvement of small, but perhaps meaningful magnitude is one from the best performing AutoEncoder to Skip-Thought variant \\((3\\%)\\), suggesting that some benefit in this training scheme may be found for sentences, too."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.474,
+                0.489,
+                0.634
+            ],
+            "angle": 0,
+            "content": "However, our experimental design limits us in drawing conclusions about what the Skip-Thought set-up benefited from specifically: the more general sentence representations may simply be an artifact of the more difficult learning task (predicting neighboring sentences), as opposed to being a result of form disambiguation via context, the way it works with words. Future work should employ methodology that separates the contribution of the competing sources of signal more clearly."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.636,
+                0.489,
+                0.699
+            ],
+            "angle": 0,
+            "content": "When compared to the obvious benefits of context on learning word representations, it is striking that the benefits for learning sentence representations appear to be very small at best."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.701,
+                0.489,
+                0.797
+            ],
+            "angle": 0,
+            "content": "This observation is suggestive of a more serious blind-spot in NLP: because context-based representations work well for words, many researchers assume that the same holds for any level of analysis. This is akin to claiming arbitrariness of the sign at any level of language processing."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.812,
+                0.246,
+                0.827
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We presented evidence that explicitly examines the differences between content-, and context-oriented unsupervised sentence representations. While results suggest that dimensionality, rather than context, is the determining factor in building good"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.085,
+                0.882,
+                0.133
+            ],
+            "angle": 0,
+            "content": "sentence representations, end-to-end approaches based on reconstructing surrounding sentences are a promising way forward."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.145,
+                0.615,
+                0.16
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.17,
+                0.884,
+                0.364
+            ],
+            "angle": 0,
+            "content": "Since this work only presents evidence from one sentence task, the conclusions drawn are task-specific. Although we believe the chosen task to be fairly general, one can easily think of tasks that would likely challenge the drawn conclusions, e.g., recognition of humor or sarcasm. However, these are rather rare cases and it is questionable, where the best position of the boundary between unsupervised and task-specific lies. Additionally, the selected range of dimensions is limited to a small amount, and at larger scales the patterns of results might change."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8788"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.085,
+                0.214,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.108,
+                0.489,
+                0.161
+            ],
+            "angle": 0,
+            "content": "Sanjeev Arora, Yingyu Liang, and Tengyu Ma. 2017. A simple but tough-to-beat baseline for sentence embeddings. In International conference on learning representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.171,
+                0.489,
+                0.21
+            ],
+            "angle": 0,
+            "content": "Yonatan Belinkov. 2022a. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.221,
+                0.489,
+                0.26
+            ],
+            "angle": 0,
+            "content": "Yonatan Belinkov. 2022b. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.27,
+                0.489,
+                0.298
+            ],
+            "angle": 0,
+            "content": "David M Blei. 2012. Probabilistic topic models. Communications of the ACM, 55(4):77-84."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.307,
+                0.489,
+                0.348
+            ],
+            "angle": 0,
+            "content": "David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993-1022."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.358,
+                0.489,
+                0.437
+            ],
+            "angle": 0,
+            "content": "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.446,
+                0.489,
+                0.539
+            ],
+            "angle": 0,
+            "content": "Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, et al. 2018. Universal sentence encoder for english. In Proceedings of the 2018 conference on empirical methods in natural language processing: system demonstrations, pages 169-174."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.549,
+                0.489,
+                0.614
+            ],
+            "angle": 0,
+            "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.625,
+                0.489,
+                0.665
+            ],
+            "angle": 0,
+            "content": "Grzegorz Chrupa and Afra Alishahi. 2019. Correlating neural and symbolic representations of language. arXiv preprint arXiv:1905.06401."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.675,
+                0.489,
+                0.806
+            ],
+            "angle": 0,
+            "content": "Yung-Sung Chuang, Rumen Dangovski, Hongyin Luo, Yang Zhang, Shiyu Chang, Marin Soljacic, Shang-Wen Li, Scott Yih, Yoon Kim, and James Glass. 2022. DiffCSE: Difference-based contrastive learning for sentence embeddings. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4207-4218, Seattle, United States. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.816,
+                0.489,
+                0.881
+            ],
+            "angle": 0,
+            "content": "Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. arXiv preprint arXiv:1705.02364."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.892,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Alexis Conneau, German Kruszewski, Guillaume Lample, Loic Barrault, and Marco Baroni. 2018. What"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.108,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.087,
+                0.882,
+                0.125
+            ],
+            "angle": 0,
+            "content": "you can cram into a single vector: Probing sentence embeddings for linguistic properties. arXiv preprint arXiv:1805.01070."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.137,
+                0.882,
+                0.164
+            ],
+            "angle": 0,
+            "content": "Ferdinand De Saussure. 1916. Nature of the linguistic sign. Course in general linguistics, 1:65-70."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.175,
+                0.883,
+                0.24
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin, Ming-Wei Chang, and Kenton Lee. 2019. Google, kt, language, ai: Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.252,
+                0.882,
+                0.305
+            ],
+            "angle": 0,
+            "content": "William Dolan, Chris Quirk, Chris Brockett, and Bill Dolan. 2004. Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.316,
+                0.883,
+                0.408
+            ],
+            "angle": 0,
+            "content": "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple contrastive learning of sentence embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6894-6910, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.419,
+                0.883,
+                0.471
+            ],
+            "angle": 0,
+            "content": "Yoav Goldberg and Omer Levy. 2014. word2vec explained: deriving mikolov et al.'s negative-sampling word-embedding method. arXiv preprint arXiv:1402.3722."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.483,
+                0.883,
+                0.537
+            ],
+            "angle": 0,
+            "content": "Michael U Gutmann and Aapo Hyvarinen. 2012. Noisecontrastive estimation of unnormalized statistical models, with applications to natural image statistics. Journal of machine learning research, 13(2)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.547,
+                0.883,
+                0.586
+            ],
+            "angle": 0,
+            "content": "Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural computation*, 9(8):1735-1780."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.598,
+                0.883,
+                0.638
+            ],
+            "angle": 0,
+            "content": "Thomas Hofmann. 2001. Unsupervised learning by probabilistic latent semantic analysis. Machine learning, 42(1-2):177."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.649,
+                0.883,
+                0.689
+            ],
+            "angle": 0,
+            "content": "Dan Jurafsky. 1997. Switchboard swbd-damsl shallow-discourse-function annotation coders manual. Technical report."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.7,
+                0.883,
+                0.74
+            ],
+            "angle": 0,
+            "content": "Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent convolutional neural networks for discourse compositionality. arXiv preprint arXiv:1306.3584."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.751,
+                0.883,
+                0.804
+            ],
+            "angle": 0,
+            "content": "Ryan Kiros, Yukun Zhu, Russ R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Skip-thought vectors. Advances in neural information processing systems, 28."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.814,
+                0.883,
+                0.868
+            ],
+            "angle": 0,
+            "content": "Thomas K Landauer and Susan T Dumais. 1997. A solution to Plato's problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological review, 104(2):211."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.879,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Thomas K Landauer, Peter W Foltz, and Darrell Laham. 1998. An introduction to latent semantic analysis. Discourse processes, 25(2-3):259-284."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.087,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8789"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Wei Li and Yunfang Wu. 2019. Multi-level gated recurrent neural network for dialog act classification. arXiv preprint arXiv:1910.01822."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.139,
+                0.49,
+                0.205
+            ],
+            "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,
+                0.219,
+                0.489,
+                0.258
+            ],
+            "angle": 0,
+            "content": "Lajanugen Logeswaran and Honglak Lee. 2018. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.271,
+                0.49,
+                0.376
+            ],
+            "angle": 0,
+            "content": "Marco Marelli, Luisa Bentivogli, Marco Baroni, Raffaella Bernardi, Stefano Menini, and Roberto Zamparelli. 2014. Semeval-2014 task 1: Evaluation of compositional distributional semantic models on full sentences through semantic relatedness and textual entailment. In Proceedings of the 8th international workshop on semantic evaluation (SemEval 2014), pages 1-8."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.389,
+                0.49,
+                0.441
+            ],
+            "angle": 0,
+            "content": "Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.455,
+                0.49,
+                0.52
+            ],
+            "angle": 0,
+            "content": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.534,
+                0.489,
+                0.562
+            ],
+            "angle": 0,
+            "content": "Minsik Oh, Jiwei Li, and Guoyin Wang. 2023. Tadse: Template-aware dialogue sentence embeddings."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.575,
+                0.49,
+                0.64
+            ],
+            "angle": 0,
+            "content": "Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.654,
+                0.489,
+                0.707
+            ],
+            "angle": 0,
+            "content": "Christian S Perone, Roberto Silveira, and Thomas S Paula. 2018. Evaluation of sentence embeddings in downstream and linguistic probing tasks. arXiv preprint arXiv:1806.06259."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.719,
+                0.489,
+                0.747
+            ],
+            "angle": 0,
+            "content": "Jordan B Pollack. 1990. Recursive distributed representations. Artificial Intelligence, 46(1-2):77-105."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.759,
+                0.489,
+                0.8
+            ],
+            "angle": 0,
+            "content": "Vipul Raheja and Joel Tetreault. 2019. Dialogue act classification with context-aware self-attention. arXiv preprint arXiv:1904.02594."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.813,
+                0.489,
+                0.866
+            ],
+            "angle": 0,
+            "content": "Juan Ramos et al. 2003. Using tfidf to determine word relevance in document queries. In Proceedings of the first instructional conference on machine learning, volume 242, pages 29-48. Citeseer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.878,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Richard Socher, Eric Huang, Jeffrey Pennin, Christopher D Manning, and Andrew Ng. 2011a. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. Advances in neural information processing systems, 24."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.162,
+                0.885,
+                0.241
+            ],
+            "angle": 0,
+            "content": "Richard Socher, Jeffrey Pennington, Eric H Huang, Andrew Y Ng, and Christopher D Manning. 2011b. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the 2011 conference on empirical methods in natural language processing, pages 151-161."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.25,
+                0.885,
+                0.342
+            ],
+            "angle": 0,
+            "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.352,
+                0.885,
+                0.405
+            ],
+            "angle": 0,
+            "content": "Kai Sheng Tai, Richard Socher, and Christopher D Manning. 2015. Improved semantic representations from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.415,
+                0.885,
+                0.481
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.491,
+                0.885,
+                0.544
+            ],
+            "angle": 0,
+            "content": "Sami Virpioja, Peter Smit, Stig-Arne Gronroos, Mikko Kurimo, et al. 2013. Morfessor 2.0: Python implementation and extensions for morfessor baseline. Technical report."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.553,
+                0.885,
+                0.646
+            ],
+            "angle": 0,
+            "content": "Chien-Sheng Wu, Steven C.H. Hoi, Richard Socher, and Caiming Xiong. 2020. TOD-BERT: Pre-trained natural language understanding for task-oriented dialogue. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 917-929, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.655,
+                0.885,
+                0.709
+            ],
+            "angle": 0,
+            "content": "Weihao Zeng, Keqing He, Yejie Wang, Chen Zeng, Jinggang Wang, Yunsen Xian, and Weiran Xu. 2023. Futuretod: Teaching future knowledge to pre-trained language model for task-oriented dialogue."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.718,
+                0.885,
+                0.824
+            ],
+            "angle": 0,
+            "content": "Zhihan Zhou, Dejiao Zhang, Wei Xiao, Nicholas Dingwall, Xiaofei Ma, Andrew Arnold, and Bing Xiang. 2022. Learning dialogue representations from consecutive utterances. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 754-768, Seattle, United States. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.824
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8790"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.115,
+                0.085,
+                0.336,
+                0.099
+            ],
+            "angle": 0,
+            "content": "A Correlation Matrices"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.34,
+                0.124,
+                0.651,
+                0.137
+            ],
+            "angle": 0,
+            "content": "Pearson correlation between features (dim=6)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.148,
+                0.258,
+                0.249
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.149,
+                0.415,
+                0.248
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.432,
+                0.149,
+                0.55,
+                0.248
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.579,
+                0.149,
+                0.698,
+                0.248
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.727,
+                0.149,
+                0.845,
+                0.248
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.251,
+                0.256,
+                0.349
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.251,
+                0.415,
+                0.349
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.432,
+                0.251,
+                0.55,
+                0.349
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.579,
+                0.251,
+                0.697,
+                0.349
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.727,
+                0.251,
+                0.844,
+                0.349
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.353,
+                0.256,
+                0.452
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.353,
+                0.415,
+                0.452
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.432,
+                0.353,
+                0.55,
+                0.452
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.579,
+                0.353,
+                0.698,
+                0.452
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.716,
+                0.353,
+                0.875,
+                0.453
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.468,
+                0.489,
+                0.497
+            ],
+            "angle": 0,
+            "content": "Figure 1: Correlation matrices of feature vectors across representation variants with dimension of 6."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.338,
+                0.529,
+                0.654,
+                0.542
+            ],
+            "angle": 0,
+            "content": "Pearson correlation between features (dim=12)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.553,
+                0.264,
+                0.653
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.554,
+                0.42,
+                0.653
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.431,
+                0.554,
+                0.552,
+                0.653
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.578,
+                0.554,
+                0.701,
+                0.653
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.725,
+                0.554,
+                0.848,
+                0.653
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.655,
+                0.263,
+                0.754
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.655,
+                0.419,
+                0.754
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.432,
+                0.655,
+                0.552,
+                0.754
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.578,
+                0.655,
+                0.701,
+                0.754
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.725,
+                0.655,
+                0.848,
+                0.754
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.757,
+                0.263,
+                0.856
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.757,
+                0.42,
+                0.856
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.431,
+                0.757,
+                0.552,
+                0.856
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.578,
+                0.757,
+                0.701,
+                0.856
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.715,
+                0.757,
+                0.875,
+                0.856
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.872,
+                0.489,
+                0.9
+            ],
+            "angle": 0,
+            "content": "Figure 2: Correlation matrices of feature vectors across representation variants with dimension of 12."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8791"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.338,
+                0.323,
+                0.654,
+                0.335
+            ],
+            "angle": 0,
+            "content": "Pearson correlation between features (dim=25)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.119,
+                0.347,
+                0.264,
+                0.446
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.45,
+                0.264,
+                0.546
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.549,
+                0.263,
+                0.648
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.347,
+                0.421,
+                0.446
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.449,
+                0.421,
+                0.546
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.549,
+                0.421,
+                0.648
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.43,
+                0.347,
+                0.555,
+                0.446
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.43,
+                0.449,
+                0.554,
+                0.546
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.43,
+                0.549,
+                0.554,
+                0.648
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.576,
+                0.347,
+                0.701,
+                0.446
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.577,
+                0.449,
+                0.701,
+                0.546
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.577,
+                0.549,
+                0.701,
+                0.648
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.723,
+                0.347,
+                0.848,
+                0.446
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.723,
+                0.449,
+                0.848,
+                0.546
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.715,
+                0.549,
+                0.875,
+                0.649
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.157,
+                0.664,
+                0.84,
+                0.68
+            ],
+            "angle": 0,
+            "content": "Figure 3: Correlation matrices of feature vectors across representation variants with dimension of 25."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8792"
+        }
+    ]
+]

2023/Role of Context in Unsupervised Sentence Representation Learning_ the Case of Dialog Act Modeling/7f70510d-2f4b-4269-8c9d-3b663839a931_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:636e10d892974c9df1fb8f6cfc932a933861df018f339b65b0287502d1056a25
+size 335591

2023/Role of Context in Unsupervised Sentence Representation Learning_ the Case of Dialog Act Modeling/full.md

@@ -0,0 +1,295 @@
+# Role of Context in Unsupervised Sentence Representation Learning: the Case of Dialog Act Tagging
+
+Rastislav Hronsky
+
+Jheronimus Academy of Data Science
+
+Sint Janssingel 92, 5211 DA 's-Hertogenbosch
+
+Netherlands
+
+r.hronsky@tue.nl
+
+Emmanuel Keuleers
+
+Tilburg University
+
+Warandelaan 2, 5037 AB Tilburg
+
+Netherlands
+
+E.A.Keuleers@tilburguniversity.edu
+
+# Abstract
+
+Unsupervised learning of word representations involves capturing the contextual information surrounding word occurrences, which can be grounded in the observation that word form is largely disconnected from word meaning. While there are fewer reasons to believe that the same holds for sentences, learning through context has been carried over to learning representations of word sequences. However, this work pays minimal to no attention to the role of context in inferring sentence representations. In this article, we present a dialog act tag probing task designed to explicitly compare content- and context-oriented sentence representations inferred on utterances of telephone conversations (SwDA). Our results suggest that there is no clear benefit of context-based sentence representations over content-based sentence representations. However, there is a very clear benefit of increasing the dimensionality of the sentence vectors in nearly all approaches.
+
+# 1 Introduction
+
+Unsupervised methods for constructing word representations vary widely in their implementation (Landauer and Dumais, 1997; Mikolov et al., 2013a; Pennington et al., 2014), but share the core property that they try to capture the contextual information surrounding word occurrences. This can be grounded in the observation that word form is largely disconnected from word meaning (De Saussure, 1916) and that, therefore, trying to extract meaning from form directly would be pointless.
+
+However, there are fewer reasons to believe that representing the meaning of a fully composed sentence requires the description of its context in the way that a word does. Still, much effort has been devoted to developing and evaluating context-oriented methods for sentence representation (Kiros et al., 2015; Logeswaran and Lee, 2018; Cer et al., 2018). Interestingly, the theoretical justification for the approach –roughly paraphrased– is
+
+that it works well for words (Kiros et al., 2015).
+
+In the present article, we examine whether a contextual signal is of any significant benefit to general-purpose sentence representation.
+
+To accommodate this goal, we infer two sets of utterance representations from telephone conversation transcripts, one by a model trained to encode the transcripts' content, and one by a model trained to encode the transcripts context in a Skip-Thought-like fashion. We evaluate the learned representations with a linear probe classifying the dialog act tags. Below, we discuss these choices more extensively.
+
+The reason for the linear probing technique is to avoid any further representation learning in the downstream task training, such that the resulting accuracy numbers solely reflect the unsupervised representations. This technique is commonly used to assess to what extent sentence representations exhibit certain linguistic properties, such as part-of-speech or syntax (Conneau et al., 2018; Chrupała and Alishahi, 2019; Perone et al., 2018; Belinkov, 2022a).
+
+We chose the Switchboard Dialog Act Corpus (SwDA) (Jurafsky, 1997) corpus for the following properties: (1) its size allows for inferring representations with both traditional, statistical learning methods and modern neural network based methods; (2) at the same time, the corpus is fairly narrow-domain, enabling small models to converge to meaningful representations<sup>1</sup>; (3) it contains utterance labels that we consider suitable for a probing task with respect to contextuality: while a dialog act tag unquestionably depends on the utterance contents, the surrounding utterances likely provide a good signal about its label too. See example utterances from the corpus in Table 1.
+
+<table><tr><td>SD^E</td><td>But I, my grandparents were looking into it before.</td></tr><tr><td>SD</td><td>So I know what they’ve said.</td></tr><tr><td>B</td><td>Uh-huh.</td></tr><tr><td>^H</td><td>Well, I’m trying to think.</td></tr></table>
+
+Table 1: Example dialog act tags and utterances from the SwDA corpus. (SD^E = Statement expanding y/n answer, SD = Statement-non-opinion, B = Backchannel, ^H = Hold before answer/agreement).
+
+Using context to enhance dialog act predictions is not a novel idea (Raheja and Tetreault, 2019; Li and Wu, 2019; Kalchbrenner and Blunsom, 2013). To ascertain the usefulness of context in this task, we checked for the reduction in entropy of a tag in a "unigram" setting vs. conditioned on a previous tag, a "bigram" setting. The results were  $H(\text{currenttag}) = 2.1$  bit and  $H(\text{currenttag|previousat}) = 1.86$  bit, meaning that such a context-enhanced baseline would, at least, increase the chance-level accuracy.
+
+For the content-oriented representation variants, the representations of utterances were inferred via classic topic modeling methods, word vector averaging, or an LSTM-based auto-encoder (see Section 3 for details). The resulting sentence vectors were then used as input for a Skip-Thought-like training regime, where the goal was to learn their context-oriented version by projecting them such that they are predictive of surrounding utterances. Additionally, we trained a version where the utterance encoding and projection are trained jointly, with an LSTM-based encoder of identical proportions with the one used for auto-encoding.
+
+# 2 Related Work
+
+In the literature review, we summarize some of the most important work in sentence representation and relate it to the issues outlined in previous paragraphs.
+
+Before the deep learning era of NLP, text-processing methods relied mainly on lexical features of sentences and were invariant to word order. Using word frequency transformations, e.g., tfidf (Ramos et al., 2003), has proven effective to base the representations on more semantically laden words. Techniques such as singular value decomposition (SVD) (Landauer and Dumais, 1997; Landauer et al., 1998), or topic modeling (Blei et al., 2003; Blei, 2012; Hofmann, 2001) were used
+
+to transform the sparse word-count matrices into dense vectors.
+
+With the subsequent success of simple neural networks such as word2vec (Mikolov et al., 2013a,b; Goldberg and Levy, 2014) or GloVe (Pennington et al., 2014) at representing words and phrases, a seemingly naive method of averaging these word vectors was shown to produce useful sentence/paragraph/document representations (Arora et al., 2017).
+
+With further success of neural networks in text processing, the research on the topic grew markedly. Among the unsupervised methods, recursive auto-encoders (RAEs) (Socher et al., 2011a,b; Pollack, 1990) were trained to reconstruct sentences via recursively encoding nodes in a dependency parse tree and performed well at a paraphrasing task (Dolan et al., 2004). Tree-LSTMs (Tai et al., 2015) later succeeded the RAEs in the same recursive set up but using an LSTM encoder (Hochreiter and Schmidhuber, 1997), and evaluated on predicting semantic relatedness between two sentences (Marelli et al., 2014) and classifying sentiment (Socher et al., 2013). The current state-of-the-art unsupervised, general-purpose sentence representation systems are based on pre-trained Transformer language models (Vaswani et al., 2017) and fine-tuned via contrastive learning (Gutmann and Hyvarinen, 2012; Chen et al., 2020), i.e., pulling representations of alternative sentence formulations closer together and those of unrelated, random sentence pairs apart (Gao et al., 2021; Chuang et al., 2022).
+
+The success of skip-gram-based word representations inspired researchers to transfer the idea to sentence representation. Skip-Thought (Kiros et al., 2015) is a large, LSTM-based (Hochreiter and Schmidhuber, 1997), sentence encoder that was trained to produce representations predictive of surrounding sentences. Quick-Thought (Logeswaran and Lee, 2018) is a similar system that differs in that it casts the problem as contrastive learning (Gutmann and Hyvarinen, 2012). The Universal Sentence Encoder (USE) is a Transformer-based (Vaswani et al., 2017) system introduced by Cer et al. (2018) that uses the Skip-Thought as one of the tasks in a multi-task learning set-up (including some supervised tasks). The task of deciding whether two sentences are consecutive or not (NSP task) was also used in training of BERT (Devlin et al., 2019). The latest dialog processing sys
+
+tems, built on top of Transformer language models (Vaswani et al., 2017), typically learn to predict contextual utterances within the contrastive learning framework and achieve superior performance on dialog tasks compared to universal representations (Wu et al., 2020; Zhou et al., 2022; Zeng et al., 2023; Oh et al., 2023).
+
+While sentence contextuality is used in all of the above-mentioned studies, there is little to no focus on assessing the necessity for using context to derive sentence representations. An ablation of the NSP task is presented by Devlin et al. (2019), suggesting a slight performance benefit over a version where the task is omitted. However, Devlin et al. (2019) only evaluate the fine-tuning pipeline, while the raw BERT representations were shown to perform poorly overall (Reimers and Gurevych, 2019). Further, results by Liu et al. (2019) are contradictory and do not credit a performance improvement to the NSP task. Lastly, Zhou et al. (2022) additionally compare the representations resulting from dropout augmentations (Gao et al., 2021) to those of consecutive utterances and found a significant benefit in favor of the latter augmentations. Their results are perplexing, however, since the dropout version they implemented performed even worse than that of SimCSE (Gao et al., 2021), which also uses dropout augmentation, and should in fact perform worse, not being tailored for dialog.
+
+Using the natural language inference (NLI) task has been effective in several general sentence representation learning systems (Conneau et al., 2017; Reimers and Gurevych, 2019; Zhou et al., 2022). However, along with machine translation models, these systems fall outside the category of unsupervised sentence representations.
+
+# 3 Method
+
+We trained several variants of content-based representations, their context-based counterparts, and an end-to-end, Skip-Thought like system, trained to encode and predict sentences.
+
+We evaluated the following variants of content-based sentence representation<sup>2</sup>:
+
+1. Latent Semantic Analysis (LSA)  $\star$  
+2. Latent Dirichlet Allocation (LDA)\* (Blei et al., 2003)
+
+3. Mean of GloVe word vectors  $\star$  (Pennington et al., 2014)  
+4. LSTM-based auto-encoder (Hochreiter and Schmidhuber, 1997)  
+5. Mean of arbitrarily assigned word vectors
+
+We also varied the final dimension of the sentence vector, namely  $D \in \{6, 12, 25\}$ , either by setting the number_of_topics parameter, setting the auto-encoder bottleneck size, or truncating GloVe vectors. The auto-encoder was based on a 2-layer, bidirectional Long Short-Term Memory (LSTM) neural network with hidden size of  $D_{hidden} = 64$ , dropout of  $p = 0.1$ . The variant 5 is one where we randomly assigned a feature vector from an 8-dimensional raster to each vocabulary entry (a procedure by which we want to emulate the usage of 1-hot vector encoding, but with dense, real-valued vectors), and created the final sentence representation by averaging corresponding word vectors.
+
+To create the context-based representations, we devised a learning regimen for 2 linear layers for each variant. One layer was used to project the sentence content vector to the context space (encode), and the other layer was used to predict the content vectors of surrounding sentences from the context vector (decode).
+
+Finally, we trained a Skip-Thought-like model, where each sentence was represented by a sequence of word vectors (obtained via the same procedure as in variant 5), optionally truncated and padded to 20 tokens, processed via an LSTM (identical with the auto-encoder in variant 4), and projected such that, conditioned on the resulting vector, a decoder-LSTM reconstructs the surrounding sentences well via teacher forcing.
+
+In all of the context-based variants, we varied the context size between 1 and 2 sentences, on each side.
+
+# 3.1 Dataset and Evaluation
+
+To infer and evaluate the representations, we used the Switchboard Dialog Act Corpus (SwDA) (Jurafsky, 1997), with exception of the GloVe vectors, where we used the pre-trained glove-twitter-25 vectors. The SwDA contains transcripts of casual telephone conversations in which the two parties were instructed to discuss a certain topic, e.g. "care of the elderly". Every utterance is assigned a label
+
+<table><tr><td></td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td>Utterances</td><td>212,145</td><td>26,518</td><td>26,519</td></tr><tr><td>Tokens</td><td>1,833M</td><td>227K</td><td>227K</td></tr><tr><td>Tokens per att.</td><td>8.64</td><td>8.56</td><td>8.32</td></tr><tr><td>SD</td><td>35.47%</td><td>33.52%</td><td>35.00%</td></tr><tr><td>%/Other</td><td>27.45%</td><td>25.77%</td><td>25.60%</td></tr><tr><td>B</td><td>18.52%</td><td>21.54%</td><td>23.02%</td></tr><tr><td>SV</td><td>13.76%</td><td>14.60%</td><td>13.73%</td></tr><tr><td>AA</td><td>4.81%</td><td>4.57%</td><td>2.65%</td></tr></table>
+
+Table 2: Summary of the SwDA dataset. In the upper section of the table, we summarize the corpus statistics. In the lower section we list the percentages of dialog act labels per corpus split (SD="Statement-non-opinion", B="Backchannel", SV="Statement-opinion", AA="Agree/Accept").
+
+describing the "role" of the act, in a similar way that part-of-speech tags describe roles of words in sentences 4.
+
+To pre-process the texts, (1) we tokenized the utterances with the reference Morfessor 2.0 (Virpioja et al., 2013) implementation, (2) long utterances were broken down into multiple shorter utterances with a sliding-window of 20 words and a step-size of 18 words, and (3) we split the corpus into a train, dev, and test split with respective proportions of  $80\%$ ,  $10\%$ , and  $10\%$  of the utterances. For an overview of the final corpus statistics, including the dialog act tags, see Table 2.
+
+For each representation type, we trained a linear ridge-regression $^{5}$ -based classifier with 10-fold cross-validation (on the training set) to predict the dialog act labels. Altogether, there were 42 distinct dialog acts. To simplify the probe, we only selected the 10 most frequent labels to classify, while grouping the remaining labels under a single category.
+
+# 4 Results
+
+The classification accuracy results are presented in Table 3. We established baseline performance as the frequency of the majority class. The only variant performing below the majority baseline was the Random variant ( $D_{hidden} = 6$ ), which used means of arbitrarily assigned word vectors. The other variants achieved scores in the range of up to  $15 - 21\%$  above the majority baseline. Except
+
+<table><tr><td></td><td></td><td>Context</td><td>1 att.</td><td>2 att.</td></tr><tr><td>Algo.</td><td>Dim.</td><td colspan="3">Accuracy [%]</td></tr><tr><td>Maj. class</td><td></td><td>35.00</td><td></td><td></td></tr><tr><td rowspan="3">GLOVE</td><td>6</td><td>51.62</td><td>50.37</td><td>50.14</td></tr><tr><td>12</td><td>52.82</td><td>52.60</td><td>52.55</td></tr><tr><td>25</td><td>57.29</td><td>57.24</td><td>56.52</td></tr><tr><td rowspan="3">LDA</td><td>6</td><td>51.10</td><td>51.16</td><td>51.05</td></tr><tr><td>12</td><td>53.32</td><td>52.99</td><td>52.73</td></tr><tr><td>25</td><td>56.82</td><td>56.36</td><td>56.13</td></tr><tr><td rowspan="3">LSA</td><td>6</td><td>51.07</td><td>50.62</td><td>50.58</td></tr><tr><td>12</td><td>51.94</td><td>51.44</td><td>51.04</td></tr><tr><td>25</td><td>52.02</td><td>52.11</td><td>52.12</td></tr><tr><td rowspan="3">Random</td><td>6</td><td>36.52</td><td>41.55</td><td>41.83</td></tr><tr><td>12</td><td>46.99</td><td>47.17</td><td>46.47</td></tr><tr><td>25</td><td>48.46</td><td>47.69</td><td>47.59</td></tr><tr><td rowspan="3">Auto-enc.</td><td>6</td><td>53.88</td><td>52.27</td><td>51.63</td></tr><tr><td>12</td><td>54.46</td><td>53.89</td><td>52.34</td></tr><tr><td>25</td><td>52.43</td><td>53.18</td><td>52.57</td></tr><tr><td rowspan="3">Skip-Th.</td><td>6</td><td></td><td>51.13</td><td>39.66</td></tr><tr><td>12</td><td></td><td>53.33</td><td>52.38</td></tr><tr><td>25</td><td></td><td>55.21</td><td>57.54</td></tr></table>
+
+Table 3: Results. The three right-most columns contain the accuracy percentages for the content-only, context of distance 1, and context of distance 2 variants.
+
+for LSA, each algorithm achieved the highest score with  $D_{hidden} = 25$ .
+
+There were no big differences and no discerning patterns in the performance of the contextually projected variants.
+
+The only case where the training regime of predicting two neighboring sentences on each side scored the best was with the end-to-end, SkipThought like variant, which also achieved the overall highest score.
+
+# 5 Discussion
+
+All tested models performed well over the majority baseline, including the averaged random word vectors which exceeded the majority baseline by a large margin when coupled with high dimensionality. One reason for this might be that, when averaging word vectors of higher dimension, there is a higher chance that some important lexical features will remain separable from the noise (also applicable to GloVe). Only in case of LSA and AutoEncoder, the effect of dimensionality was small. For the latter, using a variational or contrastive learning set-up might allow for better scaling of
+
+performance with capacity. Based on the patterns observed here, increasing the dimensionality increases performance on the probe, with or without context.
+
+We did not find either of these aspects discussed in the mainstream related work (Conneau et al., 2018; Belinkov, 2022b). The question of how far increasing the dimensionality helps performance should be explored in future work.
+
+The presented results on the dialog act tag probing task are in line with related literature in that sequence/document representations with word embeddings and classical topic modeling approaches are a tough-to-beat baseline for many neural networks (Conneau et al., 2018; Reimers and Gurevych, 2019). Our results suggest that transforming existing, content-oriented representations to context-oriented ones provides minimal to no meaningful improvement in representation. An improvement of small, but perhaps meaningful magnitude is one from the best performing AutoEncoder to Skip-Thought variant  $(3\%)$ , suggesting that some benefit in this training scheme may be found for sentences, too.
+
+However, our experimental design limits us in drawing conclusions about what the Skip-Thought set-up benefited from specifically: the more general sentence representations may simply be an artifact of the more difficult learning task (predicting neighboring sentences), as opposed to being a result of form disambiguation via context, the way it works with words. Future work should employ methodology that separates the contribution of the competing sources of signal more clearly.
+
+When compared to the obvious benefits of context on learning word representations, it is striking that the benefits for learning sentence representations appear to be very small at best.
+
+This observation is suggestive of a more serious blind-spot in NLP: because context-based representations work well for words, many researchers assume that the same holds for any level of analysis. This is akin to claiming arbitrariness of the sign at any level of language processing.
+
+# 6 Conclusion
+
+We presented evidence that explicitly examines the differences between content-, and context-oriented unsupervised sentence representations. While results suggest that dimensionality, rather than context, is the determining factor in building good
+
+sentence representations, end-to-end approaches based on reconstructing surrounding sentences are a promising way forward.
+
+# Limitations
+
+Since this work only presents evidence from one sentence task, the conclusions drawn are task-specific. Although we believe the chosen task to be fairly general, one can easily think of tasks that would likely challenge the drawn conclusions, e.g., recognition of humor or sarcasm. However, these are rather rare cases and it is questionable, where the best position of the boundary between unsupervised and task-specific lies. Additionally, the selected range of dimensions is limited to a small amount, and at larger scales the patterns of results might change.
+
+# References
+
+Sanjeev Arora, Yingyu Liang, and Tengyu Ma. 2017. A simple but tough-to-beat baseline for sentence embeddings. In International conference on learning representations.  
+Yonatan Belinkov. 2022a. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219.  
+Yonatan Belinkov. 2022b. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219.  
+David M Blei. 2012. Probabilistic topic models. Communications of the ACM, 55(4):77-84.  
+David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993-1022.  
+Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics.  
+Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, et al. 2018. Universal sentence encoder for english. In Proceedings of the 2018 conference on empirical methods in natural language processing: system demonstrations, pages 169-174.  
+Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR.  
+Grzegorz Chrupa and Afra Alishahi. 2019. Correlating neural and symbolic representations of language. arXiv preprint arXiv:1905.06401.  
+Yung-Sung Chuang, Rumen Dangovski, Hongyin Luo, Yang Zhang, Shiyu Chang, Marin Soljacic, Shang-Wen Li, Scott Yih, Yoon Kim, and James Glass. 2022. DiffCSE: Difference-based contrastive learning for sentence embeddings. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4207-4218, Seattle, United States. Association for Computational Linguistics.  
+Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. arXiv preprint arXiv:1705.02364.  
+Alexis Conneau, German Kruszewski, Guillaume Lample, Loic Barrault, and Marco Baroni. 2018. What
+
+you can cram into a single vector: Probing sentence embeddings for linguistic properties. arXiv preprint arXiv:1805.01070.  
+Ferdinand De Saussure. 1916. Nature of the linguistic sign. Course in general linguistics, 1:65-70.  
+Jacob Devlin, Ming-Wei Chang, and Kenton Lee. 2019. Google, kt, language, ai: Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186.  
+William Dolan, Chris Quirk, Chris Brockett, and Bill Dolan. 2004. Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources.  
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple contrastive learning of sentence embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6894-6910, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Yoav Goldberg and Omer Levy. 2014. word2vec explained: deriving mikolov et al.'s negative-sampling word-embedding method. arXiv preprint arXiv:1402.3722.  
+Michael U Gutmann and Aapo Hyvarinen. 2012. Noisecontrastive estimation of unnormalized statistical models, with applications to natural image statistics. Journal of machine learning research, 13(2).  
+Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural computation*, 9(8):1735-1780.  
+Thomas Hofmann. 2001. Unsupervised learning by probabilistic latent semantic analysis. Machine learning, 42(1-2):177.  
+Dan Jurafsky. 1997. Switchboard swbd-damsl shallow-discourse-function annotation coders manual. Technical report.  
+Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent convolutional neural networks for discourse compositionality. arXiv preprint arXiv:1306.3584.  
+Ryan Kiros, Yukun Zhu, Russ R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Skip-thought vectors. Advances in neural information processing systems, 28.  
+Thomas K Landauer and Susan T Dumais. 1997. A solution to Plato's problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological review, 104(2):211.  
+Thomas K Landauer, Peter W Foltz, and Darrell Laham. 1998. An introduction to latent semantic analysis. Discourse processes, 25(2-3):259-284.
+
+Wei Li and Yunfang Wu. 2019. Multi-level gated recurrent neural network for dialog act classification. arXiv preprint arXiv:1910.01822.  
+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.  
+Lajanugen Logeswaran and Honglak Lee. 2018. An efficient framework for learning sentence representations. arXiv preprint arXiv:1803.02893.  
+Marco Marelli, Luisa Bentivogli, Marco Baroni, Raffaella Bernardi, Stefano Menini, and Roberto Zamparelli. 2014. Semeval-2014 task 1: Evaluation of compositional distributional semantic models on full sentences through semantic relatedness and textual entailment. In Proceedings of the 8th international workshop on semantic evaluation (SemEval 2014), pages 1-8.  
+Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781.  
+Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26.  
+Minsik Oh, Jiwei Li, and Guoyin Wang. 2023. Tadse: Template-aware dialogue sentence embeddings.  
+Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543.  
+Christian S Perone, Roberto Silveira, and Thomas S Paula. 2018. Evaluation of sentence embeddings in downstream and linguistic probing tasks. arXiv preprint arXiv:1806.06259.  
+Jordan B Pollack. 1990. Recursive distributed representations. Artificial Intelligence, 46(1-2):77-105.  
+Vipul Raheja and Joel Tetreault. 2019. Dialogue act classification with context-aware self-attention. arXiv preprint arXiv:1904.02594.  
+Juan Ramos et al. 2003. Using tfidf to determine word relevance in document queries. In Proceedings of the first instructional conference on machine learning, volume 242, pages 29-48. Citeseer.  
+Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084.
+
+Richard Socher, Eric Huang, Jeffrey Pennin, Christopher D Manning, and Andrew Ng. 2011a. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. Advances in neural information processing systems, 24.  
+Richard Socher, Jeffrey Pennington, Eric H Huang, Andrew Y Ng, and Christopher D Manning. 2011b. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the 2011 conference on empirical methods in natural language processing, pages 151-161.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642.  
+Kai Sheng Tai, Richard Socher, and Christopher D Manning. 2015. Improved semantic representations from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075.  
+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.  
+Sami Virpioja, Peter Smit, Stig-Arne Gronroos, Mikko Kurimo, et al. 2013. Morfessor 2.0: Python implementation and extensions for morfessor baseline. Technical report.  
+Chien-Sheng Wu, Steven C.H. Hoi, Richard Socher, and Caiming Xiong. 2020. TOD-BERT: Pre-trained natural language understanding for task-oriented dialogue. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 917-929, Online. Association for Computational Linguistics.  
+Weihao Zeng, Keqing He, Yejie Wang, Chen Zeng, Jinggang Wang, Yunsen Xian, and Weiran Xu. 2023. Futuretod: Teaching future knowledge to pre-trained language model for task-oriented dialogue.  
+Zhihan Zhou, Dejiao Zhang, Wei Xiao, Nicholas Dingwall, Xiaofei Ma, Andrew Arnold, and Bing Xiang. 2022. Learning dialogue representations from consecutive utterances. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 754-768, Seattle, United States. Association for Computational Linguistics.
+
+![](images/e3bd7dab7f09a78492363b6a1387f06f781a32a9a2d1431da995acd3a7d59e8a.jpg)
+
+![](images/b2b7228fad90e69a59d321d64ac43cb69620915ca6ae02157f52ff20191a9668.jpg)  
+Pearson correlation between features (dim=6)
+
+![](images/f4dc16a9bfd9f146d249cfd0a918b088a34fefa7f600a26f3c8189f1ab1e1e62.jpg)
+
+![](images/b4aa368cbcbdcb09b1eb7f8f9d572459db95e964c62dec886e638d0c5861ca07.jpg)
+
+![](images/a1b7a46a28ae57dbbbe6c0e2e4376fe4dc32594ad8fba9383dfaf5a21b1a4af9.jpg)
+
+![](images/b8318ad01d3dc6d703407573d070942519a0b3e2105f51dc1970dbe829d44af9.jpg)
+
+![](images/406714a29b6928cbbd3cf99a5c1cf477acea5e0d3f634d05d710097bcadf6b8c.jpg)
+
+![](images/ab49c52879c9e63344ebb0b11d288945b4623c0921773c8bc1e507da73951071.jpg)
+
+![](images/18b2be039e660f5a21f72a97b37fb0712ffcabb9d570f6289638adbe73da9de6.jpg)
+
+![](images/bbae9db3c2c61be51038065c6271133afbf794e7d81fb6f6dd2c9407dcc9e875.jpg)
+
+![](images/eaaff27a405d9fe37685a0491a51dae754afaba045402119ea5d7937a27e3471.jpg)  
+Figure 1: Correlation matrices of feature vectors across representation variants with dimension of 6.
+
+![](images/3f510d7fa00621c384e8521f1580fa5bdcf8eafda028390758d19d87d1eb36f8.jpg)
+
+![](images/3fb5ed6ae184b0b7cee9a431f65693bd100a4d0695b429e6f3bb0fd8fb051bd3.jpg)
+
+![](images/9f5cf7fbe539f5feb4e41b80a7565a289b306aa39c619b848002750a27de7930.jpg)
+
+![](images/cdef12ac6c191c98ddd2445564c513d1bed3178d3106f1330f56cc272e3c1568.jpg)
+
+![](images/ba962b57c0c5246fc90ef58b76a0c12cca8d16889acbe6c156a5f6a472c20ce7.jpg)
+
+![](images/a366fdfbea52d788a92af08818565a327e98c3df86a050e41e372e00b9c010f6.jpg)  
+Pearson correlation between features (dim=12)
+
+![](images/04ac375c04e26358ccd6ea6389ef32d851582187466f15170f15e29747ddb1c6.jpg)
+
+![](images/464693b6622a338abc6c1f8268461dc4b3dc6d49ede288df8fa4cc1885724f2b.jpg)
+
+![](images/f3ba839187d7c35414596f45ae59a132d0142f60358af367f1c0a80a6f8e52eb.jpg)
+
+![](images/ebe9284b92ab7aa72dc45aead00eec0f87619fb570715e1ca389850afabbb711.jpg)
+
+![](images/72fab356a9886508898e03f65065d7c6843406057da271191f22b7c4db25ed9b.jpg)
+
+![](images/1b8161abc276285198cd56cd2876819edc5098c492e311e7ed0a6ce3071f0cf0.jpg)
+
+![](images/fd1b29c5fdd1b56235c96ab54b644854c2e4899736b1437390b8ebb0da8a0739.jpg)
+
+![](images/7559ab50de6be62847564c41822757c71d9f1024d785f2024d9f2d2b23101e08.jpg)
+
+![](images/69bdb5a9c1f7be3ac3ecea2e05d8fa810b85df666a6f76d0cabf96354bb29a86.jpg)  
+Figure 2: Correlation matrices of feature vectors across representation variants with dimension of 12.
+
+![](images/1841577c8eea6af3ff3767280aed87b54821d4228cccdf4386de82f0e4736ec4.jpg)
+
+![](images/2b0b460584ecf83a5f860bce74ea515a663989ca952fbfeccb6b2a3e242e649f.jpg)
+
+![](images/1057ade4d2ac1b65a8b24ba0436c30938bab50136818954bad3380fa91e22a9b.jpg)
+
+![](images/cd79acc52bec3506d06e3e8e662ff62d9650b35e9e7e0f21552102c3e9b4882f.jpg)
+
+# Pearson correlation between features (dim=25)
+
+![](images/4150446ab607544a5eb7c0f00ce24ce092873d5b85f8b9b72527d72fe720221a.jpg)
+
+![](images/c8cce625139a6879322e78635b20ef8a0bc1b8f4cbb939704211c3e9582da226.jpg)
+
+![](images/8930cdf5fe5c77ffa84035a1d0d737b490b30ceb6625ca733e5b5019c9527df0.jpg)  
+Figure 3: Correlation matrices of feature vectors across representation variants with dimension of 25.
+
+![](images/e47ecb0b45c1943cf4122e45fa8a6cf9b64a9cb4763ce1b7a2a826caeb3b7d5e.jpg)
+
+![](images/cf9627ac9b4e98c9ebee65949b37c162d6ef97cc4f18665daaa5047d90c22ba9.jpg)
+
+![](images/d82564006893445346d5e0d61c487f3ef32857a20dd2f3262f19758aa4ba2d86.jpg)
+
+![](images/7b20f1378f18117882eeae91bcff18b5198d07eb49123bf3a6e61543d743e9f7.jpg)
+
+![](images/b2d98184a46d3e4682108937827856479f67325ff4e3edd3677090a21c527ad5.jpg)
+
+![](images/763d9c57c03a73a650ca8d36c638af7a1a82d0472a7d623ae8c9a9f28aa33ff4.jpg)
+
+![](images/1cfd29e61a7083a3c9fd13f24a4282967ff1005d386da95ca8944188ed647fc3.jpg)
+
+![](images/247cf832c26e0ffc9c8c1c35201d0cdd38890345da7d16fa13e7991d145877fd.jpg)
+
+![](images/39a24e18f9c9208d3b5d828bb77ee788f8111bd5534b1873320dbc4a4adbd332.jpg)
+
+![](images/028d2dd7a7a963c8fe04ee18e5dbdf982408ca9235cf9a731442d6334b6aeb33.jpg)
+
+![](images/d2f0df741541b95d244eb4b1fc61eec1602867cddbe37ed93fb16e1da9e21f0d.jpg)
+
+![](images/21a5f9311fc83cebd8d0f22c68e1bd69a752bfd38b6c4e584aee0d552567b8e7.jpg)

2023/Role of Context in Unsupervised Sentence Representation Learning_ the Case of Dialog Act Modeling/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e12a1f43eab7a65b615bf92f5081a9627f48d6ccc56368df5bc212162008d7f
+size 518839

2023/Roles of Scaling and Instruction Tuning in Language Perception_ Model vs. Human Attention/4d81e6a0-662e-4a1d-8420-5bca7d9f2692_content_list.json

@@ -0,0 +1,2369 @@
+[
+    {
+        "type": "text",
+        "text": "Roles of Scaling and Instruction Tuning in Language Perception: Model vs. Human Attention",
+        "text_level": 1,
+        "bbox": [
+            166,
+            87,
+            835,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Changjiang Gao<sup>1</sup> Shujian Huang<sup>1*</sup> Jixing Li<sup>2</sup> Jiajun Chen<sup>1</sup>",
+        "bbox": [
+            226,
+            141,
+            774,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ National Key Laboratory for Novel Software Technology, Nanjing University",
+        "bbox": [
+            184,
+            159,
+            816,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Department of Linguistics and Translation, City University of Hong Kong gaocj@smail.nju.edu.cn {huangsj, chenjj}@nju.edu.cn",
+        "bbox": [
+            196,
+            175,
+            803,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "jixingli@cityu.edu.hk",
+        "bbox": [
+            394,
+            210,
+            608,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent large language models (LLMs) have revealed strong abilities to understand natural language. Since most of them share the same basic structure, i.e. the transformer block, possible contributors to their success in the training process are scaling and instruction tuning. However, how these factors affect the models' language perception is unclear. This work compares the self-attention of several existing LLMs (LLaMA, Alpaca and Vicuna) in different sizes (7B, 13B, 30B, 65B), together with eye saccade, an aspect of human reading attention, to assess the effect of scaling and instruction tuning on language perception. Results show that scaling enhances the human resemblance and improves the effective attention by reducing the trivial pattern reliance, while instruction tuning does not. However, instruction tuning significantly enhances the models' sensitivity to instructions. We also find that current LLMs are consistently closer to non-native than native speakers in attention, suggesting a suboptimal language perception of all models. Our code and data used in the analysis is available on GitHub.",
+        "bbox": [
+            144,
+            279,
+            460,
+            633
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            646,
+            258,
+            662
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs), e.g., GPT-4, ChatGPT and Vicuna, have shown nearly human-level understanding of text inputs, indicating better human-like language perception compared with their predecessors such as BERT (Devlin et al., 2019) and GPT-2 (Radford et al., 2019). However, the mechanism behind such improvement is largely unknown. One way to interpret such mechanism is comparing model computation processes to human data. For example, prior work compared language models (mainly GPT-2) with human neuroimaging data during language comprehension, suggesting that models with higher human resemblance also perform better in NLP tasks. (Hasson",
+        "bbox": [
+            112,
+            671,
+            489,
+            897
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "et al., 2020; Schrimpf et al., 2021). However, given recent breakthroughs in LLMs, it remains to be tested whether the newest LLMs still align with human perception data. Based on the training pipeline of current LLMs, there are two possible sources for the potential change: the scaled model/data size and the aligning process after pretraining, such as instruction tuning (Ouyang et al., 2022; Wei et al., 2022; Sanh et al., 2022). This paper aims to provide a further understanding of LLMs' mechanism by investigating the roles of these two factors in language understanding.",
+        "bbox": [
+            507,
+            252,
+            884,
+            445
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we analyze self-attention as a channel to the language perception ability of LLMs, because it is the key mechanism of transformer language models, and it naturally resembles the human attention mechanism in language perceiving (Merkx and Frank, 2021). The attention patterns under investigation includes those from trending open-sourced LLMs with different sizes, i.e., 7B, 13B, 30B, 65B, and different training stages, i.e., pretrained (LLaMA) and instruction-tuned (Alpaca and Vicuna) models. Given these models are all trained mainly with English data, we also compare the attention patterns of LLMs with human saccade (a form of eye-movement) data, including both native speakers (L1) and non-native learners (L2) of English (Li et al., 2019) to see if the models correlates more with L1 subjects.",
+        "bbox": [
+            507,
+            449,
+            884,
+            722
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Our analysis is three-fold. First, we compare the general attention distributions of LLMs to measure the impact of these two varying factors on the model ( $\\S 3.1$ ). Second, we perform linear regressions (LRs) between the model self-attention and human saccade matrices to see how human-like they are, and how the resemblance is affected by the two factors ( $\\S 3.2$ ). Third, we analyze the relation between the given attention and some trivial patterns, revealing the difference between different levels of language perception abilities ( $\\S 3.3$ ). The main findings are:",
+        "bbox": [
+            507,
+            726,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding author",
+        "bbox": [
+            141,
+            904,
+            294,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "13042",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13042-13055 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            210,
+            945,
+            786,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Scaling significantly affects the general attention distribution on plain text, while instruction tuning has limited effect. However, instruction tuning enhances the model's sensitivity to instructions (§5.1).",
+            "- Higher human resemblance is significantly correlated to better language modeling. Scaling improves the resemblance by scaling law (Henighan et al., 2020), while instruction tuning reduces it. All models have higher resemblance to L2 rather than to L1, suggesting further room for the improvement in language perception (§5.2).",
+            "- L2 saccade has higher dependency on trivial attention patterns than L1, indicating a more rigid way of perception. Scaling significantly lowers LLMs' reliance on trivial attention patterns, while instruction tuning does not (§5.3)."
+        ],
+        "bbox": [
+            136,
+            84,
+            487,
+            399
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            414,
+            270,
+            429
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Self-attention analysis is a common way to interpret transformer models. Several studies has shown that it is correlated to linguistic information, such as syntactic or semantic structures (Mareček and Rosa, 2018; Raganato and Tiedemann, 2018; Voita et al., 2019; Clark et al., 2019). There are also some attempts and debates to explain the models' prediction with attention (Jain and Wallace, 2019; Wegreffe and Pinter, 2019; Serrano and Smith, 2019; Vashisth et al., 2019). Different from them, we use self-attention patterns as a measurement of the language perception abilities of LLMs. Also, some new ways to interpret model self-attention have been proposed, such as Attention Flow (Abnar and Zuidema, 2020), vector norm (Kobayashi et al., 2020, 2021), ALTI (Ferrando et al., 2022), Value Zeroing (Mohebbi et al., 2023), and logit analysis (Ferrando et al., 2023). These methods can raise the interpretability of model attention w.r.t. linguistic structural information, but are not tested with human eye-tracking data. Also, some of them suffer from huge computational cost. Instead, we choose to use the raw attention scores, as it is produced in the models' original computational processes, and they already correlate the human data, demonstrating their interpretability.",
+        "bbox": [
+            112,
+            441,
+            489,
+            860
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Instruction tuned LLMs achieve better performances in performing tasks (Ouyang et al., 2022; Wei et al., 2022; Sanh et al., 2022). It seems that the",
+        "bbox": [
+            112,
+            871,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "LLMs are better at understanding human instructions. However, there is little discussion on how the instruction tuning process affects the language perception.",
+        "bbox": [
+            507,
+            84,
+            884,
+            149
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Eye-movement in the human reading process has drawn attention in the computational linguistic field. It is proposed that scanpaths, i.e. tracks of eye-movement, reveals characteristics of the parsing processes and relates to working memory during reading (von der Malsburg and Vasishth, 2013), and the irregularity of scanpaths is affected by sentence structures and age (von der Malsburg et al., 2015). Such work provides theoretical ground for our analysis on saccade data.",
+        "bbox": [
+            507,
+            160,
+            884,
+            319
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Joint study of language models and human eye-movement is also receiving growing interest. Prior studies have shown the correlation between language processing of human and artificial neural networks (Kell et al., 2018), especially pretrained language models such as GPT-2 (Caucheteux et al., 2021, 2022; Schrimpf et al., 2021; Lamarre et al., 2022). Especially, Oh and Schuler (2023a,b) explored the fit of model surprisal to human reading times, noting the effect of models' sequential memorization and dataset size. Hollenstein and Beinborn (2021) studied the correlation between relative word importance derived from models and humans, and Morger et al. (2022) extended that research to a multilingual setting, as well as analysis on model attention. Some studies also use human attention as supervision to assist specific tasks (Sood et al., 2020; Bansal et al., 2023), or use sequential models to predict eye-tracking data (Chersoni et al., 2022). Different from them, we use human attention as references for the language perception abilities.",
+        "bbox": [
+            507,
+            331,
+            884,
+            684
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            697,
+            621,
+            713
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To show the effect of scaling and instruction tuning on different models, we compare the self-attention scores of different LLMs given the same input, by viewing model attention as probability distributions and calculating the general attention divergence based on Jensen-Shannon divergence.",
+        "bbox": [
+            507,
+            725,
+            882,
+            820
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To analyze model self-attention, we take the human saccade as a reference, and design a human resemblance metric based on linear regression scores.",
+        "bbox": [
+            507,
+            822,
+            884,
+            869
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In addition, we select several trivial, context-free attention patterns, and design a trivial pattern reliance metric to help demonstrate the difference",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "13043",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "between models and human subject groups.",
+        "bbox": [
+            112,
+            84,
+            438,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 General Attention Divergence",
+        "text_level": 1,
+        "bbox": [
+            112,
+            111,
+            394,
+            126
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We assess the mean Jensen-Shannon (J-S) divergence between attention from different models on all sentences. For two attention matrices  $A$ ,  $B \\in \\mathbb{R}^{n_{\\mathrm{word}} \\times n_{\\mathrm{word}}}$ , the J-S divergence is:",
+        "bbox": [
+            112,
+            131,
+            489,
+            195
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} D _ {\\mathrm {J S}} (A, B) = \\frac {1}{2} \\sum_ {i = 1} ^ {n _ {\\text {w o r d}}} \\left[ D _ {\\mathrm {K L}} \\left(A _ {i} \\| B _ {i}\\right) \\right. \\\\ + D _ {\\mathrm {K L}} \\left(B _ {i} \\| A _ {i}\\right) ] \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            171,
+            219,
+            428,
+            280
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $D_{\\mathrm{KL}}$  is the Kullback-Leibler (K-L) divergence,  $A_{i}$  and  $B_{i}$  are the  $i$ -th rows in the two matrices. Note the attention matrices are re-normalized to keep the sum of every row to 1.0. We calculate the J-S divergence between heads-averaged attention scores per sentence, and demonstrate the mean values. For models with the same number of layers, the divergence is calculated layerwise; For models with different numbers of layers, we divide their layers into four quarters and calculate the divergence quarter-wise.",
+        "bbox": [
+            112,
+            290,
+            489,
+            467
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To identify high and low divergence, a reference value is required. We propose to use the attention divergence between two models that share the same structure, size, and training data, but have small difference in training strategies as the reference. A divergence value higher than the reference suggests relatively large change attention, while one lower than that indicates small (but not necessarily negligible) change.",
+        "bbox": [
+            112,
+            468,
+            489,
+            612
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Divergence by Scaling and Instruction Tuning Models with different scales or training stages are compared with the above divergence to show the effect of different factors.",
+        "bbox": [
+            112,
+            621,
+            485,
+            684
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Sensitivity to Instructions To evaluate the effect of instruction tuning, we measure a given model's sensitivity to instructions by the same general attention divergence metric, but on different input, i.e., plain input and input with instructions. If the divergence between them is significantly higher than reference, we say the model has high sensitivity to instructions.",
+        "bbox": [
+            112,
+            694,
+            489,
+            820
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To construct instructed text, we attach two prefixes to plain text sentences: \"Please translate this sentence into German:\", and \"Please paraphrase this sentence\":\". They are within the reported capacity of most LLMs, and require no extra information to follow. Then, we collect model attention on",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the two corpora within the original sentence spans (re-normalized) to represent non-instruction and instruction scenarios. Also, as a control group, we attached a noise prefix made of 5 randomly sampled English words \"Cigarette first steel convenience champion.\", and calculate the divergence between prefixed and plain text to account for the effect of a meaningless prefix.",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Human Resemblance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            224,
+            722,
+            237
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Human Resemblance Calculation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            244,
+            831,
+            259
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given the model self-attention and the human saccade data in matrix form for each sentence, we first extract the lower-triangle parts of the matrices (marking right-to-left attendance), because prior studies have found that such eye-movements in reading are related to working memory efficiency (Walczyk and Taylor, 1996) and parsing strategies (von der Malsburg and Vasishth, 2013), and the auto-regressive LLMs only have right-to-left self-attention. Then, we flatten the extracted parts of each sentence attention matrix, and concatenate them to get attention vectors  $\\{v_{\\mathrm{human}}^i\\}$  for each subject  $i$ , and  $\\{v_{\\mathrm{model}}^{j,k}\\}$  for head  $k$  in model layer  $j$ . We then stack  $\\{v_{\\mathrm{model}}^{j,k}\\}$  along the attention heads and get a matrix  $\\{V_{\\mathrm{model}}^j\\}$  for the  $j$ -th layer. The LR for model layer  $j$  and subject  $i$  is defined as:",
+        "bbox": [
+            507,
+            263,
+            884,
+            527
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\arg \\min _ {w, b} \\left\\| V _ {\\mathrm {m o d e l}} ^ {j \\top} w + b - v _ {\\mathrm {h u m a n}} ^ {i} \\right\\| _ {2} ^ {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            561,
+            538,
+            828,
+            569
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This means we take the linear combination of attention scores from all heads in a layer to predict the human saccade. Note that the model attention here are not re-normalized, in order to preserve the relative weight w.r.t. the  $<s>$  token. The regression scores  $R_{i,j}^{2}$  (ranging from 0 to 1, larger value means better regression) are averaged over subjects to represent the human resemblance of layer  $j$ , and the highest layer score is considered as the regression score of the whole model,  $R_{\\mathrm{model}}^{2}$ .",
+        "bbox": [
+            507,
+            581,
+            882,
+            743
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Because the human saccade pattern varies across individuals, there exists a natural upper bound for human resemblance of models, i.e., the human-human resemblance, or the inter-subject correlation. To calculate this correlation within the two language groups, L1 and L2, we take the group mean to regress over each individual data independently, and use the average regression score  $R_{\\mathrm{inter}}^2$  as the inter-subject correlation for that group. The final human resemblance score is  $R_{\\mathrm{model}}^2 / R_{\\mathrm{inter}}^2$ , ranging from 0 to 100%.",
+        "bbox": [
+            507,
+            743,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "13044",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Human Resemblance vs. Next-Token Prediction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            463,
+            115
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A previous study finds a positive correlation between the next-token predication (NTP) performance and the \"Brain Score\" of neural networks (Schrimpf et al., 2021), which supports the intuition that the more human-like models are, the better generation they output. To test this finding in our setting, we calculated the per-token prediction loss (the negative log likelihood loss normalized by sequence lengths) of the models employed in this study on the Reading Brain dataset, and calculate the Pearson's Correlation between their prediction loss and max layerwise human resemblance scores.",
+        "bbox": [
+            112,
+            120,
+            489,
+            313
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Trivial Pattern Reliance",
+        "text_level": 1,
+        "bbox": [
+            112,
+            324,
+            349,
+            338
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3.1 Trivial Pattern Reliance of Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            344,
+            443,
+            360
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We select three representative trivial patterns of transformer attention: attending to the first word in the sentence, attending to the previous word (Vig and Belinkov, 2019), and self-attending (Clark et al., 2019), and analyze the effect of scaling and instruction tuning on models' reliance on them.",
+        "bbox": [
+            112,
+            363,
+            487,
+            460
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "First, we represent the trivial patterns as binary adjacent matrices, where attending relations are marked with 1, others with 0, and use their flattened lower-triangle parts as the independent variables. The model attention is similarly processed to be the dependent variables. Then, an LR is fit between the two, and the regression scores are collected to represent the trivial pattern reliance of each model layer.",
+        "bbox": [
+            112,
+            461,
+            487,
+            605
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3.2 Trivial Pattern Reliance of Humans",
+        "text_level": 1,
+        "bbox": [
+            112,
+            615,
+            453,
+            630
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This research compares the models' resemblance to L1 and L2 humans because intuitively, L1 is better than L2 in language understanding. To test this intuition, we compare their reliance on the three trivial patterns to see whether their attention mode differs. Similar to model attention, the flattened human saccade vectors are used as the dependent variable to fit the LRs with the trivial patterns.",
+        "bbox": [
+            112,
+            634,
+            487,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            775,
+            324,
+            791
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Data",
+        "text_level": 1,
+        "bbox": [
+            112,
+            802,
+            200,
+            816
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We use the Reading Brain dataset (Li et al., 2019) for both the text input and the human data. The dataset includes 5 English STEM articles. Each article has  $29.6 \\pm 0.68$  sentences of  $10.33 \\pm 0.15$  words. It also includes human eye-tracking and fMRI data recorded synchronously during self-paced reading",
+        "bbox": [
+            112,
+            822,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/72af1df936c3ae4afd689692a37f326c1690dfeaa2151877027d8fb5cc2d467c.jpg",
+        "image_caption": [
+            "(a) Individual",
+            "Figure 1: Examples of individual and group mean saccade of the sentence \"Could humans live on Mars some day?\""
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            83,
+            692,
+            205
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/00e40331f9f9f846d20f3c9cd578170c07c200d1400a103ffaf58bb6603b4d7c.jpg",
+        "image_caption": [
+            "(b) Group Mean"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            82,
+            875,
+            205
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "on the articles. The human subjects are 52 native English speakers (L1) and 56 non-native learners of English (L2). We choose this dataset for the following reasons: (1) The text is presented to the subjects sentence by sentence instead of page by page, making in-sentence reading patterns clearer. (2) It has synchronous eye-tracking and fMRI data, which allows both behavioral- and neural-level analysis. In this article we mainly focus on the behavioral part, but we also have some preliminary results on the fMRI analysis (See Appendix D).",
+        "bbox": [
+            507,
+            306,
+            884,
+            482
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We take the saccade number, i.e. times of eye movements from a word to another, instead of saccade duration, to represent the human reading attention, because this reduces the effect of other factors such as word length. The saccade data of a human subject for a given sentence is a  $n_{\\mathrm{word}} \\times n_{\\mathrm{word}}$  matrix. Figure 1 gives examples of individual and group mean saccades.",
+        "bbox": [
+            507,
+            483,
+            882,
+            612
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 LLMs",
+        "text_level": 1,
+        "bbox": [
+            507,
+            627,
+            606,
+            640
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We employ LLaMA (Touvron et al., 2023) and its instruction-tuned versions, Alpaca (Taori et al., 2023) and Vicuna (Chiang et al., 2023) in this research. We use them because: (1) The model sizes and different instruction-tuning methods covers the majority of current open-sourced LLMs. (2) These models are being widely used and customized globally. Thus, analysis on them is representative and meaningful.",
+        "bbox": [
+            507,
+            649,
+            882,
+            793
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "LLaMA is a series of pretrained causal language models with parameter sizes 7B, 13B, 30B and 65B, trained on over 1T publicly available text tokens, and reaches state-of-the-art on most LLM benchmarks (Touvron et al., 2023). The training corpus is mainly English. We use all 4 sizes of LLaMA, which covers the sizes of most current",
+        "bbox": [
+            507,
+            806,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "13045",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "open-sourced LLMs. We also use the 774M GPT-2 Large (Radford et al., 2019) to represent smaller pretrained models in the scaling analysis.",
+        "bbox": [
+            112,
+            84,
+            487,
+            134
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Alpaca is fine-tuned from the 7B LLaMA model on 52K English instruction-following demonstrations generated by GPT-3 (Brown et al., 2020) using self-instruct (Wang et al., 2023). To analyze the effect of scaling, we also trained a 13B version of Alpaca using the official data and training strategy of Alpaca. Our 13B Alpaca model scores 43.9 and 46.0 on the MMLU dataset in zero-shot and one-shot setting respectively, proving the soundness of our fine-tuning. (The corresponding scores are 40.9 and 39.2 for the official Alpaca 7B model.)",
+        "bbox": [
+            112,
+            143,
+            489,
+            322
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Vicuna models are fine-tuned from the 7B and 13B LLaMA models on 70K user-shared conversations with ChatGPT. The data contains instruction and in-context learning samples in multiple languages, thus the Vicuna models can also be viewed as instruction-tuned.",
+        "bbox": [
+            112,
+            334,
+            489,
+            430
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 LLMs Attention Collection",
+        "text_level": 1,
+        "bbox": [
+            112,
+            444,
+            371,
+            458
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To obtain model attention, text is provided to the models sentence by sentence to perform next token predict, and we collect the attention tensors in each layer. This resembles the human self-paced task-free reading process, where only one sentence is visible at a time to the subjects, and the natural human behavior is to predict the next word (Kell et al., 2018; Schrimpf et al., 2021).",
+        "bbox": [
+            112,
+            466,
+            487,
+            594
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Because models use the SentencePiece tokenization (Kudo and Richardson, 2018), but the human saccade data is recorded in words, the attention matrices need to be reshaped. Following Clark et al. (2019) and Manning et al. (2020), we take the sum over the \"to\" tokens, and the average over the \"from\" tokens in a split word. Also, the  $<s>$  token is dropped as it is not presented to human subjects.",
+        "bbox": [
+            112,
+            596,
+            489,
+            726
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Reference of Divergence We use the attention divergence of Vicuna v0 and Vicuna v1.1 13B as the reference values. The two versions share the same training data and strategy, but have minor difference in separators and loss computation<sup>2</sup>. We use the divergence calculated on 13B Vicunas across all divergence analyses, because it has lower mean and variance than the 7B one, making our comparison fairer and stricter.",
+        "bbox": [
+            112,
+            737,
+            489,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Results",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            608,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ec644743c14c8077c18e304cd43e115de1a02fc57935327bff3bdeb62939a0c4.jpg",
+        "image_caption": [
+            "Figure 2: Mean J-S divergence between 7B and 13B model attention in each layer quarter, measured on non-instruction sentences."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            124,
+            870,
+            244
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 General Attention Divergence",
+        "text_level": 1,
+        "bbox": [
+            507,
+            332,
+            789,
+            349
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1.1 Scaling Significantly Changes the General Attention",
+        "text_level": 1,
+        "bbox": [
+            507,
+            354,
+            830,
+            385
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The general attention divergence between models in different sizes is shown in Figure 2. The results show that LLaMA, Alpaca and Vicuna all reveal divergence significantly higher than the reference values when scaled from 7B to 13B. This means scaling causes large change in the overall attention distribution.",
+        "bbox": [
+            507,
+            391,
+            882,
+            502
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1.2 Instruction Tuning Has Limited Effect on General Attention",
+        "text_level": 1,
+        "bbox": [
+            507,
+            514,
+            870,
+            545
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The general attention divergence between pretrained and instruction tuned models is shown in Figure 3. It can be observed that only Vicuna 13B brings divergence above the reference values, while all the others don't. The same result can also be observed on instruction-prefixed sentences. This suggests that instruction tuning can only bring limited change to the models' general attention distribution. This can also be verified by other results in the following parts.",
+        "bbox": [
+            507,
+            551,
+            882,
+            712
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1.3 Instruction Tuning Enhances Sensitivity to Instructions",
+        "text_level": 1,
+        "bbox": [
+            507,
+            722,
+            880,
+            752
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The divergence between the models' attention obtained on the original and instruction-attached sentences are shown in Figure 4. There are two major observations. Firstly, LLaMA, Alpaca and Vicuna all show divergence significantly above the reference values between the two types of sentences, especially in the shallower layers, which means they all have sensitivity to instructions. Secondly, while the divergence of LLaMA drops in the deeper layers, the divergence of Alpaca and Vicuna does",
+        "bbox": [
+            507,
+            758,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "Based on this GitHub repository for LLM evaluation.",
+        "bbox": [
+            134,
+            890,
+            467,
+            904
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "2According to notes in the Vicuna GitHub repository.",
+        "bbox": [
+            136,
+            904,
+            460,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "13046",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/fc5d47d7c18de76657c2fc72e97c7d6618092e97df812d52e938b4fa7a26142e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            86,
+            470,
+            204
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/85a93c8b9a28b689b6faa748f92743d88ed79f05e4da2930be71c7e7b557f76e.jpg",
+        "image_caption": [
+            "Figure 3: General attention divergence between pretrained and instruction-tuned models in 7B and 13B, measured on non-instruction sentences."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            214,
+            470,
+            332
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "not go down, but rather up, suggesting higher sensitivity to instructions in the deeper layers. This effect is not observed in the noise-prefixed scenario (See Appendix A).",
+        "bbox": [
+            112,
+            416,
+            487,
+            480
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "From these two observations, we can conclude that the pretrained LLaMA has already obtained some sensitivity to instructions, i.e. to change its attention mode when it meets instructions, and this sensitivity is further strengthened by instruction tuning, which makes this sensitivity higher in the deeper layers. Because in causal language models like LLaMA, the last few layers are related to generation, this effect of instruction tuning may bring more suitable generation for a certain instruction.",
+        "bbox": [
+            112,
+            481,
+            487,
+            642
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Human Resemblance",
+        "text_level": 1,
+        "bbox": [
+            112,
+            653,
+            327,
+            668
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2.1 Higher Human Resemblance, Lower Prediction Loss",
+        "text_level": 1,
+        "bbox": [
+            112,
+            674,
+            460,
+            705
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The per-token NTP loss and the maximum layerwise human resemblance scores are compared in Table 1 and Figure 5. One can tell from it that, within the considered range of parameter sizes, the human resemblance is negatively correlated with the NTP loss, where the resemblance goes higher nearly linearly with the loss going lower. The Pearson's correlation is  $-0.875$  for L1 ( $p < 0.002$ ) and  $-0.917$  for L2 ( $p < 0.0005$ ), supporting a significant and strong negative linear correlation.",
+        "bbox": [
+            112,
+            709,
+            487,
+            869
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This result shows that, the human resemblance we defined is positively related to the language modeling performance of the LLMs considered in",
+        "bbox": [
+            112,
+            870,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1f008220a318e85eb96f9974abe4cb35e28cfcd3055c85a6548958ea60eda23d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            86,
+            867,
+            205
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/f76cbc30a5588ac8a4301ce3303f98d0edd580660d4ba45a0c94f8532c70bab8.jpg",
+        "image_caption": [
+            "Figure 4: Attention divergence on the original and instruction-attached text of 7B and 13B models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            214,
+            867,
+            332
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "this research, giving a practical meaning to our human resemblance analysis. Also, based on this result, we know the pursuit for higher human resemblance is consistent with the target of training better LLMs. Similarly, factors negatively impacting the human resemblance could also harm the models' language modeling performance, such as instruction tuning brings lower human resemblance as well as higher NTP loss.",
+        "bbox": [
+            507,
+            403,
+            884,
+            548
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/84993aceea938241ed48581fe63c0ac38da565077937761d44e5cb5a2247baf2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Size</td><td>Name</td><td>Loss</td><td>L1(%)</td><td>L2(%)</td></tr><tr><td>774M</td><td>GPT-2</td><td>0.3264</td><td>34.99</td><td>40.03</td></tr><tr><td rowspan=\"3\">7B</td><td>LLaMA</td><td>0.2408</td><td>53.04</td><td>62.44</td></tr><tr><td>Alpaca</td><td>0.2646</td><td>52.51</td><td>61.71</td></tr><tr><td>Vicuna</td><td>0.2593</td><td>51.90</td><td>61.19</td></tr><tr><td rowspan=\"3\">13B</td><td>LLaMA</td><td>0.2406</td><td>55.66</td><td>64.20</td></tr><tr><td>Alpaca</td><td>0.2634</td><td>55.05</td><td>64.46</td></tr><tr><td>Vicuna</td><td>0.2847</td><td>54.26</td><td>61.31</td></tr><tr><td>30B</td><td>LLaMA</td><td>0.2372</td><td>63.16</td><td>69.40</td></tr><tr><td>65B</td><td>LLaMA</td><td>0.2375</td><td>64.05</td><td>70.07</td></tr></table>",
+        "bbox": [
+            539,
+            560,
+            852,
+            686
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Comparison between NTP loss and the human resemblance scores",
+        "bbox": [
+            507,
+            695,
+            880,
+            722
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2.2 Scaling Enhances Human Resemblance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            753,
+            878,
+            769
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The layerwise human resemblance of pretrained models in different scales are shown in Figure 6, where the LLaMAs are very different from GPT-2 Large in two ways. First, the layerwise human resemblance of the LLaMAs are much higher than GPT-2, but the gaps among the LLaMAs are small; Second, human resemblance of GPT-2 drop quickly in the deeper layers, while it remains high across all layers in the LLaMAs. one can refer from these",
+        "bbox": [
+            507,
+            774,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "13047",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/ea819d17b6d953235351d5a73b5a5b7e65ce7a5938cf23f98cd7b2ffcce21a37.jpg",
+        "image_caption": [
+            "Figure 5: The correlation between the per-token negative log likelihood loss and the max layerwise human resemblance of LLMs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            86,
+            470,
+            215
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "results that LLaMA has undergone a phase change in terms of human-like attention from the level of GPT-2. However, the scaling from 7B to 30B does not cause another phase change.",
+        "bbox": [
+            112,
+            300,
+            487,
+            363
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2 and Figure 7 show the max layerwise human resemblance increases linearly while the model parameter size increases exponentially, where the Pearson's correlation is 0.989 for L1  $(p < 0.002)$  and 0.964 for L2  $(p < 0.01)$ . This agrees with the scaling law of LLMs (Henighan et al., 2020).",
+        "bbox": [
+            112,
+            365,
+            487,
+            476
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "This result shows that, within the scope of this study, scaling significantly enhances the human resemblance. If the scaling law continues to take effect, the human resemblance is expected to reach  $98.80\\%$  for L2 and  $88.82\\%$  for L1 at the scale of 100B. However, because the law is expected to fail after reaching a threshold (Hestness et al., 2017), the relation between human resemblance and model scale is expected to change before reaching that scale.",
+        "bbox": [
+            112,
+            479,
+            489,
+            639
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/99999bf1715290bbd957e3d38c0341977f821304f94c56a1505d37af8accb45c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Model</td><td colspan=\"2\">Resemblance (%)</td></tr><tr><td>Name</td><td>Size</td><td>L1</td><td>L2</td></tr><tr><td rowspan=\"2\">GPT-2</td><td>774M</td><td>34.99</td><td>40.03</td></tr><tr><td>7B</td><td>53.04</td><td>62.44</td></tr><tr><td rowspan=\"3\">LLaMA</td><td>13B</td><td>55.56</td><td>64.20</td></tr><tr><td>30B</td><td>63.16</td><td>69.40</td></tr><tr><td>65B</td><td>64.05</td><td>70.07</td></tr></table>",
+        "bbox": [
+            173,
+            650,
+            428,
+            738
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Max layerwise human resemblance of models ranging from 774M to 65B.",
+        "bbox": [
+            112,
+            747,
+            485,
+            776
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2.3 Instruction Tuning Harms Human Resemblance",
+        "text_level": 1,
+        "bbox": [
+            112,
+            803,
+            443,
+            833
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3 shows the max layerwise human resemblance of the pretrained and instruction tuned models. It shows that the instruction tuned Alpaca and Vicuna models have lower max layerwise human resemblance to both L1 and L2 than LLaMA in",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1616780722649fc2e8bd4012e111df8abb9d9fbd88f15d3c1b25f44acaaab699.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            85,
+            867,
+            205
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/37368eb43a0304641eb4bde8759e4bf745ad410b101965746d6be748aa6b7ea1.jpg",
+        "image_caption": [
+            "Figure 6: Layerwise L1 and L2 human resemblance for different sized pretrained models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            212,
+            867,
+            332
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/422fa9644ff1b557cfbe55e417d82fbf8846d43f61e00df97e26a8bb9394c59b.jpg",
+        "image_caption": [
+            "Figure 7: The change of max layerwise human resemblance caused by the scaling of model sizes, where the x-axis is log scaled."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            395,
+            867,
+            526
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "both sizes, which suggests that instruction tuning causes decline in the human resemblance of attention. However, the decline is minor compared with the increase brought by scaling, which agrees with the limited effect of instruction tuning.",
+        "bbox": [
+            507,
+            612,
+            882,
+            692
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We also conduct relative t-tests to detect significant difference in layerwise human resemblance  $(p = 0.05 / n_{\\mathrm{test}})$ . The numbers of significant layers are shown in Table 4, and the complete lists of layers are in Table 6, 7 in Appendix C). The tables show that instruction tuning enhances or reduces human resemblance on about the equal numbers of layers on the 7B models, but reduces human resemblance to much more layers on the 13B models. Also, the layers with significantly lower resemblance are distributed across all layers. This result means that, while the total effect of instruction tuning is small, the drop in human resemblance caused by it is significant and widespread.",
+        "bbox": [
+            507,
+            694,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "13048",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To sum up, instruction tuning does small but significant harm to the human resemblance of LLMs, which in turn may bring potential damage to their language perception. This challenges the common assumption that instruction tuning can activate the model's ability to perceive language with human-aligned behaviors. However, this negative effect could be compensated by scaling, for scaling increases the human resemblance in a larger degree. This provides a novel reason for further scaling.",
+        "bbox": [
+            112,
+            84,
+            492,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/d780297f3a566d20e077e26dc998e4c550d1006ea9aebba740858ec1ac9d40cd.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Model</td><td colspan=\"2\">Resemblance(%)</td></tr><tr><td>Size</td><td>Name</td><td>L1</td><td>L2</td></tr><tr><td rowspan=\"3\">7B</td><td>LLaMA</td><td>53.04</td><td>62.44</td></tr><tr><td>Alpaca</td><td>52.51</td><td>61.71</td></tr><tr><td>Vicuna</td><td>51.90</td><td>61.19</td></tr><tr><td rowspan=\"3\">13B</td><td>LLaMA</td><td>55.66</td><td>64.20</td></tr><tr><td>Alpaca</td><td>55.05</td><td>63.46</td></tr><tr><td>Vicuna</td><td>54.26</td><td>61.31</td></tr></table>",
+        "bbox": [
+            179,
+            261,
+            425,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/4085e71631c817dbdffefdc4becf5988bf96ea1a67e95cf1dab6ebf17dca090b.jpg",
+        "table_caption": [
+            "Table 3: Max layerwise human resemblance scores of pretrained and instruction tuned models. This shows instruction tuning causes small decline in the max layer-wise human resemblance."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Model</td><td colspan=\"2\">Higher</td><td colspan=\"2\">Lower</td></tr><tr><td>Size</td><td>Name</td><td>L1</td><td>L2</td><td>L1</td><td>L2</td></tr><tr><td rowspan=\"2\">7B</td><td>Alpaca</td><td>3</td><td>8</td><td>4</td><td>7</td></tr><tr><td>Vicuna</td><td>9</td><td>10</td><td>8</td><td>9</td></tr><tr><td rowspan=\"2\">13B</td><td>Alpaca</td><td>4</td><td>7</td><td>16</td><td>25</td></tr><tr><td>Vicuna</td><td>5</td><td>8</td><td>14</td><td>16</td></tr></table>",
+        "bbox": [
+            169,
+            464,
+            431,
+            543
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2.4 All Models Show Higher L2 Resemblance",
+        "text_level": 1,
+        "bbox": [
+            112,
+            634,
+            394,
+            667
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Comparing the L1 and L2 human resemblance of all models (LLaMA, Alpaca and Vicuna, 7B and 13B) in Table 2 and Table 3, an advantage of L2 over L1 can be observed steadily. Independent t-test also shows that, all models considered in this research show significantly higher resemblance to L2 than to L1 ( $p = 0.05 / n_{\\mathrm{test}}$  to avoid false positives). This means the models are closer to non-native English learners than to native English speakers in attention, though they are all trained mainly on English data. Furthermore, this trend is not reversed by the scaling or instruction tuning within the scope of this research, which suggests that it is a innate feature of LLMs. We will look further into this phenomenon in the next section.",
+        "bbox": [
+            112,
+            677,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3 Trivial Pattern Reliance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            742,
+            98
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3.1 L2 Relies More on Trivial Patterns",
+        "text_level": 1,
+        "bbox": [
+            507,
+            105,
+            840,
+            120
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The results of LR scores between trivial patterns and L1 and L2 human saccade is shown in Table 5, where L2's score is higher than L1's in minimum, maximum and mean values. Independent t-test also supports a significant advantage of L2 over L1 on the regression scores  $(p < 5 \\times 10^{-8})$ . This means one significant difference between L1 and L2 English speakers is that L2 people show more trivial and fixed attention mode than L1 people while reading, which suggests a weaker language understanding ability as those patterns contain no linguistic or factual information.",
+        "bbox": [
+            505,
+            124,
+            884,
+            318
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This finding can help us interpret the difference between the models' L1 and L2 human resemblance. Note that all models considered in this research are trained mainly on English, but show higher attention resemblance to L2 subjects, this result suggests that their language perception is not ideal.",
+        "bbox": [
+            507,
+            319,
+            884,
+            431
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/ae9923d9e6ca9e111e8a4a0d6a3d06636b71a880b2e59bc9ef35edd71fcd3b5f.jpg",
+        "table_caption": [
+            "Table 4: Numbers of Alpaca and Vicuna layers with significantly higher or lower human resemblance compared with LLaMA in the same sizes."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Group</td><td>Min</td><td>Max</td><td>Mean</td><td>SE</td></tr><tr><td>L1</td><td>0.0048</td><td>0.1525</td><td>0.0473</td><td>0.0037</td></tr><tr><td>L2</td><td>0.0134</td><td>0.2165</td><td>0.0894</td><td>0.0060</td></tr></table>",
+        "bbox": [
+            539,
+            439,
+            852,
+            481
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Regression scores of human saccade on the trivial patterns, across L1 and L2 subjects. SE stands for standard errors.",
+        "bbox": [
+            507,
+            491,
+            882,
+            533
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3.2 Scaling Reduces Trivial Pattern Reliance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            563,
+            880,
+            579
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The relative difference between 13B and 7B models in trivial pattern reliance is shown in Figure 8, where the decline in the deeper layer quarters is substantial across all models. There is also a consistent increase in the second quarter, but the amplitude is small, so the total trivial pattern reliance is reduced. This phenomenon is also observed between LLaMA 13B and 30B (Figure 11 in Appendix B). This means scaling can effectively reduce the trivial pattern reliance, especially in the deeper layers, indicating a more contextualized and adaptive language perception.",
+        "bbox": [
+            505,
+            583,
+            882,
+            778
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3.3 Instruction Tuning Increases Trivial Pattern Reliance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            785,
+            853,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 9 shows the relative difference between pretrained and instruction tuned models in trivial pattern reliance, where the tuned ones show reliance gain in the deeper layers. There is also a small drop in the second quarter, but the total trivial pattern reliance is increasing. Also, total increase here is",
+        "bbox": [
+            507,
+            822,
+            884,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "13049",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/84748b7f55f7fdaea4e031031566ffb84d5ba054996d35d7c775213b9e606b90.jpg",
+        "image_caption": [
+            "Figure 8: The difference between trivial pattern reliance of 7B and 13B model attention."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            86,
+            470,
+            204
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "also smaller in amplitude than the reduction caused by scaling, again supporting the limit effect of instruction tuning.",
+        "bbox": [
+            112,
+            272,
+            487,
+            319
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Also, one can tell from Figure 8 and Figure 9 that scaling and instruction tuning changes the trivial pattern reliance in opposite directions. However, because the effect of instruction tuning is also smaller than that of scaling, as long as the model size continues to grow, this gain in trivial pattern reliance brought by instruction tuning could be compensated or even covered, offering another reason for further scaling.",
+        "bbox": [
+            112,
+            319,
+            487,
+            464
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/2ff27b9802e5b1274efa0fbe3d7d72a0c686dd78e360414feadcbe4bb61bfe6d.jpg",
+        "image_caption": [
+            "Figure 9: The difference in trivial pattern reliance of attentions in pretrained and instruction-tuned models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            126,
+            480,
+            473,
+            599
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            668,
+            247,
+            684
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This research evaluates the effects of scaling and instruction tuning on LLMs' attention. We find that scaling effectively changes the general attention distribution, enhances the human resemblance, and reduces the trivial pattern reliance, while instruction tuning does the opposite, but also increases the models' sensitivity to instructions. Furthermore, we find current open-sourced LLMs closer to non-native English speakers in language perception, which contains more trivial attention patterns. To the best of our knowledge, we are the first to analyze the effect of instruction tuning on the human resemblance of model attention. We hope this can inspire future work on analyzing LLMs.",
+        "bbox": [
+            112,
+            694,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            615,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "One key limitation of this study is that we cannot apply our current method to closed-sourced LLMs such as ChatGPT, for their layerwise attention scores are unavailable. With the rapid development of open LLMs, it is hopeful that we can examine these findings on the largest and most advanced LLMs in the near future.",
+        "bbox": [
+            507,
+            108,
+            884,
+            219
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Besides, though we demonstrate the difference in LLMs' resemblance to L1 and L2 people, and partially explained the effect by trivial pattern reliance, it is not enough to explain the underlying cause or consequences of this phenomenon. We do have some observations, such as the inter-subject correlation in the L1 group  $(0.1560 \\pm 0.0073)$  is lower than L2  $(0.2463 \\pm 0.0090)$ , and L1 subjects tend to look back to pervious words less frequently than L2 subjects. However, the underlying mechanism is still unclear. We hope to dive deeper into this non-native resemblance in our following study.",
+        "bbox": [
+            507,
+            221,
+            884,
+            414
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Another limitation of this work comes from the way we obtain model attention. We used the attention scores when predicting the next single token, which makes sense. However, the attention patterns could be different in the following decoding steps. Also, it is possible that combining attention scores across multiple layers can improve the analysis. We will investigate this in our future work.",
+        "bbox": [
+            507,
+            414,
+            882,
+            541
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In addition, the Reading Brain dataset we used in this research is monolingual, which does not reflect the language perception process in the cross-lingual setting, which is also important for the application of LLMs. We also hope to examine our findings in the cross-lingual scenario in the future.",
+        "bbox": [
+            507,
+            542,
+            882,
+            639
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            650,
+            660,
+            665
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The authors declare no competing interests. The human behavioral data we use is publicly available and does not contain personal information of the subjects.",
+        "bbox": [
+            507,
+            674,
+            880,
+            739
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            750,
+            680,
+            766
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We would like to thank the anonymous reviewers for their insightful comments. Shujian Huang is the corresponding author. We thank Yunzhe Lv and Wenhao Zhu for finetuning the Alpaca 13B model. This work is supported by National Science Foundation of China (No. 62376116, 62176120), the Liaoning Provincial Research Foundation for Basic Research (No. 2022-KF-26-02).",
+        "bbox": [
+            507,
+            774,
+            882,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "13050",
+        "bbox": [
+            477,
+            927,
+            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": [
+            "Samira Abnar and Willem Zuidema. 2020. Quantifying Attention Flow in Transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, Online. Association for Computational Linguistics.",
+            "Aakash Bansal, Bonita Sharif, and Collin McMillan. 2023. Towards modeling human attention from eye movements for neural source code summarization. Proceedings of the ACM on Human-Computer Interaction, 7(ETRA):1-19.",
+            "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.",
+            "Charlotte Caucheteux, Alexandre Gramfort, and JeanRémi King. 2021. Model-based analysis of brain activity reveals the hierarchy of language in 305 subjects. arXiv preprint arXiv:2110.06078.",
+            "Charlotte Caucheteux, Alexandre Gramfort, and JeanRémi King. 2022. Deep language algorithms predict semantic comprehension from brain activity. Scientific Reports, 12(1):16327.",
+            "Emmanuele Chersoni, Nora Hollenstein, Cassandra Jacobs, Yohei Oseki, Laurent Prévot, and Enrico Santus, editors. 2022. Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics. Association for Computational Linguistics, Dublin, Ireland.",
+            "Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An opensource chatbot impressing gpt-4 with  $90\\%$  * chatgpt quality.",
+            "Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. 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.",
+            "Javier Ferrando, Gerard I. Gállego, and Marta R. Costajussà. 2022. Measuring the Mixing of Contextual Information in the Transformer. In Proceedings of"
+        ],
+        "bbox": [
+            115,
+            105,
+            487,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8698-8714, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Javier Ferrando, Gerard I. Gallego, Ioannis Tsiamas, and Marta R. Costa-jussa. 2023. Explaining How Transformers Use Context to Build Predictions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5486-5513, Toronto, Canada. Association for Computational Linguistics.",
+            "Uri Hasson, Samuel A. Nastase, and Ariel Goldstein. 2020. Direct fit to nature: An evolutionary perspective on biological and artificial neural networks. *Neuron*, 105(3):416-434.",
+            "Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B. Brown, Prafulla Dhariwal, Scott Gray, Chris Hallacy, Benjamin Mann, Alec Radford, Aditya Ramesh, Nick Ryder, Daniel M. Ziegler, John Schulman, Dario Amodei, and Sam McCandlish. 2020. Scaling Laws for Autoregressive Generative Modeling. ArXiv:2010.14701 [cs].",
+            "Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. 2017. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409.",
+            "Nora Hollenstein and Lisa Beinborn. 2021. Relative Importance in Sentence Processing. 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 141-150, Online. Association for Computational Linguistics.",
+            "Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. 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 3543-3556, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Alexander JE Kell, Daniel LK Yamins, Erica N Shook, Sam V Norman-Haignere, and Josh H McDermott. 2018. A task-optimized neural network replicates human auditory behavior, predicts brain responses, and reveals a cortical processing hierarchy. Neuron, 98(3):630-644.",
+            "Goro Kobayashi, Tatsuki Kuribayashi, Sho Yokoi, and Kentaro Inui. 2020. Attention is Not Only a Weight: Analyzing Transformers with Vector Norms. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7057-7075, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "13051",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Goro Kobayashi, Tatsuki Kuribayashi, Sho Yokoi, and Kentaro Inui. 2021. Incorporating Residual and Normalization Layers into Analysis of Masked Language Models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4547-4568, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics.",
+            "Mathis Lamarre, Catherine Chen, and Fatma Deniz. 2022. Attention weights accurately predict language representations in the brain. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4513-4529, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Ping Li, Anya Yu, Benjamin Schloss, Chun-Ting Hsu, Lindsey Ma, Chih-Ting Chang, and Marissa Scotto. 2019. The reading brain project: An open science data-sharing initiative. In 26th Annual Meeting of the Cognitive Neuroscience Society.",
+            "Christopher D Manning, Kevin Clark, John Hewitt, Urvashi Khandelwal, and Omer Levy. 2020. Emergent linguistic structure in artificial neural networks trained by self-supervision. Proceedings of the National Academy of Sciences, 117(48):30046-30054.",
+            "David Mareček and Rudolf Rosa. 2018. Extracting syntactic trees from transformer encoder self-attentions. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 347-349, Brussels, Belgium. Association for Computational Linguistics.",
+            "Danny Merkx and Stefan L. Frank. 2021. Human sentence processing: Recurrence or attention? In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, pages 12-22, Online. Association for Computational Linguistics.",
+            "Hosein Mohebbi, Willem Zuidema, Grzegorz Chrupa\\la, and Afra Alishahi. 2023. Quantifying Context Mixing in Transformers. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3378-3400, Dubrovnik, Croatia. Association for Computational Linguistics.",
+            "Felix Morger, Stephanie Brandl, Lisa Beinborn, and Nora Hollenstein. 2022. A Cross-lingual Comparison of Human and Model Relative Word Importance. In Proceedings of the 2022 CLASP Conference on (Dis)embodiment, pages 11-23, Gothenburg, Sweden. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Byung-Doh Oh and William Schuler. 2023a. Transformer-Based LM Surprisal Predicts Human Reading Times Best with About Two Billion Training Tokens.",
+            "Byung-Doh Oh and William Schuler. 2023b. Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times? Transactions of the Association for Computational Linguistics, 11:336-350.",
+            "Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. ArXiv:2203.02155 [cs].",
+            "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners.",
+            "Alessandro Raganato and Jörg Tiedemann. 2018. An analysis of encoder representations in transformer-based machine translation. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 287-297, Brussels, Belgium. Association for Computational Linguistics.",
+            "Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, Manan Dey, M. Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevry, Jason Alan Fries, Ryan Teehan, Tali Bers, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M. Rush. 2022. Multitask Prompted Training Enables Zero-Shot Task Generalization. ArXiv:2110.08207 [cs].",
+            "Martin Schrimpf, Idan Asher Blank, Greta Tuckute, Carina Kauf, Eghbal A Hosseini, Nancy Kanwisher, Joshua B Tenenbaum, and Evelina Fedorenko. 2021. The neural architecture of language: Integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences, 118(45):e2105646118.",
+            "Sofia Serrano and Noah A. Smith. 2019. Is attention interpretable? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2931-2951, Florence, Italy. Association for Computational Linguistics.",
+            "Ekta Sood, Simon Tannert, Philipp Müller, and Andreas Bulling. 2020. Improving natural language processing tasks with human gaze-guided neural attention."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "13052",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Advances in Neural Information Processing Systems, 33:6327-6341.",
+        "bbox": [
+            131,
+            85,
+            487,
+            112
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.",
+        "bbox": [
+            114,
+            121,
+            487,
+            189
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models.",
+        "bbox": [
+            114,
+            197,
+            487,
+            278
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Shikhar Vashisth, Shyam Upadhyay, Gaurav Singh Tomar, and Manaal Faruqui. 2019. Attention interpretability across nlp tasks. arXiv preprint arXiv:1909.11218.",
+        "bbox": [
+            114,
+            286,
+            487,
+            338
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Jesse Vig and Yonatan Belinkov. 2019. Analyzing the structure of attention in a transformer language model. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 63-76, Florence, Italy. Association for Computational Linguistics.",
+        "bbox": [
+            114,
+            348,
+            487,
+            429
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Elena Voita, David Talbot, Fedor Moiseev, Rico Senrich, and Ivan Titov. 2019. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797-5808, Florence, Italy. Association for Computational Linguistics.",
+        "bbox": [
+            114,
+            437,
+            487,
+            530
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Titus von der Malsburg, Reinhold Kliegl, and Shravan Vasisith. 2015. Determinants of Scanpath Regularity in Reading. Cognitive Science, 39(7):1675-1703.",
+        "bbox": [
+            114,
+            539,
+            487,
+            580
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Titus von der Malsburg and Shravan Vasishth. 2013. Scanpaths reveal syntactic underspecification and reanalysis strategies. Language and Cognitive Processes, 28(10):1545-1578.",
+        "bbox": [
+            114,
+            588,
+            487,
+            642
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Jeffrey J. Walczyk and Robert W. Taylor. 1996. How do the efficiencies of reading subcomponents relate to looking back in text? Journal of Educational Psychology, 88(3):537-545.",
+        "bbox": [
+            114,
+            651,
+            487,
+            705
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-Instruct: Aligning Language Models with Self-Generated Instructions. ArXiv:2212.10560 [cs].",
+        "bbox": [
+            114,
+            713,
+            487,
+            780
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "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.",
+        "bbox": [
+            114,
+            789,
+            487,
+            854
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Sarah Wiegrefe and Yuval Pinter. 2019. Attention is not explanation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference",
+        "bbox": [
+            114,
+            865,
+            487,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "on Natural Language Processing (EMNLP-IJCNLP), pages 11-20, Hong Kong, China. Association for Computational Linguistics.",
+        "bbox": [
+            526,
+            85,
+            882,
+            126
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Attention divergence on plain and noise-prefix text",
+        "text_level": 1,
+        "bbox": [
+            509,
+            137,
+            840,
+            170
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 10, the divergence between plain and noise-prefixed text of Alpaca and Vicuna is not significantly higher than LLaMA in most of the layers, which further supports our findings on the higher instruction sensitivity of instruction-tuned models.",
+        "bbox": [
+            507,
+            179,
+            882,
+            275
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/7d361eaa7e2230b259f47c0ed24a4ab897787020979f47f9ce6be3b4a8349f08.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            288,
+            867,
+            407
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/ac4cc8bc037770a46c7e04cf00ef0adbaf0b2bc35b90d9bec9e403f3150e4f17.jpg",
+        "image_caption": [
+            "Figure 10: Attention divergence on the original and noise-attached text of 7B and 13B models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            416,
+            867,
+            536
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Layerwise trivial pattern reliance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            609,
+            835,
+            627
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Figure 11 shows the layerwise trivial pattern reliance of 7B and 13B models, as well as the results for LLaMA 30B. The reliance keeps decreasing as the model size scales up, especially in the deeper layers.",
+        "bbox": [
+            507,
+            634,
+            882,
+            715
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Layers with significant changes in human resemblance after instruction tuning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            727,
+            867,
+            777
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 6 shows the list of layers in 7B Alpaca and Vicuna with significant change in human resemblance after instruction tuning compared with LLaMA, and Table 7 shows the list in 13B.",
+        "bbox": [
+            507,
+            785,
+            882,
+            848
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Preliminary fMRI results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            860,
+            766,
+            878
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Some preliminary analysis on the fMRI data in the Reading Brain dataset is done, and the results com",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "13053",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/cd785dcab0e4b4d8fece282b3d909f8c5494973384c746b3858dfc41b73dd3a3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            86,
+            470,
+            204
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/5f865bb3eb9bf6a0136938d04e2bf5dfd817900983c804abd586a8c0872426d4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            214,
+            470,
+            332
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/a1e9b5c7d35de8515d3cc9eca08ad8e52b0abd225302e39b6bb068f57d512ebc.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            341,
+            470,
+            458
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/53ffc9bdd996fa8ab968753798918e4d1a2b1f9c1669e83d125ac79344337a2e.jpg",
+        "image_caption": [
+            "Figure 11: Lyerwise trivial pattern reliance of 7B and 13B models. \"RB\" stands for \"Reading Brain\"."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            468,
+            470,
+            586
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "ply with our findings on saccade data. We investigate several linguistics-related brain regions, and find that the highest correlation of fMRI signals to model attention are from the anterior left temporal gyrus (Figure 12). The fMRI data from this region is pre-processed into matrix forms as saccade data. Figure 14 gives examples of individual and group mean fMRI matrices.",
+        "bbox": [
+            112,
+            659,
+            487,
+            787
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The findings are: (1) The models' NTP loss significantly  $(p < 0.02$  for both L1 and L2) and strongly negatively  $(r = -0.777$  for L1,  $r = -0.796$  for L2) correlates with fMRI resemblance (Figure 13); (2) The models' highest layerwise fMRI resemblance significantly  $(r > 0.99$ ,  $p < 0.001$  for both L1 and L2) correlates with the model scale in logarithm (Figure 15, 16); (3) While scal",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/98f6eabe13bbbcaaa71692d9c062537ad4a582afa75d71260d1750645b61bb00.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Relation</td><td>Group</td><td>Alpaca</td><td>Vicuna</td></tr><tr><td rowspan=\"2\">Higher</td><td>L1</td><td>2, 5, 7</td><td>12, 13, 17, 18, 21, 26, 28, 29, 30</td></tr><tr><td>L2</td><td>2, 5, 6, 8, 12, 15, 21, 27</td><td>3, 13, 15, 18, 21, 23, 26, 28, 29, 30</td></tr><tr><td rowspan=\"2\">Lower</td><td>L1</td><td>4, 10, 23, 31</td><td>0, 1, 4, 5, 9, 11, 20, 31</td></tr><tr><td>L2</td><td>0, 10, 14, 19, 23, 28, 31</td><td>0, 1, 4, 5, 9, 11, 14, 20, 31</td></tr></table>",
+        "bbox": [
+            549,
+            80,
+            843,
+            252
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/3546e27fabbe3e99b06ebcb5a26abf1d1f51689a16c8fdc4516ca21f7ac151d7.jpg",
+        "table_caption": [
+            "Table 6: Layers of Alpaca and Vicuna with significantly higher or lower human resemblance compared with LLaMA in 7B."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Relation</td><td>Group</td><td>Alpaca</td><td>Vicuna</td></tr><tr><td rowspan=\"2\">Higher</td><td>L1</td><td>16, 32, 33, 36</td><td>4, 16, 27, 33, 36</td></tr><tr><td>L2</td><td>1, 16, 19, 27, 32, 33, 36</td><td>4, 16, 19, 23, 27, 33, 36, 37</td></tr><tr><td rowspan=\"2\">Lower</td><td>L1</td><td>0, 2, 3, 5, 6, 7, 8, 9, 14, 18, 24, 25, 28, 30, 35, 39</td><td>0, 1, 3, 5, 6, 9, 10, 14, 17, 18, 24, 30, 34, 39</td></tr><tr><td>L2</td><td>0, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 17, 18, 20, 24, 25, 28, 29, 30, 31, 35, 38, 39</td><td>0, 2, 3, 5, 6, 8, 9, 10, 11, 12, 14, 17, 18, 20, 21 30</td></tr></table>",
+        "bbox": [
+            544,
+            316,
+            848,
+            571
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Layers of Alpaca and Vicuna with significantly higher or lower human resemblance compared with LLaMA in 13B.",
+        "bbox": [
+            507,
+            580,
+            882,
+            621
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "ing enhances human fMRI resemblance, instruction tuning does not (Table 8); (4) Resemblance to L2 fMRI is consistently higher than that to L1.",
+        "bbox": [
+            507,
+            649,
+            882,
+            697
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "13054",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/1a4c004f619dcbabbb6a01164539845a5be779d88fd6a9573642a97e701507bc.jpg",
+        "image_caption": [
+            "Figure 12: An illustration of the anterior left temporal gyrus in the human brain."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            216,
+            99,
+            386,
+            190
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/611ee90ccf0d1d563394ecd2a127465e030f92ec3c59ad1b78e1b6be9d913769.jpg",
+        "image_caption": [
+            "Figure 13: The correlation between per-token negative log likelihood loss and max layerwise fMRI resemblance of the LLMs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            271,
+            473,
+            401
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/b86120adce5637b55705e69da3c6052c84efe7e4f688f95ff66d6ce6ab2e23b4.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Model</td><td colspan=\"2\">Resemblance(%)</td></tr><tr><td>Size</td><td>Name</td><td>L1</td><td>L2</td></tr><tr><td rowspan=\"3\">7B</td><td>LLaMA</td><td>11.02</td><td>15.45</td></tr><tr><td>Alpaca</td><td>10.99</td><td>15.61</td></tr><tr><td>Vicuna</td><td>11.35</td><td>15.44</td></tr><tr><td rowspan=\"3\">13B</td><td>LLaMA</td><td>14.69</td><td>20.19</td></tr><tr><td>Alpaca</td><td>14.67</td><td>19.44</td></tr><tr><td>Vicuna</td><td>14.02</td><td>18.57</td></tr></table>",
+        "bbox": [
+            179,
+            495,
+            423,
+            595
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Max layerwise human fMRI resemblance scores of pretrained and instruction tuned models. This shows instruction tuning causes small decline in the max layerwise human resemblance.",
+        "bbox": [
+            112,
+            605,
+            487,
+            662
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/76078bf11392e7075e08e4d424292ba9479a10db75ea03f84b13b04a08bd7b82.jpg",
+        "image_caption": [
+            "(a) Individual"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            703,
+            295,
+            826
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/3469d2471ded04cf9bd91703fbef916897abea2483e4faacbc0100147bc18e4e.jpg",
+        "image_caption": [
+            "(b) Group Mean"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            305,
+            705,
+            480,
+            826
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/05568093f276b55f2bbc25cf0fffd1bd810c844632c23eee1700f0096a937308.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            172,
+            870,
+            293
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd49a90499d2eec64f51f9d08fdb560f781e3d274b893ee0bc88ed334b62f01f.jpg",
+        "image_caption": [
+            "Figure 15: Layerwise L1 and L2 human fMRI resemblance for different sized pretrained models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            300,
+            870,
+            420
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/6578ebffd275ffd9cf5335aff9398a942e7331af325d98a028f3ca51e4ea9f49.jpg",
+        "image_caption": [
+            "Figure 14: Examples of individual and group mean fMRI signals of the sentence \"Could humans live on Mars some day?\"",
+            "Figure 16: The change of max layerwise human fMRI resemblance caused by scaling. The x-axis is log scaled."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            653,
+            870,
+            782
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "13055",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/Roles of Scaling and Instruction Tuning in Language Perception_ Model vs. Human Attention/4d81e6a0-662e-4a1d-8420-5bca7d9f2692_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c728f693932cda7c5a0ca93de291fb8662f02fa131e33a9c22961914d62169d
+size 3516051

2023/Roles of Scaling and Instruction Tuning in Language Perception_ Model vs. Human Attention/full.md

@@ -0,0 +1,420 @@
+# Roles of Scaling and Instruction Tuning in Language Perception: Model vs. Human Attention
+
+Changjiang Gao<sup>1</sup> Shujian Huang<sup>1*</sup> Jixing Li<sup>2</sup> Jiajun Chen<sup>1</sup>
+
+$^{1}$ National Key Laboratory for Novel Software Technology, Nanjing University
+
+$^{2}$ Department of Linguistics and Translation, City University of Hong Kong gaocj@smail.nju.edu.cn {huangsj, chenjj}@nju.edu.cn
+
+jixingli@cityu.edu.hk
+
+# Abstract
+
+Recent large language models (LLMs) have revealed strong abilities to understand natural language. Since most of them share the same basic structure, i.e. the transformer block, possible contributors to their success in the training process are scaling and instruction tuning. However, how these factors affect the models' language perception is unclear. This work compares the self-attention of several existing LLMs (LLaMA, Alpaca and Vicuna) in different sizes (7B, 13B, 30B, 65B), together with eye saccade, an aspect of human reading attention, to assess the effect of scaling and instruction tuning on language perception. Results show that scaling enhances the human resemblance and improves the effective attention by reducing the trivial pattern reliance, while instruction tuning does not. However, instruction tuning significantly enhances the models' sensitivity to instructions. We also find that current LLMs are consistently closer to non-native than native speakers in attention, suggesting a suboptimal language perception of all models. Our code and data used in the analysis is available on GitHub.
+
+# 1 Introduction
+
+Large language models (LLMs), e.g., GPT-4, ChatGPT and Vicuna, have shown nearly human-level understanding of text inputs, indicating better human-like language perception compared with their predecessors such as BERT (Devlin et al., 2019) and GPT-2 (Radford et al., 2019). However, the mechanism behind such improvement is largely unknown. One way to interpret such mechanism is comparing model computation processes to human data. For example, prior work compared language models (mainly GPT-2) with human neuroimaging data during language comprehension, suggesting that models with higher human resemblance also perform better in NLP tasks. (Hasson
+
+et al., 2020; Schrimpf et al., 2021). However, given recent breakthroughs in LLMs, it remains to be tested whether the newest LLMs still align with human perception data. Based on the training pipeline of current LLMs, there are two possible sources for the potential change: the scaled model/data size and the aligning process after pretraining, such as instruction tuning (Ouyang et al., 2022; Wei et al., 2022; Sanh et al., 2022). This paper aims to provide a further understanding of LLMs' mechanism by investigating the roles of these two factors in language understanding.
+
+In this paper, we analyze self-attention as a channel to the language perception ability of LLMs, because it is the key mechanism of transformer language models, and it naturally resembles the human attention mechanism in language perceiving (Merkx and Frank, 2021). The attention patterns under investigation includes those from trending open-sourced LLMs with different sizes, i.e., 7B, 13B, 30B, 65B, and different training stages, i.e., pretrained (LLaMA) and instruction-tuned (Alpaca and Vicuna) models. Given these models are all trained mainly with English data, we also compare the attention patterns of LLMs with human saccade (a form of eye-movement) data, including both native speakers (L1) and non-native learners (L2) of English (Li et al., 2019) to see if the models correlates more with L1 subjects.
+
+Our analysis is three-fold. First, we compare the general attention distributions of LLMs to measure the impact of these two varying factors on the model ( $\S 3.1$ ). Second, we perform linear regressions (LRs) between the model self-attention and human saccade matrices to see how human-like they are, and how the resemblance is affected by the two factors ( $\S 3.2$ ). Third, we analyze the relation between the given attention and some trivial patterns, revealing the difference between different levels of language perception abilities ( $\S 3.3$ ). The main findings are:
+
+- Scaling significantly affects the general attention distribution on plain text, while instruction tuning has limited effect. However, instruction tuning enhances the model's sensitivity to instructions (§5.1).  
+- Higher human resemblance is significantly correlated to better language modeling. Scaling improves the resemblance by scaling law (Henighan et al., 2020), while instruction tuning reduces it. All models have higher resemblance to L2 rather than to L1, suggesting further room for the improvement in language perception (§5.2).  
+- L2 saccade has higher dependency on trivial attention patterns than L1, indicating a more rigid way of perception. Scaling significantly lowers LLMs' reliance on trivial attention patterns, while instruction tuning does not (§5.3).
+
+# 2 Related Work
+
+Self-attention analysis is a common way to interpret transformer models. Several studies has shown that it is correlated to linguistic information, such as syntactic or semantic structures (Mareček and Rosa, 2018; Raganato and Tiedemann, 2018; Voita et al., 2019; Clark et al., 2019). There are also some attempts and debates to explain the models' prediction with attention (Jain and Wallace, 2019; Wegreffe and Pinter, 2019; Serrano and Smith, 2019; Vashisth et al., 2019). Different from them, we use self-attention patterns as a measurement of the language perception abilities of LLMs. Also, some new ways to interpret model self-attention have been proposed, such as Attention Flow (Abnar and Zuidema, 2020), vector norm (Kobayashi et al., 2020, 2021), ALTI (Ferrando et al., 2022), Value Zeroing (Mohebbi et al., 2023), and logit analysis (Ferrando et al., 2023). These methods can raise the interpretability of model attention w.r.t. linguistic structural information, but are not tested with human eye-tracking data. Also, some of them suffer from huge computational cost. Instead, we choose to use the raw attention scores, as it is produced in the models' original computational processes, and they already correlate the human data, demonstrating their interpretability.
+
+Instruction tuned LLMs achieve better performances in performing tasks (Ouyang et al., 2022; Wei et al., 2022; Sanh et al., 2022). It seems that the
+
+LLMs are better at understanding human instructions. However, there is little discussion on how the instruction tuning process affects the language perception.
+
+Eye-movement in the human reading process has drawn attention in the computational linguistic field. It is proposed that scanpaths, i.e. tracks of eye-movement, reveals characteristics of the parsing processes and relates to working memory during reading (von der Malsburg and Vasishth, 2013), and the irregularity of scanpaths is affected by sentence structures and age (von der Malsburg et al., 2015). Such work provides theoretical ground for our analysis on saccade data.
+
+Joint study of language models and human eye-movement is also receiving growing interest. Prior studies have shown the correlation between language processing of human and artificial neural networks (Kell et al., 2018), especially pretrained language models such as GPT-2 (Caucheteux et al., 2021, 2022; Schrimpf et al., 2021; Lamarre et al., 2022). Especially, Oh and Schuler (2023a,b) explored the fit of model surprisal to human reading times, noting the effect of models' sequential memorization and dataset size. Hollenstein and Beinborn (2021) studied the correlation between relative word importance derived from models and humans, and Morger et al. (2022) extended that research to a multilingual setting, as well as analysis on model attention. Some studies also use human attention as supervision to assist specific tasks (Sood et al., 2020; Bansal et al., 2023), or use sequential models to predict eye-tracking data (Chersoni et al., 2022). Different from them, we use human attention as references for the language perception abilities.
+
+# 3 Methods
+
+To show the effect of scaling and instruction tuning on different models, we compare the self-attention scores of different LLMs given the same input, by viewing model attention as probability distributions and calculating the general attention divergence based on Jensen-Shannon divergence.
+
+To analyze model self-attention, we take the human saccade as a reference, and design a human resemblance metric based on linear regression scores.
+
+In addition, we select several trivial, context-free attention patterns, and design a trivial pattern reliance metric to help demonstrate the difference
+
+between models and human subject groups.
+
+# 3.1 General Attention Divergence
+
+We assess the mean Jensen-Shannon (J-S) divergence between attention from different models on all sentences. For two attention matrices  $A$ ,  $B \in \mathbb{R}^{n_{\mathrm{word}} \times n_{\mathrm{word}}}$ , the J-S divergence is:
+
+$$
+\begin{array}{l} D _ {\mathrm {J S}} (A, B) = \frac {1}{2} \sum_ {i = 1} ^ {n _ {\text {w o r d}}} \left[ D _ {\mathrm {K L}} \left(A _ {i} \| B _ {i}\right) \right. \\ + D _ {\mathrm {K L}} \left(B _ {i} \| A _ {i}\right) ] \\ \end{array}
+$$
+
+where  $D_{\mathrm{KL}}$  is the Kullback-Leibler (K-L) divergence,  $A_{i}$  and  $B_{i}$  are the  $i$ -th rows in the two matrices. Note the attention matrices are re-normalized to keep the sum of every row to 1.0. We calculate the J-S divergence between heads-averaged attention scores per sentence, and demonstrate the mean values. For models with the same number of layers, the divergence is calculated layerwise; For models with different numbers of layers, we divide their layers into four quarters and calculate the divergence quarter-wise.
+
+To identify high and low divergence, a reference value is required. We propose to use the attention divergence between two models that share the same structure, size, and training data, but have small difference in training strategies as the reference. A divergence value higher than the reference suggests relatively large change attention, while one lower than that indicates small (but not necessarily negligible) change.
+
+Divergence by Scaling and Instruction Tuning Models with different scales or training stages are compared with the above divergence to show the effect of different factors.
+
+Sensitivity to Instructions To evaluate the effect of instruction tuning, we measure a given model's sensitivity to instructions by the same general attention divergence metric, but on different input, i.e., plain input and input with instructions. If the divergence between them is significantly higher than reference, we say the model has high sensitivity to instructions.
+
+To construct instructed text, we attach two prefixes to plain text sentences: "Please translate this sentence into German:", and "Please paraphrase this sentence":". They are within the reported capacity of most LLMs, and require no extra information to follow. Then, we collect model attention on
+
+the two corpora within the original sentence spans (re-normalized) to represent non-instruction and instruction scenarios. Also, as a control group, we attached a noise prefix made of 5 randomly sampled English words "Cigarette first steel convenience champion.", and calculate the divergence between prefixed and plain text to account for the effect of a meaningless prefix.
+
+# 3.2 Human Resemblance
+
+# 3.2.1 Human Resemblance Calculation
+
+Given the model self-attention and the human saccade data in matrix form for each sentence, we first extract the lower-triangle parts of the matrices (marking right-to-left attendance), because prior studies have found that such eye-movements in reading are related to working memory efficiency (Walczyk and Taylor, 1996) and parsing strategies (von der Malsburg and Vasishth, 2013), and the auto-regressive LLMs only have right-to-left self-attention. Then, we flatten the extracted parts of each sentence attention matrix, and concatenate them to get attention vectors  $\{v_{\mathrm{human}}^i\}$  for each subject  $i$ , and  $\{v_{\mathrm{model}}^{j,k}\}$  for head  $k$  in model layer  $j$ . We then stack  $\{v_{\mathrm{model}}^{j,k}\}$  along the attention heads and get a matrix  $\{V_{\mathrm{model}}^j\}$  for the  $j$ -th layer. The LR for model layer  $j$  and subject  $i$  is defined as:
+
+$$
+\arg \min _ {w, b} \left\| V _ {\mathrm {m o d e l}} ^ {j \top} w + b - v _ {\mathrm {h u m a n}} ^ {i} \right\| _ {2} ^ {2}
+$$
+
+This means we take the linear combination of attention scores from all heads in a layer to predict the human saccade. Note that the model attention here are not re-normalized, in order to preserve the relative weight w.r.t. the  $<s>$  token. The regression scores  $R_{i,j}^{2}$  (ranging from 0 to 1, larger value means better regression) are averaged over subjects to represent the human resemblance of layer  $j$ , and the highest layer score is considered as the regression score of the whole model,  $R_{\mathrm{model}}^{2}$ .
+
+Because the human saccade pattern varies across individuals, there exists a natural upper bound for human resemblance of models, i.e., the human-human resemblance, or the inter-subject correlation. To calculate this correlation within the two language groups, L1 and L2, we take the group mean to regress over each individual data independently, and use the average regression score  $R_{\mathrm{inter}}^2$  as the inter-subject correlation for that group. The final human resemblance score is  $R_{\mathrm{model}}^2 / R_{\mathrm{inter}}^2$ , ranging from 0 to 100%.
+
+# 3.2.2 Human Resemblance vs. Next-Token Prediction
+
+A previous study finds a positive correlation between the next-token predication (NTP) performance and the "Brain Score" of neural networks (Schrimpf et al., 2021), which supports the intuition that the more human-like models are, the better generation they output. To test this finding in our setting, we calculated the per-token prediction loss (the negative log likelihood loss normalized by sequence lengths) of the models employed in this study on the Reading Brain dataset, and calculate the Pearson's Correlation between their prediction loss and max layerwise human resemblance scores.
+
+# 3.3 Trivial Pattern Reliance
+
+# 3.3.1 Trivial Pattern Reliance of Models
+
+We select three representative trivial patterns of transformer attention: attending to the first word in the sentence, attending to the previous word (Vig and Belinkov, 2019), and self-attending (Clark et al., 2019), and analyze the effect of scaling and instruction tuning on models' reliance on them.
+
+First, we represent the trivial patterns as binary adjacent matrices, where attending relations are marked with 1, others with 0, and use their flattened lower-triangle parts as the independent variables. The model attention is similarly processed to be the dependent variables. Then, an LR is fit between the two, and the regression scores are collected to represent the trivial pattern reliance of each model layer.
+
+# 3.3.2 Trivial Pattern Reliance of Humans
+
+This research compares the models' resemblance to L1 and L2 humans because intuitively, L1 is better than L2 in language understanding. To test this intuition, we compare their reliance on the three trivial patterns to see whether their attention mode differs. Similar to model attention, the flattened human saccade vectors are used as the dependent variable to fit the LRs with the trivial patterns.
+
+# 4 Experiment Settings
+
+# 4.1 Data
+
+We use the Reading Brain dataset (Li et al., 2019) for both the text input and the human data. The dataset includes 5 English STEM articles. Each article has  $29.6 \pm 0.68$  sentences of  $10.33 \pm 0.15$  words. It also includes human eye-tracking and fMRI data recorded synchronously during self-paced reading
+
+![](images/72af1df936c3ae4afd689692a37f326c1690dfeaa2151877027d8fb5cc2d467c.jpg)  
+(a) Individual  
+Figure 1: Examples of individual and group mean saccade of the sentence "Could humans live on Mars some day?"
+
+![](images/00e40331f9f9f846d20f3c9cd578170c07c200d1400a103ffaf58bb6603b4d7c.jpg)  
+(b) Group Mean
+
+on the articles. The human subjects are 52 native English speakers (L1) and 56 non-native learners of English (L2). We choose this dataset for the following reasons: (1) The text is presented to the subjects sentence by sentence instead of page by page, making in-sentence reading patterns clearer. (2) It has synchronous eye-tracking and fMRI data, which allows both behavioral- and neural-level analysis. In this article we mainly focus on the behavioral part, but we also have some preliminary results on the fMRI analysis (See Appendix D).
+
+We take the saccade number, i.e. times of eye movements from a word to another, instead of saccade duration, to represent the human reading attention, because this reduces the effect of other factors such as word length. The saccade data of a human subject for a given sentence is a  $n_{\mathrm{word}} \times n_{\mathrm{word}}$  matrix. Figure 1 gives examples of individual and group mean saccades.
+
+# 4.2 LLMs
+
+We employ LLaMA (Touvron et al., 2023) and its instruction-tuned versions, Alpaca (Taori et al., 2023) and Vicuna (Chiang et al., 2023) in this research. We use them because: (1) The model sizes and different instruction-tuning methods covers the majority of current open-sourced LLMs. (2) These models are being widely used and customized globally. Thus, analysis on them is representative and meaningful.
+
+LLaMA is a series of pretrained causal language models with parameter sizes 7B, 13B, 30B and 65B, trained on over 1T publicly available text tokens, and reaches state-of-the-art on most LLM benchmarks (Touvron et al., 2023). The training corpus is mainly English. We use all 4 sizes of LLaMA, which covers the sizes of most current
+
+open-sourced LLMs. We also use the 774M GPT-2 Large (Radford et al., 2019) to represent smaller pretrained models in the scaling analysis.
+
+Alpaca is fine-tuned from the 7B LLaMA model on 52K English instruction-following demonstrations generated by GPT-3 (Brown et al., 2020) using self-instruct (Wang et al., 2023). To analyze the effect of scaling, we also trained a 13B version of Alpaca using the official data and training strategy of Alpaca. Our 13B Alpaca model scores 43.9 and 46.0 on the MMLU dataset in zero-shot and one-shot setting respectively, proving the soundness of our fine-tuning. (The corresponding scores are 40.9 and 39.2 for the official Alpaca 7B model.)
+
+Vicuna models are fine-tuned from the 7B and 13B LLaMA models on 70K user-shared conversations with ChatGPT. The data contains instruction and in-context learning samples in multiple languages, thus the Vicuna models can also be viewed as instruction-tuned.
+
+# 4.3 LLMs Attention Collection
+
+To obtain model attention, text is provided to the models sentence by sentence to perform next token predict, and we collect the attention tensors in each layer. This resembles the human self-paced task-free reading process, where only one sentence is visible at a time to the subjects, and the natural human behavior is to predict the next word (Kell et al., 2018; Schrimpf et al., 2021).
+
+Because models use the SentencePiece tokenization (Kudo and Richardson, 2018), but the human saccade data is recorded in words, the attention matrices need to be reshaped. Following Clark et al. (2019) and Manning et al. (2020), we take the sum over the "to" tokens, and the average over the "from" tokens in a split word. Also, the  $<s>$  token is dropped as it is not presented to human subjects.
+
+Reference of Divergence We use the attention divergence of Vicuna v0 and Vicuna v1.1 13B as the reference values. The two versions share the same training data and strategy, but have minor difference in separators and loss computation<sup>2</sup>. We use the divergence calculated on 13B Vicunas across all divergence analyses, because it has lower mean and variance than the 7B one, making our comparison fairer and stricter.
+
+# 5 Results
+
+![](images/ec644743c14c8077c18e304cd43e115de1a02fc57935327bff3bdeb62939a0c4.jpg)  
+Figure 2: Mean J-S divergence between 7B and 13B model attention in each layer quarter, measured on non-instruction sentences.
+
+# 5.1 General Attention Divergence
+
+# 5.1.1 Scaling Significantly Changes the General Attention
+
+The general attention divergence between models in different sizes is shown in Figure 2. The results show that LLaMA, Alpaca and Vicuna all reveal divergence significantly higher than the reference values when scaled from 7B to 13B. This means scaling causes large change in the overall attention distribution.
+
+# 5.1.2 Instruction Tuning Has Limited Effect on General Attention
+
+The general attention divergence between pretrained and instruction tuned models is shown in Figure 3. It can be observed that only Vicuna 13B brings divergence above the reference values, while all the others don't. The same result can also be observed on instruction-prefixed sentences. This suggests that instruction tuning can only bring limited change to the models' general attention distribution. This can also be verified by other results in the following parts.
+
+# 5.1.3 Instruction Tuning Enhances Sensitivity to Instructions
+
+The divergence between the models' attention obtained on the original and instruction-attached sentences are shown in Figure 4. There are two major observations. Firstly, LLaMA, Alpaca and Vicuna all show divergence significantly above the reference values between the two types of sentences, especially in the shallower layers, which means they all have sensitivity to instructions. Secondly, while the divergence of LLaMA drops in the deeper layers, the divergence of Alpaca and Vicuna does
+
+![](images/fc5d47d7c18de76657c2fc72e97c7d6618092e97df812d52e938b4fa7a26142e.jpg)
+
+![](images/85a93c8b9a28b689b6faa748f92743d88ed79f05e4da2930be71c7e7b557f76e.jpg)  
+Figure 3: General attention divergence between pretrained and instruction-tuned models in 7B and 13B, measured on non-instruction sentences.
+
+not go down, but rather up, suggesting higher sensitivity to instructions in the deeper layers. This effect is not observed in the noise-prefixed scenario (See Appendix A).
+
+From these two observations, we can conclude that the pretrained LLaMA has already obtained some sensitivity to instructions, i.e. to change its attention mode when it meets instructions, and this sensitivity is further strengthened by instruction tuning, which makes this sensitivity higher in the deeper layers. Because in causal language models like LLaMA, the last few layers are related to generation, this effect of instruction tuning may bring more suitable generation for a certain instruction.
+
+# 5.2 Human Resemblance
+
+# 5.2.1 Higher Human Resemblance, Lower Prediction Loss
+
+The per-token NTP loss and the maximum layerwise human resemblance scores are compared in Table 1 and Figure 5. One can tell from it that, within the considered range of parameter sizes, the human resemblance is negatively correlated with the NTP loss, where the resemblance goes higher nearly linearly with the loss going lower. The Pearson's correlation is  $-0.875$  for L1 ( $p < 0.002$ ) and  $-0.917$  for L2 ( $p < 0.0005$ ), supporting a significant and strong negative linear correlation.
+
+This result shows that, the human resemblance we defined is positively related to the language modeling performance of the LLMs considered in
+
+![](images/1f008220a318e85eb96f9974abe4cb35e28cfcd3055c85a6548958ea60eda23d.jpg)
+
+![](images/f76cbc30a5588ac8a4301ce3303f98d0edd580660d4ba45a0c94f8532c70bab8.jpg)  
+Figure 4: Attention divergence on the original and instruction-attached text of 7B and 13B models.
+
+this research, giving a practical meaning to our human resemblance analysis. Also, based on this result, we know the pursuit for higher human resemblance is consistent with the target of training better LLMs. Similarly, factors negatively impacting the human resemblance could also harm the models' language modeling performance, such as instruction tuning brings lower human resemblance as well as higher NTP loss.
+
+<table><tr><td>Size</td><td>Name</td><td>Loss</td><td>L1(%)</td><td>L2(%)</td></tr><tr><td>774M</td><td>GPT-2</td><td>0.3264</td><td>34.99</td><td>40.03</td></tr><tr><td rowspan="3">7B</td><td>LLaMA</td><td>0.2408</td><td>53.04</td><td>62.44</td></tr><tr><td>Alpaca</td><td>0.2646</td><td>52.51</td><td>61.71</td></tr><tr><td>Vicuna</td><td>0.2593</td><td>51.90</td><td>61.19</td></tr><tr><td rowspan="3">13B</td><td>LLaMA</td><td>0.2406</td><td>55.66</td><td>64.20</td></tr><tr><td>Alpaca</td><td>0.2634</td><td>55.05</td><td>64.46</td></tr><tr><td>Vicuna</td><td>0.2847</td><td>54.26</td><td>61.31</td></tr><tr><td>30B</td><td>LLaMA</td><td>0.2372</td><td>63.16</td><td>69.40</td></tr><tr><td>65B</td><td>LLaMA</td><td>0.2375</td><td>64.05</td><td>70.07</td></tr></table>
+
+Table 1: Comparison between NTP loss and the human resemblance scores
+
+# 5.2.2 Scaling Enhances Human Resemblance
+
+The layerwise human resemblance of pretrained models in different scales are shown in Figure 6, where the LLaMAs are very different from GPT-2 Large in two ways. First, the layerwise human resemblance of the LLaMAs are much higher than GPT-2, but the gaps among the LLaMAs are small; Second, human resemblance of GPT-2 drop quickly in the deeper layers, while it remains high across all layers in the LLaMAs. one can refer from these
+
+![](images/ea819d17b6d953235351d5a73b5a5b7e65ce7a5938cf23f98cd7b2ffcce21a37.jpg)  
+Figure 5: The correlation between the per-token negative log likelihood loss and the max layerwise human resemblance of LLMs.
+
+results that LLaMA has undergone a phase change in terms of human-like attention from the level of GPT-2. However, the scaling from 7B to 30B does not cause another phase change.
+
+Table 2 and Figure 7 show the max layerwise human resemblance increases linearly while the model parameter size increases exponentially, where the Pearson's correlation is 0.989 for L1  $(p < 0.002)$  and 0.964 for L2  $(p < 0.01)$ . This agrees with the scaling law of LLMs (Henighan et al., 2020).
+
+This result shows that, within the scope of this study, scaling significantly enhances the human resemblance. If the scaling law continues to take effect, the human resemblance is expected to reach  $98.80\%$  for L2 and  $88.82\%$  for L1 at the scale of 100B. However, because the law is expected to fail after reaching a threshold (Hestness et al., 2017), the relation between human resemblance and model scale is expected to change before reaching that scale.
+
+<table><tr><td colspan="2">Model</td><td colspan="2">Resemblance (%)</td></tr><tr><td>Name</td><td>Size</td><td>L1</td><td>L2</td></tr><tr><td rowspan="2">GPT-2</td><td>774M</td><td>34.99</td><td>40.03</td></tr><tr><td>7B</td><td>53.04</td><td>62.44</td></tr><tr><td rowspan="3">LLaMA</td><td>13B</td><td>55.56</td><td>64.20</td></tr><tr><td>30B</td><td>63.16</td><td>69.40</td></tr><tr><td>65B</td><td>64.05</td><td>70.07</td></tr></table>
+
+Table 2: Max layerwise human resemblance of models ranging from 774M to 65B.
+
+# 5.2.3 Instruction Tuning Harms Human Resemblance
+
+Table 3 shows the max layerwise human resemblance of the pretrained and instruction tuned models. It shows that the instruction tuned Alpaca and Vicuna models have lower max layerwise human resemblance to both L1 and L2 than LLaMA in
+
+![](images/1616780722649fc2e8bd4012e111df8abb9d9fbd88f15d3c1b25f44acaaab699.jpg)
+
+![](images/37368eb43a0304641eb4bde8759e4bf745ad410b101965746d6be748aa6b7ea1.jpg)  
+Figure 6: Layerwise L1 and L2 human resemblance for different sized pretrained models.
+
+![](images/422fa9644ff1b557cfbe55e417d82fbf8846d43f61e00df97e26a8bb9394c59b.jpg)  
+Figure 7: The change of max layerwise human resemblance caused by the scaling of model sizes, where the x-axis is log scaled.
+
+both sizes, which suggests that instruction tuning causes decline in the human resemblance of attention. However, the decline is minor compared with the increase brought by scaling, which agrees with the limited effect of instruction tuning.
+
+We also conduct relative t-tests to detect significant difference in layerwise human resemblance  $(p = 0.05 / n_{\mathrm{test}})$ . The numbers of significant layers are shown in Table 4, and the complete lists of layers are in Table 6, 7 in Appendix C). The tables show that instruction tuning enhances or reduces human resemblance on about the equal numbers of layers on the 7B models, but reduces human resemblance to much more layers on the 13B models. Also, the layers with significantly lower resemblance are distributed across all layers. This result means that, while the total effect of instruction tuning is small, the drop in human resemblance caused by it is significant and widespread.
+
+To sum up, instruction tuning does small but significant harm to the human resemblance of LLMs, which in turn may bring potential damage to their language perception. This challenges the common assumption that instruction tuning can activate the model's ability to perceive language with human-aligned behaviors. However, this negative effect could be compensated by scaling, for scaling increases the human resemblance in a larger degree. This provides a novel reason for further scaling.
+
+<table><tr><td colspan="2">Model</td><td colspan="2">Resemblance(%)</td></tr><tr><td>Size</td><td>Name</td><td>L1</td><td>L2</td></tr><tr><td rowspan="3">7B</td><td>LLaMA</td><td>53.04</td><td>62.44</td></tr><tr><td>Alpaca</td><td>52.51</td><td>61.71</td></tr><tr><td>Vicuna</td><td>51.90</td><td>61.19</td></tr><tr><td rowspan="3">13B</td><td>LLaMA</td><td>55.66</td><td>64.20</td></tr><tr><td>Alpaca</td><td>55.05</td><td>63.46</td></tr><tr><td>Vicuna</td><td>54.26</td><td>61.31</td></tr></table>
+
+Table 3: Max layerwise human resemblance scores of pretrained and instruction tuned models. This shows instruction tuning causes small decline in the max layer-wise human resemblance.  
+
+<table><tr><td colspan="2">Model</td><td colspan="2">Higher</td><td colspan="2">Lower</td></tr><tr><td>Size</td><td>Name</td><td>L1</td><td>L2</td><td>L1</td><td>L2</td></tr><tr><td rowspan="2">7B</td><td>Alpaca</td><td>3</td><td>8</td><td>4</td><td>7</td></tr><tr><td>Vicuna</td><td>9</td><td>10</td><td>8</td><td>9</td></tr><tr><td rowspan="2">13B</td><td>Alpaca</td><td>4</td><td>7</td><td>16</td><td>25</td></tr><tr><td>Vicuna</td><td>5</td><td>8</td><td>14</td><td>16</td></tr></table>
+
+# 5.2.4 All Models Show Higher L2 Resemblance
+
+Comparing the L1 and L2 human resemblance of all models (LLaMA, Alpaca and Vicuna, 7B and 13B) in Table 2 and Table 3, an advantage of L2 over L1 can be observed steadily. Independent t-test also shows that, all models considered in this research show significantly higher resemblance to L2 than to L1 ( $p = 0.05 / n_{\mathrm{test}}$  to avoid false positives). This means the models are closer to non-native English learners than to native English speakers in attention, though they are all trained mainly on English data. Furthermore, this trend is not reversed by the scaling or instruction tuning within the scope of this research, which suggests that it is a innate feature of LLMs. We will look further into this phenomenon in the next section.
+
+# 5.3 Trivial Pattern Reliance
+
+# 5.3.1 L2 Relies More on Trivial Patterns
+
+The results of LR scores between trivial patterns and L1 and L2 human saccade is shown in Table 5, where L2's score is higher than L1's in minimum, maximum and mean values. Independent t-test also supports a significant advantage of L2 over L1 on the regression scores  $(p < 5 \times 10^{-8})$ . This means one significant difference between L1 and L2 English speakers is that L2 people show more trivial and fixed attention mode than L1 people while reading, which suggests a weaker language understanding ability as those patterns contain no linguistic or factual information.
+
+This finding can help us interpret the difference between the models' L1 and L2 human resemblance. Note that all models considered in this research are trained mainly on English, but show higher attention resemblance to L2 subjects, this result suggests that their language perception is not ideal.
+
+Table 4: Numbers of Alpaca and Vicuna layers with significantly higher or lower human resemblance compared with LLaMA in the same sizes.  
+
+<table><tr><td>Group</td><td>Min</td><td>Max</td><td>Mean</td><td>SE</td></tr><tr><td>L1</td><td>0.0048</td><td>0.1525</td><td>0.0473</td><td>0.0037</td></tr><tr><td>L2</td><td>0.0134</td><td>0.2165</td><td>0.0894</td><td>0.0060</td></tr></table>
+
+Table 5: Regression scores of human saccade on the trivial patterns, across L1 and L2 subjects. SE stands for standard errors.
+
+# 5.3.2 Scaling Reduces Trivial Pattern Reliance
+
+The relative difference between 13B and 7B models in trivial pattern reliance is shown in Figure 8, where the decline in the deeper layer quarters is substantial across all models. There is also a consistent increase in the second quarter, but the amplitude is small, so the total trivial pattern reliance is reduced. This phenomenon is also observed between LLaMA 13B and 30B (Figure 11 in Appendix B). This means scaling can effectively reduce the trivial pattern reliance, especially in the deeper layers, indicating a more contextualized and adaptive language perception.
+
+# 5.3.3 Instruction Tuning Increases Trivial Pattern Reliance
+
+Figure 9 shows the relative difference between pretrained and instruction tuned models in trivial pattern reliance, where the tuned ones show reliance gain in the deeper layers. There is also a small drop in the second quarter, but the total trivial pattern reliance is increasing. Also, total increase here is
+
+![](images/84748b7f55f7fdaea4e031031566ffb84d5ba054996d35d7c775213b9e606b90.jpg)  
+Figure 8: The difference between trivial pattern reliance of 7B and 13B model attention.
+
+also smaller in amplitude than the reduction caused by scaling, again supporting the limit effect of instruction tuning.
+
+Also, one can tell from Figure 8 and Figure 9 that scaling and instruction tuning changes the trivial pattern reliance in opposite directions. However, because the effect of instruction tuning is also smaller than that of scaling, as long as the model size continues to grow, this gain in trivial pattern reliance brought by instruction tuning could be compensated or even covered, offering another reason for further scaling.
+
+![](images/2ff27b9802e5b1274efa0fbe3d7d72a0c686dd78e360414feadcbe4bb61bfe6d.jpg)  
+Figure 9: The difference in trivial pattern reliance of attentions in pretrained and instruction-tuned models.
+
+# 6 Conclusion
+
+This research evaluates the effects of scaling and instruction tuning on LLMs' attention. We find that scaling effectively changes the general attention distribution, enhances the human resemblance, and reduces the trivial pattern reliance, while instruction tuning does the opposite, but also increases the models' sensitivity to instructions. Furthermore, we find current open-sourced LLMs closer to non-native English speakers in language perception, which contains more trivial attention patterns. To the best of our knowledge, we are the first to analyze the effect of instruction tuning on the human resemblance of model attention. We hope this can inspire future work on analyzing LLMs.
+
+# Limitations
+
+One key limitation of this study is that we cannot apply our current method to closed-sourced LLMs such as ChatGPT, for their layerwise attention scores are unavailable. With the rapid development of open LLMs, it is hopeful that we can examine these findings on the largest and most advanced LLMs in the near future.
+
+Besides, though we demonstrate the difference in LLMs' resemblance to L1 and L2 people, and partially explained the effect by trivial pattern reliance, it is not enough to explain the underlying cause or consequences of this phenomenon. We do have some observations, such as the inter-subject correlation in the L1 group  $(0.1560 \pm 0.0073)$  is lower than L2  $(0.2463 \pm 0.0090)$ , and L1 subjects tend to look back to pervious words less frequently than L2 subjects. However, the underlying mechanism is still unclear. We hope to dive deeper into this non-native resemblance in our following study.
+
+Another limitation of this work comes from the way we obtain model attention. We used the attention scores when predicting the next single token, which makes sense. However, the attention patterns could be different in the following decoding steps. Also, it is possible that combining attention scores across multiple layers can improve the analysis. We will investigate this in our future work.
+
+In addition, the Reading Brain dataset we used in this research is monolingual, which does not reflect the language perception process in the cross-lingual setting, which is also important for the application of LLMs. We also hope to examine our findings in the cross-lingual scenario in the future.
+
+# Ethics Statement
+
+The authors declare no competing interests. The human behavioral data we use is publicly available and does not contain personal information of the subjects.
+
+# Acknowledgements
+
+We would like to thank the anonymous reviewers for their insightful comments. Shujian Huang is the corresponding author. We thank Yunzhe Lv and Wenhao Zhu for finetuning the Alpaca 13B model. This work is supported by National Science Foundation of China (No. 62376116, 62176120), the Liaoning Provincial Research Foundation for Basic Research (No. 2022-KF-26-02).
+
+# References
+
+Samira Abnar and Willem Zuidema. 2020. Quantifying Attention Flow in Transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, Online. Association for Computational Linguistics.  
+Aakash Bansal, Bonita Sharif, and Collin McMillan. 2023. Towards modeling human attention from eye movements for neural source code summarization. Proceedings of the ACM on Human-Computer Interaction, 7(ETRA):1-19.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Charlotte Caucheteux, Alexandre Gramfort, and JeanRémi King. 2021. Model-based analysis of brain activity reveals the hierarchy of language in 305 subjects. arXiv preprint arXiv:2110.06078.  
+Charlotte Caucheteux, Alexandre Gramfort, and JeanRémi King. 2022. Deep language algorithms predict semantic comprehension from brain activity. Scientific Reports, 12(1):16327.  
+Emmanuele Chersoni, Nora Hollenstein, Cassandra Jacobs, Yohei Oseki, Laurent Prévot, and Enrico Santus, editors. 2022. Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics. Association for Computational Linguistics, Dublin, Ireland.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An opensource chatbot impressing gpt-4 with  $90\%$  * chatgpt quality.  
+Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. 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.  
+Javier Ferrando, Gerard I. Gállego, and Marta R. Costajussà. 2022. Measuring the Mixing of Contextual Information in the Transformer. In Proceedings of
+
+the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8698-8714, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Javier Ferrando, Gerard I. Gallego, Ioannis Tsiamas, and Marta R. Costa-jussa. 2023. Explaining How Transformers Use Context to Build Predictions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5486-5513, Toronto, Canada. Association for Computational Linguistics.  
+Uri Hasson, Samuel A. Nastase, and Ariel Goldstein. 2020. Direct fit to nature: An evolutionary perspective on biological and artificial neural networks. *Neuron*, 105(3):416-434.  
+Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B. Brown, Prafulla Dhariwal, Scott Gray, Chris Hallacy, Benjamin Mann, Alec Radford, Aditya Ramesh, Nick Ryder, Daniel M. Ziegler, John Schulman, Dario Amodei, and Sam McCandlish. 2020. Scaling Laws for Autoregressive Generative Modeling. ArXiv:2010.14701 [cs].  
+Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. 2017. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409.  
+Nora Hollenstein and Lisa Beinborn. 2021. Relative Importance in Sentence Processing. 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 141-150, Online. Association for Computational Linguistics.  
+Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. 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 3543-3556, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Alexander JE Kell, Daniel LK Yamins, Erica N Shook, Sam V Norman-Haignere, and Josh H McDermott. 2018. A task-optimized neural network replicates human auditory behavior, predicts brain responses, and reveals a cortical processing hierarchy. Neuron, 98(3):630-644.  
+Goro Kobayashi, Tatsuki Kuribayashi, Sho Yokoi, and Kentaro Inui. 2020. Attention is Not Only a Weight: Analyzing Transformers with Vector Norms. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7057-7075, Online. Association for Computational Linguistics.
+
+Goro Kobayashi, Tatsuki Kuribayashi, Sho Yokoi, and Kentaro Inui. 2021. Incorporating Residual and Normalization Layers into Analysis of Masked Language Models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4547-4568, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics.  
+Mathis Lamarre, Catherine Chen, and Fatma Deniz. 2022. Attention weights accurately predict language representations in the brain. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4513-4529, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Ping Li, Anya Yu, Benjamin Schloss, Chun-Ting Hsu, Lindsey Ma, Chih-Ting Chang, and Marissa Scotto. 2019. The reading brain project: An open science data-sharing initiative. In 26th Annual Meeting of the Cognitive Neuroscience Society.  
+Christopher D Manning, Kevin Clark, John Hewitt, Urvashi Khandelwal, and Omer Levy. 2020. Emergent linguistic structure in artificial neural networks trained by self-supervision. Proceedings of the National Academy of Sciences, 117(48):30046-30054.  
+David Mareček and Rudolf Rosa. 2018. Extracting syntactic trees from transformer encoder self-attentions. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 347-349, Brussels, Belgium. Association for Computational Linguistics.  
+Danny Merkx and Stefan L. Frank. 2021. Human sentence processing: Recurrence or attention? In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, pages 12-22, Online. Association for Computational Linguistics.  
+Hosein Mohebbi, Willem Zuidema, Grzegorz Chrupa\la, and Afra Alishahi. 2023. Quantifying Context Mixing in Transformers. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3378-3400, Dubrovnik, Croatia. Association for Computational Linguistics.  
+Felix Morger, Stephanie Brandl, Lisa Beinborn, and Nora Hollenstein. 2022. A Cross-lingual Comparison of Human and Model Relative Word Importance. In Proceedings of the 2022 CLASP Conference on (Dis)embodiment, pages 11-23, Gothenburg, Sweden. Association for Computational Linguistics.
+
+Byung-Doh Oh and William Schuler. 2023a. Transformer-Based LM Surprisal Predicts Human Reading Times Best with About Two Billion Training Tokens.  
+Byung-Doh Oh and William Schuler. 2023b. Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times? Transactions of the Association for Computational Linguistics, 11:336-350.  
+Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. ArXiv:2203.02155 [cs].  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners.  
+Alessandro Raganato and Jörg Tiedemann. 2018. An analysis of encoder representations in transformer-based machine translation. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 287-297, Brussels, Belgium. Association for Computational Linguistics.  
+Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, Manan Dey, M. Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevry, Jason Alan Fries, Ryan Teehan, Tali Bers, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M. Rush. 2022. Multitask Prompted Training Enables Zero-Shot Task Generalization. ArXiv:2110.08207 [cs].  
+Martin Schrimpf, Idan Asher Blank, Greta Tuckute, Carina Kauf, Eghbal A Hosseini, Nancy Kanwisher, Joshua B Tenenbaum, and Evelina Fedorenko. 2021. The neural architecture of language: Integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences, 118(45):e2105646118.  
+Sofia Serrano and Noah A. Smith. 2019. Is attention interpretable? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2931-2951, Florence, Italy. Association for Computational Linguistics.  
+Ekta Sood, Simon Tannert, Philipp Müller, and Andreas Bulling. 2020. Improving natural language processing tasks with human gaze-guided neural attention.
+
+Advances in Neural Information Processing Systems, 33:6327-6341.
+
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.
+
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models.
+
+Shikhar Vashisth, Shyam Upadhyay, Gaurav Singh Tomar, and Manaal Faruqui. 2019. Attention interpretability across nlp tasks. arXiv preprint arXiv:1909.11218.
+
+Jesse Vig and Yonatan Belinkov. 2019. Analyzing the structure of attention in a transformer language model. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 63-76, Florence, Italy. Association for Computational Linguistics.
+
+Elena Voita, David Talbot, Fedor Moiseev, Rico Senrich, and Ivan Titov. 2019. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797-5808, Florence, Italy. Association for Computational Linguistics.
+
+Titus von der Malsburg, Reinhold Kliegl, and Shravan Vasisith. 2015. Determinants of Scanpath Regularity in Reading. Cognitive Science, 39(7):1675-1703.
+
+Titus von der Malsburg and Shravan Vasishth. 2013. Scanpaths reveal syntactic underspecification and reanalysis strategies. Language and Cognitive Processes, 28(10):1545-1578.
+
+Jeffrey J. Walczyk and Robert W. Taylor. 1996. How do the efficiencies of reading subcomponents relate to looking back in text? Journal of Educational Psychology, 88(3):537-545.
+
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-Instruct: Aligning Language Models with Self-Generated Instructions. ArXiv:2212.10560 [cs].
+
+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.
+
+Sarah Wiegrefe and Yuval Pinter. 2019. Attention is not explanation. 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 11-20, Hong Kong, China. Association for Computational Linguistics.
+
+# A Attention divergence on plain and noise-prefix text
+
+As shown in Figure 10, the divergence between plain and noise-prefixed text of Alpaca and Vicuna is not significantly higher than LLaMA in most of the layers, which further supports our findings on the higher instruction sensitivity of instruction-tuned models.
+
+![](images/7d361eaa7e2230b259f47c0ed24a4ab897787020979f47f9ce6be3b4a8349f08.jpg)
+
+![](images/ac4cc8bc037770a46c7e04cf00ef0adbaf0b2bc35b90d9bec9e403f3150e4f17.jpg)  
+Figure 10: Attention divergence on the original and noise-attached text of 7B and 13B models.
+
+# B Layerwise trivial pattern reliance
+
+Figure 11 shows the layerwise trivial pattern reliance of 7B and 13B models, as well as the results for LLaMA 30B. The reliance keeps decreasing as the model size scales up, especially in the deeper layers.
+
+# C Layers with significant changes in human resemblance after instruction tuning
+
+Table 6 shows the list of layers in 7B Alpaca and Vicuna with significant change in human resemblance after instruction tuning compared with LLaMA, and Table 7 shows the list in 13B.
+
+# D Preliminary fMRI results
+
+Some preliminary analysis on the fMRI data in the Reading Brain dataset is done, and the results com
+
+![](images/cd785dcab0e4b4d8fece282b3d909f8c5494973384c746b3858dfc41b73dd3a3.jpg)
+
+![](images/5f865bb3eb9bf6a0136938d04e2bf5dfd817900983c804abd586a8c0872426d4.jpg)
+
+![](images/a1e9b5c7d35de8515d3cc9eca08ad8e52b0abd225302e39b6bb068f57d512ebc.jpg)
+
+![](images/53ffc9bdd996fa8ab968753798918e4d1a2b1f9c1669e83d125ac79344337a2e.jpg)  
+Figure 11: Lyerwise trivial pattern reliance of 7B and 13B models. "RB" stands for "Reading Brain".
+
+ply with our findings on saccade data. We investigate several linguistics-related brain regions, and find that the highest correlation of fMRI signals to model attention are from the anterior left temporal gyrus (Figure 12). The fMRI data from this region is pre-processed into matrix forms as saccade data. Figure 14 gives examples of individual and group mean fMRI matrices.
+
+The findings are: (1) The models' NTP loss significantly  $(p < 0.02$  for both L1 and L2) and strongly negatively  $(r = -0.777$  for L1,  $r = -0.796$  for L2) correlates with fMRI resemblance (Figure 13); (2) The models' highest layerwise fMRI resemblance significantly  $(r > 0.99$ ,  $p < 0.001$  for both L1 and L2) correlates with the model scale in logarithm (Figure 15, 16); (3) While scal
+
+<table><tr><td>Relation</td><td>Group</td><td>Alpaca</td><td>Vicuna</td></tr><tr><td rowspan="2">Higher</td><td>L1</td><td>2, 5, 7</td><td>12, 13, 17, 18, 21, 26, 28, 29, 30</td></tr><tr><td>L2</td><td>2, 5, 6, 8, 12, 15, 21, 27</td><td>3, 13, 15, 18, 21, 23, 26, 28, 29, 30</td></tr><tr><td rowspan="2">Lower</td><td>L1</td><td>4, 10, 23, 31</td><td>0, 1, 4, 5, 9, 11, 20, 31</td></tr><tr><td>L2</td><td>0, 10, 14, 19, 23, 28, 31</td><td>0, 1, 4, 5, 9, 11, 14, 20, 31</td></tr></table>
+
+Table 6: Layers of Alpaca and Vicuna with significantly higher or lower human resemblance compared with LLaMA in 7B.  
+
+<table><tr><td>Relation</td><td>Group</td><td>Alpaca</td><td>Vicuna</td></tr><tr><td rowspan="2">Higher</td><td>L1</td><td>16, 32, 33, 36</td><td>4, 16, 27, 33, 36</td></tr><tr><td>L2</td><td>1, 16, 19, 27, 32, 33, 36</td><td>4, 16, 19, 23, 27, 33, 36, 37</td></tr><tr><td rowspan="2">Lower</td><td>L1</td><td>0, 2, 3, 5, 6, 7, 8, 9, 14, 18, 24, 25, 28, 30, 35, 39</td><td>0, 1, 3, 5, 6, 9, 10, 14, 17, 18, 24, 30, 34, 39</td></tr><tr><td>L2</td><td>0, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 17, 18, 20, 24, 25, 28, 29, 30, 31, 35, 38, 39</td><td>0, 2, 3, 5, 6, 8, 9, 10, 11, 12, 14, 17, 18, 20, 21 30</td></tr></table>
+
+Table 7: Layers of Alpaca and Vicuna with significantly higher or lower human resemblance compared with LLaMA in 13B.
+
+ing enhances human fMRI resemblance, instruction tuning does not (Table 8); (4) Resemblance to L2 fMRI is consistently higher than that to L1.
+
+![](images/1a4c004f619dcbabbb6a01164539845a5be779d88fd6a9573642a97e701507bc.jpg)  
+Figure 12: An illustration of the anterior left temporal gyrus in the human brain.
+
+![](images/611ee90ccf0d1d563394ecd2a127465e030f92ec3c59ad1b78e1b6be9d913769.jpg)  
+Figure 13: The correlation between per-token negative log likelihood loss and max layerwise fMRI resemblance of the LLMs.
+
+<table><tr><td colspan="2">Model</td><td colspan="2">Resemblance(%)</td></tr><tr><td>Size</td><td>Name</td><td>L1</td><td>L2</td></tr><tr><td rowspan="3">7B</td><td>LLaMA</td><td>11.02</td><td>15.45</td></tr><tr><td>Alpaca</td><td>10.99</td><td>15.61</td></tr><tr><td>Vicuna</td><td>11.35</td><td>15.44</td></tr><tr><td rowspan="3">13B</td><td>LLaMA</td><td>14.69</td><td>20.19</td></tr><tr><td>Alpaca</td><td>14.67</td><td>19.44</td></tr><tr><td>Vicuna</td><td>14.02</td><td>18.57</td></tr></table>
+
+Table 8: Max layerwise human fMRI resemblance scores of pretrained and instruction tuned models. This shows instruction tuning causes small decline in the max layerwise human resemblance.
+
+![](images/76078bf11392e7075e08e4d424292ba9479a10db75ea03f84b13b04a08bd7b82.jpg)  
+(a) Individual
+
+![](images/3469d2471ded04cf9bd91703fbef916897abea2483e4faacbc0100147bc18e4e.jpg)  
+(b) Group Mean
+
+![](images/05568093f276b55f2bbc25cf0fffd1bd810c844632c23eee1700f0096a937308.jpg)
+
+![](images/dd49a90499d2eec64f51f9d08fdb560f781e3d274b893ee0bc88ed334b62f01f.jpg)  
+Figure 15: Layerwise L1 and L2 human fMRI resemblance for different sized pretrained models.
+
+![](images/6578ebffd275ffd9cf5335aff9398a942e7331af325d98a028f3ca51e4ea9f49.jpg)  
+Figure 14: Examples of individual and group mean fMRI signals of the sentence "Could humans live on Mars some day?"  
+Figure 16: The change of max layerwise human fMRI resemblance caused by scaling. The x-axis is log scaled.

2023/Roles of Scaling and Instruction Tuning in Language Perception_ Model vs. Human Attention/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e9aa0298199286d7eddd39762bec5ccbd79a52401fe9f195560d75c366c036e
+size 655414

2023/Romanization-based Large-scale Adaptation of Multilingual Language Models/1a02cba7-8ebc-4f20-8832-cb43efde101d_content_list.json

@@ -0,0 +1,1342 @@
+[
+    {
+        "type": "text",
+        "text": "Romanization-based Large-scale Adaptation of Multilingual Language Models",
+        "text_level": 1,
+        "bbox": [
+            268,
+            79,
+            729,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sukannya Purkayastha<sup>1</sup>, Sebastian Ruder<sup>2</sup>, Jonas Pfeiffer<sup>2</sup>, Iryna Gurevych<sup>1</sup>, Ivan Vulić<sup>3</sup>",
+        "bbox": [
+            115,
+            124,
+            885,
+            142
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup> Ubiquitous Knowledge Processing Lab,",
+        "bbox": [
+            332,
+            143,
+            670,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science and Hessian Center for AI (hessian.AI),",
+        "bbox": [
+            201,
+            159,
+            801,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Technical University of Darmstadt",
+        "bbox": [
+            359,
+            175,
+            642,
+            191
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Google DeepMind",
+        "bbox": [
+            416,
+            192,
+            584,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$  Language Technology Lab, University of Cambridge",
+        "bbox": [
+            280,
+            209,
+            722,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "www.ukp.tu-darmstadt.de",
+        "bbox": [
+            359,
+            227,
+            640,
+            241
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large multilingual pretrained language models (mPLMs) have become the de facto state of the art for cross-lingual transfer in NLP. However, their large-scale deployment to many languages, besides pretraining data scarcity, is also hindered by the increase in vocabulary size and limitations in their parameter budget. In order to boost the capacity of mPLMs to deal with low-resource and unseen languages, we explore the potential of leveraging transliteration on a massive scale. In particular, we explore the UROMAN transliteration tool, which provides mappings from UTF-8 to Latin characters for all the writing systems, enabling inexpensive romanization for virtually any language. We first focus on establishing how UROMAN compares against other language-specific and manually curated transliterators for adapting multilingual PLMs. We then study and compare a plethora of data- and parameter-efficient strategies for adapting the mPLMs to romanized and non-romanized corpora of 14 diverse low-resource languages. Our results reveal that UROMAN-based transliteration can offer strong performance for many languages, with particular gains achieved in the most challenging setups: on languages with unseen scripts and with limited training data without any vocabulary augmentation. Further analyses reveal that an improved tokenizer based on romanized data can even outperform non-transliteration-based methods in the majority of languages.",
+        "bbox": [
+            144,
+            280,
+            460,
+            734
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            747,
+            258,
+            763
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Massively multilingual language models (mPLMs) such as mBERT (Devlin et al., 2019) and XLM-R (Conneau et al., 2020) have become the driving force for a variety of applications in multilingual NLP (Ponti et al., 2020; Hu et al., 2020; Moghe et al., 2023). However, guaranteeing and maintaining strong performance for a wide spectrum of low-resource languages is difficult due to two crucial problems. The first issue is the vocabulary",
+        "bbox": [
+            112,
+            774,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/8756bb25e844410f100eaba69877e4802d56b0792e642e1c5fb13ec2152b6cd6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Language</td><td>Original Text</td><td>Romanized Text</td><td>Translation</td></tr><tr><td>Bhojpuri</td><td>बाल्दियम् गार्वा</td><td>jorjiyan bhaasaas</td><td>Georgian language</td></tr><tr><td>Sinhala</td><td>सिकूल्तिूल्तिूल्तिूल्तिूल्तिूल्तिू�िूल्तिूल्तिूल्तिू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�र्था</td><td></td><td></td></tr><tr><td>Sindhi</td><td>सिकूल्तिूल्तिूल्तिू�िू�िू�िू�र्था</td><td>grahalooka</td><td>planets</td></tr><tr><td>Khmer</td><td>khμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμmu μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμum μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ µ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μµ μµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµüµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµúµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ µµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₃₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₂₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₁₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀⁻⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸��⁸⁸⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁰⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁵⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁲⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁴⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁽⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁻⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁼⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸ⁱ⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁅⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁰⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₃⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄²⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄.\n&lt;|im_start|&gt;assistant</td><td>Bhojpuri</td><td>BJJUJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJSJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ JJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ</td></tr></table>",
+        "bbox": [
+            512,
+            250,
+            873,
+            315
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: Romanization across different languages.",
+        "bbox": [
+            519,
+            325,
+            870,
+            340
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "size, as the vocabulary is bound to increase with the number of languages added if per-language performance is to be maintained (Hu et al., 2020; Artetxe et al., 2020; Pfeiffer et al., 2022). Second, pretraining mPLMs with a fixed model capacity improves cross-lingual performance up to a point after which it starts to decrease; this is the phenomenon termed the curse of multilinguality (Conneau et al., 2020).",
+        "bbox": [
+            507,
+            366,
+            884,
+            495
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Transliteration refers to the process of converting language represented in one writing system to another (Wellisch et al., 1978). Latin script-centered transliteration or romanization is the most common form of transliteration (Lin et al., 2018; Amrhein and Sennrich, 2020; Demirsin et al., 2022) as the Latin/Roman script is by far the most widely adopted writing script in the world (Daniels and Bright, 1996; van Esch et al., 2022). Adapting mPLMs via transliteration can address the two aforementioned critical issues. 1) Since the Latin script covers a dominant portion of the mPLM's vocabulary (e.g.,  $77\\%$  in case of mBERT, see Ács), 'romanizing' the remaining part of the vocabulary might mitigate the vocabulary size issue and boost vocabulary sharing. 2) Since no new tokens are added during the romanization process, reusing pretrained embeddings from the mPLM's embedding matrix helps reuse the information already present within the mPLM, thereby allocating the model's parameter budget more efficiently.",
+        "bbox": [
+            507,
+            495,
+            884,
+            833
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, the main drawback of transliteration seems to be the expensive process of creating effective language-specific transliterations, as they",
+        "bbox": [
+            507,
+            834,
+            882,
+            882
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>According to Encyclopedia Britannica, up to 70% of the world population is employing the Latin script.",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "7996",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7996-8005",
+        "bbox": [
+            216,
+            945,
+            779,
+            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": "text",
+        "text": "typically require language expertise to curate dictionaries that map tokens from one language and script to another. Therefore, previous attempts at mPLM adaptation to unseen languages via transliteration (Muller et al., 2021; Chau and Smith, 2021; Dhamecha et al., 2021; Moosa et al., 2023) were constrained to a handful of languages due to the limited availability of language-specific translit-. rators, or were applied only to languages that have 'language siblings' with developed translitators.",
+        "bbox": [
+            110,
+            84,
+            489,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, unlike previous work, we propose to use and then evaluate the usefulness of a universal romanization tool, UROMAN (Hermjakob et al., 2018), for quick, large-scale and effective adaptation of mPLMs to low-resource languages. The UROMAN tool disposes of language-specific curated dictionaries and maps any UTF-8 character to the Latin script, increasing the portability of romanization, with some examples in Figure 1.",
+        "bbox": [
+            110,
+            246,
+            489,
+            391
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We analyze language adaptation on a massive scale via UROMAN-based romanization on a set of 14 diverse low-resource languages. We conduct experiments within the standard parameter-efficient adapter-based cross-lingual transfer setup on two tasks: Named Entity Recognition (NER) on the WikiANN dataset (Pan et al., 2017; Rahimi et al., 2019), and Dependency Parsing (DP) with Universal Dependencies v2.7 (Nivre et al., 2020). Our key results suggest that UROMAN-based transliteration can offer strong performance on par or even outperforming adaptation with language-specific transliterrators, setting up the basis for wider use of transliteration-based mPLM adaptation techniques in future work. The gains with romanization-based adaptation over standard adaptation baselines are particularly pronounced for languages with unseen scripts ( $\\sim$ 8-22 performance points) without any vocabulary augmentation.[2]",
+        "bbox": [
+            110,
+            392,
+            489,
+            697
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Background",
+        "text_level": 1,
+        "bbox": [
+            112,
+            712,
+            257,
+            728
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Why UROMAN-Based Romanization? UROMAN-based romanization is not always fully reversible, and its usage for transliteration has thus been limited in the literature. However, due to its high portability, UROMAN can help scale the process of transliteration massively and as such benefit low-resource scenarios and wider adaptation of mPLMs. The main idea, as hinted in §1, is to (learn to) map",
+        "bbox": [
+            112,
+            740,
+            489,
+            869
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "any UTF-8 character to the Latin script, without the use of any external language-specific dictionaries (see Hermjakob et al. (2018) for technical details).",
+        "bbox": [
+            507,
+            84,
+            884,
+            134
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Cross-Lingual Transfer to Low-Resource Languages. Parameter-efficient and modular finetuning methods (Pfeiffer et al., 2023) such as adapters (Houlsby et al., 2019; Pfeiffer et al., 2020b) have been used for cross-lingual transfer, putting a particular focus on enabling transfer to low-resource languages and scenarios, including languages with scripts 'unseen' by the base mPLM (Pfeiffer et al., 2021). Adapters are small lightweight components stitched into the base mPLM, and then trained for particular languages and tasks while keeping the parameters of the original mPLM frozen. This circumvents the issues of catastrophic forgetting and interference (McCloskey and Cohen, 1989) within the mPLM, and allows for extending its reach also to unseen languages (Pfeiffer et al., 2021; Ansell et al., 2021).",
+        "bbox": [
+            507,
+            137,
+            884,
+            412
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For our main empirical analyses, we adopt a state-of-the-art modular method for cross-lingual transfer: MAD-X (Pfeiffer et al., 2020b). In short, MAD-X is based on language adapters (LA), task adapters (TA), and invertible adapters (INV). While LAs are trained for specific languages relying on masked language modeling, TAs are trained with high-resource languages relying on task-annotated data and task-specific objectives. At inference, the source LA is replaced with the target LA while the TA is kept. In order to do parameter-efficient learning for the token-level embeddings across different languages and to deal with the vocabulary mismatch between source and target languages, Pfeiffer et al. (2020b) also propose INV adapters: they are placed on top of the embedding layer and their inverses precede the output embedding layer.<sup>3</sup> We adopt the better-performing MAD-X 2.0 setup (Pfeiffer et al., 2021) where the adapters in the last Transformer layer are dropped at inference.<sup>4</sup>",
+        "bbox": [
+            507,
+            413,
+            884,
+            736
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Experiments and Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            749,
+            761,
+            766
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As the main means of analyzing the impact of transliteration in general and UROMAN-based romanization in particular, we train different variants of language adapters within the MAD-X framework, based on transliterated and non-transliterated",
+        "bbox": [
+            507,
+            777,
+            884,
+            857
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ Code and data available at https://github.com/UKPLab/emnlp23_romanization_based_adaptation",
+        "bbox": [
+            112,
+            879,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "3They are trained together with the LAs while the rest of the mPLM is kept frozen.",
+        "bbox": [
+            507,
+            866,
+            882,
+            892
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "4We refer the reader to the original papers for further technical details regarding the MAD-X framework.",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "7997",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/d42bbd3ecbeebbeac8ab08be9e8d19a45d4ee91791b211e956e606f6ffc629c0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Language</td><td>Family</td><td>Script</td><td># Sentences</td></tr><tr><td>Bhojpuri (bh)</td><td>Indo-European</td><td>Devanagari</td><td>35,983</td></tr><tr><td>Buryat (bxr)</td><td>Mongolic</td><td>Cyrillic</td><td>41,692</td></tr><tr><td>Erzya (myv)</td><td>Uralic</td><td>Cyrillic</td><td>42,575</td></tr><tr><td>Meadow Mari (mhr)</td><td>Uralic</td><td>Cyrillic</td><td>144,529</td></tr><tr><td>Min Dong (cdo)</td><td>Sino-Tibetan</td><td>Chinese</td><td>33,978</td></tr><tr><td>Mingrelian (xmf)</td><td>Kartvelian</td><td>Georgian</td><td>63,032</td></tr><tr><td>Sindhi (sd)</td><td>Indo-European</td><td>Arabic</td><td>86,176</td></tr><tr><td>Sorani Kurdish (ckb)</td><td>Indo-European</td><td>Arabic</td><td>459,475</td></tr><tr><td>Uyghur (ug)</td><td>Turkic</td><td>Arabic</td><td>149,813</td></tr><tr><td>Amháric (am)</td><td>Afro-Asiatic</td><td>Ge&#x27;ez</td><td>88,320</td></tr><tr><td>Divehi (dv)</td><td>Indo-European</td><td>Thaana</td><td>34,779</td></tr><tr><td>Khmer (km)</td><td>Austroasiatic</td><td>Khmer</td><td>139,704</td></tr><tr><td>Sinhala (si)</td><td>Indo-European</td><td>Sinhala</td><td>219,866</td></tr><tr><td>Tibetan (bo)</td><td>Sino-Tibetan</td><td>Tibetan</td><td>131,362</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            485,
+            239
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Languages with their ISO 639-3 codes used in our evaluation, along with their script, language family, and number of sentences available for pretraining. The dashed line separates languages with unseen scripts, placed in the bottom part of the table.",
+        "bbox": [
+            112,
+            250,
+            489,
+            322
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "versions of target language data, outlined here.",
+        "bbox": [
+            112,
+            351,
+            460,
+            366
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Variants with Non-Transliterated Data. For the Non-Trans $_{\\mathrm{LA} + \\mathrm{INV}}$  variant, we train LAs and INV adapters together. This variant serves to examine the extent to which mPLMs can adapt to unseen languages without any vocabulary extension. $^5$  We compare this to Non-Trans $_{\\mathrm{LA} + \\mathrm{Emb}_{\\mathrm{Lex}}}$ , which trains a new tokenizer for the target language (Pfeiffer et al., 2021): the so-called 'lexically overlapping' tokens are initialized with mPLM's trained embeddings, while the remaining embeddings are initialized randomly. All these embeddings (Emb $_{\\mathrm{Lex}}$ ) are fine-tuned along with LAs.",
+        "bbox": [
+            112,
+            374,
+            489,
+            567
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Variants with Transliterated Data. We evaluate a  $\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{INV}}$  variant, which uses the same setup as Non-Trans  $\\mathrm{LA} + \\mathrm{INV}$  but now with transliterated data. We again note that in this efficient setup, we do not extend the vocabulary size, and use the fewest trainable parameters. In the  $\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{mPLM}_{\\mathrm{ft}}}$  variant, we train LAs along with fine-tuning the pretrained embeddings of mPLM (mPLMft). This further enhances the model capacity by fine-tuning the embedding layer instead of using invertible adapters. For both variants, transliterated data can be produced via different transliterations: (i) language-specific ones; (ii) the ones from 'language siblings' (e.g., using a Georgian transliterator for Mingrelian), or (iii) UROMAN.",
+        "bbox": [
+            112,
+            575,
+            489,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            714,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Data, Languages and Tasks. Following Pfeiffer et al. (2021), we select mBERT as our base mPLM. We experiment with 14 typologically diverse low-resource languages that are not part of mBERT's pretraining corpora, with 5/14 languages written in distinct scripts (see Table 1 for details). For LA training, we use Wikipedia dumps for the target languages, which we also transliterate (using different translitators). Evaluation is conducted on two standard cross-lingual transfer tasks in zero-shot setups: 1) the WikiAnn NER dataset (Pan et al., 2017) with the train, dev, and test splits from (Rahimi et al., 2019); 2) for dependency parsing, we rely on the UD Dataset v2.7 (Nivre et al., 2020).",
+        "bbox": [
+            505,
+            105,
+            884,
+            330
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "LAs and TAs. English is the source language in all experiments, and is used for training TAs. The English LA is obtained directly from Adapterhub.ml (Pfeiffer et al., 2020a), LAs and embeddings (when needed) are only trained for target languages.",
+        "bbox": [
+            507,
+            335,
+            882,
+            416
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Finally, for the Non-Trans $_{\\mathrm{LA + Emb_{Lex}}}$  variant, we train a WordPiece tokenizer on the target language data with a vocabulary size of 10K.",
+        "bbox": [
+            507,
+            417,
+            880,
+            463
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Training of Language and Task Adapters. We train all the language adapters for 50 epochs or  $\\sim 50K$  update steps based on the corpus size. The batch size is set to 64 and the learning rate is  $1e - 4$ .",
+        "bbox": [
+            507,
+            470,
+            882,
+            533
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We train English task adapters following the setup from (Pfeiffer et al., 2020b). For NER, we directly obtain the task adapter from Adapterhub.ml which is trained with a learning rate of  $1e - 4$  for 10 epochs. For DP, we train a Transformer-based (Glavaš and Vulić, 2021) biaffine attention dependency parser (Dozat and Manning, 2017). We use a learning rate of  $5e - 4$  and train for 10 epochs as in (Pfeiffer et al., 2021).",
+        "bbox": [
+            505,
+            533,
+            882,
+            678
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "All the reported results in both tasks (NER and DP) are reported as averages over 6 random seeds. All the models have been trained on A100 or V100 GPUs. None of the training methods consumed more than 36 hours. As the main means of analyzing the impact of transliteration in general and UROMAN-based romanization in particular, we train different variants of language adapters within the MAD-X framework, based on transliterated and non-transliterated versions of target language data, outlined here.",
+        "bbox": [
+            507,
+            678,
+            884,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Results and Discussion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            734,
+            881
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "UROMAN versus Other Transliterators and Transliteration Strategies. In order to estab",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>5</sup>Since LAs without INV typically perform worse than with INV (Pfeiffer et al., 2020b), also confirmed in our preliminary experiments, we do not ablate to the setup without INV.",
+        "bbox": [
+            112,
+            832,
+            487,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>6</sup>We do not have this setup for non-transliterated data since, for languages with unseen scripts, most of the tokens are replaced by the generic 'UNK' token, and fine-tuning embeddings hardly benefit downstream performance.",
+        "bbox": [
+            112,
+            869,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "7998",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/17f500782286371555ad76c5e55f4b14d1005ea0f0102048de75e8b0959449dc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Task</td><td>Transliterator</td><td>am</td><td>ar</td><td>ka</td><td>ru</td><td>hi</td><td>sd</td><td>avg</td></tr><tr><td rowspan=\"2\">NER (Macro F1)</td><td>UROMAN</td><td>25.6</td><td>24.8</td><td>61.4</td><td>66.5</td><td>48.6</td><td>35.3</td><td>43.7</td></tr><tr><td>Other</td><td>25.5</td><td>23.7</td><td>57.3</td><td>63.9</td><td>56.7</td><td>35.9</td><td>43.8</td></tr><tr><td rowspan=\"2\">UD (UAS / LAS)</td><td>UROMAN</td><td>36.17 / 6.6</td><td>33.0 / 19.8</td><td>-</td><td>47.3 / 32.4</td><td>33.8 / 17.8</td><td>-</td><td>37.5 / 19.1</td></tr><tr><td>Other</td><td>29.9 / 5.4</td><td>32.6 / 19.9</td><td>-</td><td>45.0 / 19.9</td><td>33.2 / 17.9</td><td>-</td><td>35.2 / 15.8</td></tr></table>",
+        "bbox": [
+            181,
+            80,
+            811,
+            140
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/7a6de9526591995d744bec8a2012e9907ad2062ab23b5394436279dfda87716b.jpg",
+        "table_caption": [
+            "Table 2: Comparison of UROMAN with language-specific transliterations."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"7\">Seen Script</td><td colspan=\"6\">Unseen Script</td><td>avg</td></tr><tr><td>bh</td><td>cdo</td><td>ckb</td><td>mhr</td><td>sd</td><td>ug</td><td>xmf</td><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td><td></td><td></td></tr><tr><td>UROMAN</td><td>32.59</td><td>27.34</td><td>67.73</td><td>64.68</td><td>35.33</td><td>28.10</td><td>52.58</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td><td>26.89</td><td></td></tr><tr><td>BORROW</td><td>53.42 (hi)</td><td>-</td><td>12.46 (ar)</td><td>45.86 (ru)</td><td>16.79 (ar)</td><td>12.85 (ar)</td><td>24.77 (ru)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td></tr><tr><td>RAND</td><td>25.42</td><td>19.51</td><td>53.55</td><td>42.02</td><td>27.20</td><td>25.18</td><td>35.82</td><td>18.00</td><td>18.95</td><td>21.19</td><td>32.75</td><td>20.01</td><td>21.59</td><td></td></tr></table>",
+        "bbox": [
+            115,
+            168,
+            843,
+            227
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/1915c6cb2e399899f3761feb50827c5a44936950714f360f4af54d6dee007c18.jpg",
+        "table_caption": [
+            "Table 3: Comparison of various transliteration strategies on the NER task (Macro-F1)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Variant</td><td colspan=\"7\">Seen Script</td><td colspan=\"5\">Unseen Script</td><td rowspan=\"2\">avg</td></tr><tr><td>bh</td><td>cdo</td><td>sd</td><td>xmf</td><td>mhr</td><td>ckb</td><td>ug</td><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td></tr><tr><td>Non-TransLA+INV</td><td>55.14</td><td>24.19</td><td>31.31</td><td>49.74</td><td>70.31</td><td>45.54</td><td>33.53</td><td>3.26</td><td>19.86</td><td>18.72</td><td>13.81</td><td>23.14</td><td>18.39</td></tr><tr><td>TransLA+INV</td><td>32.59</td><td>27.34</td><td>35.33</td><td>52.58</td><td>64.68</td><td>67.73</td><td>28.10</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td><td>26.89</td></tr><tr><td>Non-TransLA+EmblEx</td><td>60.00</td><td>28.91</td><td>42.47</td><td>51.99</td><td>61.05</td><td>79.12</td><td>50.42</td><td>47.60</td><td>40.96</td><td>31.21</td><td>53.94</td><td>45.89</td><td>49.01</td></tr><tr><td>TransLA + mPLMrl</td><td>49.05</td><td>36.92</td><td>39.16</td><td>57.99</td><td>69.85</td><td>73.92</td><td>33.43</td><td>37.09</td><td>33.82</td><td>40.40</td><td>52.39</td><td>45.24</td><td>47.44</td></tr></table>",
+        "bbox": [
+            147,
+            253,
+            843,
+            325
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/5960a8e7cb8c66d7f4380408ebf24c27dadb9a542c45095d3d3386872bd2e792.jpg",
+        "table_caption": [
+            "Table 4: Results (Macro-F1 scores) on WikiAnn NER averaged over 6 random seeds."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Variant</td><td colspan=\"4\">Seen Script</td><td colspan=\"2\">Unseen Script</td></tr><tr><td>bh</td><td>myv</td><td>ug</td><td>bxr</td><td>am</td><td>avg</td></tr><tr><td>Non-TransLA+INV</td><td>28.46 / 11.53</td><td>45.28 / 26.27</td><td>33.44 / 15.28</td><td>39.75 / 19.77</td><td>19.08 / 1.85</td><td>33.20 / 10.81</td></tr><tr><td>TransLA+INV</td><td>25.12 / 10.17</td><td>45.74 / 26.64</td><td>32.30 / 15.10</td><td>37.92 / 17.23</td><td>36.07 / 7.58</td><td>35.43 / 12.41</td></tr><tr><td>Non-TransLA+EmblEx</td><td>26.68 / 10.10</td><td>48.34 / 25.34</td><td>41.20 / 22.81</td><td>39.51 / 16.02</td><td>36.4778.39</td><td>38.44 / 12.20</td></tr><tr><td>TransLA + mPLMrl</td><td>28.04 / 11.13</td><td>41.97 / 20.29</td><td>50.89 / 16.56</td><td>35.03 / 20.29</td><td>39.10 / 9.00</td><td>39.01 / 14.65</td></tr></table>",
+        "bbox": [
+            181,
+            353,
+            811,
+            425
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Results (UAS / LAS scores) in the DP task with UD, averaged over 6 random seeds.",
+        "bbox": [
+            183,
+            430,
+            810,
+            445
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "lish the utility of UROMAN as a viable transliterator, especially for low-resource languages, we compare its performance with transliteration options using the TransLA + INV setup as the most efficient scenario. First, we compare UROMAN with language-specific transliterations available for selected languages: amseg (Yimam et al., 2021) for Amharic, ai4bharat-transliteration (Madhani et al., 2022) for Hindi and Sindhi, lang-trans for Arabic, and transliterate for Russian and Georgian<sup>7</sup>. The transliterations used in this work are outlined in Table 6. The results are provided in Table 2. On average, UROMAN performs better or comparable to the language-specific transliterations. This provides justification to use UROMAN for massive transliteration at scale.",
+        "bbox": [
+            110,
+            463,
+            489,
+            719
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Second, we compare UROMAN to two other transliteration strategies. (i) BORROW refers to borrowing transliterators from languages within the same language family and written in the same script. Since building transliterators are costly, this gives us an estimate of whether it is possible to rely on the related transliterators when we do not have a language-specific one at hand. (ii) RAND",
+        "bbox": [
+            112,
+            722,
+            489,
+            852
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "refers to a random setting where we associate any non-ASCII character with any ASCII character, giving us an estimate of whether we actually need knowledge of the language to build transliterations. The results are provided in Table 3: UROMAN is largely and consistently outperforming both BORROW and RAND, where the single exception is BORROW (from Hindi to Bhojpuri). Surprisingly, RAND also yields reasonable performance and on average even outperforms the Non-TransLA+INV variant with non-transliterated data (21.59 vs 18.39 in Table 4 later). This provides further evidence towards the utility of transliteration in general and UROMAN-based romanization in particular to assist and improve language adaptation.",
+        "bbox": [
+            507,
+            464,
+            884,
+            705
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Performance on Low-Resource Languages is summarized in Table 4 and Table 5. We note that  $\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{INV}}$  outperforms Non-Trans  $\\mathrm{LA} + \\mathrm{INV}$  for all the languages with unseen scripts, and achieves that with huge margins ( $\\sim$  8-22 points for NER and  $\\sim$  17 points in UAS scores). We observe similar trends for some of the languages with seen scripts such as Min Dong (cdo), Sindhi (sd), Mingrelian",
+        "bbox": [
+            507,
+            712,
+            882,
+            841
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>9</sup>We also compare the performance of these methods with the standard cross-lingual transfer setup for finetuning an out-of-the-box mBERT for languages with unseen scripts in Appendix B. All the adapter-based methods massively outperform an out-of-the-box mBERT in this scenario.",
+        "bbox": [
+            507,
+            857,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "For reproducibility, the links to the language-specific translitators are available in Appendix A",
+        "bbox": [
+            112,
+            866,
+            485,
+            891
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{8}$ E.g., a Hindi transliterator can be borrowed for Bhojpuri since the two are related and written in Devanagari.",
+        "bbox": [
+            112,
+            892,
+            485,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "7999",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/a30d48cb3ddb4d2c393c30a8025c0d311b6c789c062f842d428fb4c9ea4a8963.jpg",
+        "image_caption": [
+            "Figure 2: Sample efficiency in the NER task."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            84,
+            472,
+            206
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/1c950205100aab4509e107caa10d380e39c2abd245e0aba617d30a97a05dd5fc.jpg",
+        "image_caption": [
+            "Figure 3: Sample efficiency in the DP task."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            244,
+            443,
+            355
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "(xmf) on NER tasks and Erzya (myv) on DP. The less efficient  $\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{mPLM_{ft}}}$ , as expected, further improves the performance for all the languages except for Tibetan (bo).<sup>10</sup> Non-Trans $_{\\mathrm{LA} + \\mathrm{Emb}_{\\mathrm{Lex}}}$ , however, now outperforms UROMAN-based methods for a majority of the languages. This observation can be attributed to various factors related to mBERT's tokenizer, and we provide an in-depth analysis later in Appendix C. Nonetheless, we observe strong and competitive performance of  $\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{mPLM_{ft}}}$  in both tasks, again indicating that more attention should be put on transliteration-based language adaptation in future work.",
+        "bbox": [
+            112,
+            401,
+            489,
+            609
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Sample Efficiency. Finally, we simulate a few-shot setup to study the effectiveness of using transliterated versus non-transliterated data in data-scarce scenarios. For NER, we evaluate performance on all the languages and on languages with unseen scripts; for DP, we evaluate on all the languages. Figure 2 indicates that  $\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{INV}}$  on average performs better than all the other methods at sample sizes 100 (i.e., 100 sentences in the target language) and 1,000. However, from 10,000 sentences onward, Non-Trans  $\\mathrm{LA} + \\mathrm{Emb}_{\\mathrm{Lex}}$  takes the lead. We observe similar trends in the DP task (see Fig 3). This establishes the utility of transliteration for (extremely) low-resource scenarios.",
+        "bbox": [
+            112,
+            620,
+            489,
+            845
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            640,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this work, we have systematically analyzed and confirmed the potential of romanization, implemented via the UROMAN tool, to help with adaptation of multilingual pretrained language models. Given (i) its broad applicability and (ii) strong performance overall and for languages with unseen scripts, we hope our study will inspire more work on transliteration-based adaptation.",
+        "bbox": [
+            507,
+            109,
+            885,
+            237
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            249,
+            613,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we work with UROMAN (Hermjakob et al., 2018) which is an unsupervised romanization tool. While it is an effective tool for romanization at scale, it still has potential drawbacks. Since it is only based on lexical substitution, its transliterations may not semantically or phonetically align with the source content and may differ from transliterations preferred by native speakers. Moreover, UROMAN is not invertible—as we have highlighted—and may thus be less appealing when text in the original script needs to be exactly reproduced. Our proposed method, while it is parameter-efficient and effective—particularly for low-resource languages—still underperforms language-specific tokenizer-based non-transliteration methods. Future work may focus on developing an improved and more efficient tokenizer for transliteration-based methods as we highlight in the Appendix.",
+        "bbox": [
+            507,
+            275,
+            884,
+            580
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "While there is now a growing body of available evaluation resources for low-resource languages (Ebrahimimi et al., 2022; Mhaske et al., 2023; Winata et al., 2023, among others), our final selection of tasks, resources and languages has been driven and constrained by the specific concrete goal of our short paper: studying and evaluating if and how transliteration/romanization can help with adaptation of languages with scripts unseen by the pretrained multilingual language model. We thus closely follow the experimental setup of Pfeiffer et al. (2021) which used the same set of tasks and languages with unseen scripts.",
+        "bbox": [
+            507,
+            581,
+            882,
+            789
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Finally, romanization can be seen as a step towards providing more universal, or rather language-agnostic, input text representation. Full-fledged comparisons against other approaches that aim to strike language independence at the input or feature level, such as byte-level models (e.g., ByT5 (Xue et al., 2022)) and pixel-based models (e.g., PILXEL (Rust et al., 2023)) go beyond",
+        "bbox": [
+            507,
+            791,
+            884,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "10 For Tibetan, longer words are composed using shorter words separated by tsek (\":\") which is not a valid space delimiter for the mBERT tokenizer; the number of produced subwords is thus much higher than for the other languages.",
+        "bbox": [
+            112,
+            868,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8000",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "the scope of this particular work, but we point out to this as a very interesting future research avenue. Moreover, the integration of these language-agnostic representations with 'romanization'-based approaches might yield additional benefits, and should also be attested in future research.",
+        "bbox": [
+            112,
+            84,
+            489,
+            180
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            193,
+            285,
+            209
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This work has been funded by the German Research Foundation (DFG) as part of the Research Training Group KRITIS No. GRK 2222. The work of Ivan Vulić has been supported by a personal Royal Society University Research Fellowship (no 221137; 2022-).",
+        "bbox": [
+            112,
+            219,
+            487,
+            313
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We thank Indraneil Paul, Yongxin Huang, Ivan Habernal, and Massimo Nicosia for their valuable feedback and suggestions on a draft of this paper.",
+        "bbox": [
+            112,
+            316,
+            485,
+            365
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            391,
+            213,
+            407
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Chantal Amrhein and Rico Sennrich. 2020. On Romanization for model transfer between scripts in neural machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 2461-2469, Online. Association for Computational Linguistics.",
+            "Alan Ansell, Edoardo Maria Ponti, Jonas Pfeiffer, Sebastian Ruder, Goran Glavaš, Ivan Vulić, and Anna Korhonen. 2021. MAD-G: Multilingual adapter generation for efficient cross-lingual transfer. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 4762–4781, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics.",
+            "Ethan C. Chau and Noah A. Smith. 2021. Specializing multilingual language models: An empirical study. In Proceedings of the 1st Workshop on Multilingual Representation Learning, pages 51-61, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            414,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Peter T. Daniels and William Bright. 1996. The world's writing systems. Oxford University Press.",
+            "Isin Demirsahin, Cibu Johny, Alexander Gutkin, and Brian Roark. 2022. Criteria for useful automatic Romanization in South Asian languages. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6662-6673, Marseille, France. European Language Resources Association.",
+            "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.",
+            "Tejas Dhamecha, Rudra Murthy, Samarth Bharadwaj, Karthik Sankaranarayanan, and Pushpak Bhattacharyya. 2021. Role of Language Relatedness in Multilingual Fine-tuning of Language Models: A Case Study in Indo-Aryan Languages. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8584-8595, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency parsing. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. Open-Review.net.",
+            "Abteen Ebrahimi, Manuel Mager, Arturo Oncevay, Vishrav Chaudhary, Luis Chiruzzo, Angela Fan, John Ortega, Ricardo Ramos, Annette Rios, Ivan Vladimir Meza Ruiz, Gustavo Giménez-Lugo, Elisabeth Mager, Graham Neubig, Alexis Palmer, Rolando Coto-Solano, Thang Vu, and Katharina Kann. 2022. AmericasNLI: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6279-6299, Dublin, Ireland. Association for Computational Linguistics.",
+            "Goran Glavaš and Ivan Vulić. 2021. Is supervised syntactic parsing beneficial for language understanding tasks? an empirical investigation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3090-3104, Online. Association for Computational Linguistics.",
+            "Ulf Hermjakob, Jonathan May, and Kevin Knight. 2018. Out-of-the-box universal Romanization tool uroman. In Proceedings of ACL 2018, System Demonstrations, pages 13-18, Melbourne, Australia. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            84,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8001",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR.",
+            "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalisation. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 4411-4421. PMLR.",
+            "Ying Lin, Cash Costello, Boliang Zhang, Di Lu, Heng Ji, James Mayfield, and Paul McNamee. 2018. Platforms for non-speakers annotating names in any language. In Proceedings of ACL 2018, System Demonstrations, pages 1-6, Melbourne, Australia. Association for Computational Linguistics.",
+            "Yash Madhani, Sushane Parthan, Priyanka A. Bedekar, Ruchi Khapra, Vivek Seshadri, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. Aksharantar: Towards building open transliteration tools for the next billion users. arXiv preprint.",
+            "Michael McCloskey and Neal J. Cohen. 1989. Catastrophic interference in connectionist networks: The sequential learning problem. volume 24 of Psychology of Learning and Motivation, pages 109-165. Academic Press.",
+            "Arnav Mhaske, Harshit Kedia, Sumanth Doddapaneni, Mitesh M. Khapra, Pratyush Kumar, Rudra Murthy, and Anoop Kunchukuttan. 2023. Naamapadam: A large-scale named entity annotated data for Indic languages. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10441-10456, Toronto, Canada. Association for Computational Linguistics.",
+            "Nikita Moghe, Evgeniia Razumovskaia, Liane Guillou, Ivan Vulic, Anna Korhonen, and Alexandra Birch. 2023. Multi3NLU++: A multilingual, multi-intent, multi-domain dataset for natural language understanding in task-oriented dialogue. In Findings of the Association for Computational Linguistics: ACL 2023, pages 3732-3755, Toronto, Canada. Association for Computational Linguistics.",
+            "Ibraheem Muhammad Moosa, Mahmud Elahi Akhter, and Ashfia Binte Habib. 2023. Does transliteration help multilingual language modeling? In *Findings of the Association for Computational Linguistics: EACL* 2023, pages 670–685, Dubrovnik, Croatia. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Benjamin Muller, Antonios Anastasopoulos, Benoit Sagot, and Djamé Seddah. 2021. When being unseen from mBERT is just the beginning: Handling new languages with multilingual language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 448-462, Online. Association for Computational Linguistics.",
+            "Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Jan Hajic, Christopher D. Manning, Sampo Pyysalo, Sebastian Schuster, Francis Tyers, and Daniel Zeman. 2020. Universal Dependencies v2: An evergrowing multilingual treebank collection. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4034-4043, Marseille, France. European Language Resources Association.",
+            "Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1946–1958, Vancouver, Canada. Association for Computational Linguistics.",
+            "Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3479-3495, Seattle, United States. Association for Computational Linguistics.",
+            "Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020a. AdapterHub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics.",
+            "Jonas Pfeiffer, Sebastian Ruder, Ivan Vulić, and Edoardo Maria Ponti. 2023. Modular deep learning. arXiv preprint.",
+            "Jonas Pfeiffer, Ivan Vulić, Iryna Gurevych, and Sebastian Ruder. 2020b. MAD-X: An Adapter-Based Framework for Multi-Task Cross-Linguual Transfer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7654–7673, Online. Association for Computational Linguistics.",
+            "Jonas Pfeiffer, Ivan Vulić, Iryna Gurevych, and Sebastian Ruder. 2021. UNKs everywhere: Adapting multilingual language models to new scripts. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10186-10203, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8002",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Edoardo Maria Ponti, Goran Glavaš, Olga Majewska, Qianchu Liu, Ivan Vulić, and Anna Korhonen. 2020. XCOPA: A multilingual dataset for causal commonsense reasoning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362-2376, Online. Association for Computational Linguistics.",
+            "Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 151-164, Florence, Italy. Association for Computational Linguistics.",
+            "Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. 2023. Language modelling with pixels. In *The Eleventh International Conference on Learning Representations*, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.",
+            "Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? on the monolingual performance of multilingual language models. 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 3118-3135, Online. Association for Computational Linguistics.",
+            "Daan van Esch, Tamar Lucassen, Sebastian Ruder, Isaac Caswell, and Clara Rivera. 2022. Writing system and speaker metadata for  $2,800+$  language varieties. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 5035-5046, Marseille, France. European Language Resources Association.",
+            "Hans H Wellisch, Lee Breuer, Richard Foreman, and Robert Wilson. 1978. The conversion of scripts, its nature, history, and utilization. New York: Toronto: Wiley.",
+            "Genta Indra Winata, Alham Fikri Aji, Samuel Cahyawijaya, Rahmad Mahendra, Fajri Koto, Ade Romadhony, Kemal Kurniawan, David Moeljadi, Radityo Eko Prasojo, Pascale Fung, Timothy Baldwin, Jey Han Lau, Rico Sennrich, and Sebastian Ruder. 2023. NusaX: Multilingual parallel sentiment dataset for 10 Indonesian local languages. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 815-834, Dubrovnik, Croatia. Association for Computational Linguistics.",
+            "Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. 2022. ByT5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics, 10:291-306.",
+            "Seid Muhie Yimam, Abinew Ali Ayele, Gopalakrishnan Venkatesh, Ibrahim Gashaw, and Chris Biemann. 2021. Introducing various semantic models"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "ref_text",
+        "text": "for amharic: Experimentation and evaluation with multiple tasks and datasets. Future Internet, 13(11).",
+        "bbox": [
+            527,
+            85,
+            880,
+            112
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Judit Ács. Exploring bert's vocabulary. Blog Post.",
+        "bbox": [
+            510,
+            122,
+            852,
+            137
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8003",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A Transliterators in Evaluation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            406,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Besides UROMAN, we also employ various language-specific transliterations which are publicly available. We list them in Table 6.",
+        "bbox": [
+            112,
+            112,
+            487,
+            159
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B Performance comparison of mBERT",
+        "text_level": 1,
+        "bbox": [
+            114,
+            173,
+            467,
+            191
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We adapt the standard cross-lingual transfer setup for mBERT. The model is finetuned on the task data for a source language (high-resource) and is used to perform inference on the target language (low-resource). We report the performance comparison of the standard cross-lingual transfer setup for mBERT on the NER task for languages with unseen scripts with the adapter-based methods in Table 7. We observe that the adapter-based methods outperform mBERT by huge margins.",
+        "bbox": [
+            112,
+            200,
+            489,
+            361
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "C Further Analyses",
+        "text_level": 1,
+        "bbox": [
+            112,
+            375,
+            305,
+            392
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Following previous work (Ács; Rust et al., 2021; Moosa et al., 2023), we further analyze tokenization quality of the mBERT tokenizer using the following established metrics: 1) % of \"UNK\"s measures the % of \"UNK\" tokens produced by the tokenizer, and our aim is to compare their rate before and after transliteration; 2) fertility measures the number of subwords that are produced per tokenized word; 3) proportion of continued subwords measures the proportion of words for which the tokenized word is split across at least two subwords (denoted by the symbol ##).",
+        "bbox": [
+            112,
+            401,
+            489,
+            595
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "From the results summarized in Figure 4, it is apparent that transliteration drastically reduces  $\\%$  of UNKs. However, mBERT's tokenizer underperforms as compared to monolingual tokenizers based on fertility and the proportion of continued subwords (Rust et al., 2021). Transliteration performs better for some languages where the quality of the mBERT tokenizer is similar to the monolingual tokenizer such as for dv, km, and cdo. On the other hand, transliteration methods perform worse on languages where the quality of the underlying mBERT tokenizer is relatively poor.",
+        "bbox": [
+            112,
+            596,
+            489,
+            789
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In order to test the hypothesis that the tokenizer quality might be the principal reason for the performance gap for the transliteration-based methods in comparison to the non-transliteration based methods, we carried out an additional experiment. For the experiment, we adapt the Non-Trans $_{\\mathrm{LA + Emb_{Lex}}}$  to operate on transliterated data, and call this variant Trans $_{\\mathrm{LA + Emb_{Lex}}}$ . Here, we train a new tokenizer",
+        "bbox": [
+            112,
+            791,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "on the transliterated data and initialize lexically overlapping embeddings with mBERT's pretrained embeddings.",
+        "bbox": [
+            507,
+            84,
+            882,
+            131
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We plot the performance in Figure 5. The new method,  $\\mathrm{Trans_{LA + Emb_{Lex}}}$  now outperforms the nontransliteration-based variant on 8/12 languages and also on average. Consequently, this validates our hypothesis and is in line with the previous work (Moosa et al., 2023). However, we found a drop in performance in the case of mhr (-10.71) and cdo (-10.14) when compared to  $\\mathrm{Trans_{LA} + mPLM_{ft}}$ . These drops may be attributed to the lower degree of lexical overlap with mBERT's vocabulary, and consequently a higher number of randomly initialized embeddings for those target languages.",
+        "bbox": [
+            507,
+            133,
+            884,
+            326
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8004",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/b801220320630fbbe81d4b101ae56e278240e06979cf58e131cb98a4654f2a8e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Transliterator</td><td>Used for languages</td><td>Available at</td></tr><tr><td>UROMAN</td><td>All</td><td>github.com/isi-nlp/uroman</td></tr><tr><td>amseg</td><td>am</td><td>pypi.org/project/amseg/</td></tr><tr><td>transliterate</td><td>ru, ka</td><td>pypi.org/project/transliterate/</td></tr><tr><td>ai4bharat-transliteration</td><td>hi, sd</td><td>pypi.org/project/ai4bharat-transliteration/</td></tr><tr><td>lang-trans</td><td>ar</td><td>pypi.org/project/lang-trans/</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            884,
+            171
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "image",
+        "img_path": "images/8309dd1ed12a59dc1f0528edf5e64d0557b4f8d3dbec0880e76b532cd845f7ca.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            208,
+            349,
+            278
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "image",
+        "img_path": "images/47711266c4017b2aa9f669ce98d4d7d7d04173802500672272aea02a88d75ca3.jpg",
+        "image_caption": [
+            "(b)",
+            "Figure 4: Tokenizer quality analysis. a) % of UNKs before and after transliteration, b) Fertility, and c) Proportion of continued subwords for mBERT vs monolingual tokenizer."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            381,
+            208,
+            616,
+            278
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "image",
+        "img_path": "images/fb3ee83c84d87592bf4ee3c75d4c4ae5a5a1aa7ed161b08c329155e3eb3efc12.jpg",
+        "image_caption": [
+            "(c)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            648,
+            206,
+            880,
+            278
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/f32a117937df02c309221b0fc5c6eaacf022d84aedff963fd480dd7bdd40a24c.jpg",
+        "table_caption": [
+            "Table 6: Transliterators used in this work."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Variant</td><td colspan=\"5\">Unseen Scripts</td></tr><tr><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td></tr><tr><td>mBERT</td><td>0.91</td><td>17.40</td><td>1.30</td><td>10.71</td><td>2.50</td></tr><tr><td>Non-TransLA+INV</td><td>3.26</td><td>19.86</td><td>18.72</td><td>13.81</td><td>23.14</td></tr><tr><td>TransLA + INV</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td></tr><tr><td>Non-TransLA+EmblEx</td><td>47.60</td><td>40.96</td><td>31.21</td><td>53.94</td><td>45.89</td></tr><tr><td>TransLA + mPLMft</td><td>37.09</td><td>33.82</td><td>40.40</td><td>52.39</td><td>45.24</td></tr></table>",
+        "bbox": [
+            132,
+            426,
+            470,
+            501
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Performance Comparison of mBERT with adapter-based finetuning methods for unseen scripts on the NER task.",
+        "bbox": [
+            112,
+            510,
+            487,
+            552
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "image",
+        "img_path": "images/f862f3a59498858d80d8303dd287b78b2713cb4ca14a6614907596786428fec4.jpg",
+        "image_caption": [
+            "Figure 5: Comparison of Non-Trans $_{\\mathrm{LA + Emb_{Lex}}}$  with Trans $_{\\mathrm{LA + Emb_{Lex}}}$  on NER (left) and DP (right)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            700,
+            487,
+            807
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8005",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    }
+]

2023/Romanization-based Large-scale Adaptation of Multilingual Language Models/1a02cba7-8ebc-4f20-8832-cb43efde101d_model.json

@@ -0,0 +1,1815 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.269,
+                0.08,
+                0.731,
+                0.12
+            ],
+            "angle": 0,
+            "content": "Romanization-based Large-scale Adaptation of Multilingual Language Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.125,
+                0.886,
+                0.143
+            ],
+            "angle": 0,
+            "content": "Sukannya Purkayastha<sup>1</sup>, Sebastian Ruder<sup>2</sup>, Jonas Pfeiffer<sup>2</sup>, Iryna Gurevych<sup>1</sup>, Ivan Vulić<sup>3</sup>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.334,
+                0.144,
+                0.671,
+                0.159
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup> Ubiquitous Knowledge Processing Lab,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.202,
+                0.16,
+                0.802,
+                0.176
+            ],
+            "angle": 0,
+            "content": "Department of Computer Science and Hessian Center for AI (hessian.AI),"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.36,
+                0.177,
+                0.643,
+                0.192
+            ],
+            "angle": 0,
+            "content": "Technical University of Darmstadt"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.418,
+                0.193,
+                0.586,
+                0.209
+            ],
+            "angle": 0,
+            "content": "Google DeepMind"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.281,
+                0.21,
+                0.723,
+                0.226
+            ],
+            "angle": 0,
+            "content": "\\(^{3}\\) Language Technology Lab, University of Cambridge"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.361,
+                0.228,
+                0.642,
+                0.242
+            ],
+            "angle": 0,
+            "content": "www.ukp.tu-darmstadt.de"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.281,
+                0.461,
+                0.736
+            ],
+            "angle": 0,
+            "content": "Large multilingual pretrained language models (mPLMs) have become the de facto state of the art for cross-lingual transfer in NLP. However, their large-scale deployment to many languages, besides pretraining data scarcity, is also hindered by the increase in vocabulary size and limitations in their parameter budget. In order to boost the capacity of mPLMs to deal with low-resource and unseen languages, we explore the potential of leveraging transliteration on a massive scale. In particular, we explore the UROMAN transliteration tool, which provides mappings from UTF-8 to Latin characters for all the writing systems, enabling inexpensive romanization for virtually any language. We first focus on establishing how UROMAN compares against other language-specific and manually curated transliterators for adapting multilingual PLMs. We then study and compare a plethora of data- and parameter-efficient strategies for adapting the mPLMs to romanized and non-romanized corpora of 14 diverse low-resource languages. Our results reveal that UROMAN-based transliteration can offer strong performance for many languages, with particular gains achieved in the most challenging setups: on languages with unseen scripts and with limited training data without any vocabulary augmentation. Further analyses reveal that an improved tokenizer based on romanized data can even outperform non-transliteration-based methods in the majority of languages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.749,
+                0.26,
+                0.764
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.775,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Massively multilingual language models (mPLMs) such as mBERT (Devlin et al., 2019) and XLM-R (Conneau et al., 2020) have become the driving force for a variety of applications in multilingual NLP (Ponti et al., 2020; Hu et al., 2020; Moghe et al., 2023). However, guaranteeing and maintaining strong performance for a wide spectrum of low-resource languages is difficult due to two crucial problems. The first issue is the vocabulary"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.513,
+                0.251,
+                0.875,
+                0.316
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Language</td><td>Original Text</td><td>Romanized Text</td><td>Translation</td></tr><tr><td>Bhojpuri</td><td>बाल्दियम् गार्वा</td><td>jorjiyan bhaasaas</td><td>Georgian language</td></tr><tr><td>Sinhala</td><td>सिकूल्तिूल्तिूल्तिूल्तिूल्तिूल्तिू�िूल्तिूल्तिूल्तिू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�र्था</td><td></td><td></td></tr><tr><td>Sindhi</td><td>सिकूल्तिूल्तिूल्तिू�िू�िू�िू�र्था</td><td>grahalooka</td><td>planets</td></tr><tr><td>Khmer</td><td>khμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμmu μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμum μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ µ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μµ μµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµüµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµúµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ µµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₃₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₂₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₁₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀���₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀⁻⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁰⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁵⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁲⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁴⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁽⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁻⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁼⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸ⁱ⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁅⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁰⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₃⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄²⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄.\n&lt;|im_start|&gt;assistant</td><td>Bhojpuri</td><td>BJJUJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJSJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ JJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ</td></tr></table>"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.521,
+                0.326,
+                0.871,
+                0.341
+            ],
+            "angle": 0,
+            "content": "Figure 1: Romanization across different languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.367,
+                0.885,
+                0.496
+            ],
+            "angle": 0,
+            "content": "size, as the vocabulary is bound to increase with the number of languages added if per-language performance is to be maintained (Hu et al., 2020; Artetxe et al., 2020; Pfeiffer et al., 2022). Second, pretraining mPLMs with a fixed model capacity improves cross-lingual performance up to a point after which it starts to decrease; this is the phenomenon termed the curse of multilinguality (Conneau et al., 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.497,
+                0.885,
+                0.834
+            ],
+            "angle": 0,
+            "content": "Transliteration refers to the process of converting language represented in one writing system to another (Wellisch et al., 1978). Latin script-centered transliteration or romanization is the most common form of transliteration (Lin et al., 2018; Amrhein and Sennrich, 2020; Demirsin et al., 2022) as the Latin/Roman script is by far the most widely adopted writing script in the world (Daniels and Bright, 1996; van Esch et al., 2022). Adapting mPLMs via transliteration can address the two aforementioned critical issues. 1) Since the Latin script covers a dominant portion of the mPLM's vocabulary (e.g., \\(77\\%\\) in case of mBERT, see Ács), 'romanizing' the remaining part of the vocabulary might mitigate the vocabulary size issue and boost vocabulary sharing. 2) Since no new tokens are added during the romanization process, reusing pretrained embeddings from the mPLM's embedding matrix helps reuse the information already present within the mPLM, thereby allocating the model's parameter budget more efficiently."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.835,
+                0.884,
+                0.883
+            ],
+            "angle": 0,
+            "content": "However, the main drawback of transliteration seems to be the expensive process of creating effective language-specific transliterations, as they"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.893,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>According to Encyclopedia Britannica, up to 70% of the world population is employing the Latin script."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7996"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.78,
+                0.959
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7996-8005"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.278,
+                0.959,
+                0.72,
+                0.972
+            ],
+            "angle": 0,
+            "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.085,
+                0.49,
+                0.246
+            ],
+            "angle": 0,
+            "content": "typically require language expertise to curate dictionaries that map tokens from one language and script to another. Therefore, previous attempts at mPLM adaptation to unseen languages via transliteration (Muller et al., 2021; Chau and Smith, 2021; Dhamecha et al., 2021; Moosa et al., 2023) were constrained to a handful of languages due to the limited availability of language-specific translit-. rators, or were applied only to languages that have 'language siblings' with developed translitators."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.247,
+                0.49,
+                0.392
+            ],
+            "angle": 0,
+            "content": "In this work, unlike previous work, we propose to use and then evaluate the usefulness of a universal romanization tool, UROMAN (Hermjakob et al., 2018), for quick, large-scale and effective adaptation of mPLMs to low-resource languages. The UROMAN tool disposes of language-specific curated dictionaries and maps any UTF-8 character to the Latin script, increasing the portability of romanization, with some examples in Figure 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.393,
+                0.49,
+                0.699
+            ],
+            "angle": 0,
+            "content": "We analyze language adaptation on a massive scale via UROMAN-based romanization on a set of 14 diverse low-resource languages. We conduct experiments within the standard parameter-efficient adapter-based cross-lingual transfer setup on two tasks: Named Entity Recognition (NER) on the WikiANN dataset (Pan et al., 2017; Rahimi et al., 2019), and Dependency Parsing (DP) with Universal Dependencies v2.7 (Nivre et al., 2020). Our key results suggest that UROMAN-based transliteration can offer strong performance on par or even outperforming adaptation with language-specific transliterrators, setting up the basis for wider use of transliteration-based mPLM adaptation techniques in future work. The gains with romanization-based adaptation over standard adaptation baselines are particularly pronounced for languages with unseen scripts (\\(\\sim\\)8-22 performance points) without any vocabulary augmentation.[2]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.713,
+                0.258,
+                0.73
+            ],
+            "angle": 0,
+            "content": "2 Background"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.741,
+                0.49,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Why UROMAN-Based Romanization? UROMAN-based romanization is not always fully reversible, and its usage for transliteration has thus been limited in the literature. However, due to its high portability, UROMAN can help scale the process of transliteration massively and as such benefit low-resource scenarios and wider adaptation of mPLMs. The main idea, as hinted in §1, is to (learn to) map"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.135
+            ],
+            "angle": 0,
+            "content": "any UTF-8 character to the Latin script, without the use of any external language-specific dictionaries (see Hermjakob et al. (2018) for technical details)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.139,
+                0.885,
+                0.413
+            ],
+            "angle": 0,
+            "content": "Cross-Lingual Transfer to Low-Resource Languages. Parameter-efficient and modular finetuning methods (Pfeiffer et al., 2023) such as adapters (Houlsby et al., 2019; Pfeiffer et al., 2020b) have been used for cross-lingual transfer, putting a particular focus on enabling transfer to low-resource languages and scenarios, including languages with scripts 'unseen' by the base mPLM (Pfeiffer et al., 2021). Adapters are small lightweight components stitched into the base mPLM, and then trained for particular languages and tasks while keeping the parameters of the original mPLM frozen. This circumvents the issues of catastrophic forgetting and interference (McCloskey and Cohen, 1989) within the mPLM, and allows for extending its reach also to unseen languages (Pfeiffer et al., 2021; Ansell et al., 2021)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.414,
+                0.885,
+                0.737
+            ],
+            "angle": 0,
+            "content": "For our main empirical analyses, we adopt a state-of-the-art modular method for cross-lingual transfer: MAD-X (Pfeiffer et al., 2020b). In short, MAD-X is based on language adapters (LA), task adapters (TA), and invertible adapters (INV). While LAs are trained for specific languages relying on masked language modeling, TAs are trained with high-resource languages relying on task-annotated data and task-specific objectives. At inference, the source LA is replaced with the target LA while the TA is kept. In order to do parameter-efficient learning for the token-level embeddings across different languages and to deal with the vocabulary mismatch between source and target languages, Pfeiffer et al. (2020b) also propose INV adapters: they are placed on top of the embedding layer and their inverses precede the output embedding layer.<sup>3</sup> We adopt the better-performing MAD-X 2.0 setup (Pfeiffer et al., 2021) where the adapters in the last Transformer layer are dropped at inference.<sup>4</sup>"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.75,
+                0.762,
+                0.767
+            ],
+            "angle": 0,
+            "content": "3 Experiments and Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.778,
+                0.885,
+                0.858
+            ],
+            "angle": 0,
+            "content": "As the main means of analyzing the impact of transliteration in general and UROMAN-based romanization in particular, we train different variants of language adapters within the MAD-X framework, based on transliterated and non-transliterated"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.881,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)Code and data available at https://github.com/UKPLab/emnlp23_romanization_based_adaptation"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.868,
+                0.883,
+                0.894
+            ],
+            "angle": 0,
+            "content": "3They are trained together with the LAs while the rest of the mPLM is kept frozen."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.894,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "4We refer the reader to the original papers for further technical details regarding the MAD-X framework."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.509,
+                0.868,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7997"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.119,
+                0.082,
+                0.486,
+                0.24
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Language</td><td>Family</td><td>Script</td><td># Sentences</td></tr><tr><td>Bhojpuri (bh)</td><td>Indo-European</td><td>Devanagari</td><td>35,983</td></tr><tr><td>Buryat (bxr)</td><td>Mongolic</td><td>Cyrillic</td><td>41,692</td></tr><tr><td>Erzya (myv)</td><td>Uralic</td><td>Cyrillic</td><td>42,575</td></tr><tr><td>Meadow Mari (mhr)</td><td>Uralic</td><td>Cyrillic</td><td>144,529</td></tr><tr><td>Min Dong (cdo)</td><td>Sino-Tibetan</td><td>Chinese</td><td>33,978</td></tr><tr><td>Mingrelian (xmf)</td><td>Kartvelian</td><td>Georgian</td><td>63,032</td></tr><tr><td>Sindhi (sd)</td><td>Indo-European</td><td>Arabic</td><td>86,176</td></tr><tr><td>Sorani Kurdish (ckb)</td><td>Indo-European</td><td>Arabic</td><td>459,475</td></tr><tr><td>Uyghur (ug)</td><td>Turkic</td><td>Arabic</td><td>149,813</td></tr><tr><td>Amháric (am)</td><td>Afro-Asiatic</td><td>Ge&#x27;ez</td><td>88,320</td></tr><tr><td>Divehi (dv)</td><td>Indo-European</td><td>Thaana</td><td>34,779</td></tr><tr><td>Khmer (km)</td><td>Austroasiatic</td><td>Khmer</td><td>139,704</td></tr><tr><td>Sinhala (si)</td><td>Indo-European</td><td>Sinhala</td><td>219,866</td></tr><tr><td>Tibetan (bo)</td><td>Sino-Tibetan</td><td>Tibetan</td><td>131,362</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.251,
+                0.49,
+                0.323
+            ],
+            "angle": 0,
+            "content": "Table 1: Languages with their ISO 639-3 codes used in our evaluation, along with their script, language family, and number of sentences available for pretraining. The dashed line separates languages with unseen scripts, placed in the bottom part of the table."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.352,
+                0.462,
+                0.367
+            ],
+            "angle": 0,
+            "content": "versions of target language data, outlined here."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.375,
+                0.49,
+                0.568
+            ],
+            "angle": 0,
+            "content": "Variants with Non-Transliterated Data. For the Non-Trans\\(_{\\mathrm{LA} + \\mathrm{INV}}\\) variant, we train LAs and INV adapters together. This variant serves to examine the extent to which mPLMs can adapt to unseen languages without any vocabulary extension.\\(^5\\) We compare this to Non-Trans\\(_{\\mathrm{LA} + \\mathrm{Emb}_{\\mathrm{Lex}}}\\), which trains a new tokenizer for the target language (Pfeiffer et al., 2021): the so-called 'lexically overlapping' tokens are initialized with mPLM's trained embeddings, while the remaining embeddings are initialized randomly. All these embeddings (Emb\\(_{\\mathrm{Lex}}\\)) are fine-tuned along with LAs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.576,
+                0.49,
+                0.817
+            ],
+            "angle": 0,
+            "content": "Variants with Transliterated Data. We evaluate a \\(\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{INV}}\\) variant, which uses the same setup as Non-Trans \\(\\mathrm{LA} + \\mathrm{INV}\\) but now with transliterated data. We again note that in this efficient setup, we do not extend the vocabulary size, and use the fewest trainable parameters. In the \\(\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{mPLM}_{\\mathrm{ft}}}\\) variant, we train LAs along with fine-tuning the pretrained embeddings of mPLM (mPLMft). This further enhances the model capacity by fine-tuning the embedding layer instead of using invertible adapters. For both variants, transliterated data can be produced via different transliterations: (i) language-specific ones; (ii) the ones from 'language siblings' (e.g., using a Georgian transliterator for Mingrelian), or (iii) UROMAN."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.085,
+                0.715,
+                0.101
+            ],
+            "angle": 0,
+            "content": "3.1 Experimental Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.106,
+                0.885,
+                0.331
+            ],
+            "angle": 0,
+            "content": "Data, Languages and Tasks. Following Pfeiffer et al. (2021), we select mBERT as our base mPLM. We experiment with 14 typologically diverse low-resource languages that are not part of mBERT's pretraining corpora, with 5/14 languages written in distinct scripts (see Table 1 for details). For LA training, we use Wikipedia dumps for the target languages, which we also transliterate (using different translitators). Evaluation is conducted on two standard cross-lingual transfer tasks in zero-shot setups: 1) the WikiAnn NER dataset (Pan et al., 2017) with the train, dev, and test splits from (Rahimi et al., 2019); 2) for dependency parsing, we rely on the UD Dataset v2.7 (Nivre et al., 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.336,
+                0.884,
+                0.417
+            ],
+            "angle": 0,
+            "content": "LAs and TAs. English is the source language in all experiments, and is used for training TAs. The English LA is obtained directly from Adapterhub.ml (Pfeiffer et al., 2020a), LAs and embeddings (when needed) are only trained for target languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.418,
+                0.882,
+                0.464
+            ],
+            "angle": 0,
+            "content": "Finally, for the Non-Trans\\(_{\\mathrm{LA + Emb_{Lex}}}\\) variant, we train a WordPiece tokenizer on the target language data with a vocabulary size of 10K."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.471,
+                0.884,
+                0.534
+            ],
+            "angle": 0,
+            "content": "Training of Language and Task Adapters. We train all the language adapters for 50 epochs or \\(\\sim 50K\\) update steps based on the corpus size. The batch size is set to 64 and the learning rate is \\(1e - 4\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.535,
+                0.884,
+                0.679
+            ],
+            "angle": 0,
+            "content": "We train English task adapters following the setup from (Pfeiffer et al., 2020b). For NER, we directly obtain the task adapter from Adapterhub.ml which is trained with a learning rate of \\(1e - 4\\) for 10 epochs. For DP, we train a Transformer-based (Glavaš and Vulić, 2021) biaffine attention dependency parser (Dozat and Manning, 2017). We use a learning rate of \\(5e - 4\\) and train for 10 epochs as in (Pfeiffer et al., 2021)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.68,
+                0.885,
+                0.856
+            ],
+            "angle": 0,
+            "content": "All the reported results in both tasks (NER and DP) are reported as averages over 6 random seeds. All the models have been trained on A100 or V100 GPUs. None of the training methods consumed more than 36 hours. As the main means of analyzing the impact of transliteration in general and UROMAN-based romanization in particular, we train different variants of language adapters within the MAD-X framework, based on transliterated and non-transliterated versions of target language data, outlined here."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.867,
+                0.736,
+                0.882
+            ],
+            "angle": 0,
+            "content": "3.2 Results and Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.888,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "UROMAN versus Other Transliterators and Transliteration Strategies. In order to estab"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.833,
+                0.488,
+                0.87
+            ],
+            "angle": 0,
+            "content": "<sup>5</sup>Since LAs without INV typically perform worse than with INV (Pfeiffer et al., 2020b), also confirmed in our preliminary experiments, we do not ablate to the setup without INV."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.87,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>6</sup>We do not have this setup for non-transliterated data since, for languages with unseen scripts, most of the tokens are replaced by the generic 'UNK' token, and fine-tuning embeddings hardly benefit downstream performance."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.113,
+                0.833,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7998"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.182,
+                0.082,
+                0.813,
+                0.141
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Task</td><td>Transliterator</td><td>am</td><td>ar</td><td>ka</td><td>ru</td><td>hi</td><td>sd</td><td>avg</td></tr><tr><td rowspan=\"2\">NER (Macro F1)</td><td>UROMAN</td><td>25.6</td><td>24.8</td><td>61.4</td><td>66.5</td><td>48.6</td><td>35.3</td><td>43.7</td></tr><tr><td>Other</td><td>25.5</td><td>23.7</td><td>57.3</td><td>63.9</td><td>56.7</td><td>35.9</td><td>43.8</td></tr><tr><td rowspan=\"2\">UD (UAS / LAS)</td><td>UROMAN</td><td>36.17 / 6.6</td><td>33.0 / 19.8</td><td>-</td><td>47.3 / 32.4</td><td>33.8 / 17.8</td><td>-</td><td>37.5 / 19.1</td></tr><tr><td>Other</td><td>29.9 / 5.4</td><td>32.6 / 19.9</td><td>-</td><td>45.0 / 19.9</td><td>33.2 / 17.9</td><td>-</td><td>35.2 / 15.8</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.253,
+                0.147,
+                0.741,
+                0.161
+            ],
+            "angle": 0,
+            "content": "Table 2: Comparison of UROMAN with language-specific transliterations."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.169,
+                0.844,
+                0.228
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"7\">Seen Script</td><td colspan=\"6\">Unseen Script</td><td>avg</td></tr><tr><td>bh</td><td>cdo</td><td>ckb</td><td>mhr</td><td>sd</td><td>ug</td><td>xmf</td><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td><td></td><td></td></tr><tr><td>UROMAN</td><td>32.59</td><td>27.34</td><td>67.73</td><td>64.68</td><td>35.33</td><td>28.10</td><td>52.58</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td><td>26.89</td><td></td></tr><tr><td>BORROW</td><td>53.42 (hi)</td><td>-</td><td>12.46 (ar)</td><td>45.86 (ru)</td><td>16.79 (ar)</td><td>12.85 (ar)</td><td>24.77 (ru)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td></tr><tr><td>RAND</td><td>25.42</td><td>19.51</td><td>53.55</td><td>42.02</td><td>27.20</td><td>25.18</td><td>35.82</td><td>18.00</td><td>18.95</td><td>21.19</td><td>32.75</td><td>20.01</td><td>21.59</td><td></td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.206,
+                0.234,
+                0.788,
+                0.248
+            ],
+            "angle": 0,
+            "content": "Table 3: Comparison of various transliteration strategies on the NER task (Macro-F1)."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.149,
+                0.254,
+                0.844,
+                0.326
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Variant</td><td colspan=\"7\">Seen Script</td><td colspan=\"5\">Unseen Script</td><td rowspan=\"2\">avg</td></tr><tr><td>bh</td><td>cdo</td><td>sd</td><td>xmf</td><td>mhr</td><td>ckb</td><td>ug</td><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td></tr><tr><td>Non-TransLA+INV</td><td>55.14</td><td>24.19</td><td>31.31</td><td>49.74</td><td>70.31</td><td>45.54</td><td>33.53</td><td>3.26</td><td>19.86</td><td>18.72</td><td>13.81</td><td>23.14</td><td>18.39</td></tr><tr><td>TransLA+INV</td><td>32.59</td><td>27.34</td><td>35.33</td><td>52.58</td><td>64.68</td><td>67.73</td><td>28.10</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td><td>26.89</td></tr><tr><td>Non-TransLA+EmblEx</td><td>60.00</td><td>28.91</td><td>42.47</td><td>51.99</td><td>61.05</td><td>79.12</td><td>50.42</td><td>47.60</td><td>40.96</td><td>31.21</td><td>53.94</td><td>45.89</td><td>49.01</td></tr><tr><td>TransLA + mPLMrl</td><td>49.05</td><td>36.92</td><td>39.16</td><td>57.99</td><td>69.85</td><td>73.92</td><td>33.43</td><td>37.09</td><td>33.82</td><td>40.40</td><td>52.39</td><td>45.24</td><td>47.44</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.21,
+                0.332,
+                0.785,
+                0.346
+            ],
+            "angle": 0,
+            "content": "Table 4: Results (Macro-F1 scores) on WikiAnn NER averaged over 6 random seeds."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.182,
+                0.354,
+                0.812,
+                0.426
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Variant</td><td colspan=\"4\">Seen Script</td><td colspan=\"2\">Unseen Script</td></tr><tr><td>bh</td><td>myv</td><td>ug</td><td>bxr</td><td>am</td><td>avg</td></tr><tr><td>Non-TransLA+INV</td><td>28.46 / 11.53</td><td>45.28 / 26.27</td><td>33.44 / 15.28</td><td>39.75 / 19.77</td><td>19.08 / 1.85</td><td>33.20 / 10.81</td></tr><tr><td>TransLA+INV</td><td>25.12 / 10.17</td><td>45.74 / 26.64</td><td>32.30 / 15.10</td><td>37.92 / 17.23</td><td>36.07 / 7.58</td><td>35.43 / 12.41</td></tr><tr><td>Non-TransLA+EmblEx</td><td>26.68 / 10.10</td><td>48.34 / 25.34</td><td>41.20 / 22.81</td><td>39.51 / 16.02</td><td>36.4778.39</td><td>38.44 / 12.20</td></tr><tr><td>TransLA + mPLMrl</td><td>28.04 / 11.13</td><td>41.97 / 20.29</td><td>50.89 / 16.56</td><td>35.03 / 20.29</td><td>39.10 / 9.00</td><td>39.01 / 14.65</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.184,
+                0.431,
+                0.811,
+                0.446
+            ],
+            "angle": 0,
+            "content": "Table 5: Results (UAS / LAS scores) in the DP task with UD, averaged over 6 random seeds."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.464,
+                0.49,
+                0.72
+            ],
+            "angle": 0,
+            "content": "lish the utility of UROMAN as a viable transliterator, especially for low-resource languages, we compare its performance with transliteration options using the TransLA + INV setup as the most efficient scenario. First, we compare UROMAN with language-specific transliterations available for selected languages: amseg (Yimam et al., 2021) for Amharic, ai4bharat-transliteration (Madhani et al., 2022) for Hindi and Sindhi, lang-trans for Arabic, and transliterate for Russian and Georgian<sup>7</sup>. The transliterations used in this work are outlined in Table 6. The results are provided in Table 2. On average, UROMAN performs better or comparable to the language-specific transliterations. This provides justification to use UROMAN for massive transliteration at scale."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.724,
+                0.49,
+                0.853
+            ],
+            "angle": 0,
+            "content": "Second, we compare UROMAN to two other transliteration strategies. (i) BORROW refers to borrowing transliterators from languages within the same language family and written in the same script. Since building transliterators are costly, this gives us an estimate of whether it is possible to rely on the related transliterators when we do not have a language-specific one at hand. (ii) RAND"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.465,
+                0.885,
+                0.706
+            ],
+            "angle": 0,
+            "content": "refers to a random setting where we associate any non-ASCII character with any ASCII character, giving us an estimate of whether we actually need knowledge of the language to build transliterations. The results are provided in Table 3: UROMAN is largely and consistently outperforming both BORROW and RAND, where the single exception is BORROW (from Hindi to Bhojpuri). Surprisingly, RAND also yields reasonable performance and on average even outperforms the Non-TransLA+INV variant with non-transliterated data (21.59 vs 18.39 in Table 4 later). This provides further evidence towards the utility of transliteration in general and UROMAN-based romanization in particular to assist and improve language adaptation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.713,
+                0.884,
+                0.842
+            ],
+            "angle": 0,
+            "content": "Performance on Low-Resource Languages is summarized in Table 4 and Table 5. We note that \\(\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{INV}}\\) outperforms Non-Trans \\(\\mathrm{LA} + \\mathrm{INV}\\) for all the languages with unseen scripts, and achieves that with huge margins (\\(\\sim\\) 8-22 points for NER and \\(\\sim\\) 17 points in UAS scores). We observe similar trends for some of the languages with seen scripts such as Min Dong (cdo), Sindhi (sd), Mingrelian"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.858,
+                0.884,
+                0.918
+            ],
+            "angle": 0,
+            "content": "<sup>9</sup>We also compare the performance of these methods with the standard cross-lingual transfer setup for finetuning an out-of-the-box mBERT for languages with unseen scripts in Appendix B. All the adapter-based methods massively outperform an out-of-the-box mBERT in this scenario."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.868,
+                0.486,
+                0.892
+            ],
+            "angle": 0,
+            "content": "For reproducibility, the links to the language-specific translitators are available in Appendix A"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.893,
+                0.486,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(^{8}\\)E.g., a Hindi transliterator can be borrowed for Bhojpuri since the two are related and written in Devanagari."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.114,
+                0.868,
+                0.486,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7999"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.129,
+                0.085,
+                0.473,
+                0.208
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.147,
+                0.219,
+                0.454,
+                0.234
+            ],
+            "angle": 0,
+            "content": "Figure 2: Sample efficiency in the NER task."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.122,
+                0.245,
+                0.445,
+                0.356
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.154,
+                0.368,
+                0.446,
+                0.383
+            ],
+            "angle": 0,
+            "content": "Figure 3: Sample efficiency in the DP task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.402,
+                0.49,
+                0.611
+            ],
+            "angle": 0,
+            "content": "(xmf) on NER tasks and Erzya (myv) on DP. The less efficient \\(\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{mPLM_{ft}}}\\), as expected, further improves the performance for all the languages except for Tibetan (bo).<sup>10</sup> Non-Trans\\(_{\\mathrm{LA} + \\mathrm{Emb}_{\\mathrm{Lex}}}\\), however, now outperforms UROMAN-based methods for a majority of the languages. This observation can be attributed to various factors related to mBERT's tokenizer, and we provide an in-depth analysis later in Appendix C. Nonetheless, we observe strong and competitive performance of \\(\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{mPLM_{ft}}}\\) in both tasks, again indicating that more attention should be put on transliteration-based language adaptation in future work."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.621,
+                0.49,
+                0.846
+            ],
+            "angle": 0,
+            "content": "Sample Efficiency. Finally, we simulate a few-shot setup to study the effectiveness of using transliterated versus non-transliterated data in data-scarce scenarios. For NER, we evaluate performance on all the languages and on languages with unseen scripts; for DP, we evaluate on all the languages. Figure 2 indicates that \\(\\mathrm{Trans}_{\\mathrm{LA} + \\mathrm{INV}}\\) on average performs better than all the other methods at sample sizes 100 (i.e., 100 sentences in the target language) and 1,000. However, from 10,000 sentences onward, Non-Trans \\(\\mathrm{LA} + \\mathrm{Emb}_{\\mathrm{Lex}}\\) takes the lead. We observe similar trends in the DP task (see Fig 3). This establishes the utility of transliteration for (extremely) low-resource scenarios."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.084,
+                0.642,
+                0.099
+            ],
+            "angle": 0,
+            "content": "4 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.11,
+                0.886,
+                0.239
+            ],
+            "angle": 0,
+            "content": "In this work, we have systematically analyzed and confirmed the potential of romanization, implemented via the UROMAN tool, to help with adaptation of multilingual pretrained language models. Given (i) its broad applicability and (ii) strong performance overall and for languages with unseen scripts, we hope our study will inspire more work on transliteration-based adaptation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.25,
+                0.615,
+                0.265
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.276,
+                0.885,
+                0.581
+            ],
+            "angle": 0,
+            "content": "In this paper, we work with UROMAN (Hermjakob et al., 2018) which is an unsupervised romanization tool. While it is an effective tool for romanization at scale, it still has potential drawbacks. Since it is only based on lexical substitution, its transliterations may not semantically or phonetically align with the source content and may differ from transliterations preferred by native speakers. Moreover, UROMAN is not invertible—as we have highlighted—and may thus be less appealing when text in the original script needs to be exactly reproduced. Our proposed method, while it is parameter-efficient and effective—particularly for low-resource languages—still underperforms language-specific tokenizer-based non-transliteration methods. Future work may focus on developing an improved and more efficient tokenizer for transliteration-based methods as we highlight in the Appendix."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.582,
+                0.884,
+                0.79
+            ],
+            "angle": 0,
+            "content": "While there is now a growing body of available evaluation resources for low-resource languages (Ebrahimimi et al., 2022; Mhaske et al., 2023; Winata et al., 2023, among others), our final selection of tasks, resources and languages has been driven and constrained by the specific concrete goal of our short paper: studying and evaluating if and how transliteration/romanization can help with adaptation of languages with scripts unseen by the pretrained multilingual language model. We thus closely follow the experimental setup of Pfeiffer et al. (2021) which used the same set of tasks and languages with unseen scripts."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.792,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Finally, romanization can be seen as a step towards providing more universal, or rather language-agnostic, input text representation. Full-fledged comparisons against other approaches that aim to strike language independence at the input or feature level, such as byte-level models (e.g., ByT5 (Xue et al., 2022)) and pixel-based models (e.g., PILXEL (Rust et al., 2023)) go beyond"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.869,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "10 For Tibetan, longer words are composed using shorter words separated by tsek (\":\") which is not a valid space delimiter for the mBERT tokenizer; the number of produced subwords is thus much higher than for the other languages."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8000"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.181
+            ],
+            "angle": 0,
+            "content": "the scope of this particular work, but we point out to this as a very interesting future research avenue. Moreover, the integration of these language-agnostic representations with 'romanization'-based approaches might yield additional benefits, and should also be attested in future research."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.194,
+                0.287,
+                0.21
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.22,
+                0.488,
+                0.315
+            ],
+            "angle": 0,
+            "content": "This work has been funded by the German Research Foundation (DFG) as part of the Research Training Group KRITIS No. GRK 2222. The work of Ivan Vulić has been supported by a personal Royal Society University Research Fellowship (no 221137; 2022-)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.317,
+                0.487,
+                0.366
+            ],
+            "angle": 0,
+            "content": "We thank Indraneil Paul, Yongxin Huang, Ivan Habernal, and Massimo Nicosia for their valuable feedback and suggestions on a draft of this paper."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.392,
+                0.214,
+                0.408
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.415,
+                0.49,
+                0.496
+            ],
+            "angle": 0,
+            "content": "Chantal Amrhein and Rico Sennrich. 2020. On Romanization for model transfer between scripts in neural machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 2461-2469, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.505,
+                0.49,
+                0.611
+            ],
+            "angle": 0,
+            "content": "Alan Ansell, Edoardo Maria Ponti, Jonas Pfeiffer, Sebastian Ruder, Goran Glavaš, Ivan Vulić, and Anna Korhonen. 2021. MAD-G: Multilingual adapter generation for efficient cross-lingual transfer. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 4762–4781, Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.621,
+                0.49,
+                0.701
+            ],
+            "angle": 0,
+            "content": "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.711,
+                0.49,
+                0.79
+            ],
+            "angle": 0,
+            "content": "Ethan C. Chau and Noah A. Smith. 2021. Specializing multilingual language models: An empirical study. In Proceedings of the 1st Workshop on Multilingual Representation Learning, pages 51-61, Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.801,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.415,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.085,
+                0.882,
+                0.114
+            ],
+            "angle": 0,
+            "content": "Peter T. Daniels and William Bright. 1996. The world's writing systems. Oxford University Press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.124,
+                0.885,
+                0.205
+            ],
+            "angle": 0,
+            "content": "Isin Demirsahin, Cibu Johny, Alexander Gutkin, and Brian Roark. 2022. Criteria for useful automatic Romanization in South Asian languages. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6662-6673, Marseille, France. European Language Resources Association."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.216,
+                0.885,
+                0.334
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.346,
+                0.885,
+                0.464
+            ],
+            "angle": 0,
+            "content": "Tejas Dhamecha, Rudra Murthy, Samarth Bharadwaj, Karthik Sankaranarayanan, and Pushpak Bhattacharyya. 2021. Role of Language Relatedness in Multilingual Fine-tuning of Language Models: A Case Study in Indo-Aryan Languages. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8584-8595, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.476,
+                0.885,
+                0.554
+            ],
+            "angle": 0,
+            "content": "Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency parsing. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. Open-Review.net."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.566,
+                0.885,
+                0.737
+            ],
+            "angle": 0,
+            "content": "Abteen Ebrahimi, Manuel Mager, Arturo Oncevay, Vishrav Chaudhary, Luis Chiruzzo, Angela Fan, John Ortega, Ricardo Ramos, Annette Rios, Ivan Vladimir Meza Ruiz, Gustavo Giménez-Lugo, Elisabeth Mager, Graham Neubig, Alexis Palmer, Rolando Coto-Solano, Thang Vu, and Katharina Kann. 2022. AmericasNLI: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6279-6299, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.749,
+                0.885,
+                0.842
+            ],
+            "angle": 0,
+            "content": "Goran Glavaš and Ivan Vulić. 2021. Is supervised syntactic parsing beneficial for language understanding tasks? an empirical investigation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3090-3104, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.852,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Ulf Hermjakob, Jonathan May, and Kevin Knight. 2018. Out-of-the-box universal Romanization tool uroman. In Proceedings of ACL 2018, System Demonstrations, pages 13-18, Melbourne, Australia. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.085,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8001"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.218,
+                0.49,
+                0.335
+            ],
+            "angle": 0,
+            "content": "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalisation. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 4411-4421. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.35,
+                0.49,
+                0.429
+            ],
+            "angle": 0,
+            "content": "Ying Lin, Cash Costello, Boliang Zhang, Di Lu, Heng Ji, James Mayfield, and Paul McNamee. 2018. Platforms for non-speakers annotating names in any language. In Proceedings of ACL 2018, System Demonstrations, pages 1-6, Melbourne, Australia. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.442,
+                0.49,
+                0.509
+            ],
+            "angle": 0,
+            "content": "Yash Madhani, Sushane Parthan, Priyanka A. Bedekar, Ruchi Khapra, Vivek Seshadri, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. Aksharantar: Towards building open transliteration tools for the next billion users. arXiv preprint."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.522,
+                0.49,
+                0.587
+            ],
+            "angle": 0,
+            "content": "Michael McCloskey and Neal J. Cohen. 1989. Catastrophic interference in connectionist networks: The sequential learning problem. volume 24 of Psychology of Learning and Motivation, pages 109-165. Academic Press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.602,
+                0.49,
+                0.707
+            ],
+            "angle": 0,
+            "content": "Arnav Mhaske, Harshit Kedia, Sumanth Doddapaneni, Mitesh M. Khapra, Pratyush Kumar, Rudra Murthy, and Anoop Kunchukuttan. 2023. Naamapadam: A large-scale named entity annotated data for Indic languages. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10441-10456, Toronto, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.721,
+                0.49,
+                0.826
+            ],
+            "angle": 0,
+            "content": "Nikita Moghe, Evgeniia Razumovskaia, Liane Guillou, Ivan Vulic, Anna Korhonen, and Alexandra Birch. 2023. Multi3NLU++: A multilingual, multi-intent, multi-domain dataset for natural language understanding in task-oriented dialogue. In Findings of the Association for Computational Linguistics: ACL 2023, pages 3732-3755, Toronto, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.84,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Ibraheem Muhammad Moosa, Mahmud Elahi Akhter, and Ashfia Binte Habib. 2023. Does transliteration help multilingual language modeling? In *Findings of the Association for Computational Linguistics: EACL* 2023, pages 670–685, Dubrovnik, Croatia. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.204
+            ],
+            "angle": 0,
+            "content": "Benjamin Muller, Antonios Anastasopoulos, Benoit Sagot, and Djamé Seddah. 2021. When being unseen from mBERT is just the beginning: Handling new languages with multilingual language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 448-462, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.214,
+                0.885,
+                0.32
+            ],
+            "angle": 0,
+            "content": "Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Jan Hajic, Christopher D. Manning, Sampo Pyysalo, Sebastian Schuster, Francis Tyers, and Daniel Zeman. 2020. Universal Dependencies v2: An evergrowing multilingual treebank collection. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4034-4043, Marseille, France. European Language Resources Association."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.33,
+                0.885,
+                0.421
+            ],
+            "angle": 0,
+            "content": "Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1946–1958, Vancouver, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.431,
+                0.885,
+                0.55
+            ],
+            "angle": 0,
+            "content": "Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3479-3495, Seattle, United States. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.56,
+                0.885,
+                0.665
+            ],
+            "angle": 0,
+            "content": "Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020a. AdapterHub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.675,
+                0.885,
+                0.715
+            ],
+            "angle": 0,
+            "content": "Jonas Pfeiffer, Sebastian Ruder, Ivan Vulić, and Edoardo Maria Ponti. 2023. Modular deep learning. arXiv preprint."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.724,
+                0.885,
+                0.817
+            ],
+            "angle": 0,
+            "content": "Jonas Pfeiffer, Ivan Vulić, Iryna Gurevych, and Sebastian Ruder. 2020b. MAD-X: An Adapter-Based Framework for Multi-Task Cross-Linguual Transfer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7654–7673, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.826,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Jonas Pfeiffer, Ivan Vulić, Iryna Gurevych, and Sebastian Ruder. 2021. UNKs everywhere: Adapting multilingual language models to new scripts. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10186-10203, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8002"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.179
+            ],
+            "angle": 0,
+            "content": "Edoardo Maria Ponti, Goran Glavaš, Olga Majewska, Qianchu Liu, Ivan Vulić, and Anna Korhonen. 2020. XCOPA: A multilingual dataset for causal commonsense reasoning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362-2376, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.187,
+                0.49,
+                0.254
+            ],
+            "angle": 0,
+            "content": "Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 151-164, Florence, Italy. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.261,
+                0.49,
+                0.341
+            ],
+            "angle": 0,
+            "content": "Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. 2023. Language modelling with pixels. In *The Eleventh International Conference on Learning Representations*, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.349,
+                0.49,
+                0.468
+            ],
+            "angle": 0,
+            "content": "Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? on the monolingual performance of multilingual language models. 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 3118-3135, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.476,
+                0.49,
+                0.567
+            ],
+            "angle": 0,
+            "content": "Daan van Esch, Tamar Lucassen, Sebastian Ruder, Isaac Caswell, and Clara Rivera. 2022. Writing system and speaker metadata for \\(2,800+\\) language varieties. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 5035-5046, Marseille, France. European Language Resources Association."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.577,
+                0.49,
+                0.63
+            ],
+            "angle": 0,
+            "content": "Hans H Wellisch, Lee Breuer, Richard Foreman, and Robert Wilson. 1978. The conversion of scripts, its nature, history, and utilization. New York: Toronto: Wiley."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.638,
+                0.49,
+                0.783
+            ],
+            "angle": 0,
+            "content": "Genta Indra Winata, Alham Fikri Aji, Samuel Cahyawijaya, Rahmad Mahendra, Fajri Koto, Ade Romadhony, Kemal Kurniawan, David Moeljadi, Radityo Eko Prasojo, Pascale Fung, Timothy Baldwin, Jey Han Lau, Rico Sennrich, and Sebastian Ruder. 2023. NusaX: Multilingual parallel sentiment dataset for 10 Indonesian local languages. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 815-834, Dubrovnik, Croatia. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.791,
+                0.49,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. 2022. ByT5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics, 10:291-306."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.878,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Seid Muhie Yimam, Abinew Ali Ayele, Gopalakrishnan Venkatesh, Ibrahim Gashaw, and Chris Biemann. 2021. Introducing various semantic models"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.114
+            ],
+            "angle": 0,
+            "content": "for amharic: Experimentation and evaluation with multiple tasks and datasets. Future Internet, 13(11)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.123,
+                0.853,
+                0.138
+            ],
+            "angle": 0,
+            "content": "Judit Ács. Exploring bert's vocabulary. Blog Post."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8003"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.407,
+                0.1
+            ],
+            "angle": 0,
+            "content": "A Transliterators in Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.113,
+                0.488,
+                0.16
+            ],
+            "angle": 0,
+            "content": "Besides UROMAN, we also employ various language-specific transliterations which are publicly available. We list them in Table 6."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.174,
+                0.468,
+                0.192
+            ],
+            "angle": 0,
+            "content": "B Performance comparison of mBERT"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.201,
+                0.49,
+                0.362
+            ],
+            "angle": 0,
+            "content": "We adapt the standard cross-lingual transfer setup for mBERT. The model is finetuned on the task data for a source language (high-resource) and is used to perform inference on the target language (low-resource). We report the performance comparison of the standard cross-lingual transfer setup for mBERT on the NER task for languages with unseen scripts with the adapter-based methods in Table 7. We observe that the adapter-based methods outperform mBERT by huge margins."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.376,
+                0.306,
+                0.393
+            ],
+            "angle": 0,
+            "content": "C Further Analyses"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.403,
+                0.49,
+                0.596
+            ],
+            "angle": 0,
+            "content": "Following previous work (Ács; Rust et al., 2021; Moosa et al., 2023), we further analyze tokenization quality of the mBERT tokenizer using the following established metrics: 1) % of \"UNK\"s measures the % of \"UNK\" tokens produced by the tokenizer, and our aim is to compare their rate before and after transliteration; 2) fertility measures the number of subwords that are produced per tokenized word; 3) proportion of continued subwords measures the proportion of words for which the tokenized word is split across at least two subwords (denoted by the symbol ##)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.598,
+                0.49,
+                0.79
+            ],
+            "angle": 0,
+            "content": "From the results summarized in Figure 4, it is apparent that transliteration drastically reduces \\(\\%\\) of UNKs. However, mBERT's tokenizer underperforms as compared to monolingual tokenizers based on fertility and the proportion of continued subwords (Rust et al., 2021). Transliteration performs better for some languages where the quality of the mBERT tokenizer is similar to the monolingual tokenizer such as for dv, km, and cdo. On the other hand, transliteration methods perform worse on languages where the quality of the underlying mBERT tokenizer is relatively poor."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.792,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "In order to test the hypothesis that the tokenizer quality might be the principal reason for the performance gap for the transliteration-based methods in comparison to the non-transliteration based methods, we carried out an additional experiment. For the experiment, we adapt the Non-Trans\\(_{\\mathrm{LA + Emb_{Lex}}}\\) to operate on transliterated data, and call this variant Trans\\(_{\\mathrm{LA + Emb_{Lex}}}\\). Here, we train a new tokenizer"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.883,
+                0.133
+            ],
+            "angle": 0,
+            "content": "on the transliterated data and initialize lexically overlapping embeddings with mBERT's pretrained embeddings."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.134,
+                0.885,
+                0.327
+            ],
+            "angle": 0,
+            "content": "We plot the performance in Figure 5. The new method, \\(\\mathrm{Trans_{LA + Emb_{Lex}}}\\) now outperforms the nontransliteration-based variant on 8/12 languages and also on average. Consequently, this validates our hypothesis and is in line with the previous work (Moosa et al., 2023). However, we found a drop in performance in the case of mhr (-10.71) and cdo (-10.14) when compared to \\(\\mathrm{Trans_{LA} + mPLM_{ft}}\\). These drops may be attributed to the lower degree of lexical overlap with mBERT's vocabulary, and consequently a higher number of randomly initialized embeddings for those target languages."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8004"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.115,
+                0.082,
+                0.885,
+                0.172
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Transliterator</td><td>Used for languages</td><td>Available at</td></tr><tr><td>UROMAN</td><td>All</td><td>github.com/isi-nlp/uroman</td></tr><tr><td>amseg</td><td>am</td><td>pypi.org/project/amseg/</td></tr><tr><td>transliterate</td><td>ru, ka</td><td>pypi.org/project/transliterate/</td></tr><tr><td>ai4bharat-transliteration</td><td>hi, sd</td><td>pypi.org/project/ai4bharat-transliteration/</td></tr><tr><td>lang-trans</td><td>ar</td><td>pypi.org/project/lang-trans/</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.356,
+                0.181,
+                0.642,
+                0.195
+            ],
+            "angle": 0,
+            "content": "Table 6: Transliterators used in this work."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.116,
+                0.209,
+                0.351,
+                0.279
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.223,
+                0.285,
+                0.242,
+                0.297
+            ],
+            "angle": 0,
+            "content": "(a)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.382,
+                0.209,
+                0.617,
+                0.279
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.489,
+                0.284,
+                0.509,
+                0.296
+            ],
+            "angle": 0,
+            "content": "(b)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.649,
+                0.208,
+                0.882,
+                0.279
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.756,
+                0.284,
+                0.775,
+                0.296
+            ],
+            "angle": 0,
+            "content": "(c)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.308,
+                0.884,
+                0.339
+            ],
+            "angle": 0,
+            "content": "Figure 4: Tokenizer quality analysis. a) % of UNKs before and after transliteration, b) Fertility, and c) Proportion of continued subwords for mBERT vs monolingual tokenizer."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.134,
+                0.428,
+                0.472,
+                0.502
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Variant</td><td colspan=\"5\">Unseen Scripts</td></tr><tr><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td></tr><tr><td>mBERT</td><td>0.91</td><td>17.40</td><td>1.30</td><td>10.71</td><td>2.50</td></tr><tr><td>Non-TransLA+INV</td><td>3.26</td><td>19.86</td><td>18.72</td><td>13.81</td><td>23.14</td></tr><tr><td>TransLA + INV</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td></tr><tr><td>Non-TransLA+EmblEx</td><td>47.60</td><td>40.96</td><td>31.21</td><td>53.94</td><td>45.89</td></tr><tr><td>TransLA + mPLMft</td><td>37.09</td><td>33.82</td><td>40.40</td><td>52.39</td><td>45.24</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.511,
+                0.489,
+                0.553
+            ],
+            "angle": 0,
+            "content": "Table 7: Performance Comparison of mBERT with adapter-based finetuning methods for unseen scripts on the NER task."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.117,
+                0.701,
+                0.489,
+                0.808
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.817,
+                0.489,
+                0.847
+            ],
+            "angle": 0,
+            "content": "Figure 5: Comparison of Non-Trans\\(_{\\mathrm{LA + Emb_{Lex}}}\\) with Trans\\(_{\\mathrm{LA + Emb_{Lex}}}\\) on NER (left) and DP (right)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8005"
+        }
+    ]
+]

2023/Romanization-based Large-scale Adaptation of Multilingual Language Models/1a02cba7-8ebc-4f20-8832-cb43efde101d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e554fd4d8acd4454e4e2cde2ab422d0a9a43e13a35adf7cd4596e23a8a8f1f9
+size 486379

2023/Romanization-based Large-scale Adaptation of Multilingual Language Models/full.md

@@ -0,0 +1,226 @@
+# Romanization-based Large-scale Adaptation of Multilingual Language Models
+
+Sukannya Purkayastha<sup>1</sup>, Sebastian Ruder<sup>2</sup>, Jonas Pfeiffer<sup>2</sup>, Iryna Gurevych<sup>1</sup>, Ivan Vulić<sup>3</sup>
+
+<sup>1</sup> Ubiquitous Knowledge Processing Lab,
+
+Department of Computer Science and Hessian Center for AI (hessian.AI),
+
+Technical University of Darmstadt
+
+Google DeepMind
+
+$^{3}$  Language Technology Lab, University of Cambridge
+
+www.ukp.tu-darmstadt.de
+
+# Abstract
+
+Large multilingual pretrained language models (mPLMs) have become the de facto state of the art for cross-lingual transfer in NLP. However, their large-scale deployment to many languages, besides pretraining data scarcity, is also hindered by the increase in vocabulary size and limitations in their parameter budget. In order to boost the capacity of mPLMs to deal with low-resource and unseen languages, we explore the potential of leveraging transliteration on a massive scale. In particular, we explore the UROMAN transliteration tool, which provides mappings from UTF-8 to Latin characters for all the writing systems, enabling inexpensive romanization for virtually any language. We first focus on establishing how UROMAN compares against other language-specific and manually curated transliterators for adapting multilingual PLMs. We then study and compare a plethora of data- and parameter-efficient strategies for adapting the mPLMs to romanized and non-romanized corpora of 14 diverse low-resource languages. Our results reveal that UROMAN-based transliteration can offer strong performance for many languages, with particular gains achieved in the most challenging setups: on languages with unseen scripts and with limited training data without any vocabulary augmentation. Further analyses reveal that an improved tokenizer based on romanized data can even outperform non-transliteration-based methods in the majority of languages.
+
+# 1 Introduction
+
+Massively multilingual language models (mPLMs) such as mBERT (Devlin et al., 2019) and XLM-R (Conneau et al., 2020) have become the driving force for a variety of applications in multilingual NLP (Ponti et al., 2020; Hu et al., 2020; Moghe et al., 2023). However, guaranteeing and maintaining strong performance for a wide spectrum of low-resource languages is difficult due to two crucial problems. The first issue is the vocabulary
+
+<table><tr><td>Language</td><td>Original Text</td><td>Romanized Text</td><td>Translation</td></tr><tr><td>Bhojpuri</td><td>बाल्दियम् गार्वा</td><td>jorjiyan bhaasaas</td><td>Georgian language</td></tr><tr><td>Sinhala</td><td>सिकूल्तिूल्तिूल्तिूल्तिूल्तिूल्तिू�िूल्तिूल्तिूल्तिू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�र्था</td><td></td><td></td></tr><tr><td>Sindhi</td><td>सिकूल्तिूल्तिूल्तिू�िू�िू�िू�र्था</td><td>grahalooka</td><td>planets</td></tr><tr><td>Khmer</td><td>khμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμmu μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμum μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ µ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μµ μµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµüµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµúµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ µµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₃₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₂₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₁₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀⁻⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁰⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁵⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁲⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁴⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁽⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁻⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁼⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸ⁱ⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁅⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁰⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₃⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄²⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂���₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄­⁄₂⁄.
+&lt;|im_start|&gt;assistant</td><td>Bhojpuri</td><td>BJJUJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJSJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ JJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ</td></tr></table>
+
+Figure 1: Romanization across different languages.
+
+size, as the vocabulary is bound to increase with the number of languages added if per-language performance is to be maintained (Hu et al., 2020; Artetxe et al., 2020; Pfeiffer et al., 2022). Second, pretraining mPLMs with a fixed model capacity improves cross-lingual performance up to a point after which it starts to decrease; this is the phenomenon termed the curse of multilinguality (Conneau et al., 2020).
+
+Transliteration refers to the process of converting language represented in one writing system to another (Wellisch et al., 1978). Latin script-centered transliteration or romanization is the most common form of transliteration (Lin et al., 2018; Amrhein and Sennrich, 2020; Demirsin et al., 2022) as the Latin/Roman script is by far the most widely adopted writing script in the world (Daniels and Bright, 1996; van Esch et al., 2022). Adapting mPLMs via transliteration can address the two aforementioned critical issues. 1) Since the Latin script covers a dominant portion of the mPLM's vocabulary (e.g.,  $77\%$  in case of mBERT, see Ács), 'romanizing' the remaining part of the vocabulary might mitigate the vocabulary size issue and boost vocabulary sharing. 2) Since no new tokens are added during the romanization process, reusing pretrained embeddings from the mPLM's embedding matrix helps reuse the information already present within the mPLM, thereby allocating the model's parameter budget more efficiently.
+
+However, the main drawback of transliteration seems to be the expensive process of creating effective language-specific transliterations, as they
+
+typically require language expertise to curate dictionaries that map tokens from one language and script to another. Therefore, previous attempts at mPLM adaptation to unseen languages via transliteration (Muller et al., 2021; Chau and Smith, 2021; Dhamecha et al., 2021; Moosa et al., 2023) were constrained to a handful of languages due to the limited availability of language-specific translit-. rators, or were applied only to languages that have 'language siblings' with developed translitators.
+
+In this work, unlike previous work, we propose to use and then evaluate the usefulness of a universal romanization tool, UROMAN (Hermjakob et al., 2018), for quick, large-scale and effective adaptation of mPLMs to low-resource languages. The UROMAN tool disposes of language-specific curated dictionaries and maps any UTF-8 character to the Latin script, increasing the portability of romanization, with some examples in Figure 1.
+
+We analyze language adaptation on a massive scale via UROMAN-based romanization on a set of 14 diverse low-resource languages. We conduct experiments within the standard parameter-efficient adapter-based cross-lingual transfer setup on two tasks: Named Entity Recognition (NER) on the WikiANN dataset (Pan et al., 2017; Rahimi et al., 2019), and Dependency Parsing (DP) with Universal Dependencies v2.7 (Nivre et al., 2020). Our key results suggest that UROMAN-based transliteration can offer strong performance on par or even outperforming adaptation with language-specific transliterrators, setting up the basis for wider use of transliteration-based mPLM adaptation techniques in future work. The gains with romanization-based adaptation over standard adaptation baselines are particularly pronounced for languages with unseen scripts ( $\sim$ 8-22 performance points) without any vocabulary augmentation.[2]
+
+# 2 Background
+
+Why UROMAN-Based Romanization? UROMAN-based romanization is not always fully reversible, and its usage for transliteration has thus been limited in the literature. However, due to its high portability, UROMAN can help scale the process of transliteration massively and as such benefit low-resource scenarios and wider adaptation of mPLMs. The main idea, as hinted in §1, is to (learn to) map
+
+any UTF-8 character to the Latin script, without the use of any external language-specific dictionaries (see Hermjakob et al. (2018) for technical details).
+
+Cross-Lingual Transfer to Low-Resource Languages. Parameter-efficient and modular finetuning methods (Pfeiffer et al., 2023) such as adapters (Houlsby et al., 2019; Pfeiffer et al., 2020b) have been used for cross-lingual transfer, putting a particular focus on enabling transfer to low-resource languages and scenarios, including languages with scripts 'unseen' by the base mPLM (Pfeiffer et al., 2021). Adapters are small lightweight components stitched into the base mPLM, and then trained for particular languages and tasks while keeping the parameters of the original mPLM frozen. This circumvents the issues of catastrophic forgetting and interference (McCloskey and Cohen, 1989) within the mPLM, and allows for extending its reach also to unseen languages (Pfeiffer et al., 2021; Ansell et al., 2021).
+
+For our main empirical analyses, we adopt a state-of-the-art modular method for cross-lingual transfer: MAD-X (Pfeiffer et al., 2020b). In short, MAD-X is based on language adapters (LA), task adapters (TA), and invertible adapters (INV). While LAs are trained for specific languages relying on masked language modeling, TAs are trained with high-resource languages relying on task-annotated data and task-specific objectives. At inference, the source LA is replaced with the target LA while the TA is kept. In order to do parameter-efficient learning for the token-level embeddings across different languages and to deal with the vocabulary mismatch between source and target languages, Pfeiffer et al. (2020b) also propose INV adapters: they are placed on top of the embedding layer and their inverses precede the output embedding layer.<sup>3</sup> We adopt the better-performing MAD-X 2.0 setup (Pfeiffer et al., 2021) where the adapters in the last Transformer layer are dropped at inference.<sup>4</sup>
+
+# 3 Experiments and Results
+
+As the main means of analyzing the impact of transliteration in general and UROMAN-based romanization in particular, we train different variants of language adapters within the MAD-X framework, based on transliterated and non-transliterated
+
+<table><tr><td>Language</td><td>Family</td><td>Script</td><td># Sentences</td></tr><tr><td>Bhojpuri (bh)</td><td>Indo-European</td><td>Devanagari</td><td>35,983</td></tr><tr><td>Buryat (bxr)</td><td>Mongolic</td><td>Cyrillic</td><td>41,692</td></tr><tr><td>Erzya (myv)</td><td>Uralic</td><td>Cyrillic</td><td>42,575</td></tr><tr><td>Meadow Mari (mhr)</td><td>Uralic</td><td>Cyrillic</td><td>144,529</td></tr><tr><td>Min Dong (cdo)</td><td>Sino-Tibetan</td><td>Chinese</td><td>33,978</td></tr><tr><td>Mingrelian (xmf)</td><td>Kartvelian</td><td>Georgian</td><td>63,032</td></tr><tr><td>Sindhi (sd)</td><td>Indo-European</td><td>Arabic</td><td>86,176</td></tr><tr><td>Sorani Kurdish (ckb)</td><td>Indo-European</td><td>Arabic</td><td>459,475</td></tr><tr><td>Uyghur (ug)</td><td>Turkic</td><td>Arabic</td><td>149,813</td></tr><tr><td>Amháric (am)</td><td>Afro-Asiatic</td><td>Ge&#x27;ez</td><td>88,320</td></tr><tr><td>Divehi (dv)</td><td>Indo-European</td><td>Thaana</td><td>34,779</td></tr><tr><td>Khmer (km)</td><td>Austroasiatic</td><td>Khmer</td><td>139,704</td></tr><tr><td>Sinhala (si)</td><td>Indo-European</td><td>Sinhala</td><td>219,866</td></tr><tr><td>Tibetan (bo)</td><td>Sino-Tibetan</td><td>Tibetan</td><td>131,362</td></tr></table>
+
+Table 1: Languages with their ISO 639-3 codes used in our evaluation, along with their script, language family, and number of sentences available for pretraining. The dashed line separates languages with unseen scripts, placed in the bottom part of the table.
+
+versions of target language data, outlined here.
+
+Variants with Non-Transliterated Data. For the Non-Trans $_{\mathrm{LA} + \mathrm{INV}}$  variant, we train LAs and INV adapters together. This variant serves to examine the extent to which mPLMs can adapt to unseen languages without any vocabulary extension. $^5$  We compare this to Non-Trans $_{\mathrm{LA} + \mathrm{Emb}_{\mathrm{Lex}}}$ , which trains a new tokenizer for the target language (Pfeiffer et al., 2021): the so-called 'lexically overlapping' tokens are initialized with mPLM's trained embeddings, while the remaining embeddings are initialized randomly. All these embeddings (Emb $_{\mathrm{Lex}}$ ) are fine-tuned along with LAs.
+
+Variants with Transliterated Data. We evaluate a  $\mathrm{Trans}_{\mathrm{LA} + \mathrm{INV}}$  variant, which uses the same setup as Non-Trans  $\mathrm{LA} + \mathrm{INV}$  but now with transliterated data. We again note that in this efficient setup, we do not extend the vocabulary size, and use the fewest trainable parameters. In the  $\mathrm{Trans}_{\mathrm{LA} + \mathrm{mPLM}_{\mathrm{ft}}}$  variant, we train LAs along with fine-tuning the pretrained embeddings of mPLM (mPLMft). This further enhances the model capacity by fine-tuning the embedding layer instead of using invertible adapters. For both variants, transliterated data can be produced via different transliterations: (i) language-specific ones; (ii) the ones from 'language siblings' (e.g., using a Georgian transliterator for Mingrelian), or (iii) UROMAN.
+
+# 3.1 Experimental Setup
+
+Data, Languages and Tasks. Following Pfeiffer et al. (2021), we select mBERT as our base mPLM. We experiment with 14 typologically diverse low-resource languages that are not part of mBERT's pretraining corpora, with 5/14 languages written in distinct scripts (see Table 1 for details). For LA training, we use Wikipedia dumps for the target languages, which we also transliterate (using different translitators). Evaluation is conducted on two standard cross-lingual transfer tasks in zero-shot setups: 1) the WikiAnn NER dataset (Pan et al., 2017) with the train, dev, and test splits from (Rahimi et al., 2019); 2) for dependency parsing, we rely on the UD Dataset v2.7 (Nivre et al., 2020).
+
+LAs and TAs. English is the source language in all experiments, and is used for training TAs. The English LA is obtained directly from Adapterhub.ml (Pfeiffer et al., 2020a), LAs and embeddings (when needed) are only trained for target languages.
+
+Finally, for the Non-Trans $_{\mathrm{LA + Emb_{Lex}}}$  variant, we train a WordPiece tokenizer on the target language data with a vocabulary size of 10K.
+
+Training of Language and Task Adapters. We train all the language adapters for 50 epochs or  $\sim 50K$  update steps based on the corpus size. The batch size is set to 64 and the learning rate is  $1e - 4$ .
+
+We train English task adapters following the setup from (Pfeiffer et al., 2020b). For NER, we directly obtain the task adapter from Adapterhub.ml which is trained with a learning rate of  $1e - 4$  for 10 epochs. For DP, we train a Transformer-based (Glavaš and Vulić, 2021) biaffine attention dependency parser (Dozat and Manning, 2017). We use a learning rate of  $5e - 4$  and train for 10 epochs as in (Pfeiffer et al., 2021).
+
+All the reported results in both tasks (NER and DP) are reported as averages over 6 random seeds. All the models have been trained on A100 or V100 GPUs. None of the training methods consumed more than 36 hours. As the main means of analyzing the impact of transliteration in general and UROMAN-based romanization in particular, we train different variants of language adapters within the MAD-X framework, based on transliterated and non-transliterated versions of target language data, outlined here.
+
+# 3.2 Results and Discussion
+
+UROMAN versus Other Transliterators and Transliteration Strategies. In order to estab
+
+<table><tr><td>Task</td><td>Transliterator</td><td>am</td><td>ar</td><td>ka</td><td>ru</td><td>hi</td><td>sd</td><td>avg</td></tr><tr><td rowspan="2">NER (Macro F1)</td><td>UROMAN</td><td>25.6</td><td>24.8</td><td>61.4</td><td>66.5</td><td>48.6</td><td>35.3</td><td>43.7</td></tr><tr><td>Other</td><td>25.5</td><td>23.7</td><td>57.3</td><td>63.9</td><td>56.7</td><td>35.9</td><td>43.8</td></tr><tr><td rowspan="2">UD (UAS / LAS)</td><td>UROMAN</td><td>36.17 / 6.6</td><td>33.0 / 19.8</td><td>-</td><td>47.3 / 32.4</td><td>33.8 / 17.8</td><td>-</td><td>37.5 / 19.1</td></tr><tr><td>Other</td><td>29.9 / 5.4</td><td>32.6 / 19.9</td><td>-</td><td>45.0 / 19.9</td><td>33.2 / 17.9</td><td>-</td><td>35.2 / 15.8</td></tr></table>
+
+Table 2: Comparison of UROMAN with language-specific transliterations.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="7">Seen Script</td><td colspan="6">Unseen Script</td><td>avg</td></tr><tr><td>bh</td><td>cdo</td><td>ckb</td><td>mhr</td><td>sd</td><td>ug</td><td>xmf</td><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td><td></td><td></td></tr><tr><td>UROMAN</td><td>32.59</td><td>27.34</td><td>67.73</td><td>64.68</td><td>35.33</td><td>28.10</td><td>52.58</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td><td>26.89</td><td></td></tr><tr><td>BORROW</td><td>53.42 (hi)</td><td>-</td><td>12.46 (ar)</td><td>45.86 (ru)</td><td>16.79 (ar)</td><td>12.85 (ar)</td><td>24.77 (ru)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td></tr><tr><td>RAND</td><td>25.42</td><td>19.51</td><td>53.55</td><td>42.02</td><td>27.20</td><td>25.18</td><td>35.82</td><td>18.00</td><td>18.95</td><td>21.19</td><td>32.75</td><td>20.01</td><td>21.59</td><td></td></tr></table>
+
+Table 3: Comparison of various transliteration strategies on the NER task (Macro-F1).  
+
+<table><tr><td rowspan="2">Variant</td><td colspan="7">Seen Script</td><td colspan="5">Unseen Script</td><td rowspan="2">avg</td></tr><tr><td>bh</td><td>cdo</td><td>sd</td><td>xmf</td><td>mhr</td><td>ckb</td><td>ug</td><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td></tr><tr><td>Non-TransLA+INV</td><td>55.14</td><td>24.19</td><td>31.31</td><td>49.74</td><td>70.31</td><td>45.54</td><td>33.53</td><td>3.26</td><td>19.86</td><td>18.72</td><td>13.81</td><td>23.14</td><td>18.39</td></tr><tr><td>TransLA+INV</td><td>32.59</td><td>27.34</td><td>35.33</td><td>52.58</td><td>64.68</td><td>67.73</td><td>28.10</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td><td>26.89</td></tr><tr><td>Non-TransLA+EmblEx</td><td>60.00</td><td>28.91</td><td>42.47</td><td>51.99</td><td>61.05</td><td>79.12</td><td>50.42</td><td>47.60</td><td>40.96</td><td>31.21</td><td>53.94</td><td>45.89</td><td>49.01</td></tr><tr><td>TransLA + mPLMrl</td><td>49.05</td><td>36.92</td><td>39.16</td><td>57.99</td><td>69.85</td><td>73.92</td><td>33.43</td><td>37.09</td><td>33.82</td><td>40.40</td><td>52.39</td><td>45.24</td><td>47.44</td></tr></table>
+
+Table 4: Results (Macro-F1 scores) on WikiAnn NER averaged over 6 random seeds.  
+
+<table><tr><td rowspan="2">Variant</td><td colspan="4">Seen Script</td><td colspan="2">Unseen Script</td></tr><tr><td>bh</td><td>myv</td><td>ug</td><td>bxr</td><td>am</td><td>avg</td></tr><tr><td>Non-TransLA+INV</td><td>28.46 / 11.53</td><td>45.28 / 26.27</td><td>33.44 / 15.28</td><td>39.75 / 19.77</td><td>19.08 / 1.85</td><td>33.20 / 10.81</td></tr><tr><td>TransLA+INV</td><td>25.12 / 10.17</td><td>45.74 / 26.64</td><td>32.30 / 15.10</td><td>37.92 / 17.23</td><td>36.07 / 7.58</td><td>35.43 / 12.41</td></tr><tr><td>Non-TransLA+EmblEx</td><td>26.68 / 10.10</td><td>48.34 / 25.34</td><td>41.20 / 22.81</td><td>39.51 / 16.02</td><td>36.4778.39</td><td>38.44 / 12.20</td></tr><tr><td>TransLA + mPLMrl</td><td>28.04 / 11.13</td><td>41.97 / 20.29</td><td>50.89 / 16.56</td><td>35.03 / 20.29</td><td>39.10 / 9.00</td><td>39.01 / 14.65</td></tr></table>
+
+Table 5: Results (UAS / LAS scores) in the DP task with UD, averaged over 6 random seeds.
+
+lish the utility of UROMAN as a viable transliterator, especially for low-resource languages, we compare its performance with transliteration options using the TransLA + INV setup as the most efficient scenario. First, we compare UROMAN with language-specific transliterations available for selected languages: amseg (Yimam et al., 2021) for Amharic, ai4bharat-transliteration (Madhani et al., 2022) for Hindi and Sindhi, lang-trans for Arabic, and transliterate for Russian and Georgian<sup>7</sup>. The transliterations used in this work are outlined in Table 6. The results are provided in Table 2. On average, UROMAN performs better or comparable to the language-specific transliterations. This provides justification to use UROMAN for massive transliteration at scale.
+
+Second, we compare UROMAN to two other transliteration strategies. (i) BORROW refers to borrowing transliterators from languages within the same language family and written in the same script. Since building transliterators are costly, this gives us an estimate of whether it is possible to rely on the related transliterators when we do not have a language-specific one at hand. (ii) RAND
+
+refers to a random setting where we associate any non-ASCII character with any ASCII character, giving us an estimate of whether we actually need knowledge of the language to build transliterations. The results are provided in Table 3: UROMAN is largely and consistently outperforming both BORROW and RAND, where the single exception is BORROW (from Hindi to Bhojpuri). Surprisingly, RAND also yields reasonable performance and on average even outperforms the Non-TransLA+INV variant with non-transliterated data (21.59 vs 18.39 in Table 4 later). This provides further evidence towards the utility of transliteration in general and UROMAN-based romanization in particular to assist and improve language adaptation.
+
+Performance on Low-Resource Languages is summarized in Table 4 and Table 5. We note that  $\mathrm{Trans}_{\mathrm{LA} + \mathrm{INV}}$  outperforms Non-Trans  $\mathrm{LA} + \mathrm{INV}$  for all the languages with unseen scripts, and achieves that with huge margins ( $\sim$  8-22 points for NER and  $\sim$  17 points in UAS scores). We observe similar trends for some of the languages with seen scripts such as Min Dong (cdo), Sindhi (sd), Mingrelian
+
+![](images/a30d48cb3ddb4d2c393c30a8025c0d311b6c789c062f842d428fb4c9ea4a8963.jpg)  
+Figure 2: Sample efficiency in the NER task.
+
+![](images/1c950205100aab4509e107caa10d380e39c2abd245e0aba617d30a97a05dd5fc.jpg)  
+Figure 3: Sample efficiency in the DP task.
+
+(xmf) on NER tasks and Erzya (myv) on DP. The less efficient  $\mathrm{Trans}_{\mathrm{LA} + \mathrm{mPLM_{ft}}}$ , as expected, further improves the performance for all the languages except for Tibetan (bo).<sup>10</sup> Non-Trans $_{\mathrm{LA} + \mathrm{Emb}_{\mathrm{Lex}}}$ , however, now outperforms UROMAN-based methods for a majority of the languages. This observation can be attributed to various factors related to mBERT's tokenizer, and we provide an in-depth analysis later in Appendix C. Nonetheless, we observe strong and competitive performance of  $\mathrm{Trans}_{\mathrm{LA} + \mathrm{mPLM_{ft}}}$  in both tasks, again indicating that more attention should be put on transliteration-based language adaptation in future work.
+
+Sample Efficiency. Finally, we simulate a few-shot setup to study the effectiveness of using transliterated versus non-transliterated data in data-scarce scenarios. For NER, we evaluate performance on all the languages and on languages with unseen scripts; for DP, we evaluate on all the languages. Figure 2 indicates that  $\mathrm{Trans}_{\mathrm{LA} + \mathrm{INV}}$  on average performs better than all the other methods at sample sizes 100 (i.e., 100 sentences in the target language) and 1,000. However, from 10,000 sentences onward, Non-Trans  $\mathrm{LA} + \mathrm{Emb}_{\mathrm{Lex}}$  takes the lead. We observe similar trends in the DP task (see Fig 3). This establishes the utility of transliteration for (extremely) low-resource scenarios.
+
+# 4 Conclusion
+
+In this work, we have systematically analyzed and confirmed the potential of romanization, implemented via the UROMAN tool, to help with adaptation of multilingual pretrained language models. Given (i) its broad applicability and (ii) strong performance overall and for languages with unseen scripts, we hope our study will inspire more work on transliteration-based adaptation.
+
+# Limitations
+
+In this paper, we work with UROMAN (Hermjakob et al., 2018) which is an unsupervised romanization tool. While it is an effective tool for romanization at scale, it still has potential drawbacks. Since it is only based on lexical substitution, its transliterations may not semantically or phonetically align with the source content and may differ from transliterations preferred by native speakers. Moreover, UROMAN is not invertible—as we have highlighted—and may thus be less appealing when text in the original script needs to be exactly reproduced. Our proposed method, while it is parameter-efficient and effective—particularly for low-resource languages—still underperforms language-specific tokenizer-based non-transliteration methods. Future work may focus on developing an improved and more efficient tokenizer for transliteration-based methods as we highlight in the Appendix.
+
+While there is now a growing body of available evaluation resources for low-resource languages (Ebrahimimi et al., 2022; Mhaske et al., 2023; Winata et al., 2023, among others), our final selection of tasks, resources and languages has been driven and constrained by the specific concrete goal of our short paper: studying and evaluating if and how transliteration/romanization can help with adaptation of languages with scripts unseen by the pretrained multilingual language model. We thus closely follow the experimental setup of Pfeiffer et al. (2021) which used the same set of tasks and languages with unseen scripts.
+
+Finally, romanization can be seen as a step towards providing more universal, or rather language-agnostic, input text representation. Full-fledged comparisons against other approaches that aim to strike language independence at the input or feature level, such as byte-level models (e.g., ByT5 (Xue et al., 2022)) and pixel-based models (e.g., PILXEL (Rust et al., 2023)) go beyond
+
+the scope of this particular work, but we point out to this as a very interesting future research avenue. Moreover, the integration of these language-agnostic representations with 'romanization'-based approaches might yield additional benefits, and should also be attested in future research.
+
+# Acknowledgements
+
+This work has been funded by the German Research Foundation (DFG) as part of the Research Training Group KRITIS No. GRK 2222. The work of Ivan Vulić has been supported by a personal Royal Society University Research Fellowship (no 221137; 2022-).
+
+We thank Indraneil Paul, Yongxin Huang, Ivan Habernal, and Massimo Nicosia for their valuable feedback and suggestions on a draft of this paper.
+
+# References
+
+Chantal Amrhein and Rico Sennrich. 2020. On Romanization for model transfer between scripts in neural machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 2461-2469, Online. Association for Computational Linguistics.  
+Alan Ansell, Edoardo Maria Ponti, Jonas Pfeiffer, Sebastian Ruder, Goran Glavaš, Ivan Vulić, and Anna Korhonen. 2021. MAD-G: Multilingual adapter generation for efficient cross-lingual transfer. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 4762–4781, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics.  
+Ethan C. Chau and Noah A. Smith. 2021. Specializing multilingual language models: An empirical study. In Proceedings of the 1st Workshop on Multilingual Representation Learning, pages 51-61, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics.
+
+Peter T. Daniels and William Bright. 1996. The world's writing systems. Oxford University Press.  
+Isin Demirsahin, Cibu Johny, Alexander Gutkin, and Brian Roark. 2022. Criteria for useful automatic Romanization in South Asian languages. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6662-6673, Marseille, France. European Language Resources Association.  
+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.  
+Tejas Dhamecha, Rudra Murthy, Samarth Bharadwaj, Karthik Sankaranarayanan, and Pushpak Bhattacharyya. 2021. Role of Language Relatedness in Multilingual Fine-tuning of Language Models: A Case Study in Indo-Aryan Languages. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8584-8595, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency parsing. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. Open-Review.net.  
+Abteen Ebrahimi, Manuel Mager, Arturo Oncevay, Vishrav Chaudhary, Luis Chiruzzo, Angela Fan, John Ortega, Ricardo Ramos, Annette Rios, Ivan Vladimir Meza Ruiz, Gustavo Giménez-Lugo, Elisabeth Mager, Graham Neubig, Alexis Palmer, Rolando Coto-Solano, Thang Vu, and Katharina Kann. 2022. AmericasNLI: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6279-6299, Dublin, Ireland. Association for Computational Linguistics.  
+Goran Glavaš and Ivan Vulić. 2021. Is supervised syntactic parsing beneficial for language understanding tasks? an empirical investigation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3090-3104, Online. Association for Computational Linguistics.  
+Ulf Hermjakob, Jonathan May, and Kevin Knight. 2018. Out-of-the-box universal Romanization tool uroman. In Proceedings of ACL 2018, System Demonstrations, pages 13-18, Melbourne, Australia. Association for Computational Linguistics.
+
+Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR.  
+Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalisation. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 4411-4421. PMLR.  
+Ying Lin, Cash Costello, Boliang Zhang, Di Lu, Heng Ji, James Mayfield, and Paul McNamee. 2018. Platforms for non-speakers annotating names in any language. In Proceedings of ACL 2018, System Demonstrations, pages 1-6, Melbourne, Australia. Association for Computational Linguistics.  
+Yash Madhani, Sushane Parthan, Priyanka A. Bedekar, Ruchi Khapra, Vivek Seshadri, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. Aksharantar: Towards building open transliteration tools for the next billion users. arXiv preprint.  
+Michael McCloskey and Neal J. Cohen. 1989. Catastrophic interference in connectionist networks: The sequential learning problem. volume 24 of Psychology of Learning and Motivation, pages 109-165. Academic Press.  
+Arnav Mhaske, Harshit Kedia, Sumanth Doddapaneni, Mitesh M. Khapra, Pratyush Kumar, Rudra Murthy, and Anoop Kunchukuttan. 2023. Naamapadam: A large-scale named entity annotated data for Indic languages. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10441-10456, Toronto, Canada. Association for Computational Linguistics.  
+Nikita Moghe, Evgeniia Razumovskaia, Liane Guillou, Ivan Vulic, Anna Korhonen, and Alexandra Birch. 2023. Multi3NLU++: A multilingual, multi-intent, multi-domain dataset for natural language understanding in task-oriented dialogue. In Findings of the Association for Computational Linguistics: ACL 2023, pages 3732-3755, Toronto, Canada. Association for Computational Linguistics.  
+Ibraheem Muhammad Moosa, Mahmud Elahi Akhter, and Ashfia Binte Habib. 2023. Does transliteration help multilingual language modeling? In *Findings of the Association for Computational Linguistics: EACL* 2023, pages 670–685, Dubrovnik, Croatia. Association for Computational Linguistics.
+
+Benjamin Muller, Antonios Anastasopoulos, Benoit Sagot, and Djamé Seddah. 2021. When being unseen from mBERT is just the beginning: Handling new languages with multilingual language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 448-462, Online. Association for Computational Linguistics.  
+Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Jan Hajic, Christopher D. Manning, Sampo Pyysalo, Sebastian Schuster, Francis Tyers, and Daniel Zeman. 2020. Universal Dependencies v2: An evergrowing multilingual treebank collection. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4034-4043, Marseille, France. European Language Resources Association.  
+Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1946–1958, Vancouver, Canada. Association for Computational Linguistics.  
+Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3479-3495, Seattle, United States. Association for Computational Linguistics.  
+Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020a. AdapterHub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics.  
+Jonas Pfeiffer, Sebastian Ruder, Ivan Vulić, and Edoardo Maria Ponti. 2023. Modular deep learning. arXiv preprint.  
+Jonas Pfeiffer, Ivan Vulić, Iryna Gurevych, and Sebastian Ruder. 2020b. MAD-X: An Adapter-Based Framework for Multi-Task Cross-Linguual Transfer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7654–7673, Online. Association for Computational Linguistics.  
+Jonas Pfeiffer, Ivan Vulić, Iryna Gurevych, and Sebastian Ruder. 2021. UNKs everywhere: Adapting multilingual language models to new scripts. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10186-10203, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
+
+Edoardo Maria Ponti, Goran Glavaš, Olga Majewska, Qianchu Liu, Ivan Vulić, and Anna Korhonen. 2020. XCOPA: A multilingual dataset for causal commonsense reasoning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2362-2376, Online. Association for Computational Linguistics.  
+Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 151-164, Florence, Italy. Association for Computational Linguistics.  
+Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. 2023. Language modelling with pixels. In *The Eleventh International Conference on Learning Representations*, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.  
+Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? on the monolingual performance of multilingual language models. 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 3118-3135, Online. Association for Computational Linguistics.  
+Daan van Esch, Tamar Lucassen, Sebastian Ruder, Isaac Caswell, and Clara Rivera. 2022. Writing system and speaker metadata for  $2,800+$  language varieties. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 5035-5046, Marseille, France. European Language Resources Association.  
+Hans H Wellisch, Lee Breuer, Richard Foreman, and Robert Wilson. 1978. The conversion of scripts, its nature, history, and utilization. New York: Toronto: Wiley.  
+Genta Indra Winata, Alham Fikri Aji, Samuel Cahyawijaya, Rahmad Mahendra, Fajri Koto, Ade Romadhony, Kemal Kurniawan, David Moeljadi, Radityo Eko Prasojo, Pascale Fung, Timothy Baldwin, Jey Han Lau, Rico Sennrich, and Sebastian Ruder. 2023. NusaX: Multilingual parallel sentiment dataset for 10 Indonesian local languages. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 815-834, Dubrovnik, Croatia. Association for Computational Linguistics.  
+Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. 2022. ByT5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics, 10:291-306.  
+Seid Muhie Yimam, Abinew Ali Ayele, Gopalakrishnan Venkatesh, Ibrahim Gashaw, and Chris Biemann. 2021. Introducing various semantic models
+
+for amharic: Experimentation and evaluation with multiple tasks and datasets. Future Internet, 13(11).
+
+Judit Ács. Exploring bert's vocabulary. Blog Post.
+
+# A Transliterators in Evaluation
+
+Besides UROMAN, we also employ various language-specific transliterations which are publicly available. We list them in Table 6.
+
+# B Performance comparison of mBERT
+
+We adapt the standard cross-lingual transfer setup for mBERT. The model is finetuned on the task data for a source language (high-resource) and is used to perform inference on the target language (low-resource). We report the performance comparison of the standard cross-lingual transfer setup for mBERT on the NER task for languages with unseen scripts with the adapter-based methods in Table 7. We observe that the adapter-based methods outperform mBERT by huge margins.
+
+# C Further Analyses
+
+Following previous work (Ács; Rust et al., 2021; Moosa et al., 2023), we further analyze tokenization quality of the mBERT tokenizer using the following established metrics: 1) % of "UNK"s measures the % of "UNK" tokens produced by the tokenizer, and our aim is to compare their rate before and after transliteration; 2) fertility measures the number of subwords that are produced per tokenized word; 3) proportion of continued subwords measures the proportion of words for which the tokenized word is split across at least two subwords (denoted by the symbol ##).
+
+From the results summarized in Figure 4, it is apparent that transliteration drastically reduces  $\%$  of UNKs. However, mBERT's tokenizer underperforms as compared to monolingual tokenizers based on fertility and the proportion of continued subwords (Rust et al., 2021). Transliteration performs better for some languages where the quality of the mBERT tokenizer is similar to the monolingual tokenizer such as for dv, km, and cdo. On the other hand, transliteration methods perform worse on languages where the quality of the underlying mBERT tokenizer is relatively poor.
+
+In order to test the hypothesis that the tokenizer quality might be the principal reason for the performance gap for the transliteration-based methods in comparison to the non-transliteration based methods, we carried out an additional experiment. For the experiment, we adapt the Non-Trans $_{\mathrm{LA + Emb_{Lex}}}$  to operate on transliterated data, and call this variant Trans $_{\mathrm{LA + Emb_{Lex}}}$ . Here, we train a new tokenizer
+
+on the transliterated data and initialize lexically overlapping embeddings with mBERT's pretrained embeddings.
+
+We plot the performance in Figure 5. The new method,  $\mathrm{Trans_{LA + Emb_{Lex}}}$  now outperforms the nontransliteration-based variant on 8/12 languages and also on average. Consequently, this validates our hypothesis and is in line with the previous work (Moosa et al., 2023). However, we found a drop in performance in the case of mhr (-10.71) and cdo (-10.14) when compared to  $\mathrm{Trans_{LA} + mPLM_{ft}}$ . These drops may be attributed to the lower degree of lexical overlap with mBERT's vocabulary, and consequently a higher number of randomly initialized embeddings for those target languages.
+
+<table><tr><td>Transliterator</td><td>Used for languages</td><td>Available at</td></tr><tr><td>UROMAN</td><td>All</td><td>github.com/isi-nlp/uroman</td></tr><tr><td>amseg</td><td>am</td><td>pypi.org/project/amseg/</td></tr><tr><td>transliterate</td><td>ru, ka</td><td>pypi.org/project/transliterate/</td></tr><tr><td>ai4bharat-transliteration</td><td>hi, sd</td><td>pypi.org/project/ai4bharat-transliteration/</td></tr><tr><td>lang-trans</td><td>ar</td><td>pypi.org/project/lang-trans/</td></tr></table>
+
+![](images/8309dd1ed12a59dc1f0528edf5e64d0557b4f8d3dbec0880e76b532cd845f7ca.jpg)  
+(a)
+
+![](images/47711266c4017b2aa9f669ce98d4d7d7d04173802500672272aea02a88d75ca3.jpg)  
+(b)  
+Figure 4: Tokenizer quality analysis. a) % of UNKs before and after transliteration, b) Fertility, and c) Proportion of continued subwords for mBERT vs monolingual tokenizer.
+
+![](images/fb3ee83c84d87592bf4ee3c75d4c4ae5a5a1aa7ed161b08c329155e3eb3efc12.jpg)  
+(c)
+
+Table 6: Transliterators used in this work.  
+
+<table><tr><td rowspan="2">Variant</td><td colspan="5">Unseen Scripts</td></tr><tr><td>am</td><td>bo</td><td>dv</td><td>km</td><td>si</td></tr><tr><td>mBERT</td><td>0.91</td><td>17.40</td><td>1.30</td><td>10.71</td><td>2.50</td></tr><tr><td>Non-TransLA+INV</td><td>3.26</td><td>19.86</td><td>18.72</td><td>13.81</td><td>23.14</td></tr><tr><td>TransLA + INV</td><td>25.69</td><td>35.95</td><td>29.99</td><td>41.76</td><td>31.83</td></tr><tr><td>Non-TransLA+EmblEx</td><td>47.60</td><td>40.96</td><td>31.21</td><td>53.94</td><td>45.89</td></tr><tr><td>TransLA + mPLMft</td><td>37.09</td><td>33.82</td><td>40.40</td><td>52.39</td><td>45.24</td></tr></table>
+
+Table 7: Performance Comparison of mBERT with adapter-based finetuning methods for unseen scripts on the NER task.
+
+![](images/f862f3a59498858d80d8303dd287b78b2713cb4ca14a6614907596786428fec4.jpg)  
+Figure 5: Comparison of Non-Trans $_{\mathrm{LA + Emb_{Lex}}}$  with Trans $_{\mathrm{LA + Emb_{Lex}}}$  on NER (left) and DP (right).

2023/Romanization-based Large-scale Adaptation of Multilingual Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b42d07edd087aea94eec0bf231af1a2bb4c1a32fa16b0e8372003400e3034c4b
+size 331264

2023/SAC3_ Reliable Hallucination Detection in Black-Box Language Models via Semantic-aware Cross-check Consistency/4a5107be-dc68-42ed-bca9-6c454b388710_content_list.json

@@ -0,0 +1,1942 @@
+[
+    {
+        "type": "text",
+        "text": "SAC<sup>3</sup>: Reliable Hallucination Detection in Black-Box Language Models via Semantic-aware Cross-check Consistency",
+        "text_level": 1,
+        "bbox": [
+            115,
+            91,
+            880,
+            131
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jiaxin Zhang $^{1}$ , Zhuohang Li $^{2}$ , Kamalika Das $^{1}$ , Bradley Malin $^{2,3}$ , Sricharan Kumar $^{1}$",
+        "bbox": [
+            129,
+            149,
+            870,
+            167
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "\\(^{1}\\)Intuit AI Research \\(^{2}\\)Vanderbilt University \\(^{3}\\)Vanderbilt University Medical Center \\(\\{jiaxin_zhang, kamalika_das, sricharan_kumar\\} @intuit.com \\(\\{zhuohang.li, b.malin\\} @\\)vanderbilt.edu",
+        "bbox": [
+            149,
+            168,
+            852,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hallucination detection is a critical step toward understanding the trustworthiness of modern language models (LMs). To achieve this goal, we re-examine existing detection approaches based on the self-consistency of LMs and uncover two types of hallucinations resulting from 1) question-level and 2) model-level, which cannot be effectively identified through self-consistency check alone. Building upon this discovery, we propose a novel sampling-based method, i.e., semantic-aware cross-check consistency  $(\\mathrm{SAC}^3)$  that expands on the principle of self-consistency checking. Our  $\\mathrm{SAC}^3$  approach incorporates additional mechanisms to detect both question-level and model-level hallucinations by leveraging advances including semantically equivalent question perturbation and cross-model response consistency checking. Through extensive and systematic empirical analysis, we demonstrate that  $\\mathrm{SAC}^3$  outperforms the state of the art in detecting both nonfactual and factual statements across multiple question-answering and open-domain generation benchmarks. $^1$",
+        "bbox": [
+            142,
+            279,
+            460,
+            620
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            633,
+            260,
+            648
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large-scale pre-trained language models (LMs) have demonstrated exceptional adaptability across a diverse array of natural language tasks that require generating open-ended responses based on user prompt comprehension (Zhao et al., 2023). However, prominent LMs like GPT (Brown et al., 2020) and PaLM (Chowdhery et al., 2022), often exhibit a tendency to produce exceedingly confident, yet erroneous, assertions commonly referred to as hallucinations. This phenomenon significantly impedes their applicability in domains where factual accuracy is of utmost importance.",
+        "bbox": [
+            112,
+            658,
+            489,
+            851
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hallucinations can be detected through the assistance of metrics that capture the uncertainty about",
+        "bbox": [
+            112,
+            852,
+            489,
+            884
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the output sequences. However, these metrics require access to token-level log probabilities, which are not available in commercial black-box LMs like ChatGPT or Bard that only offer limited API access. To overcome this limitation, recent studies explore sampling-based approaches for approximating uncertainty estimation (Lin et al., 2023) through establishing a connection between confidence and self-consistency (Manakul et al., 2023; Mündler et al., 2023). The underlying premise of this principle is that LMs are more inclined to generate consistent responses when high probabilities are assigned to tokens in the answer, which, in turn, implies a level of factuality. In contrast, inconsistent responses are more likely to contain hallucinations. To operationalize this concept, current approaches are designed to sample multiple responses from the LMs for a given question and then compose a hallucination score for each sentence based on the level of response consistency.",
+        "bbox": [
+            507,
+            253,
+            885,
+            575
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Key Observations. In this work, we investigate the relationship between the self-consistency of LMs and the occurrence of hallucinations in a diverse range of tasks. Our investigation indicates that while self-inconsistency in LMs often coincides with hallucination, self-consistency does not necessarily guarantee factual answers, as shown in Figure 1. Our findings challenge the notion that self-consistency alone can serve as a reliable indicator of veracity, as it is demonstrated that LMs can exhibit various tiers of hallucination that elude detection through self-consistency checks. One such tier is question-level hallucination, where LMs consistently generate incorrect answers in response to specific questions (e.g., Table 1). We reveal that by reformulating the questions, it is possible to mitigate such instances of hallucinations. Additionally, our work further reveals the existence of model-level hallucinations, whereby different LMs show discrepancies in their propensity for halluci",
+        "bbox": [
+            507,
+            596,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "1All resources are available at https://github.com/intuit/sac3.",
+        "bbox": [
+            114,
+            891,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "15445",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15445-15458 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            210,
+            945,
+            786,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/cde3c1ba6e9e918f2090a7d724e82f080588c4ee0694871f9be7c53ca9787078.jpg",
+        "image_caption": [
+            "Figure 1: Key observation: solely checking the self-consistency of LLMs is not sufficient for deciding factuality. Left: generated responses to the same question may be consistent but non-factual. Right: generated responses may be inconsistent with the original answer that is factually correct."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            80,
+            877,
+            385
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "nation. Surprisingly, we even observe cases where smaller LMs are capable of correctly answering questions for which larger LMs hallucinate. Together, these findings accentuate the need to consider model-specific characteristics when assessing the occurrence of hallucinations.",
+        "bbox": [
+            112,
+            464,
+            487,
+            557
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Proposed Approach. Motivated by these observations, we introduce SAC<sup>3</sup>, a new sampling-based approach utilizing semantic-aware cross-check consistency to improve the detection of hallucinations in black-box LMs. An overview of our approach is provided in Fig. 2. To address question-level hallucination, we introduce a mechanism that perturbs semantically equivalent questions to evaluate the consistency of LMs' responses across variants of the same question. By examining the generated answers to these perturbed questions, we are able to identify cases where the LM consistently provides incorrect responses to a specific question, which is indicative of a question-level hallucination. Furthermore, we address model-level hallucination by introducing cross-model response consistency checking, which involves comparing the responses of different LMs to the same set of questions. By identifying discrepancies between the responses of different models, we are able to pinpoint cases where certain models exhibit hallucinations while",
+        "bbox": [
+            112,
+            581,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "others provide correct answers. Integrating these cross-checking extensions into our approach significantly improves its ability to detect hallucinations that go beyond self-consistency, thereby providing a more comprehensive assessment of the presence of question-level and model-level hallucinations.",
+        "bbox": [
+            507,
+            464,
+            882,
+            558
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of SAC<sup>3</sup>. We demonstrate the effectiveness of SAC<sup>3</sup> in multiple classification and open-domain generation QA datasets. In the classification QA tasks, our approach achieves a high AUROC score of  $99.4\\%$  and  $97.0\\%$  respectively, which significantly outperforms the self-consistency baseline. In the case of open-domain generation QA tasks, our approach achieves an AUROC score of  $88.0\\%$  and  $77.2\\%$ , which substantially improves upon the self-consistency baseline by a large margin  $(+13.8\\%, \\text{and} +6.7\\%$  respectively).",
+        "bbox": [
+            507,
+            571,
+            884,
+            749
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            763,
+            665,
+            778
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Hallucination in LMs. The issue of hallucination in language models (LMs) has gained significant attention due to its negative impact on performance and the risks it introduces in various natural language processing (NLP) tasks, such as machine translation (Zhou et al., 2020), summarization (Cao et al., 2022), dialogue generation (Das et al., 2023), and question answering (Zhang et al., 2023a; Zheng",
+        "bbox": [
+            507,
+            790,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "15446",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/c8effc0ea87c8227a54a94191fc60794e90503ec05260b451af0cf01e54a4c2c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Input</td><td>Model</td><td>Output</td><td>Consistency</td><td>Factuality</td></tr><tr><td rowspan=\"3\">Self-consistency</td><td>Q: &quot;Is 3691 a prime number?&quot;</td><td>Target</td><td>&quot;No, 3691 is not a prime number as it is divisible by 7 and 13.&quot;</td><td rowspan=\"3\">✓</td><td rowspan=\"3\">✗</td></tr><tr><td>Q: &quot;Is 3691 a prime number?&quot;</td><td>Target</td><td>&quot;No, 3691 is not a prime number as it is divisible by 7 and 13.&quot;</td></tr><tr><td>Q: &quot;Is 3691 a prime number?&quot;</td><td>Target</td><td>&quot;No, 3691 is not a prime number. It can be divided evenly by 7 and 19.&quot;</td></tr><tr><td rowspan=\"3\">Cross-Q consistency</td><td>Q1: &quot;Is the number 3691 prime?&quot;</td><td>Target</td><td>&quot;Yes, the number 3691 is prime.&quot;</td><td rowspan=\"3\">✗</td><td rowspan=\"3\">✓</td></tr><tr><td>Q2: &quot;Are the only factors of 3691 itself and one?&quot;</td><td>Target</td><td>&quot;Yes, the only factors of 3691 are itself and one.&quot;</td></tr><tr><td>Q3: &quot;Can 3691 only be divided by 1 and 3691?&quot;</td><td>Target</td><td>&quot;Yes, 3691 is a prime number and can only be divided by 1 and 3691.&quot;</td></tr><tr><td>Cross-M consistency</td><td>Q: &quot;Is 3691 a prime number?&quot;</td><td>Verifier</td><td>&quot;Yes, 3691 is a prime number.&quot;</td><td>✗</td><td>✓</td></tr><tr><td rowspan=\"3\">Cross-M &amp; Q consistency</td><td>Q1: &quot;Is the number 3691 prime?&quot;</td><td>Verifier</td><td>&quot;Yes.&quot;</td><td rowspan=\"3\">✗</td><td rowspan=\"3\">✓</td></tr><tr><td>Q2: &quot;Are the only factors of 3691 itself and one?&quot;</td><td>Verifier</td><td>&quot;Yes, the only factors of 3691 are 1 and itself.&quot;</td></tr><tr><td>Q3: &quot;Can 3691 only be divided by 1 and 3691?&quot;</td><td>Verifier</td><td>&quot;Yes, 3691 can only be divided by 1 and 3691.&quot;</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            884,
+            286
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: An illustrative example of self-consistency, cross-question consistency, and cross-model consistency check. The original question and answer are \"Is 3691 a prime number?\" and \"No, 3691 is not a prime number. It is divisible by 7 and 13\", respectively. Each row presents a set of sampled QA pairs along with its consistency regarding the original answer, and the predicted factuality of the original answer.",
+        "bbox": [
+            112,
+            294,
+            884,
+            354
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "et al., 2023b; Dhuliawala et al., 2023). Recent survey (Ji et al., 2023; Zhang et al., 2023c; Ye et al., 2023) and evaluation benchmarks (Liu et al., 2021; Li et al., 2023a; Yang et al., 2023) have highlighted the importance of addressing this issue. Previous research has explored hallucination evaluation using confidence-based approaches (Xiao and Wang, 2021; Varshney et al., 2023; Chen and Mueller, 2023) that require access to token-level log probability (Kuhn et al., 2023; Cole et al., 2023) or supervised tuning (Agrawal et al., 2023; Li et al., 2023b) that relies on internal states of the LM. However, these methods may not be applicable when only API access to the LM is available (Agrawal et al., 2023). Another approach involves retrieving knowledge from external databases to tackle hallucinations (Ji et al., 2022; Zheng et al., 2023a; Peng et al., 2023; Zhang et al., 2023b). In contrast to these studies, our work focuses on detecting hallucinations in open-domain QA tasks using black-box LMs, without relying on external resources.",
+        "bbox": [
+            115,
+            378,
+            490,
+            715
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Consistency Evaluation of LMs. An essential characteristic of logically valid intelligent systems is self-consistency, which entails that no two statements provided by the system contradict each other. Self-consistency is defined by Elazar et al. (2021) as the invariance of an LM's responses across various types of semantics-preserving prompt transformations. This definition is further enriched by multiple other consistency categories proposed by Jang et al. (2022). Wang et al. (2022) demonstrates that self-consistency can significantly enhance the chain of thought reasoning in LMs. Without self",
+        "bbox": [
+            112,
+            726,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "consistency, it becomes challenging to regard LMs as reliable or trustworthy systems. Recent studies employ self-consistency to detect hallucinations based on pretrained LMs (Manakul et al., 2023) and instruction-tuned LMs (Mündler et al., 2023). Although these methods exhibit promising accuracy on several specific tasks, potential failures (Chen et al., 2023) of self-consistency are overlooked in the current settings, as existing LMs frequently provide inconsistent responses to questions (Mitchell et al., 2022) and factual knowledge inquiries (Elazar et al., 2021; Tam et al., 2023; Gekhman et al., 2023). Our work addresses these concerns by introducing a cross-check consistency approach, aiming to bridge the gap between self-consistency and factual assessment.",
+        "bbox": [
+            507,
+            378,
+            884,
+            634
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Self-consistency Limitations in Factuality Assessment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            649,
+            808,
+            682
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The essential assumption of self-consistency in factuality assessment is that if the LM has the knowledge of the concept, responses sampled from its output distribution, should be similar and consistent; conversely, if the LM lacks corresponding knowledge, the sampled responses would contain hallucinated facts that are diverged and contradictory. Although this assumption may seem reasonable, it does not always hold in practice (more details are provided in the Appendix A). Specifically, we argue that solely checking the LM's self-consistency is insufficient for detecting hallucination or verifying factuality under the following two circumstances:",
+        "bbox": [
+            507,
+            694,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "15447",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "1. LMs may produce consistently hallucinated facts. We observe that for certain questions, LMs may output consistently wrong answers. For instance, as shown in Fig. 1, when prompted with the question \"Is pi smaller than 3.2?\", ChatGPT consistently generates incorrect answers. In this case, where the generated responses are consistent but non-factual, solely relying on self-consistency checking of a single model would yield false negative hallucination detection results.",
+        "bbox": [
+            112,
+            84,
+            489,
+            243
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2. Even in cases when LMs generate factual statements in their original response, the stochastic sampled responses may lack veracity. For example, the original answer (Answer 1) of ChatGPT under zero temperature is correct regarding the senator search question as shown in Fig. 1. However, when sampled with a higher temperature, ChatGPT generates multiple incorrect responses (Answer 2 and Answer  $m$ ). In this scenario, where the sampled responses are inconsistent and disagree with the original response which itself is factually correct, methods that rely solely on model self-checking would produce false positives.",
+        "bbox": [
+            110,
+            250,
+            489,
+            460
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In summary, although the inconsistency of sampled responses has been empirically demonstrated to be correlated with hallucinated facts on certain tasks, in general, self-consistency is neither necessary nor sufficient to verify the veracity of large LMs' statements. Therefore, methods based solely on self-consistency checking may not be able to accurately detect hallucinations in complex QA and open-domain generation tasks, which motivates us to design a more reliable and robust factuality assessment method that extends this idea.",
+        "bbox": [
+            112,
+            461,
+            489,
+            636
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 SAC<sup>3</sup>: Semantic-Aware Cross-check Consistency",
+        "text_level": 1,
+        "bbox": [
+            112,
+            650,
+            457,
+            683
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This section describes the proposed semantic-aware cross-check consistency approach, a high-level overview of which is provided in Fig. 1. Additionally, an illustrative example of each component is presented in Table 1. Here, we walk through each component in detail.",
+        "bbox": [
+            112,
+            693,
+            489,
+            789
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Stage 1: Question-level Cross-checking via Semantically Equivalent Perturbations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            801,
+            487,
+            832
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Contrary to existing techniques that assess semantic equivalence through entailment or paraphrasing, our approach involves rephrasing the input query by generating alternative inputs that preserve semantic equivalence, i.e., semantically equivalent input per",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "turbation. To achieve this, we leverage advances in LLM prompting. Starting with a queried input  $\\pmb{x}_0$ , we acquire a set of  $k$  semantically equivalent inputs  $\\{x_1, x_2, \\dots, x_k\\}$  through the prompt: \"For the question [queried Question], provide  $k$  semantically equivalent questions\".",
+        "bbox": [
+            507,
+            84,
+            882,
+            180
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To ensure the quality of the generated inputs in this step, we further double-check the semantic equivalence between the generated inputs  $\\{x_{1}, x_{2}, \\ldots, x_{k}\\}$  and the queried input  $x_{0}$  in a pair-wise manner using the prompt \"Are the following two inputs semantically equivalent? [queried INPUT] [generated INPUT]\" and filtering out the inputs that do not share the same semantic meaning as the original input. The complete prompt templates used in this work are provided in the Appendix B.",
+        "bbox": [
+            507,
+            187,
+            884,
+            363
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Stage 2: Model-level Cross-check with Additional Verifier LM",
+        "text_level": 1,
+        "bbox": [
+            507,
+            398,
+            855,
+            429
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Let  $s_0$  denote the original response from a target LM  $\\mathcal{T}$  based on a given query  $x_0$ . Our objective is to detect whether  $s_0$  is hallucinated by sampling responses from the predictive distribution of  $\\mathcal{T}$ . To avoid model-level hallucination, we introduce an additional verifier LM denoted as  $\\mathcal{V}$  for model-level cross-checking. We define the responses from both models as:",
+        "bbox": [
+            507,
+            448,
+            882,
+            575
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {s} _ {\\mathcal {T} _ {j}} = \\mathcal {T} (\\boldsymbol {x} _ {j}), \\boldsymbol {s} _ {\\mathcal {V} _ {j}} = \\mathcal {V} (\\boldsymbol {x} _ {j}), j = 1, \\dots , k, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            527,
+            582,
+            882,
+            600
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $k$  is the length of the generated semantically equivalent inputs  $\\{x_{1}, x_{2}, \\ldots, x_{k}\\}$  in stage 1. To assess the factuality of  $x_{0}$ , the self-checking mechanism operates by drawing a set of  $n_{s}$  stochastic response samples from the target LM:  $\\mathcal{S}_{\\mathcal{T}_0} = \\{s_{\\mathcal{T}_0}^1, s_{\\mathcal{T}_0}^2, \\ldots, s_{\\mathcal{T}_0}^{n_s}\\}$ . Similarly, we can apply the same self-checking mechanism to the verifier LM to generate another set of  $n_m$  responses:  $\\mathcal{S}_{\\mathcal{V}_0} = \\{s_{\\mathcal{V}_0}^1, s_{\\mathcal{V}_0}^2, \\ldots, s_{\\mathcal{V}_0}^{n_m}\\}$ . To perform question-level cross-check, for each perturbed input  $x_k$ , we generate  $n_q$  sampled response sequences  $\\mathcal{S}_{\\mathcal{T}_k} = \\{s_{\\mathcal{T}_k}^1, s_{\\mathcal{T}_k}^2, \\ldots, s_{\\mathcal{T}_k}^{n_q}\\}$  from the target LM  $\\mathcal{T}$  and  $n_{qm}$  sampled responses  $\\mathcal{S}_{\\mathcal{V}_k} = \\{s_{\\mathcal{V}_k}^1, s_{\\mathcal{V}_k}^2, \\ldots, s_{\\mathcal{V}_k}^{n_{qm}}\\}$  from the verifier LM  $\\mathcal{V}$ .",
+        "bbox": [
+            507,
+            604,
+            882,
+            831
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, we collect the total sampled sets  $S = \\{S_{\\mathcal{T}_0}, S_{\\mathcal{V}_0}, S_{\\mathcal{T}_k}, S_{\\mathcal{V}_k}\\}$  by combining all samples drawn from self-checking and cross-checking, which will be used next for calculating a consistency score.",
+        "bbox": [
+            507,
+            839,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "15448",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f6d756e06591d2dccc325d3f6a940be0fc0e7ce85e367dda14dbb9edd9739fb7.jpg",
+        "image_caption": [
+            "Figure 2: Overview of the proposed semantic-aware cross-check consistency  $(\\mathrm{SAC}^3)$  method."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            85,
+            872,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Stage 3: Consistency Score Calculation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            235,
+            465,
+            250
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This stage uses the generated sample sets in all previous stages to calculate a numerical consistency score that captures the question-level and model-level cross-checking paradigm.",
+        "bbox": [
+            112,
+            260,
+            489,
+            325
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3.1 Semantic-aware Consistency Check of QA Pairs",
+        "text_level": 1,
+        "bbox": [
+            112,
+            340,
+            475,
+            370
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Most of the existing works mainly focus on examining the consistency of LM outputs while ignoring the effect of the inputs. However, in QA tasks, it is important to consider both inputs and outputs when measuring semantic equivalence, as the same question can be rephrased in many ways. Although the answers to these questions (e.g., \"no\" and \"yes\") may not be lexically equivalent, the QA pairs as a whole can be semantically equivalent. In light of this, we propose to check the semantic consistency of the QA pairs instead of the answer only.",
+        "bbox": [
+            112,
+            379,
+            489,
+            556
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3.2 Self-checking Consistency (SC²) Score",
+        "text_level": 1,
+        "bbox": [
+            112,
+            571,
+            470,
+            588
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Let  $\\mathcal{C}(\\cdot, \\cdot)$  denote a semantic equivalence checking operator that takes two QA pairs as inputs. The operator  $\\mathcal{C}$  returns \"Yes\" if the two QA pairs are semantically equivalent, and \"No\" otherwise. This operator should be reflexive, symmetric, and transitive. We implement the checking operator using an LM by leveraging the prompt: \"Are the following two Question-Answering (QA) pairs semantically equivalent? [QA PAIR 1] [QA PAIR 2]. We then map the best guess to a numerical semantic equivalent score:  $\\{\\text{\"Yes\"} \\rightarrow 0.0, \\text{\"No\"} \\rightarrow 1.0\\}$ . We use  $\\mathcal{P}_0 = (\\pmb{x}_0, \\pmb{s}_0)$  to denote the original QA pair. The self-checking score  $\\mathcal{Z}_{\\mathrm{SC}^2}$  of the target LM  $\\mathcal{T}$  can be calculated by",
+        "bbox": [
+            112,
+            595,
+            489,
+            820
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {Z} _ {\\mathrm {S C} ^ {2}} = \\frac {1}{n _ {s}} \\sum_ {i = 1} ^ {n _ {s}} \\mathcal {C} \\left(\\mathcal {P} _ {0}, \\mathcal {P} _ {\\mathcal {S} _ {\\tau_ {0}}} ^ {i}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            825,
+            487,
+            865
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{P}_{\\mathcal{S}_{\\mathcal{T}_0}} = \\{(x_0, s_{\\mathcal{T}_0}^1), \\dots, (x_0, s_{\\mathcal{T}_0}^{n_s})\\}$  represents the QA pairs generated in the self-checking scenario.",
+        "bbox": [
+            112,
+            870,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3.3 Question-level Consistency (SAC $^3$ -Q) Score",
+        "text_level": 1,
+        "bbox": [
+            507,
+            235,
+            850,
+            266
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Besides self-checking the original question  $\\pmb{x}_0$ , SAC<sup>3</sup> further assesses cross-check consistency of perturbed questions  $\\{x_1, x_2, \\dots, x_k\\}$ . The corresponding QA pairs compose a two-dimensional matrix, where each row corresponds to a perturbed question ( $k$  in total), and each column corresponds to a sampled response ( $n_q$  in total):",
+        "bbox": [
+            507,
+            281,
+            884,
+            394
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {P} _ {\\mathcal {S} _ {\\mathcal {T} _ {j}}} ^ {i} = \\left[ \\begin{array}{c c c} \\left(\\boldsymbol {x} _ {1}, \\mathcal {S} _ {\\mathcal {T} _ {1}} ^ {1}\\right) & \\dots & \\left(\\boldsymbol {x} _ {1}, \\mathcal {S} _ {\\mathcal {T} _ {1}} ^ {n _ {q}}\\right) \\\\ \\dots & \\dots & \\dots \\\\ \\left(\\boldsymbol {x} _ {k}, \\mathcal {S} _ {\\mathcal {T} _ {k}} ^ {1}\\right) & \\dots & \\left(\\boldsymbol {x} _ {k}, \\mathcal {S} _ {\\mathcal {T} _ {k}} ^ {n _ {q}}\\right) \\end{array} \\right]. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            549,
+            399,
+            882,
+            451
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Therefore, the question-level cross-checking consistency score  $\\mathcal{Z}_{\\mathrm{SAC}^3 -\\mathrm{Q}}$  can be obtained by",
+        "bbox": [
+            507,
+            455,
+            882,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {Z} _ {\\mathrm {S A C} ^ {3} - \\mathrm {Q}} = \\frac {1}{n _ {q} \\cdot k} \\sum_ {i = 1} ^ {n _ {q}} \\sum_ {j = 1} ^ {k} \\mathcal {C} \\left(\\mathcal {P} _ {0}, \\mathcal {P} _ {\\mathcal {S} _ {\\mathcal {T} _ {j}}} ^ {i}\\right). \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            551,
+            494,
+            882,
+            538
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3.4 Model-level Consistency (SAC $^3$ -M & SAC $^3$ -QM) Score",
+        "text_level": 1,
+        "bbox": [
+            507,
+            565,
+            843,
+            596
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In addition to the question-level score, a model-level cross-check score is calculated by performing cross-model checking and cross-question checking using the verifier LM  $\\nu$ . Specifically, for the original question  $x_0$ , the model-level cross-checking consistency score  $\\mathcal{Z}_{\\mathrm{SAC}^3 -\\mathsf{M}}$  is computed by",
+        "bbox": [
+            507,
+            612,
+            882,
+            709
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {Z} _ {\\mathrm {S A C} ^ {3} - \\mathrm {M}} = \\frac {1}{n _ {m}} \\sum_ {i = 1} ^ {n _ {m}} \\mathcal {C} \\left(\\mathcal {P} _ {0}, \\mathcal {P} _ {\\mathcal {S} _ {\\mathcal {V} _ {0}}} ^ {i}\\right), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            574,
+            714,
+            882,
+            753
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{P}_{\\mathcal{S}_{\\mathcal{V}_0}} = \\{(x_0,s_{\\mathcal{V}_0}^1),\\ldots ,(x_0,s_{\\mathcal{V}_0}^{n_m})\\}$  is the QA pairs generated by the verified LM  $\\nu$",
+        "bbox": [
+            507,
+            758,
+            880,
+            791
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The cross-question consistency score on the verifier LM is computed on the QA pairs produced by  $\\nu$ :",
+        "bbox": [
+            507,
+            797,
+            882,
+            844
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {P} _ {\\mathcal {S} _ {\\mathcal {V} _ {j}}} ^ {i} = \\left[ \\begin{array}{c c c} \\left(\\boldsymbol {x} _ {1}, \\mathcal {S} _ {\\mathcal {V} _ {1}} ^ {1}\\right) & \\dots & \\left(\\boldsymbol {x} _ {1}, \\mathcal {S} _ {\\mathcal {V} _ {1}} ^ {n _ {q m}}\\right) \\\\ \\dots & \\dots & \\dots \\\\ \\left(\\boldsymbol {x} _ {k}, \\mathcal {S} _ {\\mathcal {V} _ {k}} ^ {1}\\right) & \\dots & \\left(\\boldsymbol {x} _ {k}, \\mathcal {S} _ {\\mathcal {V} _ {k}} ^ {n _ {q m}}\\right) \\end{array} \\right]. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            532,
+            846,
+            882,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The cross-model cross-question consistency score",
+        "bbox": [
+            507,
+            903,
+            880,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "15449",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "can thus be obtained through",
+        "bbox": [
+            112,
+            84,
+            331,
+            99
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {Z} _ {\\mathrm {S A C} ^ {3} - \\mathrm {Q M}} = \\frac {1}{n _ {q m} \\cdot k} \\sum_ {i = 1} ^ {n _ {q m}} \\sum_ {j = 1} ^ {k} \\mathcal {C} \\left(\\mathcal {P} _ {0}, \\mathcal {P} _ {\\mathcal {S} _ {\\mathcal {V} _ {j}}} ^ {i}\\right). \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            136,
+            105,
+            487,
+            149
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3.5 Final Score and Model Confidence",
+        "text_level": 1,
+        "bbox": [
+            112,
+            159,
+            443,
+            174
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The different variants of SAC<sup>3</sup> capture different aspects of the uncertainty about the original response and should complement each other. We thus consider a combination of all variants including SAC<sup>3</sup>-Q, SAC<sup>3</sup>-M, SAC<sup>3</sup>-QM as the final score:",
+        "bbox": [
+            112,
+            181,
+            487,
+            261
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {Z} _ {\\mathrm {S A C} ^ {3} - \\mathrm {a l l}} = \\mathcal {Z} _ {\\mathrm {S A C} ^ {3} - \\mathrm {Q}} + \\lambda \\left(\\mathcal {Z} _ {\\mathrm {S A C} ^ {3} - \\mathrm {M}} + \\mathcal {Z} _ {\\mathrm {S A C} ^ {3} - \\mathrm {Q M}}\\right), \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            124,
+            267,
+            487,
+            286
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda$  is a weight factor for the verifier LM. Unless mentioned otherwise, we use  $\\lambda = 1$  by default in our experiments. In practice, as the computation of each component is independent, they can be computed in parallel to reduce latency. The detection prediction is made by comparing the final score with a preset threshold. In addition to the computed score, we also ask the target LM to generate a verbalized confidence score (Tian et al., 2023) along with its prediction when checking the semantic equivalence of QA pairs. More discussions are offered in the Appendix C.2.",
+        "bbox": [
+            112,
+            288,
+            489,
+            482
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Data and Annotation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            499,
+            331,
+            514
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our hallucination detection approach on two categories of QA tasks, namely, classification QA and generation QA, with each category containing two datasets. Following prior work (Zhang et al., 2023a), we use the following two binary classification datasets for evaluation on the classification QA task:",
+        "bbox": [
+            112,
+            527,
+            487,
+            639
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Prime number: this dataset contains 500 questions that query the primality of a randomly chosen prime number between 1,000 and 20,000, where the factual answer is always \"Yes\". The synthesized hallucinated answers are \"No, it is not a prime number\".",
+            "- Senator search: the dataset consists of 500 questions that follow the following template: \"Was there ever a US senator that represented the state of [US STATE NAME] and whose alma mater was [US COLLEGE NAME]?\" The factual answer is always \"No\". We also generate hallucinated answers: \"Yes, there was a US senator that represented the state of [US STATE NAME] and whose alma mater was [US COLLEGE NAME].\"."
+        ],
+        "bbox": [
+            115,
+            652,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As for the generation QA tasks, we take questions from the following two open-domain QA datasets and generate answers using LLMs. Then we manually annotate the factuality of the answers following previous work (Li et al., 2023a).",
+        "bbox": [
+            507,
+            84,
+            882,
+            164
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- HotpotQA-halu: We randomly sample 250 examples from the training set of HotpotQA (Yang et al., 2018) and generate hallucinated answers drawn from gpt-3.5-turbo. Then we manually annotate the answers by comparing the ground truth and knowledge.",
+            "- NQ-open-halu: Natural Questions (NQ)-open (Lee et al., 2019) is a more challenging open domain QA benchmark (Kwiatkowski et al., 2019). We use the same setting as HotpotQA-halu to create a small-scale dataset that consists of 250 non-factual and factual examples with manual annotations."
+        ],
+        "bbox": [
+            510,
+            175,
+            882,
+            393
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Please find more relevant details about data annotations in the Appendix C.3.",
+        "bbox": [
+            507,
+            406,
+            882,
+            437
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            453,
+            653,
+            469
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.1 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            481,
+            712,
+            495
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Models. We use gpt-3.5-turbo from OpenAI as the target LM for our experiment. The verifier LM is chosen from the following two models: (1) Falcon-7b-instruct (Al-mazrouei et al., 2023): an open-source causal decoder-only model built by TII that is trained on 1,500B tokens of RefinedWeb (Penedo et al., 2023) and further enhanced using the curated corpora; and (2) Guanaco-33b: an open-source instruction-following models through QLoRA (Dettmers et al., 2023) tuning of LLaMA (Touvron et al., 2023) base model on the OASST1 dataset.",
+        "bbox": [
+            507,
+            504,
+            882,
+            697
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. The evaluation is conducted using Azure OpenAI API. When performing semantic perturbations and consistency checking, we set the temperature to 0.0 to get deterministic high-quality outputs. Given a specific input query, we generate  $k = 10$  semantically equivalent inputs using the prompt described in Section 4.1. For the self-checking-based method  $\\mathsf{SC}^2$ , we follow prior work (Manakul et al., 2023) to set the temperature to 1.0 and generate  $n_s = 10$  stochastic samples. For  $\\mathrm{SAC}^3 -\\mathrm{Q}$  and  $\\mathrm{SAC}^3 -\\mathrm{QM}$ , we set  $n_q = n_{qm} = 1$  to reduce computational cost. To further reduce the inference cost, we set  $n_m = 1$  by default and",
+        "bbox": [
+            507,
+            709,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "15450",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "combine SAC $^{3}$ -M with SAC $^{3}$ -QM to report the model-level results. We use hallucination detection accuracy and area under the ROC curve (AUROC) to evaluate the performance. In addition to the estimated hallucination score, we also show the verbalized probabilities (Tian et al., 2023) from the target LM for comparison. We execute all experiments on 8 NVIDIA V100 32G GPUs.",
+        "bbox": [
+            112,
+            83,
+            489,
+            212
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/6ed84477cda719e05a673d3d5dec9e5879f251737169523c62dcdd42988b99c4.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Prime number</td><td colspan=\"2\">Senator search</td></tr><tr><td>Score</td><td>Confidence</td><td>Score</td><td>Confidence</td></tr><tr><td>SC2(gpt-3.5-turbo)</td><td>65.9</td><td>67.5</td><td>56.1</td><td>53.1</td></tr><tr><td>SAC3-Q(gpt-3.5-turbo)</td><td>99.4</td><td>99.7</td><td>99.7</td><td>99.7</td></tr></table>",
+        "bbox": [
+            117,
+            225,
+            487,
+            288
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2 Evaluation Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            356,
+            309,
+            370
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2.1 Classification QA",
+        "text_level": 1,
+        "bbox": [
+            112,
+            380,
+            312,
+            395
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Balanced Dataset. We first experiment on balanced datasets with  $50\\%$  hallucinated samples and  $50\\%$  factual samples. Table 2 compares the detection performance of  $\\mathsf{SC^2}$  and  $\\mathrm{SAC}^3 -\\mathrm{Q}$  in terms of AUROC and verbalized confidence score. We observe that self-checking  $(\\mathrm{SC^2})$  performs poorly on both datasets, with a low AUROC of  $65.9\\%$  and  $56.1\\%$ , respectively. Our question-level cross-checking  $(\\mathrm{SAC}^3 -\\mathrm{Q})$  significantly outperforms the  $\\mathrm{SC^2}$  baseline achieving  $>99\\%$  AUROC on both datasets and is in line with the verbalized confidence score, confirming the effectiveness of cross-checking.",
+        "bbox": [
+            112,
+            404,
+            489,
+            612
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Unbalanced Dataset. We further evaluate our method in a more challenging scenario where the dataset only contains hallucinated samples. Table 3 presents the accuracy of detecting hallucinated samples using a preset threshold of 0.5. In this case, the performance of self-check drops significantly to  $48.2\\%$  and  $29.6\\%$  respectively.  $\\mathrm{SAC}^3 -\\mathrm{Q}$  still outperforms  $\\mathrm{SC}^2$  by a large margin. The model-level cross-check with verifier LMs performs well in the prime number dataset but fails to accurately detect hallucination in the senator search dataset. This is because both verifier LMs refuse to answer a large portion of the questions on this dataset due to a lack of sufficient information. By combining the target LM,  $\\mathrm{SAC}^3$ -a11 with Guanaco-33b achieves the highest detection accuracy compared to other baselines  $\\mathrm{SC}^2 (+51.2\\%)$ ,  $\\mathrm{SAC}^3 -\\mathrm{Q} (+6.2\\%)$ , and  $\\mathrm{SAC}^3 -\\mathrm{QM} (+5.0\\%)$ .",
+        "bbox": [
+            112,
+            629,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/067aa9bed5614005ec3ac13930727da8ce390d978ac354577ab92842607f7beb.jpg",
+        "table_caption": [
+            "Table 2: AUROC on classification QA tasks with  $50\\%$  hallucinated samples and  $50\\%$  factual samples."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Prime number</td><td colspan=\"2\">Senator search</td></tr><tr><td>Score</td><td>Confidence</td><td>Score</td><td>Confidence</td></tr><tr><td>SC2(gpt-3.5-turbo)</td><td>48.2</td><td>51.0</td><td>29.6</td><td>30.6</td></tr><tr><td>SAC3-Q(gpt-3.5-turbo)</td><td>93.2</td><td>96.4</td><td>97.0</td><td>97.4</td></tr><tr><td>SAC3-QM (Falcon-7b)</td><td>89.8</td><td>91.2</td><td>21.0</td><td>21.8</td></tr><tr><td>SAC3-QM (Guanaco-33b)</td><td>94.4</td><td>96.3</td><td>45.6</td><td>46.2</td></tr><tr><td>SAC3-all (Falcon-7b)</td><td>97.8</td><td>98.0</td><td>84.6</td><td>85.6</td></tr><tr><td>SAC3-all (Guanaco-33b)</td><td>99.4</td><td>99.4</td><td>85.6</td><td>86.7</td></tr></table>",
+        "bbox": [
+            512,
+            80,
+            884,
+            198
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Accuracy on classification QA tasks with  $100\\%$  hallucinated samples (with the threshold set to 0.5).",
+        "bbox": [
+            507,
+            206,
+            882,
+            235
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of Threshold. Since the choice of threshold has a significant impact on the detection accuracy in the case where the dataset only contains positive (hallucinated) samples, we further experiment with different thresholds and present the results in Fig. 3. We observe that  $\\mathrm{SAC}^3$ -Q and  $\\mathrm{SAC}^3$ -a11 with Guanaco-33b are more robust against large threshold values and outperform  $\\mathrm{SC}^2$  in most cases.",
+        "bbox": [
+            507,
+            261,
+            884,
+            388
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/cb4fcdc3651af384934237e97e162eff29423d4337e3513da6b0c376c5f628b2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            401,
+            704,
+            537
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/820cd6e634bb353c1cd856328097e6327074baccb6ceda00df50174331febc01.jpg",
+        "image_caption": [
+            "Figure 3: Impact of threshold on detection accuracy."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            541,
+            875,
+            564
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8065ea4ffe0083104ee89f4f65ac9292fb538e4c4e5913d2ab0bab281503a2fd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            705,
+            401,
+            878,
+            537
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Why Does Self-checking Fail? To further understand why self-checking methods fail to detect some hallucinated responses, we visualize the distribution of consistency scores using histogram plots in Fig. 4. We observe that for  $\\mathsf{SC}^2$ , a significant portion of hallucinated samples received highly consistent predictions. In other words, the target LM made consistently wrong predictions due to a lack of question and model diversity, which aligns with our analysis in Section 3. On the other hand, benefiting from the semantically equivalent question perturbation,  $\\mathrm{SAC}^3 -Q$ ’s scores are more spread out in the inconsistent region, which helps to improve the effectiveness of detecting hallucinations by choosing a proper threshold.",
+        "bbox": [
+            505,
+            615,
+            884,
+            858
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2.2 Open-domain Generation QA",
+        "text_level": 1,
+        "bbox": [
+            507,
+            868,
+            803,
+            883
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Compared to the classification QA tasks, detecting hallucinations in open-domain generation QA",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "15451",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7471c5546dd0753b1290c087ea06bc1198a17e4d4ba20a56b83114986c4d020f.jpg",
+        "image_caption": [
+            "Figure 4: Histogram of hallucination score."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            82,
+            484,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tasks is more challenging. As shown in Table 4,  $\\mathrm{SAC}^3 -\\mathrm{Q}$  exhibits better AUROC than  $\\mathrm{SC}^2$ $(+7\\%)$  in both datasets. Compared to  $\\mathrm{SAC}^3 -\\mathrm{Q}$ ,  $\\mathrm{SAC}^3 -\\mathrm{QM}$  shows  $6.7\\%$  improvement in the HotpotQA-halu dataset but is slightly worse in the NQ open dataset.  $\\mathrm{SAC}^3$ -all leverages the advantages of question-level and model-level cross-checking and achieves consistently good performance in both datasets.",
+        "bbox": [
+            112,
+            269,
+            489,
+            399
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/89afbc02bf8df0b55518f2341eb97c2ae2b3ab82af5fea5c14f701331b18c95a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Method</td><td colspan=\"2\">HotpotQA-halu</td><td colspan=\"2\">NQ-open-halu</td></tr><tr><td>Score</td><td>Confidence</td><td>Score</td><td>Confidence</td></tr><tr><td colspan=\"2\">SC2(gpt-3.5-turbo)</td><td>74.2</td><td>77.0</td><td>70.5</td><td>72.7</td></tr><tr><td colspan=\"2\">SAC3-Q(gpt-3.5-turbo)</td><td>81.3</td><td>81.4</td><td>77.2</td><td>78.5</td></tr><tr><td colspan=\"2\">SAC3-QM (Falcon-7b)</td><td>83.0</td><td>79.5</td><td>67.5</td><td>62.0</td></tr><tr><td colspan=\"2\">SAC3-QM (Guanaco-33b)</td><td>88.0</td><td>85.2</td><td>72.7</td><td>72.7</td></tr><tr><td colspan=\"2\">SAC3-all (Falcon-7b)</td><td>84.5</td><td>84.5</td><td>77.1</td><td>77.2</td></tr><tr><td colspan=\"2\">SAC3-all (Guanaco-33b)</td><td>87.0</td><td>86.8</td><td>77.2</td><td>77.8</td></tr></table>",
+        "bbox": [
+            117,
+            412,
+            487,
+            527
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: AUROC on open-domain generation QA tasks.",
+        "bbox": [
+            112,
+            537,
+            485,
+            551
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of Verifier LM Weight. In our previous experiments, we assigned equal importance to the consistency score computed by the target and verifier LM. However, typically, the target LM and the verifier LM have different architectures and scales such that the user may have different levels of trust in their output truthfulness. This difference in trust can be incorporated by introducing a weight  $\\lambda$  to the consistency score produced by the verifier LM. For instance, if the goal is to detect hallucination in a specialized domain and the verifier LM is a domain-specific model developed for this domain, we can assign a large weight to its scores (e.g.,  $\\lambda > 1.0$ ). In the general case, where the verifier LM is a small-sized open-source model, we can apply a small weight value (e.g.,  $\\lambda < 1.0$ ) to discount the influence of the verifier LM in the final score. Fig. 5 visualizes the effect of various weight factors on the generation QA datasets. We observe that SAC<sup>3</sup>-all with a higher weight would result in a larger advantage over SAC<sup>3</sup>-Q in",
+        "bbox": [
+            112,
+            580,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the HotpotQA-halu task, where the verifier LM outperforms the target LM. On the contrary, in the NQ open task where the target LM shows competitive performance, a smaller weight would yield better results.",
+        "bbox": [
+            507,
+            84,
+            884,
+            162
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/632cf857050e3c680125f3242f79e5862c24e71c180d80dd33329837d1a060a8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            510,
+            172,
+            878,
+            284
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b0a467b8ed3b0981a44326b99a869ea9e6f029687c6fce8000e5a4d5038bdd4f.jpg",
+        "image_caption": [
+            "Figure 5: Effect of verifier LM weight on AUROC."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            286,
+            878,
+            398
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of the Number of Perturbed Questions. The performance of sampling-based methods is expected to improve as the sample size increases, at the cost of higher latency and computational cost. In Fig. 6, we study this trade-off by varying the number of perturbed questions  $k$  from 2 to 10. We observe that the performance of SAC³ increases as more question samples are used but the performance gain gradually diminishes after using more than 5 question samples. This suggests that in practice we could use 2-4 question samples to achieve reasonably good performance at a low computational cost.",
+        "bbox": [
+            507,
+            455,
+            884,
+            664
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/336aa453fc86d1c725fd409a10cd694bff47b5f4bbd3dbec3e273b9a91f77460.jpg",
+        "image_caption": [
+            "Figure 6: Performance of SAC<sup>3</sup> with varying number of perturbed questions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            677,
+            882,
+            802
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of the Model Type. Our proposed SAC<sup>3</sup> framework does not restrict the type of LLM employed and can be naturally extended to various",
+        "bbox": [
+            507,
+            870,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "15452",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "types of target LLMs. To verify this, in addition to GPT-3.5 (gpt-3.5-turbo), we conduct experiments using GPT-4 (gpt-4) and PaLM 2 (chat-bison) on the considered four datasets in the setting of the balanced dataset. The experimental results of comparing the proposed  $\\mathrm{SAC}^3 -\\mathrm{Q}$  with the  $\\mathrm{SC}^2$  baseline are summarized in Table 5. We observe that the proposed  $\\mathrm{SAC}^3 -\\mathrm{Q}$  consistently outperforms the  $\\mathrm{SC}^2$  baseline across all LLM variants.",
+        "bbox": [
+            112,
+            84,
+            492,
+            231
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/4ab45557ea4d0d7b50d38914a0b626aef063d310863e9e5d9f99e30cea9fc4e1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Prime Number</td><td>Senator Search</td><td>HotpotQA -halu</td><td>NQ-open -halu</td></tr><tr><td>SC2(gpt-3.5-turbo)</td><td>48.2</td><td>29.6</td><td>74.2</td><td>70.5</td></tr><tr><td>SC2(gpt-4)</td><td>38.3</td><td>18.4</td><td>79.7</td><td>76.3</td></tr><tr><td>SC2(chat-bison)</td><td>26.9</td><td>19.2</td><td>75.8</td><td>67.9</td></tr><tr><td>SAC3-Q (gpt-3.5-turbo)</td><td>93.2</td><td>97.0</td><td>81.3</td><td>77.2</td></tr><tr><td>SAC3-Q (gpt-4)</td><td>91.1</td><td>61.6</td><td>87.2</td><td>82.9</td></tr><tr><td>SAC3-Q (chat-bison)</td><td>90.3</td><td>66.3</td><td>82.8</td><td>72.7</td></tr></table>",
+        "bbox": [
+            117,
+            241,
+            489,
+            344
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Accuracy of different LLMs (GPT-3.5, GPT-4, and PaLM 2) on classification and generation QA tasks.",
+        "bbox": [
+            112,
+            351,
+            489,
+            384
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Computational Cost. We monitor the computational cost of our approach based on the number of model evaluations consumed by OpenAI API and open-source LLMs inference. Assuming the number of samples equals the number of perturbed questions, i.e.,  $n_s = n_m = n_q = n_{qm}$ , the cost of  $\\mathsf{SC^2}$  is  $n_s$  API calls, and our  $\\mathsf{SAC^3}$ -all needs  $n_s$  target LM calls plus  $2 \\times n_s$  verifier LM calls. Beyond the model evaluations,  $\\mathsf{SAC^3}$  may have additional costs from question perturbations and semantic equivalence checking via prompting. Additional discussions can be found in the Appendix C.1.",
+        "bbox": [
+            112,
+            410,
+            489,
+            605
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion and Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            618,
+            379,
+            634
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We investigate the relationship between the self-consistency of LM and the factuality of the response and propose SAC<sup>3</sup> as a robust hallucination detection approach for black-box LMs. Through extensive empirical analysis, our work highlights several findings. First, self-consistency checking alone is insufficient to effectively detect question-level and model-level hallucinations, where LMs generate consistently wrong responses to certain questions. Second, cross-checking between semantically equivalent questions can reduce the occurrence of persistent hallucinations, potentially by reducing question-level ambiguity. Third, there exists a model-level disparity in hallucinations, which we attribute to the inherent differences in LM capabilities originating from different training procedures and data. Thus, the verifier LM can be selected",
+        "bbox": [
+            112,
+            645,
+            492,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "according to specific tasks to maximize detection accuracy. We believe that our work is an important step towards building reliable LLMs.",
+        "bbox": [
+            507,
+            84,
+            884,
+            134
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            143,
+            662,
+            159
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper studies hallucination detection in LMs, which has significant broader impacts in the field of natural language processing (NLP) and helps to address ethical considerations regarding trustworthiness and reliability. The research outcome may contribute to the development of more accurate and reliable LMs by mitigating the risks of misinformation and biased outputs and promoting accountability and trust in AI systems.",
+        "bbox": [
+            507,
+            168,
+            885,
+            313
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            324,
+            626,
+            338
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our current experiments focus on the question-answering setting. Further research is needed to assess the generalizability of the proposed framework and the accuracy of semantic equivalence checks on more complex tasks such as conversational or dialogue-based prompting. Additionally, it would be interesting to investigate the efficiency-utility trade-off: we expect increasing sample sizes to improve detection accuracy but may introduce additional cost and latency. Speeding up the implementation through parallelization is also worth exploring.",
+        "bbox": [
+            507,
+            349,
+            885,
+            542
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            568,
+            608,
+            583
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ayush Agrawal, Lester Mackey, and Adam Tauman Kalai. 2023. Do language models know when they're hallucinating references? arXiv preprint arXiv:2305.18248.",
+            "Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. 2023. Falcon-40B: an open large language model with state-of-the-art performance.",
+            "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.",
+            "Meng Cao, Yue Dong, and Jackie Chi Kit Cheung. 2022. Hallucinated but factual! inspecting the factuality of hallucinations in abstractive summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3340-3354."
+        ],
+        "bbox": [
+            509,
+            589,
+            885,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "15453",
+        "bbox": [
+            477,
+            927,
+            524,
+            941
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Angelica Chen, Jason Phang, Alicia Parrish, Vishakh Padmakumar, Chen Zhao, Samuel R Bowman, and Kyunghyun Cho. 2023. Two failures of self-consistency in the multi-step reasoning of llms. arXiv preprint arXiv:2305.14279.",
+            "Jiuhai Chen and Jonas Mueller. 2023. Quantifying uncertainty in answers from any language model via intrinsic and extrinsic confidence assessment. arXiv preprint arXiv:2308.16175.",
+            "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311.",
+            "Jeremy R Cole, Michael JQ Zhang, Daniel Gillick, Julian Martin Eisenschlos, Bhuwan Dhingra, and Jacob Eisenstein. 2023. Selectively answering ambiguous questions. arXiv preprint arXiv:2305.14613.",
+            "Souvik Das, Sougata Saha, and Rohini K Srihari. 2023. Diving deep into modes of fact hallucinations in dialogue systems. arXiv preprint arXiv:2301.04449.",
+            "Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314.",
+            "Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023. Chain-of-verification reduces hallucination in large language models. arXiv preprint arXiv:2309.11495.",
+            "Yanai Elazar, Nora Kassner, Shauli Ravfogel, Abhilasha Ravichander, Eduard Hovy, Hinrich Schütze, and Yoav Goldberg. 2021. Measuring and improving consistency in pretrained language models. Transactions of the Association for Computational Linguistics, 9:1012-1031.",
+            "Zorik Gekhman, Jonathan Herzig, Roee Aharoni, Chen Elkind, and Idan Szpektor. 2023. Trueteacher: Learning factual consistency evaluation with large language models. arXiv preprint arXiv:2305.11171.",
+            "Myeongjun Jang, Deuk Sin Kwon, and Thomas Lukasiewicz. 2022. Becel: Benchmark for consistency evaluation of language models. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3680-3696.",
+            "Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38.",
+            "Ziwei Ji, Zihan Liu, Nayeon Lee, Tiezheng Yu, Bryan Wilie, Min Zeng, and Pascale Fung. 2022. Rho: Reducing hallucination in open-domain dialogues with knowledge grounding. arXiv preprint arXiv:2212.01588."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv preprint arXiv:2302.09664.",
+            "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453-466.",
+            "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, pages 6086-6096.",
+            "Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023a. Halueval: A largescale hallucination evaluation benchmark for large language models. arXiv e-prints, pages arXiv-2305.",
+            "Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023b. Inference-time intervention: Eliciting truthful answers from a language model. arXiv preprint arXiv:2306.03341.",
+            "Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2023. Generating with confidence: Uncertainty quantification for black-box large language models. arXiv preprint arXiv:2305.19187.",
+            "Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao, Zhifang Sui, Weizhu Chen, and Bill Dolan. 2021. A token-level reference-free hallucination detection benchmark for free-form text generation. arXiv preprint arXiv:2104.08704.",
+            "Potsawee Manakul, Adian Liusie, and Mark JF Gales. 2023. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896.",
+            "Eric Mitchell, Joseph J Noh, Siyan Li, William S Armstrong, Ananth Agarwal, Patrick Liu, Chelsea Finn, and Christopher D Manning. 2022. Enhancing self-consistency and performance of pre-trained language models through natural language inference. arXiv preprint arXiv:2211.11875.",
+            "Niels Mündler, Jingxuan He, Slobodan Jenko, and Martin Vechev. 2023. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. arXiv preprint arXiv:2305.15852.",
+            "Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. 2023. The RefinedWeb dataset for Falcon LLM: outperforming curated corpora with web data, and web data only. arXiv preprint arXiv:2306.01116."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "15454",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. 2023. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813.",
+            "Derek Tam, Anisha Mascarenhas, Shiyue Zhang, Sarah Kwan, Mohit Bansal, and Colin Raffel. 2023. Evaluating the factual consistency of large language models through news summarization. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 5220-5255.",
+            "Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. 2023. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. arXiv preprint arXiv:2305.14975.",
+            "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.",
+            "Neeraj Varshney, Wenlin Yao, Hongming Zhang, Jianshu Chen, and Dong Yu. 2023. A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation. arXiv preprint arXiv:2307.03987.",
+            "Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.",
+            "Yijun Xiao and William Yang Wang. 2021. On hallucination and predictive uncertainty in conditional language generation. arXiv preprint arXiv:2103.15025.",
+            "Shiping Yang, Renliang Sun, and Xiaojun Wan. 2023. A new benchmark and reverse validation method for passage-level hallucination detection. arXiv preprint arXiv:2310.06498.",
+            "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.",
+            "Hongbin Ye, Tong Liu, Aijia Zhang, Wei Hua, and Weiqiang Jia. 2023. Cognitive mirage: A review of hallucinations in large language models. arXiv preprint arXiv:2309.06794.",
+            "Muru Zhang, Ofir Press, William Merrill, Alisa Liu, and Noah A Smith. 2023a. How language model hallucinations can snowball. arXiv preprint arXiv:2305.13534."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            916
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Shuo Zhang, Liangming Pan, Junzhou Zhao, and William Yang Wang. 2023b. Mitigating language model hallucination with interactive question-knowledge alignment. arXiv preprint arXiv:2305.13669.",
+            "Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023c. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219.",
+            "Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223.",
+            "Shen Zheng, Jie Huang, and Kevin Chen-Chuan Chang. 2023a. Why does chatgpt fall short in providing truthful answers. ArXiv preprint, abs/2304.10513.",
+            "Shen Zheng, Jie Huang, and Kevin Chen-Chuan Chang. 2023b. Why does chatgpt fall short in providing truthful answers? arXiv preprint arXiv:2304.10513.",
+            "Chunting Zhou, Graham Neubig, Jiatao Gu, Mona Diab, Paco Guzman, Luke Zettlemoyer, and Marjan Ghazvininejad. 2020. Detecting hallucinated content in conditional neural sequence generation. arXiv preprint arXiv:2011.02593."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            476
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "15455",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Factuality Assessment of LLMs",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            426,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "This section first introduces the background of LM factuality assessment and reviews the key components of existing black-box factuality assessment approaches.",
+        "bbox": [
+            112,
+            110,
+            489,
+            175
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.1 Background",
+        "text_level": 1,
+        "bbox": [
+            112,
+            187,
+            263,
+            202
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Existing studies (Manakul et al., 2023) showed that the factuality of LMs' statements can be assessed by capturing the output uncertainty. If permitted with complete white-box access, hallucinations of LMs can be detected through uncertainty metrics such as entropy, which models the uncertainty of the output distribution as  $\\mathcal{H}(\\boldsymbol{s}) = -\\sum_{i=1}^{N} p(s_i | s_{<i}) \\log p(s_i | s_{<i})$  for a sequence of  $N$  tokens  $\\boldsymbol{s} = \\{s_i\\}_{i=1}^N$ . The insight is that factual sentences are likely to be composed of high-probability tokens and thus should have less entropy, whereas non-factual statements are likely to be associated with lower probability and higher entropy.",
+        "bbox": [
+            112,
+            208,
+            489,
+            434
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Although these methods provide an accurate estimate of the LM's uncertainty, computation of such metrics involves token-level output probability distribution, which is typically not accessible for prominent LMs like ChatGPT that are only available through black-box interactions.",
+        "bbox": [
+            112,
+            434,
+            489,
+            531
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.2 Black-box Assessment via Checking Consistency in Sampled Responses",
+        "text_level": 1,
+        "bbox": [
+            112,
+            543,
+            448,
+            575
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "One natural way to extend the above white-box factuality assessment methods to restrictive black-box settings is to approximate the uncertainty metrics by sampling LM responses. Existing approaches (Manakul et al., 2023) mostly focus on assessing factuality through inspecting LM's self-consistency. The typical process of self-consistency check is as follows. First, for a given user-queried prompt  $\\pmb{x}$ , let the LM generate a candidate response  $s$  with a low (default) temperature. Next, to assess the factuality of the candidate statements, a higher temperature value is used to stochastically sample a diverse set of supporting responses  $\\{s^1, s^2, \\dots, s^m\\}$  from the LM. Finally, a hallucination score  $S \\in [0,1]$  is calculated by measuring the consistency between the candidate response and the supporting responses:  $S \\rightarrow 0$  if the supporting responses agree with the candidate response and  $S \\rightarrow 1$  if they disagree. The consistency between two responses is typically measured by checking the textual entailment or lexical simi",
+        "bbox": [
+            112,
+            581,
+            489,
+            919
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "larity using an additional pre-trained NLU model such as BERT (Manakul et al., 2023; Kuhn et al., 2023).",
+        "bbox": [
+            507,
+            84,
+            884,
+            131
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Complete Prompts",
+        "text_level": 1,
+        "bbox": [
+            509,
+            143,
+            710,
+            161
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The prompt templates that we used are provided in Table 7, specifically the question perturbations using semantically equivalent prompts and semantic equivalence checks between two QA pairs.",
+        "bbox": [
+            507,
+            168,
+            884,
+            233
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Additional Details and Discussions",
+        "text_level": 1,
+        "bbox": [
+            507,
+            244,
+            845,
+            260
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C.1 Computational Cost",
+        "text_level": 1,
+        "bbox": [
+            507,
+            271,
+            721,
+            286
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "High computational cost is a common research challenge faced in many areas of black-box LLM studies, including self-consistency (Wang et al., 2022) and self-check (Manakul et al., 2023). Compared with these existing works, our framework enables model-level parallel execution of LLM inferences (either target or verifier LM) in the form of parallel API calls, which helps to mitigate the overhead. Moreover, we may further employ more advanced prompt strategies for semantic consistency checking that condense the pairwise comparisons to a single inference call to substantially reduce the time complexity. Finally, our framework also offers the flexibility for the user to adjust the balance between computational cost and accuracy: users may choose to trade off accuracy for efficiency in scenarios where low overhead is more valued than high detection precision.",
+        "bbox": [
+            507,
+            291,
+            884,
+            580
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To achieve a more balanced trade-off, we primarily recommend two strategies: (1) Utilizing a smaller sample size, such as the number of question perturbations and self-evaluations. Our ablation experiments demonstrate that the performance of  $\\mathrm{SAC^3}$  improves with an increase in question samples, but the performance gain plateaus after exceeding five question samples. This indicates that using 2-4 question samples in practice could yield satisfactory results with minimal computational expense. (2) Implementing an optimized prompt strategy. With the original input question, we prompt the GPT model to generate multiple paraphrases in a single API call, i.e.,  $\\mathcal{O}(1)$  for question paraphrasing. Although semantic consistency checking via a pairwise method can be slightly burdensome, we can devise an advanced prompt strategy to parallelize the pairwise consistency check, i.e., through one API call instead of multiple calls, effectively reducing the inference complexity from  $\\mathcal{O}(n)$  to  $\\mathcal{O}(1)$ . These strategies empower users",
+        "bbox": [
+            507,
+            581,
+            884,
+            919
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "15456",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/5812de1bae3c9ae0a5b0e2ece2ddcbc69795dd5e79cabe5b83a8ebd37a87e563.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Input</td><td>Model</td><td>Output</td><td>Semantic Consistency</td></tr><tr><td rowspan=\"2\">SC2</td><td>Q</td><td>Target</td><td>Yes, there was a US senator that represented the state of North Dakota and whose alma mater was MIT. His name was Quentin N. Burdick.</td><td>✓</td></tr><tr><td>Q</td><td>Target</td><td>Yes, there was a US senator who represented the state of North Dakota and whose alma mater was MIT. His name was Quentin N. Burdick.</td><td>✓</td></tr><tr><td rowspan=\"2\">SAC3-Q</td><td>Q1</td><td>Target</td><td>As an AI language model, I do not have access to the most recent updates or news. However, as of my last update, there is no North Dakota US senator who graduated from MIT.</td><td>✗</td></tr><tr><td>Q2</td><td>Target</td><td>As an AI language model, I do not have access to real-time information or current events. However, as of my last update, I am not aware of any US senators from North Dakota who were also MIT graduates.</td><td>✗</td></tr><tr><td>SAC3-M</td><td>Q</td><td>Verifier</td><td>No, there has never been a US senator who represented the state of North Dakota and attended MIT.</td><td>✗</td></tr><tr><td rowspan=\"2\">SAC3-QM</td><td>Q1</td><td>Verifier</td><td>No, a North Dakota US senator did not have MIT listed as their alma mater.</td><td>✗</td></tr><tr><td>Q2</td><td>Verifier</td><td>There are no known US senators from North Dakota who are also MIT graduates.</td><td>✗</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            884,
+            305
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/98f85bfb11ef790b9fc7ab1bc7b489c493103f4ac666e1d7f9edad1d9a1a62bd.jpg",
+        "table_caption": [
+            "Table 6: Input Question (Q) and Answer (A):  $\\mathrm{Q} =$  \"Was there ever a US senator that represented the state of North Dakota and whose alma mater was MIT?\" A = \"Yes, there was a US senator who represented the state of North Dakota and whose alma mater was MIT. His name was Quentin Burdick.\" Semantically rephrased questions: Q1 = \"Did a North Dakota US senator have MIT listed as their alma mater?\" Q2 = \"Were there any US senators from North Dakota who were also MIT graduates?"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Objective</td><td>Template</td></tr><tr><td>Semantic question perturbation</td><td>For the question [QUERYED QUESTION], provide ${k} semantically equivalent questions \\nAre the following two inputs semantically equivalent? \\n[QUERYED INPUT] \\n[GENERATED INPUT]</td></tr><tr><td>Semantic equivalence check</td><td>Are the following two Question-Answering (QA) pairs semantically equivalent? Provide your best guess and the probability that it is correct (0.0 to 1.0). Given ONLY the guess (Yes or No) and probability, no other words or explanation. For example: \\nGuess: &lt;most likely guess, as short as possible; not a complete sentence, just the guess!&gt;\\nProbability: &lt;the probability between 0.0 and 1.0 that your guess is correct, without any extra commentary whatsoever; just the probability! \\nThe first QA pair is: \\nQ: ${THE QUESTION} \\nA: ${THE ANSWER} \\nThe second QA pair is: \\nQ: ${THE QUESTION} \\nA: ${THE ANSWER}</td></tr></table>",
+        "bbox": [
+            115,
+            399,
+            882,
+            565
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Prompt templates for different objectives, including semantically equivalent perturbations of input questions and semantic equivalent check of QA pairs.",
+        "bbox": [
+            112,
+            574,
+            882,
+            605
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "to optimally balance efficiency (light version) and accuracy (performance version), facilitating more informed decisions in practice.",
+        "bbox": [
+            112,
+            629,
+            487,
+            677
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C.2 Semantic Consistency Checking",
+        "text_level": 1,
+        "bbox": [
+            112,
+            688,
+            415,
+            703
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The goal of our research is to provide a flexible framework for the effective detection of hallucinations in black-box LLMs. Compared to previous approaches (Manakul et al., 2023) based on similarity metrics such as BERTScore, employing LLM for semantic consistency checking not only offers better accuracy but also eliminates the need for involving additional models. We note that our design is not dependent on any specific type of LLM and the users are allowed to choose freely for each component. Specifically, we chose GPT as the instantiation in our experiments due to its well-acclaimed ability to follow human instructions which is cru",
+        "bbox": [
+            112,
+            709,
+            489,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "cial for evaluation. In practice, the user may choose any open-sourced LLMs that have been aligned to follow instructions.",
+        "bbox": [
+            507,
+            629,
+            882,
+            675
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C.3 Data Annotations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            688,
+            697,
+            703
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "For the two classification QA tasks (Prime Number and Senator Search), we use the synthesized hallucinated answers following prior work (Zhang et al., 2023a). For the task of prime number, the factual/true answer is always \"Yes\", i.e., all the testing numbers are prime numbers. The synthesized hallucinated response is \"No\". The factual/true answer of the senator search is always \"No\" and the synthesized hallucinated answer is \"Yes\". Such an experimental setting on the binary classification tasks is realistic since we have verified that on these datasets most of the responses generated by gpt-3.5-turbo are indeed \"Yes\" or \"No\" which",
+        "bbox": [
+            507,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "15457",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "align with the synthesized response.",
+        "bbox": [
+            112,
+            84,
+            383,
+            99
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "For the generation QA tasks (HotpotQA-halu and NQ-Open-halu), we used answers generated by LLM (gpt-3.5-turbo) for experiments, which do not have pre-defined factual/non-factual labels. Therefore, we manually annotated these LLM-generated answers by comparing them with the ground truth.",
+        "bbox": [
+            112,
+            99,
+            487,
+            211
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "We would like to note that in practice, a more versatile hallucination detection approach should be able to evaluate the factuality of a sample regardless of its origin (e.g., synthesized or generated by itself / other LLMs). In our framework, this is achieved through semantically equivalent question perturbation and cross-model response consistency checking. We plan to release the annotated datasets to facilitate future research.",
+        "bbox": [
+            112,
+            212,
+            487,
+            356
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D Additional Examples",
+        "text_level": 1,
+        "bbox": [
+            112,
+            369,
+            334,
+            386
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 6 provides another illustrative example to explain our methodology in the senator search dataset.",
+        "bbox": [
+            112,
+            395,
+            487,
+            442
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "15458",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/SAC3_ Reliable Hallucination Detection in Black-Box Language Models via Semantic-aware Cross-check Consistency/4a5107be-dc68-42ed-bca9-6c454b388710_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac1d4fec2a8f6156b5d119fac652b75915c455a9b70df01ad72c035fcaf6674d
+size 778916

2023/SAC3_ Reliable Hallucination Detection in Black-Box Language Models via Semantic-aware Cross-check Consistency/full.md

@@ -0,0 +1,358 @@
+# SAC<sup>3</sup>: Reliable Hallucination Detection in Black-Box Language Models via Semantic-aware Cross-check Consistency
+
+Jiaxin Zhang $^{1}$ , Zhuohang Li $^{2}$ , Kamalika Das $^{1}$ , Bradley Malin $^{2,3}$ , Sricharan Kumar $^{1}$
+
+\(^{1}\)Intuit AI Research \(^{2}\)Vanderbilt University \(^{3}\)Vanderbilt University Medical Center \(\{jiaxin_zhang, kamalika_das, sricharan_kumar\} @intuit.com \(\{zhuohang.li, b.malin\} @\)vanderbilt.edu
+
+# Abstract
+
+Hallucination detection is a critical step toward understanding the trustworthiness of modern language models (LMs). To achieve this goal, we re-examine existing detection approaches based on the self-consistency of LMs and uncover two types of hallucinations resulting from 1) question-level and 2) model-level, which cannot be effectively identified through self-consistency check alone. Building upon this discovery, we propose a novel sampling-based method, i.e., semantic-aware cross-check consistency  $(\mathrm{SAC}^3)$  that expands on the principle of self-consistency checking. Our  $\mathrm{SAC}^3$  approach incorporates additional mechanisms to detect both question-level and model-level hallucinations by leveraging advances including semantically equivalent question perturbation and cross-model response consistency checking. Through extensive and systematic empirical analysis, we demonstrate that  $\mathrm{SAC}^3$  outperforms the state of the art in detecting both nonfactual and factual statements across multiple question-answering and open-domain generation benchmarks. $^1$
+
+# 1 Introduction
+
+Large-scale pre-trained language models (LMs) have demonstrated exceptional adaptability across a diverse array of natural language tasks that require generating open-ended responses based on user prompt comprehension (Zhao et al., 2023). However, prominent LMs like GPT (Brown et al., 2020) and PaLM (Chowdhery et al., 2022), often exhibit a tendency to produce exceedingly confident, yet erroneous, assertions commonly referred to as hallucinations. This phenomenon significantly impedes their applicability in domains where factual accuracy is of utmost importance.
+
+Hallucinations can be detected through the assistance of metrics that capture the uncertainty about
+
+the output sequences. However, these metrics require access to token-level log probabilities, which are not available in commercial black-box LMs like ChatGPT or Bard that only offer limited API access. To overcome this limitation, recent studies explore sampling-based approaches for approximating uncertainty estimation (Lin et al., 2023) through establishing a connection between confidence and self-consistency (Manakul et al., 2023; Mündler et al., 2023). The underlying premise of this principle is that LMs are more inclined to generate consistent responses when high probabilities are assigned to tokens in the answer, which, in turn, implies a level of factuality. In contrast, inconsistent responses are more likely to contain hallucinations. To operationalize this concept, current approaches are designed to sample multiple responses from the LMs for a given question and then compose a hallucination score for each sentence based on the level of response consistency.
+
+Key Observations. In this work, we investigate the relationship between the self-consistency of LMs and the occurrence of hallucinations in a diverse range of tasks. Our investigation indicates that while self-inconsistency in LMs often coincides with hallucination, self-consistency does not necessarily guarantee factual answers, as shown in Figure 1. Our findings challenge the notion that self-consistency alone can serve as a reliable indicator of veracity, as it is demonstrated that LMs can exhibit various tiers of hallucination that elude detection through self-consistency checks. One such tier is question-level hallucination, where LMs consistently generate incorrect answers in response to specific questions (e.g., Table 1). We reveal that by reformulating the questions, it is possible to mitigate such instances of hallucinations. Additionally, our work further reveals the existence of model-level hallucinations, whereby different LMs show discrepancies in their propensity for halluci
+
+![](images/cde3c1ba6e9e918f2090a7d724e82f080588c4ee0694871f9be7c53ca9787078.jpg)  
+Figure 1: Key observation: solely checking the self-consistency of LLMs is not sufficient for deciding factuality. Left: generated responses to the same question may be consistent but non-factual. Right: generated responses may be inconsistent with the original answer that is factually correct.
+
+nation. Surprisingly, we even observe cases where smaller LMs are capable of correctly answering questions for which larger LMs hallucinate. Together, these findings accentuate the need to consider model-specific characteristics when assessing the occurrence of hallucinations.
+
+Proposed Approach. Motivated by these observations, we introduce SAC<sup>3</sup>, a new sampling-based approach utilizing semantic-aware cross-check consistency to improve the detection of hallucinations in black-box LMs. An overview of our approach is provided in Fig. 2. To address question-level hallucination, we introduce a mechanism that perturbs semantically equivalent questions to evaluate the consistency of LMs' responses across variants of the same question. By examining the generated answers to these perturbed questions, we are able to identify cases where the LM consistently provides incorrect responses to a specific question, which is indicative of a question-level hallucination. Furthermore, we address model-level hallucination by introducing cross-model response consistency checking, which involves comparing the responses of different LMs to the same set of questions. By identifying discrepancies between the responses of different models, we are able to pinpoint cases where certain models exhibit hallucinations while
+
+others provide correct answers. Integrating these cross-checking extensions into our approach significantly improves its ability to detect hallucinations that go beyond self-consistency, thereby providing a more comprehensive assessment of the presence of question-level and model-level hallucinations.
+
+Effectiveness of SAC<sup>3</sup>. We demonstrate the effectiveness of SAC<sup>3</sup> in multiple classification and open-domain generation QA datasets. In the classification QA tasks, our approach achieves a high AUROC score of  $99.4\%$  and  $97.0\%$  respectively, which significantly outperforms the self-consistency baseline. In the case of open-domain generation QA tasks, our approach achieves an AUROC score of  $88.0\%$  and  $77.2\%$ , which substantially improves upon the self-consistency baseline by a large margin  $(+13.8\%, \text{and} +6.7\%$  respectively).
+
+# 2 Related Work
+
+Hallucination in LMs. The issue of hallucination in language models (LMs) has gained significant attention due to its negative impact on performance and the risks it introduces in various natural language processing (NLP) tasks, such as machine translation (Zhou et al., 2020), summarization (Cao et al., 2022), dialogue generation (Das et al., 2023), and question answering (Zhang et al., 2023a; Zheng
+
+<table><tr><td>Method</td><td>Input</td><td>Model</td><td>Output</td><td>Consistency</td><td>Factuality</td></tr><tr><td rowspan="3">Self-consistency</td><td>Q: &quot;Is 3691 a prime number?&quot;</td><td>Target</td><td>&quot;No, 3691 is not a prime number as it is divisible by 7 and 13.&quot;</td><td rowspan="3">✓</td><td rowspan="3">✗</td></tr><tr><td>Q: &quot;Is 3691 a prime number?&quot;</td><td>Target</td><td>&quot;No, 3691 is not a prime number as it is divisible by 7 and 13.&quot;</td></tr><tr><td>Q: &quot;Is 3691 a prime number?&quot;</td><td>Target</td><td>&quot;No, 3691 is not a prime number. It can be divided evenly by 7 and 19.&quot;</td></tr><tr><td rowspan="3">Cross-Q consistency</td><td>Q1: &quot;Is the number 3691 prime?&quot;</td><td>Target</td><td>&quot;Yes, the number 3691 is prime.&quot;</td><td rowspan="3">✗</td><td rowspan="3">✓</td></tr><tr><td>Q2: &quot;Are the only factors of 3691 itself and one?&quot;</td><td>Target</td><td>&quot;Yes, the only factors of 3691 are itself and one.&quot;</td></tr><tr><td>Q3: &quot;Can 3691 only be divided by 1 and 3691?&quot;</td><td>Target</td><td>&quot;Yes, 3691 is a prime number and can only be divided by 1 and 3691.&quot;</td></tr><tr><td>Cross-M consistency</td><td>Q: &quot;Is 3691 a prime number?&quot;</td><td>Verifier</td><td>&quot;Yes, 3691 is a prime number.&quot;</td><td>✗</td><td>✓</td></tr><tr><td rowspan="3">Cross-M &amp; Q consistency</td><td>Q1: &quot;Is the number 3691 prime?&quot;</td><td>Verifier</td><td>&quot;Yes.&quot;</td><td rowspan="3">✗</td><td rowspan="3">✓</td></tr><tr><td>Q2: &quot;Are the only factors of 3691 itself and one?&quot;</td><td>Verifier</td><td>&quot;Yes, the only factors of 3691 are 1 and itself.&quot;</td></tr><tr><td>Q3: &quot;Can 3691 only be divided by 1 and 3691?&quot;</td><td>Verifier</td><td>&quot;Yes, 3691 can only be divided by 1 and 3691.&quot;</td></tr></table>
+
+Table 1: An illustrative example of self-consistency, cross-question consistency, and cross-model consistency check. The original question and answer are "Is 3691 a prime number?" and "No, 3691 is not a prime number. It is divisible by 7 and 13", respectively. Each row presents a set of sampled QA pairs along with its consistency regarding the original answer, and the predicted factuality of the original answer.
+
+et al., 2023b; Dhuliawala et al., 2023). Recent survey (Ji et al., 2023; Zhang et al., 2023c; Ye et al., 2023) and evaluation benchmarks (Liu et al., 2021; Li et al., 2023a; Yang et al., 2023) have highlighted the importance of addressing this issue. Previous research has explored hallucination evaluation using confidence-based approaches (Xiao and Wang, 2021; Varshney et al., 2023; Chen and Mueller, 2023) that require access to token-level log probability (Kuhn et al., 2023; Cole et al., 2023) or supervised tuning (Agrawal et al., 2023; Li et al., 2023b) that relies on internal states of the LM. However, these methods may not be applicable when only API access to the LM is available (Agrawal et al., 2023). Another approach involves retrieving knowledge from external databases to tackle hallucinations (Ji et al., 2022; Zheng et al., 2023a; Peng et al., 2023; Zhang et al., 2023b). In contrast to these studies, our work focuses on detecting hallucinations in open-domain QA tasks using black-box LMs, without relying on external resources.
+
+Consistency Evaluation of LMs. An essential characteristic of logically valid intelligent systems is self-consistency, which entails that no two statements provided by the system contradict each other. Self-consistency is defined by Elazar et al. (2021) as the invariance of an LM's responses across various types of semantics-preserving prompt transformations. This definition is further enriched by multiple other consistency categories proposed by Jang et al. (2022). Wang et al. (2022) demonstrates that self-consistency can significantly enhance the chain of thought reasoning in LMs. Without self
+
+consistency, it becomes challenging to regard LMs as reliable or trustworthy systems. Recent studies employ self-consistency to detect hallucinations based on pretrained LMs (Manakul et al., 2023) and instruction-tuned LMs (Mündler et al., 2023). Although these methods exhibit promising accuracy on several specific tasks, potential failures (Chen et al., 2023) of self-consistency are overlooked in the current settings, as existing LMs frequently provide inconsistent responses to questions (Mitchell et al., 2022) and factual knowledge inquiries (Elazar et al., 2021; Tam et al., 2023; Gekhman et al., 2023). Our work addresses these concerns by introducing a cross-check consistency approach, aiming to bridge the gap between self-consistency and factual assessment.
+
+# 3 Self-consistency Limitations in Factuality Assessment
+
+The essential assumption of self-consistency in factuality assessment is that if the LM has the knowledge of the concept, responses sampled from its output distribution, should be similar and consistent; conversely, if the LM lacks corresponding knowledge, the sampled responses would contain hallucinated facts that are diverged and contradictory. Although this assumption may seem reasonable, it does not always hold in practice (more details are provided in the Appendix A). Specifically, we argue that solely checking the LM's self-consistency is insufficient for detecting hallucination or verifying factuality under the following two circumstances:
+
+1. LMs may produce consistently hallucinated facts. We observe that for certain questions, LMs may output consistently wrong answers. For instance, as shown in Fig. 1, when prompted with the question "Is pi smaller than 3.2?", ChatGPT consistently generates incorrect answers. In this case, where the generated responses are consistent but non-factual, solely relying on self-consistency checking of a single model would yield false negative hallucination detection results.
+
+2. Even in cases when LMs generate factual statements in their original response, the stochastic sampled responses may lack veracity. For example, the original answer (Answer 1) of ChatGPT under zero temperature is correct regarding the senator search question as shown in Fig. 1. However, when sampled with a higher temperature, ChatGPT generates multiple incorrect responses (Answer 2 and Answer  $m$ ). In this scenario, where the sampled responses are inconsistent and disagree with the original response which itself is factually correct, methods that rely solely on model self-checking would produce false positives.
+
+In summary, although the inconsistency of sampled responses has been empirically demonstrated to be correlated with hallucinated facts on certain tasks, in general, self-consistency is neither necessary nor sufficient to verify the veracity of large LMs' statements. Therefore, methods based solely on self-consistency checking may not be able to accurately detect hallucinations in complex QA and open-domain generation tasks, which motivates us to design a more reliable and robust factuality assessment method that extends this idea.
+
+# 4 SAC<sup>3</sup>: Semantic-Aware Cross-check Consistency
+
+This section describes the proposed semantic-aware cross-check consistency approach, a high-level overview of which is provided in Fig. 1. Additionally, an illustrative example of each component is presented in Table 1. Here, we walk through each component in detail.
+
+# 4.1 Stage 1: Question-level Cross-checking via Semantically Equivalent Perturbations
+
+Contrary to existing techniques that assess semantic equivalence through entailment or paraphrasing, our approach involves rephrasing the input query by generating alternative inputs that preserve semantic equivalence, i.e., semantically equivalent input per
+
+turbation. To achieve this, we leverage advances in LLM prompting. Starting with a queried input  $\pmb{x}_0$ , we acquire a set of  $k$  semantically equivalent inputs  $\{x_1, x_2, \dots, x_k\}$  through the prompt: "For the question [queried Question], provide  $k$  semantically equivalent questions".
+
+To ensure the quality of the generated inputs in this step, we further double-check the semantic equivalence between the generated inputs  $\{x_{1}, x_{2}, \ldots, x_{k}\}$  and the queried input  $x_{0}$  in a pair-wise manner using the prompt "Are the following two inputs semantically equivalent? [queried INPUT] [generated INPUT]" and filtering out the inputs that do not share the same semantic meaning as the original input. The complete prompt templates used in this work are provided in the Appendix B.
+
+# 4.2 Stage 2: Model-level Cross-check with Additional Verifier LM
+
+Let  $s_0$  denote the original response from a target LM  $\mathcal{T}$  based on a given query  $x_0$ . Our objective is to detect whether  $s_0$  is hallucinated by sampling responses from the predictive distribution of  $\mathcal{T}$ . To avoid model-level hallucination, we introduce an additional verifier LM denoted as  $\mathcal{V}$  for model-level cross-checking. We define the responses from both models as:
+
+$$
+\boldsymbol {s} _ {\mathcal {T} _ {j}} = \mathcal {T} (\boldsymbol {x} _ {j}), \boldsymbol {s} _ {\mathcal {V} _ {j}} = \mathcal {V} (\boldsymbol {x} _ {j}), j = 1, \dots , k, \tag {1}
+$$
+
+where  $k$  is the length of the generated semantically equivalent inputs  $\{x_{1}, x_{2}, \ldots, x_{k}\}$  in stage 1. To assess the factuality of  $x_{0}$ , the self-checking mechanism operates by drawing a set of  $n_{s}$  stochastic response samples from the target LM:  $\mathcal{S}_{\mathcal{T}_0} = \{s_{\mathcal{T}_0}^1, s_{\mathcal{T}_0}^2, \ldots, s_{\mathcal{T}_0}^{n_s}\}$ . Similarly, we can apply the same self-checking mechanism to the verifier LM to generate another set of  $n_m$  responses:  $\mathcal{S}_{\mathcal{V}_0} = \{s_{\mathcal{V}_0}^1, s_{\mathcal{V}_0}^2, \ldots, s_{\mathcal{V}_0}^{n_m}\}$ . To perform question-level cross-check, for each perturbed input  $x_k$ , we generate  $n_q$  sampled response sequences  $\mathcal{S}_{\mathcal{T}_k} = \{s_{\mathcal{T}_k}^1, s_{\mathcal{T}_k}^2, \ldots, s_{\mathcal{T}_k}^{n_q}\}$  from the target LM  $\mathcal{T}$  and  $n_{qm}$  sampled responses  $\mathcal{S}_{\mathcal{V}_k} = \{s_{\mathcal{V}_k}^1, s_{\mathcal{V}_k}^2, \ldots, s_{\mathcal{V}_k}^{n_{qm}}\}$  from the verifier LM  $\mathcal{V}$ .
+
+Finally, we collect the total sampled sets  $S = \{S_{\mathcal{T}_0}, S_{\mathcal{V}_0}, S_{\mathcal{T}_k}, S_{\mathcal{V}_k}\}$  by combining all samples drawn from self-checking and cross-checking, which will be used next for calculating a consistency score.
+
+![](images/f6d756e06591d2dccc325d3f6a940be0fc0e7ce85e367dda14dbb9edd9739fb7.jpg)  
+Figure 2: Overview of the proposed semantic-aware cross-check consistency  $(\mathrm{SAC}^3)$  method.
+
+# 4.3 Stage 3: Consistency Score Calculation
+
+This stage uses the generated sample sets in all previous stages to calculate a numerical consistency score that captures the question-level and model-level cross-checking paradigm.
+
+# 4.3.1 Semantic-aware Consistency Check of QA Pairs
+
+Most of the existing works mainly focus on examining the consistency of LM outputs while ignoring the effect of the inputs. However, in QA tasks, it is important to consider both inputs and outputs when measuring semantic equivalence, as the same question can be rephrased in many ways. Although the answers to these questions (e.g., "no" and "yes") may not be lexically equivalent, the QA pairs as a whole can be semantically equivalent. In light of this, we propose to check the semantic consistency of the QA pairs instead of the answer only.
+
+# 4.3.2 Self-checking Consistency (SC²) Score
+
+Let  $\mathcal{C}(\cdot, \cdot)$  denote a semantic equivalence checking operator that takes two QA pairs as inputs. The operator  $\mathcal{C}$  returns "Yes" if the two QA pairs are semantically equivalent, and "No" otherwise. This operator should be reflexive, symmetric, and transitive. We implement the checking operator using an LM by leveraging the prompt: "Are the following two Question-Answering (QA) pairs semantically equivalent? [QA PAIR 1] [QA PAIR 2]. We then map the best guess to a numerical semantic equivalent score:  $\{\text{"Yes"} \rightarrow 0.0, \text{"No"} \rightarrow 1.0\}$ . We use  $\mathcal{P}_0 = (\pmb{x}_0, \pmb{s}_0)$  to denote the original QA pair. The self-checking score  $\mathcal{Z}_{\mathrm{SC}^2}$  of the target LM  $\mathcal{T}$  can be calculated by
+
+$$
+\mathcal {Z} _ {\mathrm {S C} ^ {2}} = \frac {1}{n _ {s}} \sum_ {i = 1} ^ {n _ {s}} \mathcal {C} \left(\mathcal {P} _ {0}, \mathcal {P} _ {\mathcal {S} _ {\tau_ {0}}} ^ {i}\right), \tag {2}
+$$
+
+where  $\mathcal{P}_{\mathcal{S}_{\mathcal{T}_0}} = \{(x_0, s_{\mathcal{T}_0}^1), \dots, (x_0, s_{\mathcal{T}_0}^{n_s})\}$  represents the QA pairs generated in the self-checking scenario.
+
+# 4.3.3 Question-level Consistency (SAC $^3$ -Q) Score
+
+Besides self-checking the original question  $\pmb{x}_0$ , SAC<sup>3</sup> further assesses cross-check consistency of perturbed questions  $\{x_1, x_2, \dots, x_k\}$ . The corresponding QA pairs compose a two-dimensional matrix, where each row corresponds to a perturbed question ( $k$  in total), and each column corresponds to a sampled response ( $n_q$  in total):
+
+$$
+\mathcal {P} _ {\mathcal {S} _ {\mathcal {T} _ {j}}} ^ {i} = \left[ \begin{array}{c c c} \left(\boldsymbol {x} _ {1}, \mathcal {S} _ {\mathcal {T} _ {1}} ^ {1}\right) & \dots & \left(\boldsymbol {x} _ {1}, \mathcal {S} _ {\mathcal {T} _ {1}} ^ {n _ {q}}\right) \\ \dots & \dots & \dots \\ \left(\boldsymbol {x} _ {k}, \mathcal {S} _ {\mathcal {T} _ {k}} ^ {1}\right) & \dots & \left(\boldsymbol {x} _ {k}, \mathcal {S} _ {\mathcal {T} _ {k}} ^ {n _ {q}}\right) \end{array} \right]. \tag {3}
+$$
+
+Therefore, the question-level cross-checking consistency score  $\mathcal{Z}_{\mathrm{SAC}^3 -\mathrm{Q}}$  can be obtained by
+
+$$
+\mathcal {Z} _ {\mathrm {S A C} ^ {3} - \mathrm {Q}} = \frac {1}{n _ {q} \cdot k} \sum_ {i = 1} ^ {n _ {q}} \sum_ {j = 1} ^ {k} \mathcal {C} \left(\mathcal {P} _ {0}, \mathcal {P} _ {\mathcal {S} _ {\mathcal {T} _ {j}}} ^ {i}\right). \tag {4}
+$$
+
+# 4.3.4 Model-level Consistency (SAC $^3$ -M & SAC $^3$ -QM) Score
+
+In addition to the question-level score, a model-level cross-check score is calculated by performing cross-model checking and cross-question checking using the verifier LM  $\nu$ . Specifically, for the original question  $x_0$ , the model-level cross-checking consistency score  $\mathcal{Z}_{\mathrm{SAC}^3 -\mathsf{M}}$  is computed by
+
+$$
+\mathcal {Z} _ {\mathrm {S A C} ^ {3} - \mathrm {M}} = \frac {1}{n _ {m}} \sum_ {i = 1} ^ {n _ {m}} \mathcal {C} \left(\mathcal {P} _ {0}, \mathcal {P} _ {\mathcal {S} _ {\mathcal {V} _ {0}}} ^ {i}\right), \tag {5}
+$$
+
+where  $\mathcal{P}_{\mathcal{S}_{\mathcal{V}_0}} = \{(x_0,s_{\mathcal{V}_0}^1),\ldots ,(x_0,s_{\mathcal{V}_0}^{n_m})\}$  is the QA pairs generated by the verified LM  $\nu$
+
+The cross-question consistency score on the verifier LM is computed on the QA pairs produced by  $\nu$ :
+
+$$
+\mathcal {P} _ {\mathcal {S} _ {\mathcal {V} _ {j}}} ^ {i} = \left[ \begin{array}{c c c} \left(\boldsymbol {x} _ {1}, \mathcal {S} _ {\mathcal {V} _ {1}} ^ {1}\right) & \dots & \left(\boldsymbol {x} _ {1}, \mathcal {S} _ {\mathcal {V} _ {1}} ^ {n _ {q m}}\right) \\ \dots & \dots & \dots \\ \left(\boldsymbol {x} _ {k}, \mathcal {S} _ {\mathcal {V} _ {k}} ^ {1}\right) & \dots & \left(\boldsymbol {x} _ {k}, \mathcal {S} _ {\mathcal {V} _ {k}} ^ {n _ {q m}}\right) \end{array} \right]. \tag {6}
+$$
+
+The cross-model cross-question consistency score
+
+can thus be obtained through
+
+$$
+\mathcal {Z} _ {\mathrm {S A C} ^ {3} - \mathrm {Q M}} = \frac {1}{n _ {q m} \cdot k} \sum_ {i = 1} ^ {n _ {q m}} \sum_ {j = 1} ^ {k} \mathcal {C} \left(\mathcal {P} _ {0}, \mathcal {P} _ {\mathcal {S} _ {\mathcal {V} _ {j}}} ^ {i}\right). \tag {7}
+$$
+
+# 4.3.5 Final Score and Model Confidence
+
+The different variants of SAC<sup>3</sup> capture different aspects of the uncertainty about the original response and should complement each other. We thus consider a combination of all variants including SAC<sup>3</sup>-Q, SAC<sup>3</sup>-M, SAC<sup>3</sup>-QM as the final score:
+
+$$
+\mathcal {Z} _ {\mathrm {S A C} ^ {3} - \mathrm {a l l}} = \mathcal {Z} _ {\mathrm {S A C} ^ {3} - \mathrm {Q}} + \lambda \left(\mathcal {Z} _ {\mathrm {S A C} ^ {3} - \mathrm {M}} + \mathcal {Z} _ {\mathrm {S A C} ^ {3} - \mathrm {Q M}}\right), \tag {8}
+$$
+
+where  $\lambda$  is a weight factor for the verifier LM. Unless mentioned otherwise, we use  $\lambda = 1$  by default in our experiments. In practice, as the computation of each component is independent, they can be computed in parallel to reduce latency. The detection prediction is made by comparing the final score with a preset threshold. In addition to the computed score, we also ask the target LM to generate a verbalized confidence score (Tian et al., 2023) along with its prediction when checking the semantic equivalence of QA pairs. More discussions are offered in the Appendix C.2.
+
+# 5 Data and Annotation
+
+We evaluate our hallucination detection approach on two categories of QA tasks, namely, classification QA and generation QA, with each category containing two datasets. Following prior work (Zhang et al., 2023a), we use the following two binary classification datasets for evaluation on the classification QA task:
+
+- Prime number: this dataset contains 500 questions that query the primality of a randomly chosen prime number between 1,000 and 20,000, where the factual answer is always "Yes". The synthesized hallucinated answers are "No, it is not a prime number".  
+- Senator search: the dataset consists of 500 questions that follow the following template: "Was there ever a US senator that represented the state of [US STATE NAME] and whose alma mater was [US COLLEGE NAME]?" The factual answer is always "No". We also generate hallucinated answers: "Yes, there was a US senator that represented the state of [US STATE NAME] and whose alma mater was [US COLLEGE NAME].".
+
+As for the generation QA tasks, we take questions from the following two open-domain QA datasets and generate answers using LLMs. Then we manually annotate the factuality of the answers following previous work (Li et al., 2023a).
+
+- HotpotQA-halu: We randomly sample 250 examples from the training set of HotpotQA (Yang et al., 2018) and generate hallucinated answers drawn from gpt-3.5-turbo. Then we manually annotate the answers by comparing the ground truth and knowledge.  
+- NQ-open-halu: Natural Questions (NQ)-open (Lee et al., 2019) is a more challenging open domain QA benchmark (Kwiatkowski et al., 2019). We use the same setting as HotpotQA-halu to create a small-scale dataset that consists of 250 non-factual and factual examples with manual annotations.
+
+Please find more relevant details about data annotations in the Appendix C.3.
+
+# 6 Experiments
+
+# 6.1 Experimental Setup
+
+Evaluation Models. We use gpt-3.5-turbo from OpenAI as the target LM for our experiment. The verifier LM is chosen from the following two models: (1) Falcon-7b-instruct (Al-mazrouei et al., 2023): an open-source causal decoder-only model built by TII that is trained on 1,500B tokens of RefinedWeb (Penedo et al., 2023) and further enhanced using the curated corpora; and (2) Guanaco-33b: an open-source instruction-following models through QLoRA (Dettmers et al., 2023) tuning of LLaMA (Touvron et al., 2023) base model on the OASST1 dataset.
+
+Implementation Details. The evaluation is conducted using Azure OpenAI API. When performing semantic perturbations and consistency checking, we set the temperature to 0.0 to get deterministic high-quality outputs. Given a specific input query, we generate  $k = 10$  semantically equivalent inputs using the prompt described in Section 4.1. For the self-checking-based method  $\mathsf{SC}^2$ , we follow prior work (Manakul et al., 2023) to set the temperature to 1.0 and generate  $n_s = 10$  stochastic samples. For  $\mathrm{SAC}^3 -\mathrm{Q}$  and  $\mathrm{SAC}^3 -\mathrm{QM}$ , we set  $n_q = n_{qm} = 1$  to reduce computational cost. To further reduce the inference cost, we set  $n_m = 1$  by default and
+
+combine SAC $^{3}$ -M with SAC $^{3}$ -QM to report the model-level results. We use hallucination detection accuracy and area under the ROC curve (AUROC) to evaluate the performance. In addition to the estimated hallucination score, we also show the verbalized probabilities (Tian et al., 2023) from the target LM for comparison. We execute all experiments on 8 NVIDIA V100 32G GPUs.
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Prime number</td><td colspan="2">Senator search</td></tr><tr><td>Score</td><td>Confidence</td><td>Score</td><td>Confidence</td></tr><tr><td>SC2(gpt-3.5-turbo)</td><td>65.9</td><td>67.5</td><td>56.1</td><td>53.1</td></tr><tr><td>SAC3-Q(gpt-3.5-turbo)</td><td>99.4</td><td>99.7</td><td>99.7</td><td>99.7</td></tr></table>
+
+# 6.2 Evaluation Results
+
+# 6.2.1 Classification QA
+
+Balanced Dataset. We first experiment on balanced datasets with  $50\%$  hallucinated samples and  $50\%$  factual samples. Table 2 compares the detection performance of  $\mathsf{SC^2}$  and  $\mathrm{SAC}^3 -\mathrm{Q}$  in terms of AUROC and verbalized confidence score. We observe that self-checking  $(\mathrm{SC^2})$  performs poorly on both datasets, with a low AUROC of  $65.9\%$  and  $56.1\%$ , respectively. Our question-level cross-checking  $(\mathrm{SAC}^3 -\mathrm{Q})$  significantly outperforms the  $\mathrm{SC^2}$  baseline achieving  $>99\%$  AUROC on both datasets and is in line with the verbalized confidence score, confirming the effectiveness of cross-checking.
+
+Unbalanced Dataset. We further evaluate our method in a more challenging scenario where the dataset only contains hallucinated samples. Table 3 presents the accuracy of detecting hallucinated samples using a preset threshold of 0.5. In this case, the performance of self-check drops significantly to  $48.2\%$  and  $29.6\%$  respectively.  $\mathrm{SAC}^3 -\mathrm{Q}$  still outperforms  $\mathrm{SC}^2$  by a large margin. The model-level cross-check with verifier LMs performs well in the prime number dataset but fails to accurately detect hallucination in the senator search dataset. This is because both verifier LMs refuse to answer a large portion of the questions on this dataset due to a lack of sufficient information. By combining the target LM,  $\mathrm{SAC}^3$ -a11 with Guanaco-33b achieves the highest detection accuracy compared to other baselines  $\mathrm{SC}^2 (+51.2\%)$ ,  $\mathrm{SAC}^3 -\mathrm{Q} (+6.2\%)$ , and  $\mathrm{SAC}^3 -\mathrm{QM} (+5.0\%)$ .
+
+Table 2: AUROC on classification QA tasks with  $50\%$  hallucinated samples and  $50\%$  factual samples.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Prime number</td><td colspan="2">Senator search</td></tr><tr><td>Score</td><td>Confidence</td><td>Score</td><td>Confidence</td></tr><tr><td>SC2(gpt-3.5-turbo)</td><td>48.2</td><td>51.0</td><td>29.6</td><td>30.6</td></tr><tr><td>SAC3-Q(gpt-3.5-turbo)</td><td>93.2</td><td>96.4</td><td>97.0</td><td>97.4</td></tr><tr><td>SAC3-QM (Falcon-7b)</td><td>89.8</td><td>91.2</td><td>21.0</td><td>21.8</td></tr><tr><td>SAC3-QM (Guanaco-33b)</td><td>94.4</td><td>96.3</td><td>45.6</td><td>46.2</td></tr><tr><td>SAC3-all (Falcon-7b)</td><td>97.8</td><td>98.0</td><td>84.6</td><td>85.6</td></tr><tr><td>SAC3-all (Guanaco-33b)</td><td>99.4</td><td>99.4</td><td>85.6</td><td>86.7</td></tr></table>
+
+Table 3: Accuracy on classification QA tasks with  $100\%$  hallucinated samples (with the threshold set to 0.5).
+
+Impact of Threshold. Since the choice of threshold has a significant impact on the detection accuracy in the case where the dataset only contains positive (hallucinated) samples, we further experiment with different thresholds and present the results in Fig. 3. We observe that  $\mathrm{SAC}^3$ -Q and  $\mathrm{SAC}^3$ -a11 with Guanaco-33b are more robust against large threshold values and outperform  $\mathrm{SC}^2$  in most cases.
+
+![](images/cb4fcdc3651af384934237e97e162eff29423d4337e3513da6b0c376c5f628b2.jpg)
+
+![](images/820cd6e634bb353c1cd856328097e6327074baccb6ceda00df50174331febc01.jpg)  
+Figure 3: Impact of threshold on detection accuracy.
+
+![](images/8065ea4ffe0083104ee89f4f65ac9292fb538e4c4e5913d2ab0bab281503a2fd.jpg)
+
+Why Does Self-checking Fail? To further understand why self-checking methods fail to detect some hallucinated responses, we visualize the distribution of consistency scores using histogram plots in Fig. 4. We observe that for  $\mathsf{SC}^2$ , a significant portion of hallucinated samples received highly consistent predictions. In other words, the target LM made consistently wrong predictions due to a lack of question and model diversity, which aligns with our analysis in Section 3. On the other hand, benefiting from the semantically equivalent question perturbation,  $\mathrm{SAC}^3 -Q$ ’s scores are more spread out in the inconsistent region, which helps to improve the effectiveness of detecting hallucinations by choosing a proper threshold.
+
+# 6.2.2 Open-domain Generation QA
+
+Compared to the classification QA tasks, detecting hallucinations in open-domain generation QA
+
+![](images/7471c5546dd0753b1290c087ea06bc1198a17e4d4ba20a56b83114986c4d020f.jpg)  
+Figure 4: Histogram of hallucination score.
+
+tasks is more challenging. As shown in Table 4,  $\mathrm{SAC}^3 -\mathrm{Q}$  exhibits better AUROC than  $\mathrm{SC}^2$ $(+7\%)$  in both datasets. Compared to  $\mathrm{SAC}^3 -\mathrm{Q}$ ,  $\mathrm{SAC}^3 -\mathrm{QM}$  shows  $6.7\%$  improvement in the HotpotQA-halu dataset but is slightly worse in the NQ open dataset.  $\mathrm{SAC}^3$ -all leverages the advantages of question-level and model-level cross-checking and achieves consistently good performance in both datasets.
+
+<table><tr><td rowspan="2" colspan="2">Method</td><td colspan="2">HotpotQA-halu</td><td colspan="2">NQ-open-halu</td></tr><tr><td>Score</td><td>Confidence</td><td>Score</td><td>Confidence</td></tr><tr><td colspan="2">SC2(gpt-3.5-turbo)</td><td>74.2</td><td>77.0</td><td>70.5</td><td>72.7</td></tr><tr><td colspan="2">SAC3-Q(gpt-3.5-turbo)</td><td>81.3</td><td>81.4</td><td>77.2</td><td>78.5</td></tr><tr><td colspan="2">SAC3-QM (Falcon-7b)</td><td>83.0</td><td>79.5</td><td>67.5</td><td>62.0</td></tr><tr><td colspan="2">SAC3-QM (Guanaco-33b)</td><td>88.0</td><td>85.2</td><td>72.7</td><td>72.7</td></tr><tr><td colspan="2">SAC3-all (Falcon-7b)</td><td>84.5</td><td>84.5</td><td>77.1</td><td>77.2</td></tr><tr><td colspan="2">SAC3-all (Guanaco-33b)</td><td>87.0</td><td>86.8</td><td>77.2</td><td>77.8</td></tr></table>
+
+Table 4: AUROC on open-domain generation QA tasks.
+
+Effect of Verifier LM Weight. In our previous experiments, we assigned equal importance to the consistency score computed by the target and verifier LM. However, typically, the target LM and the verifier LM have different architectures and scales such that the user may have different levels of trust in their output truthfulness. This difference in trust can be incorporated by introducing a weight  $\lambda$  to the consistency score produced by the verifier LM. For instance, if the goal is to detect hallucination in a specialized domain and the verifier LM is a domain-specific model developed for this domain, we can assign a large weight to its scores (e.g.,  $\lambda > 1.0$ ). In the general case, where the verifier LM is a small-sized open-source model, we can apply a small weight value (e.g.,  $\lambda < 1.0$ ) to discount the influence of the verifier LM in the final score. Fig. 5 visualizes the effect of various weight factors on the generation QA datasets. We observe that SAC<sup>3</sup>-all with a higher weight would result in a larger advantage over SAC<sup>3</sup>-Q in
+
+the HotpotQA-halu task, where the verifier LM outperforms the target LM. On the contrary, in the NQ open task where the target LM shows competitive performance, a smaller weight would yield better results.
+
+![](images/632cf857050e3c680125f3242f79e5862c24e71c180d80dd33329837d1a060a8.jpg)
+
+![](images/b0a467b8ed3b0981a44326b99a869ea9e6f029687c6fce8000e5a4d5038bdd4f.jpg)  
+Figure 5: Effect of verifier LM weight on AUROC.
+
+Effect of the Number of Perturbed Questions. The performance of sampling-based methods is expected to improve as the sample size increases, at the cost of higher latency and computational cost. In Fig. 6, we study this trade-off by varying the number of perturbed questions  $k$  from 2 to 10. We observe that the performance of SAC³ increases as more question samples are used but the performance gain gradually diminishes after using more than 5 question samples. This suggests that in practice we could use 2-4 question samples to achieve reasonably good performance at a low computational cost.
+
+![](images/336aa453fc86d1c725fd409a10cd694bff47b5f4bbd3dbec3e273b9a91f77460.jpg)  
+Figure 6: Performance of SAC<sup>3</sup> with varying number of perturbed questions.
+
+Effect of the Model Type. Our proposed SAC<sup>3</sup> framework does not restrict the type of LLM employed and can be naturally extended to various
+
+types of target LLMs. To verify this, in addition to GPT-3.5 (gpt-3.5-turbo), we conduct experiments using GPT-4 (gpt-4) and PaLM 2 (chat-bison) on the considered four datasets in the setting of the balanced dataset. The experimental results of comparing the proposed  $\mathrm{SAC}^3 -\mathrm{Q}$  with the  $\mathrm{SC}^2$  baseline are summarized in Table 5. We observe that the proposed  $\mathrm{SAC}^3 -\mathrm{Q}$  consistently outperforms the  $\mathrm{SC}^2$  baseline across all LLM variants.
+
+<table><tr><td>Method</td><td>Prime Number</td><td>Senator Search</td><td>HotpotQA -halu</td><td>NQ-open -halu</td></tr><tr><td>SC2(gpt-3.5-turbo)</td><td>48.2</td><td>29.6</td><td>74.2</td><td>70.5</td></tr><tr><td>SC2(gpt-4)</td><td>38.3</td><td>18.4</td><td>79.7</td><td>76.3</td></tr><tr><td>SC2(chat-bison)</td><td>26.9</td><td>19.2</td><td>75.8</td><td>67.9</td></tr><tr><td>SAC3-Q (gpt-3.5-turbo)</td><td>93.2</td><td>97.0</td><td>81.3</td><td>77.2</td></tr><tr><td>SAC3-Q (gpt-4)</td><td>91.1</td><td>61.6</td><td>87.2</td><td>82.9</td></tr><tr><td>SAC3-Q (chat-bison)</td><td>90.3</td><td>66.3</td><td>82.8</td><td>72.7</td></tr></table>
+
+Table 5: Accuracy of different LLMs (GPT-3.5, GPT-4, and PaLM 2) on classification and generation QA tasks.
+
+Computational Cost. We monitor the computational cost of our approach based on the number of model evaluations consumed by OpenAI API and open-source LLMs inference. Assuming the number of samples equals the number of perturbed questions, i.e.,  $n_s = n_m = n_q = n_{qm}$ , the cost of  $\mathsf{SC^2}$  is  $n_s$  API calls, and our  $\mathsf{SAC^3}$ -all needs  $n_s$  target LM calls plus  $2 \times n_s$  verifier LM calls. Beyond the model evaluations,  $\mathsf{SAC^3}$  may have additional costs from question perturbations and semantic equivalence checking via prompting. Additional discussions can be found in the Appendix C.1.
+
+# 7 Conclusion and Discussion
+
+We investigate the relationship between the self-consistency of LM and the factuality of the response and propose SAC<sup>3</sup> as a robust hallucination detection approach for black-box LMs. Through extensive empirical analysis, our work highlights several findings. First, self-consistency checking alone is insufficient to effectively detect question-level and model-level hallucinations, where LMs generate consistently wrong responses to certain questions. Second, cross-checking between semantically equivalent questions can reduce the occurrence of persistent hallucinations, potentially by reducing question-level ambiguity. Third, there exists a model-level disparity in hallucinations, which we attribute to the inherent differences in LM capabilities originating from different training procedures and data. Thus, the verifier LM can be selected
+
+according to specific tasks to maximize detection accuracy. We believe that our work is an important step towards building reliable LLMs.
+
+# Ethics Statement
+
+This paper studies hallucination detection in LMs, which has significant broader impacts in the field of natural language processing (NLP) and helps to address ethical considerations regarding trustworthiness and reliability. The research outcome may contribute to the development of more accurate and reliable LMs by mitigating the risks of misinformation and biased outputs and promoting accountability and trust in AI systems.
+
+# Limitations
+
+Our current experiments focus on the question-answering setting. Further research is needed to assess the generalizability of the proposed framework and the accuracy of semantic equivalence checks on more complex tasks such as conversational or dialogue-based prompting. Additionally, it would be interesting to investigate the efficiency-utility trade-off: we expect increasing sample sizes to improve detection accuracy but may introduce additional cost and latency. Speeding up the implementation through parallelization is also worth exploring.
+
+# References
+
+Ayush Agrawal, Lester Mackey, and Adam Tauman Kalai. 2023. Do language models know when they're hallucinating references? arXiv preprint arXiv:2305.18248.  
+Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. 2023. Falcon-40B: an open large language model with state-of-the-art performance.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Meng Cao, Yue Dong, and Jackie Chi Kit Cheung. 2022. Hallucinated but factual! inspecting the factuality of hallucinations in abstractive summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3340-3354.
+
+Angelica Chen, Jason Phang, Alicia Parrish, Vishakh Padmakumar, Chen Zhao, Samuel R Bowman, and Kyunghyun Cho. 2023. Two failures of self-consistency in the multi-step reasoning of llms. arXiv preprint arXiv:2305.14279.  
+Jiuhai Chen and Jonas Mueller. 2023. Quantifying uncertainty in answers from any language model via intrinsic and extrinsic confidence assessment. arXiv preprint arXiv:2308.16175.  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311.  
+Jeremy R Cole, Michael JQ Zhang, Daniel Gillick, Julian Martin Eisenschlos, Bhuwan Dhingra, and Jacob Eisenstein. 2023. Selectively answering ambiguous questions. arXiv preprint arXiv:2305.14613.  
+Souvik Das, Sougata Saha, and Rohini K Srihari. 2023. Diving deep into modes of fact hallucinations in dialogue systems. arXiv preprint arXiv:2301.04449.  
+Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314.  
+Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023. Chain-of-verification reduces hallucination in large language models. arXiv preprint arXiv:2309.11495.  
+Yanai Elazar, Nora Kassner, Shauli Ravfogel, Abhilasha Ravichander, Eduard Hovy, Hinrich Schütze, and Yoav Goldberg. 2021. Measuring and improving consistency in pretrained language models. Transactions of the Association for Computational Linguistics, 9:1012-1031.  
+Zorik Gekhman, Jonathan Herzig, Roee Aharoni, Chen Elkind, and Idan Szpektor. 2023. Trueteacher: Learning factual consistency evaluation with large language models. arXiv preprint arXiv:2305.11171.  
+Myeongjun Jang, Deuk Sin Kwon, and Thomas Lukasiewicz. 2022. Becel: Benchmark for consistency evaluation of language models. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3680-3696.  
+Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38.  
+Ziwei Ji, Zihan Liu, Nayeon Lee, Tiezheng Yu, Bryan Wilie, Min Zeng, and Pascale Fung. 2022. Rho: Reducing hallucination in open-domain dialogues with knowledge grounding. arXiv preprint arXiv:2212.01588.
+
+Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv preprint arXiv:2302.09664.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453-466.  
+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, pages 6086-6096.  
+Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023a. Halueval: A largescale hallucination evaluation benchmark for large language models. arXiv e-prints, pages arXiv-2305.  
+Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023b. Inference-time intervention: Eliciting truthful answers from a language model. arXiv preprint arXiv:2306.03341.  
+Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2023. Generating with confidence: Uncertainty quantification for black-box large language models. arXiv preprint arXiv:2305.19187.  
+Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao, Zhifang Sui, Weizhu Chen, and Bill Dolan. 2021. A token-level reference-free hallucination detection benchmark for free-form text generation. arXiv preprint arXiv:2104.08704.  
+Potsawee Manakul, Adian Liusie, and Mark JF Gales. 2023. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896.  
+Eric Mitchell, Joseph J Noh, Siyan Li, William S Armstrong, Ananth Agarwal, Patrick Liu, Chelsea Finn, and Christopher D Manning. 2022. Enhancing self-consistency and performance of pre-trained language models through natural language inference. arXiv preprint arXiv:2211.11875.  
+Niels Mündler, Jingxuan He, Slobodan Jenko, and Martin Vechev. 2023. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. arXiv preprint arXiv:2305.15852.  
+Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. 2023. The RefinedWeb dataset for Falcon LLM: outperforming curated corpora with web data, and web data only. arXiv preprint arXiv:2306.01116.
+
+Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. 2023. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813.  
+Derek Tam, Anisha Mascarenhas, Shiyue Zhang, Sarah Kwan, Mohit Bansal, and Colin Raffel. 2023. Evaluating the factual consistency of large language models through news summarization. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 5220-5255.  
+Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. 2023. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. arXiv preprint arXiv:2305.14975.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Neeraj Varshney, Wenlin Yao, Hongming Zhang, Jianshu Chen, and Dong Yu. 2023. A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation. arXiv preprint arXiv:2307.03987.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.  
+Yijun Xiao and William Yang Wang. 2021. On hallucination and predictive uncertainty in conditional language generation. arXiv preprint arXiv:2103.15025.  
+Shiping Yang, Renliang Sun, and Xiaojun Wan. 2023. A new benchmark and reverse validation method for passage-level hallucination detection. arXiv preprint arXiv:2310.06498.  
+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.  
+Hongbin Ye, Tong Liu, Aijia Zhang, Wei Hua, and Weiqiang Jia. 2023. Cognitive mirage: A review of hallucinations in large language models. arXiv preprint arXiv:2309.06794.  
+Muru Zhang, Ofir Press, William Merrill, Alisa Liu, and Noah A Smith. 2023a. How language model hallucinations can snowball. arXiv preprint arXiv:2305.13534.
+
+Shuo Zhang, Liangming Pan, Junzhou Zhao, and William Yang Wang. 2023b. Mitigating language model hallucination with interactive question-knowledge alignment. arXiv preprint arXiv:2305.13669.  
+Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023c. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219.  
+Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223.  
+Shen Zheng, Jie Huang, and Kevin Chen-Chuan Chang. 2023a. Why does chatgpt fall short in providing truthful answers. ArXiv preprint, abs/2304.10513.  
+Shen Zheng, Jie Huang, and Kevin Chen-Chuan Chang. 2023b. Why does chatgpt fall short in providing truthful answers? arXiv preprint arXiv:2304.10513.  
+Chunting Zhou, Graham Neubig, Jiatao Gu, Mona Diab, Paco Guzman, Luke Zettlemoyer, and Marjan Ghazvininejad. 2020. Detecting hallucinated content in conditional neural sequence generation. arXiv preprint arXiv:2011.02593.
+
+# A Factuality Assessment of LLMs
+
+This section first introduces the background of LM factuality assessment and reviews the key components of existing black-box factuality assessment approaches.
+
+# A.1 Background
+
+Existing studies (Manakul et al., 2023) showed that the factuality of LMs' statements can be assessed by capturing the output uncertainty. If permitted with complete white-box access, hallucinations of LMs can be detected through uncertainty metrics such as entropy, which models the uncertainty of the output distribution as  $\mathcal{H}(\boldsymbol{s}) = -\sum_{i=1}^{N} p(s_i | s_{<i}) \log p(s_i | s_{<i})$  for a sequence of  $N$  tokens  $\boldsymbol{s} = \{s_i\}_{i=1}^N$ . The insight is that factual sentences are likely to be composed of high-probability tokens and thus should have less entropy, whereas non-factual statements are likely to be associated with lower probability and higher entropy.
+
+Although these methods provide an accurate estimate of the LM's uncertainty, computation of such metrics involves token-level output probability distribution, which is typically not accessible for prominent LMs like ChatGPT that are only available through black-box interactions.
+
+# A.2 Black-box Assessment via Checking Consistency in Sampled Responses
+
+One natural way to extend the above white-box factuality assessment methods to restrictive black-box settings is to approximate the uncertainty metrics by sampling LM responses. Existing approaches (Manakul et al., 2023) mostly focus on assessing factuality through inspecting LM's self-consistency. The typical process of self-consistency check is as follows. First, for a given user-queried prompt  $\pmb{x}$ , let the LM generate a candidate response  $s$  with a low (default) temperature. Next, to assess the factuality of the candidate statements, a higher temperature value is used to stochastically sample a diverse set of supporting responses  $\{s^1, s^2, \dots, s^m\}$  from the LM. Finally, a hallucination score  $S \in [0,1]$  is calculated by measuring the consistency between the candidate response and the supporting responses:  $S \rightarrow 0$  if the supporting responses agree with the candidate response and  $S \rightarrow 1$  if they disagree. The consistency between two responses is typically measured by checking the textual entailment or lexical simi
+
+larity using an additional pre-trained NLU model such as BERT (Manakul et al., 2023; Kuhn et al., 2023).
+
+# B Complete Prompts
+
+The prompt templates that we used are provided in Table 7, specifically the question perturbations using semantically equivalent prompts and semantic equivalence checks between two QA pairs.
+
+# C Additional Details and Discussions
+
+# C.1 Computational Cost
+
+High computational cost is a common research challenge faced in many areas of black-box LLM studies, including self-consistency (Wang et al., 2022) and self-check (Manakul et al., 2023). Compared with these existing works, our framework enables model-level parallel execution of LLM inferences (either target or verifier LM) in the form of parallel API calls, which helps to mitigate the overhead. Moreover, we may further employ more advanced prompt strategies for semantic consistency checking that condense the pairwise comparisons to a single inference call to substantially reduce the time complexity. Finally, our framework also offers the flexibility for the user to adjust the balance between computational cost and accuracy: users may choose to trade off accuracy for efficiency in scenarios where low overhead is more valued than high detection precision.
+
+To achieve a more balanced trade-off, we primarily recommend two strategies: (1) Utilizing a smaller sample size, such as the number of question perturbations and self-evaluations. Our ablation experiments demonstrate that the performance of  $\mathrm{SAC^3}$  improves with an increase in question samples, but the performance gain plateaus after exceeding five question samples. This indicates that using 2-4 question samples in practice could yield satisfactory results with minimal computational expense. (2) Implementing an optimized prompt strategy. With the original input question, we prompt the GPT model to generate multiple paraphrases in a single API call, i.e.,  $\mathcal{O}(1)$  for question paraphrasing. Although semantic consistency checking via a pairwise method can be slightly burdensome, we can devise an advanced prompt strategy to parallelize the pairwise consistency check, i.e., through one API call instead of multiple calls, effectively reducing the inference complexity from  $\mathcal{O}(n)$  to  $\mathcal{O}(1)$ . These strategies empower users
+
+<table><tr><td>Method</td><td>Input</td><td>Model</td><td>Output</td><td>Semantic Consistency</td></tr><tr><td rowspan="2">SC2</td><td>Q</td><td>Target</td><td>Yes, there was a US senator that represented the state of North Dakota and whose alma mater was MIT. His name was Quentin N. Burdick.</td><td>✓</td></tr><tr><td>Q</td><td>Target</td><td>Yes, there was a US senator who represented the state of North Dakota and whose alma mater was MIT. His name was Quentin N. Burdick.</td><td>✓</td></tr><tr><td rowspan="2">SAC3-Q</td><td>Q1</td><td>Target</td><td>As an AI language model, I do not have access to the most recent updates or news. However, as of my last update, there is no North Dakota US senator who graduated from MIT.</td><td>✗</td></tr><tr><td>Q2</td><td>Target</td><td>As an AI language model, I do not have access to real-time information or current events. However, as of my last update, I am not aware of any US senators from North Dakota who were also MIT graduates.</td><td>✗</td></tr><tr><td>SAC3-M</td><td>Q</td><td>Verifier</td><td>No, there has never been a US senator who represented the state of North Dakota and attended MIT.</td><td>✗</td></tr><tr><td rowspan="2">SAC3-QM</td><td>Q1</td><td>Verifier</td><td>No, a North Dakota US senator did not have MIT listed as their alma mater.</td><td>✗</td></tr><tr><td>Q2</td><td>Verifier</td><td>There are no known US senators from North Dakota who are also MIT graduates.</td><td>✗</td></tr></table>
+
+Table 6: Input Question (Q) and Answer (A):  $\mathrm{Q} =$  "Was there ever a US senator that represented the state of North Dakota and whose alma mater was MIT?" A = "Yes, there was a US senator who represented the state of North Dakota and whose alma mater was MIT. His name was Quentin Burdick." Semantically rephrased questions: Q1 = "Did a North Dakota US senator have MIT listed as their alma mater?" Q2 = "Were there any US senators from North Dakota who were also MIT graduates?  
+
+<table><tr><td>Objective</td><td>Template</td></tr><tr><td>Semantic question perturbation</td><td>For the question [QUERYED QUESTION], provide ${k} semantically equivalent questions \nAre the following two inputs semantically equivalent? \n[QUERYED INPUT] \n[GENERATED INPUT]</td></tr><tr><td>Semantic equivalence check</td><td>Are the following two Question-Answering (QA) pairs semantically equivalent? Provide your best guess and the probability that it is correct (0.0 to 1.0). Given ONLY the guess (Yes or No) and probability, no other words or explanation. For example: \nGuess: &lt;most likely guess, as short as possible; not a complete sentence, just the guess!&gt;\nProbability: &lt;the probability between 0.0 and 1.0 that your guess is correct, without any extra commentary whatsoever; just the probability! \nThe first QA pair is: \nQ: ${THE QUESTION} \nA: ${THE ANSWER} \nThe second QA pair is: \nQ: ${THE QUESTION} \nA: ${THE ANSWER}</td></tr></table>
+
+Table 7: Prompt templates for different objectives, including semantically equivalent perturbations of input questions and semantic equivalent check of QA pairs.
+
+to optimally balance efficiency (light version) and accuracy (performance version), facilitating more informed decisions in practice.
+
+# C.2 Semantic Consistency Checking
+
+The goal of our research is to provide a flexible framework for the effective detection of hallucinations in black-box LLMs. Compared to previous approaches (Manakul et al., 2023) based on similarity metrics such as BERTScore, employing LLM for semantic consistency checking not only offers better accuracy but also eliminates the need for involving additional models. We note that our design is not dependent on any specific type of LLM and the users are allowed to choose freely for each component. Specifically, we chose GPT as the instantiation in our experiments due to its well-acclaimed ability to follow human instructions which is cru
+
+cial for evaluation. In practice, the user may choose any open-sourced LLMs that have been aligned to follow instructions.
+
+# C.3 Data Annotations
+
+For the two classification QA tasks (Prime Number and Senator Search), we use the synthesized hallucinated answers following prior work (Zhang et al., 2023a). For the task of prime number, the factual/true answer is always "Yes", i.e., all the testing numbers are prime numbers. The synthesized hallucinated response is "No". The factual/true answer of the senator search is always "No" and the synthesized hallucinated answer is "Yes". Such an experimental setting on the binary classification tasks is realistic since we have verified that on these datasets most of the responses generated by gpt-3.5-turbo are indeed "Yes" or "No" which
+
+align with the synthesized response.
+
+For the generation QA tasks (HotpotQA-halu and NQ-Open-halu), we used answers generated by LLM (gpt-3.5-turbo) for experiments, which do not have pre-defined factual/non-factual labels. Therefore, we manually annotated these LLM-generated answers by comparing them with the ground truth.
+
+We would like to note that in practice, a more versatile hallucination detection approach should be able to evaluate the factuality of a sample regardless of its origin (e.g., synthesized or generated by itself / other LLMs). In our framework, this is achieved through semantically equivalent question perturbation and cross-model response consistency checking. We plan to release the annotated datasets to facilitate future research.
+
+# D Additional Examples
+
+Table 6 provides another illustrative example to explain our methodology in the senator search dataset.

2023/SAC3_ Reliable Hallucination Detection in Black-Box Language Models via Semantic-aware Cross-check Consistency/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0c8a359f7c12184d7b3ce6b3135276e4ec9de4ae88ab1f966950785b5c2dce1
+size 732443

2023/SDOH-NLI_ a Dataset for Inferring Social Determinants of Health from Clinical Notes/a4f22374-5104-49b6-a9f1-43a434ee9541_content_list.json

@@ -0,0 +1,1476 @@
+[
+    {
+        "type": "text",
+        "text": "SDOH-NLI: a Dataset for Inferring Social Determinants of Health from Clinical Notes",
+        "text_level": 1,
+        "bbox": [
+            205,
+            89,
+            794,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Adam D. Lelkes<sup>1*</sup>, Eric Loreaux<sup>2†</sup>, Tal Schuster<sup>1</sup>, Ming-Jun Chen<sup>1‡</sup>, Alvin Rajkomar<sup>1‡</sup>",
+        "bbox": [
+            137,
+            153,
+            865,
+            173
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>Google Research, <sup>2</sup>Curai Health",
+        "bbox": [
+            364,
+            187,
+            638,
+            205
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            342,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Social and behavioral determinants of health (SDOH) play a significant role in shaping health outcomes, and extracting these determinants from clinical notes is a first step to help healthcare providers systematically identify opportunities to provide appropriate care and address disparities. Progress on using NLP methods for this task has been hindered by the lack of high-quality publicly available labeled data, largely due to the privacy and regulatory constraints on the use of real patients' information. This paper introduces a new dataset, SDOH-NLI, that is based on publicly available notes and which we release publicly. We formulate SDOH extraction as a natural language inference (NLI) task, and provide binary textual entailment labels obtained from human raters for a cross product of a set of social history snippets as premises and SDOH factors as hypotheses. Our dataset differs from standard NLI benchmarks in that our premises and hypotheses are obtained independently. We evaluate both \"off-the-shelf\" entailment models as well as models fine-tuned on our data, and highlight the ways in which our dataset appears more challenging than commonly used NLI datasets.",
+        "bbox": [
+            142,
+            273,
+            460,
+            657
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            664,
+            260,
+            678
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "There has been growing recognition that social and behavioral determinants of health (SDOH) play a significant role in shaping health outcomes for individuals and populations. The ability to accurately identify and extract social and behavioral determinants of health from clinical notes can provide valuable insights that can enable healthcare providers to better understand and address the underlying determinants of health that contribute to poor health outcomes and health disparities.",
+        "bbox": [
+            112,
+            688,
+            490,
+            848
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Social determinants of health are frequently recorded in clinical notes as unstructured text, so natural language processing (NLP) can be a valuable tool for extracting actionable insights for care teams. However, research in this area often uses patient records from private health systems' electronic health records (EHRs), which makes it difficult to compare results from other health systems or even replicate the studies. The development and release of high-quality publicly available datasets could enable more reproducible research in this area.",
+        "bbox": [
+            507,
+            253,
+            885,
+            444
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this work, we introduce a new, public SDOH dataset based on MTSamples.com, an online collection of transcribed medical reports. Our setup is motivated by the use cases of slicing patient populations along social determinant dimensions for population analytics, and retrieving patients with certain social determinants of health to allow for more targeted outreach and intervention. Given a large set of social determinant factors, our goal is to make binary determinations for each patient about whether that patient's notes imply a particular SDOH factor. In other words, for example, we want to be able to find all patients who lack access to transportation, as opposed to just tagging transportation-related spans in their notes, as done in some previous work.",
+        "bbox": [
+            507,
+            448,
+            885,
+            705
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To achieve this goal, we formulate the task as a textual entailment problem, with patient note snippets as the premises, SDOH factors as the hypotheses, and binary entailment labels. We use human annotators to label 1,398 social history snippets according to a curated list of 60 SDOH statements, resulting in a dataset of 29,635 labeled premise-hypothesis examples after some filtering (see Section 3 for details). We release this dataset publicly. We also evaluate state-of-the-art publicly available large language models on our data in a range of different settings (see Section 4).",
+        "bbox": [
+            507,
+            707,
+            885,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A notable feature of our entailment dataset is",
+        "bbox": [
+            527,
+            903,
+            880,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author. lelkes@google.com",
+        "bbox": [
+            139,
+            854,
+            413,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$\\dagger$  Work completed at Google.",
+        "bbox": [
+            142,
+            866,
+            319,
+            879
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "Equal leadership.",
+        "bbox": [
+            142,
+            881,
+            257,
+            892
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "<https://github.com/google-research-datasets/SDOH-NLI>",
+        "bbox": [
+            115,
+            892,
+            478,
+            916
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "4789",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4789-4798 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            779,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "that unlike other entailment datasets, our premises and hypotheses were obtained independently, and we label the full cross product of premises and hypotheses. In traditional entailment datasets, the hypotheses are constructed to be tied to a particular premise; however, in our formulation, the hypotheses are drawn from a large set of SDOH factors that may or may not be discussed in a particular premise (drawn from a clinical note). Since all our text comes from the same domain, we still have a non-negligible fraction of positive entailment labels (albeit with a much larger label imbalance than in standard NLI benchmark datasets).",
+        "bbox": [
+            112,
+            84,
+            489,
+            292
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This requires NLI methods to understand both the premise and the hypothesis, and defeats typical shortcuts that have been observed to work for other entailment datasets, such as guessing the label based on the hypothesis alone, or relying on simple syntactic clues such as the presence of negations (see Section 2.2). Indeed, even though our task does not require domain-specific knowledge, we observe that state-of-the-art models struggle to generalize from common NLI benchmarks to our dataset, and highlight typical failure cases (see Section 4).",
+        "bbox": [
+            115,
+            294,
+            489,
+            485
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We evaluate both off-the-shelf and fine-tuned models in different setups: dual encoders, zero/few-shot prompting, and binary classification. We show that state-of-the-art off-the-shelf models, even if they were fine-tuned on various NLI datasets, do not reliably solve our problem; on the other hand, models fine-tuned on our training set robustly generalize both to unseen notes and to unseen factors, providing evidence for the usefulness of our dataset for model fine-tuning and evaluation.",
+        "bbox": [
+            112,
+            487,
+            489,
+            648
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            661,
+            268,
+            676
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Social and Behavioral Determinants of Health",
+        "text_level": 1,
+        "bbox": [
+            112,
+            688,
+            465,
+            718
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "There has been a lot of interest in using NLP techniques for SDOH extraction; see Patra et al. (2021) for a recent survey. The range and granularity of SDOH factors vary considerably across different papers. There has also been a range of methods used, from rule-based heuristics, to n-grams, to fine-tuning pretrained Transformer models.",
+        "bbox": [
+            112,
+            725,
+            489,
+            837
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Many previous research studies on SDOH extraction were performed on EHR data from particular health systems and are not released publicly. Exceptions include the i2b2 NLP Smoking Challenge (Uzuner et al., 2008), which classified 502",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "deidentified medical discharge records for smoking status only, and small number of datasets based on MIMIC-III (Johnson et al., 2016), a large publicly available database of deidentified health records of patients who stayed in critical care units of Beth Israel Deaconess Medical Center between 2001 and 2012. For example, Gehrmann et al. (2018) annotated MIMIC-III with binary labels for certain \"phenotypes\" including alcohol and substance abuse, and Lybarger et al. (2021) annotated note spans with SDOH information.",
+        "bbox": [
+            507,
+            84,
+            884,
+            261
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We are aware of four other previous SDOH-related papers which used MTSamples data (Wang et al., 2015; Winden et al., 2017; Yetisgen et al., 2016; Yetisgen and Vanderwende, 2017), all of which focused on extracting and tagging SDOH-related spans from social history sections.",
+        "bbox": [
+            507,
+            267,
+            884,
+            363
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Among previous papers, the one methodologically closest to ours is Lituiev et al. (2023) which also formulated SDOH extraction as an entailment problem and evaluated RoBERTa (Liu et al., 2019) fine-tuned on ANLI (Nie et al., 2020) on clinical notes from UCSF, without any in-domain fine-tuning experiments.",
+        "bbox": [
+            507,
+            370,
+            884,
+            483
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Natural Language Inference",
+        "text_level": 1,
+        "bbox": [
+            507,
+            514,
+            779,
+            531
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Natural language inference (NLI), also called recognizing textual entailment (RTE), has been a very well-studied NLP task; see e.g. Storks et al. (2019); Poliak (2020) for recent surveys. Many standard NLI datasets, such as SNLI (Bowman et al., 2015) or MultiNLI (Williams et al., 2018), are obtained by automatically collecting premises and, for each premise and target entailment label, having human annotators write a hypothesis with the specified entailment relation to the premise. (Even some of the datasets specifically designed to address these datasets' shortcomings, such as ANLI (Nie et al., 2020), follow a similar setup.) It has been observed that this leads to annotation artifacts that lets models do well on these tasks without requiring true understanding of entailment, including by using simple syntactic heuristics (McCoy et al., 2019) or by completely ignoring the premise, and considering only the hypothesis (Gururangan et al., 2018; Poliak et al., 2018). ContractNLI (Koreeda and Manning, 2021) is an example of a previous dataset which used the same fixed set of hypotheses for all the premises.",
+        "bbox": [
+            507,
+            549,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "4790",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Dataset Construction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            332,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We scraped all 5,003 medical reports from MTSamples. Within these reports, we obtained 1,030 note sections related to social history by searching for a collection of note section titles identified as synonyms of social history. We then split each note section into sentences, resulting in 3,281 sentences. Many sentences (such as \"He is married\") appear in multiple notes; after dedduplication, we have 1,398 unique sentences.",
+        "bbox": [
+            112,
+            111,
+            489,
+            256
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We manually curated a collection of SDOH factors primarily from two sources, the AHRQ Social Determinants of Health Database (for Healthcare Research and Quality, 2020) and UCSF SIREN's Compendium of medical terminology codes for social risk factors (Arons et al., 2018). We rephrased each factor as a full English sentence stating a fact about a person; e.g. \"The person is employed.\" For binary factors, we included a statement and its negation; e.g. \"The person currently drinks alcohol\" and \"The person currently does not drink alcohol.\" For factors with multiple potential values (such as housing status), we aimed to list all the common options. We grouped these 60 statements into 10 categories: smoking, alcohol use, drug use, employment, housing, food, transportation, health insurance, social support, and financial situation. See the full list in Appendix A.",
+        "bbox": [
+            115,
+            256,
+            489,
+            546
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For each social history sentence, we asked human raters to select all the relevant categories and, within each category, all the statements that are supported by the social history snippet. Each snippet was rated by at least three raters. For each (snippet, statement) pair, we took the majority vote of raters to get binary entailment labels. Rater agreement was high, with a Krippendorff's alpha of 0.97 (computed over the binary entailment labels provided by different raters for (premise, hypothesis) pairs).",
+        "bbox": [
+            112,
+            548,
+            489,
+            709
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We removed all SDOH statements which were not entailed by any of the social history snippets as well as all note sentences that were not relevant to any SDOH categories. Finally, after inspecting the dataset, we removed three SDOH factors (\"The person is stably housed.\", \"The person has social support.\", \"The person does not have social support.\") because raters weren't able to consistently give correct ratings. That left us with 787 unique sentences and 38 SDOH factors. Since each snippet would typically only entail one factor or a small number of factors, the resulting dataset is heavily imbalanced: only  $4.6\\%$  of the labels are positive.",
+        "bbox": [
+            112,
+            709,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We split the dataset along the snippets into training, validation, and test sets with a 70:15:15 ratio, with the following modification: we remove a single pair of factors, \"The person lives alone\" and \"The person does not live alone,\" from the training and validation sets, in order to evaluate fine-tuned models' ability to generalize to unseen factors. In other words, the training, validation, and test sets have disjoint note snippets but the same SDOH factors, except for the test set which also contains an additional two factors that are not present in the training or validation sets.",
+        "bbox": [
+            507,
+            84,
+            884,
+            277
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Model Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            291,
+            697,
+            307
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Since the typical use case is retrieving patients with a particular SDOH factor, and because of the heavy label imbalance in our dataset, we use precision, recall, and F1 score as our evaluation metrics.",
+        "bbox": [
+            507,
+            318,
+            882,
+            381
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We evaluate state-of-the-art public models in four different setups:",
+        "bbox": [
+            507,
+            382,
+            882,
+            414
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Treating the problem as a retrieval task with SDOH factors as queries and note snippets as documents. We evaluate Sentence-T5 (Ni et al., 2022a) and GTR (Ni et al., 2022b), two state-of-the-art dual encoder models. We select the cosine similarity threshold which maximizes F1 score on the training set.",
+            "- A general-purpose NLI model. We evaluate the SENTLI model (Schuster et al., 2022), a state-of-the-art NLI model (T5 large fine-tuned on the SNLI (Bowman et al., 2015), MNLI (Williams et al., 2018), ANLI (Nie et al., 2020), Fever (Thorne et al., 2018), and VitaminC (Schuster et al., 2021) datasets). To help the model adapt to the label imbalance in our dataset, we also evaluate it with re-tuning the threshold for positive prediction to maximize F1 score on our training set.",
+            "- Zero/few-shot prompting. We evaluate FlanT5 XXL (Chung et al., 2022) and FlanUL2 (Tay et al., 2023) in both a zero- and a few-shot setting. These models instruction tuned for a large set of tasks, including NLI tasks. (See Appendix B for details.)",
+            "- Fine-tuning experiments. We fine-tune T5 and Flan-T5 on our dataset (SDOH-NLI), on ANLI, and on a mixture of both at various model sizes. (See Appendix C for details.)"
+        ],
+        "bbox": [
+            531,
+            429,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "4791",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/f5d45ef0a0275ce0faad0f2cd9d2443f549cb383f5a0e095f15c5a815cd17c99.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td>Sentence-T5 11B</td><td>.2826</td><td>.523</td><td>.3669</td></tr><tr><td>GTR XXL</td><td>.3148</td><td>.4885</td><td>.3829</td></tr><tr><td>SENTLI</td><td>.4438</td><td>.8161</td><td>.5749</td></tr><tr><td>SENTLI, threshold tuned</td><td>.5721</td><td>.7299</td><td>.6414</td></tr><tr><td>Flan-T5 XXL 0-shot</td><td>.6867</td><td>.6552</td><td>.6706</td></tr><tr><td>Flan-T5 XXL 5-shot</td><td>.6297</td><td>.6839</td><td>.6556</td></tr><tr><td>Flan-UL2 0-shot</td><td>.5922</td><td>.7011</td><td>.6421</td></tr><tr><td>Flan-UL2 5-shot</td><td>.5536</td><td>.7126</td><td>.6231</td></tr><tr><td>Flan-T5 XXL finetuned on ANLI</td><td>.356</td><td>.8736</td><td>.5058</td></tr><tr><td>Flan-T5 XXL finetuned on SDOH-NLI</td><td>.9295</td><td>.8333</td><td>.8788</td></tr><tr><td>Flan-T5 XXL finetuned on ANLI + SDOH-NLI</td><td>.8765</td><td>.8563</td><td>.8663</td></tr></table>",
+        "bbox": [
+            194,
+            80,
+            801,
+            310
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Scores of retrieval-based, general-purpose NLI, in-context, and fine-tuned models on the SDOH test set. See Appendix C for additional fine-tuning experiments.",
+        "bbox": [
+            112,
+            319,
+            882,
+            350
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            375,
+            218,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1 shows selected results (see Appendix D for more model fine-tuning metrics). First, we observe that while our problem can naturally be framed as an information retrieval problem, even state-of-the-art retrieval models perform poorly on it; formulating the problem as natural language inference yields dramatically better results.",
+        "bbox": [
+            112,
+            395,
+            485,
+            508
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "However, even powerful models fine-tuned for NLI, either on NLI datasets alone or as part of the much larger Flan collection, do not reliably solve our problem, with an F1 score of at most .67 on the test set. Even T5 small fine-tuned on our dataset outperforms the largest off-the-shelf models, highlighting the added value of our dataset.",
+        "bbox": [
+            112,
+            511,
+            487,
+            624
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "When prompting instruction-tuned models, including few-shot examples does not appear to add any value compared to a zero-shot setup. We conjecture that this is because the models are already familiar with the NLI task and the prompt format from instruction tuning, and presenting them with a small number of additional examples is not sufficient for teaching them the difference between this task and the NLI benchmark datasets they have seen during fine-tuning.",
+        "bbox": [
+            112,
+            627,
+            487,
+            788
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For fine-tuned models below XXL size, when fine-tuning on the SDOH-NLI training set only, we observed poor generalization to the held-out SDOH factors in the test set (see Appendix D for detailed metrics). Because of this, why we experimented with fine-tuning on a combination of our data and ANLI R1, a challenging general-purpose NLI dataset of a similar size. For smaller models,",
+        "bbox": [
+            112,
+            790,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "fine-tuning on this mixture enabled robust generalization to unseen factors without sacrificing overall test performance.",
+        "bbox": [
+            507,
+            375,
+            882,
+            423
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Discussion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            453,
+            638,
+            467
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "What makes our dataset challenging for models fine-tuned on standard NLI datasets? We emphasize that it is not that it requires any specialized (e.g., medical) knowledge: most of our examples describe everyday situations in plain English (e.g., \"The patient is retired on disability due to her knee replacements\"); human raters without any domain-specific training had no difficulty understanding them. We conjecture that the main difference is that the hypotheses of typical NLI datasets are written to satisfy a given entailment label for a given premise, whereas ours are obtained independently from the premises, and therefore their entailment status can be more subtle and ambiguous. These models also struggled with distinguishing between statements about the present and the past; e.g., Flan-UL2 erroneously predicting that \"He is not a smoker\" entails \"The person wasn't a smoker in the past.\" Also, our dataset contains a lot of irrelevant hypotheses and requires the model to correctly classify all of those; we see off-the-shelf models occasionally giving them positive labels (e.g. predicting that \"She was living alone and is now living in assisted living.\" implies \"The person drank alcohol in the past.), hurting precision. As an example, Flan-T5 fine-tuned on ANLI had comparable recall to our best models, but much worse precision.",
+        "bbox": [
+            507,
+            483,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "4792",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            247,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this work, we introduced SDOH-NLI, a new entailment dataset containing social history snippets as premises and social and behavioral determinants of health as hypotheses. Our dataset was designed both to reflect realistic use cases of SDOH extraction in clinical settings as well as to provide a high quality entailment dataset to support broader NLI research. We evaluated baseline methods using state-of-the-art public models in a variety of setups, and highlighted novel and challenging features of our dataset.",
+        "bbox": [
+            112,
+            109,
+            492,
+            287
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            300,
+            278,
+            317
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The authors would like to thank Jonas B. Kemp, Von Nguyen, Birju Patel, Martin G. Seneviratne, Andy Strunk, and Vinh Q. Tran for their valuable feedback and discussions.",
+        "bbox": [
+            112,
+            326,
+            489,
+            390
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            401,
+            218,
+            418
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our dataset is English-only, and reflects the American healthcare system. While a lot of the social and behavioral determinants of health mentioned in the data could apply elsewhere, too, their distribution and the language used to describe them could reflect U.S. norms. Also, since the dataset is based on transcription samples, the text can be cleaner than in some other settings (such as notes in EHRs), where the task could be more challenging due to typos and abbreviations that do not appear in our dataset. In such settings, performance could be improved by first using the methods of Rajkumar et al. (2022) to decode abbreviations, but we have not included this in our evaluations. Finally, our dataset only contains short note snippets, and we have not evaluated the models' ability to reconcile contradictory statements or reason about the chronology of information in longer patient records. For longer contexts, especially if the social history sections don't fit in the Transformer model's context window, we recommend evaluating the methods of Schuster et al. (2022).",
+        "bbox": [
+            112,
+            429,
+            490,
+            783
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            796,
+            265,
+            810
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Machine learning research in the clinical domain has many ethical considerations. Given the nature of this work is to identify opportunities for care teams to improve the health outcomes of patients due to factors commonly not addressed, we think it contributes to human well-being and avoids harms.",
+        "bbox": [
+            112,
+            822,
+            490,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The use of public, non-identifiable data that is released for the NLP community helps balance the need to have reproducible (i.e., honest and trustworthy) data to enable technical advances while limiting the need for sensitive, private medical records for research. We acknowledge that the purpose of the work is to identify SDOH to provide additional help and services to patients, and we warn against any use to deny patients care or services based on any factor that can be identified. Although we picked a large number of SDOH factors to test our method, we acknowledge that there may be additional factors that are important for specific patients and populations, so we encourage researchers to reflect on those possible factors and create datasets to help others study them, as well.",
+        "bbox": [
+            507,
+            84,
+            884,
+            342
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            368,
+            608,
+            382
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Abigail Arons, Sarah DeSilvey, Caroline Fichtenberg, and Laura Gottlieb. 2018. Compendium of medical terminology codes for social risk factors.",
+            "Samuel Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642.",
+            "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416.",
+            "Agency for Healthcare Research and MD. Quality, Rockville. 2020. Social determinants of health database.",
+            "Sebastian Gehrmann, Franck Dernoncourt, Yeran Li, Eric T Carlson, Joy T Wu, Jonathan Welt, John Foote Jr, Edward T Moseley, David W Grant, Patrick D Tyler, et al. 2018. Comparing deep learning and concept extraction based methods for patient phenotyping from clinical narratives. *PloS one*, 13(2):e0192360.",
+            "Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. 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 107-112.",
+            "Alistair EW Johnson, Tom J Pollard, Lu Shen, Liwei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-iii,"
+        ],
+        "bbox": [
+            509,
+            388,
+            885,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "4793",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "a freely accessible critical care database. Scientific data, 3(1):1-9.",
+            "Yuta Koreeda and Christopher D Manning. 2021. Contractnli: A dataset for document-level natural language inference for contracts. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1907-1919.",
+            "Dmytro S Lituiev, Benjamin Lacar, Sang Pak, Peter L Abramowitsch, Emilia H De Marchis, and Thomas A Peterson. 2023. Automatic extraction of social determinants of health from medical notes of chronic lower back pain patients. J. Am. Med. Inform. Assoc.",
+            "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.",
+            "Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. 2023. The flan collection: Designing data and methods for effective instruction tuning. arXiv preprint arXiv:2301.13688.",
+            "Kevin Lybarger, Mari Ostendorf, and Meliha Yetisgen. 2021. Annotating social determinants of health using active learning, and characterizing determinants using neural event extraction. Journal of Biomedical Informatics, 113:103631.",
+            "Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428-3448.",
+            "Jianmo Ni, Gustavo Hernandez Abrego, Noah Constant, Ji Ma, Keith Hall, Daniel Cer, and Yinfei Yang. 2022a. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 1864–1874.",
+            "Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith Hall, Ming-Wei Chang, and Yinfei Yang. 2022b. Large dual encoders are generalizable retrievers. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9844–9855, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial nli: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901.",
+            "Braja G Patra, Mohit M Sharma, Veer Vekaria, Prakash Adekkanattu, Olga V Patterson, Benjamin Glicksberg, Lauren A Lepow, Euijung Ryu, Joanna M Biernacka, Al'ona Furmanchuk, et al. 2021. Extracting"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "social determinants of health from electronic health records using natural language processing: a systematic review. Journal of the American Medical Informatics Association, 28(12):2716-2727.",
+            "Adam Poliak. 2020. A survey on recognizing textual entailment as an nlp evaluation. In Proceedings of the First Workshop on Evaluation and Comparison of NLP Systems, pages 92-109.",
+            "Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 180-191.",
+            "Alvin Rajkomar, Eric Loreaux, Yuchen Liu, Jonas Kemp, Benny Li, Ming-Jun Chen, Yi Zhang, Afroz Mohiuddin, and Juraj Gottweis. 2022. Deciphering clinical abbreviations with a privacy protecting machine learning system. Nature Communications, 13(1):7456.",
+            "Adam Roberts, Hyung Won Chung, Anselm Levskaya, Gaurav Mishra, James Bradbury, Daniel Andor, Sharan Narang, Brian Lester, Colin Gaffney, Afroz Mohiuddin, et al. 2022. Scaling up models and data with t5x and seqio. arXiv preprint arXiv:2203.17189.",
+            "Tal Schuster, Sihao Chen, Senaka Buthpitiya, Alex Fabrikant, and Donald Metzler. 2022. Stretching sentence-pair NLI models to reason over long documents and clusters. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 394-412, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Tal Schuster, Adam Fisch, and Regina Barzilay. 2021. Get your vitamin c! robust fact verification with contrastive evidence. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 624-643.",
+            "Shane Storks, Qiaozi Gao, and Joyce Y Chai. 2019. Recent advances in natural language inference: A survey of benchmarks, resources, and approaches. arXiv preprint arXiv:1904.01172.",
+            "Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Steven Zheng, et al. 2023. UI2: Unifying language learning paradigms. In The Eleventh International Conference on Learning Representations.",
+            "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"
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "4794",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics.",
+        "bbox": [
+            132,
+            85,
+            487,
+            112
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Özlem Uzuner, Ira Goldstein, Yuan Luo, and Isaac Kohane. 2008. Identifying patient smoking status from medical discharge records. Journal of the American Medical Informatics Association, 15(1):14-24.",
+        "bbox": [
+            114,
+            121,
+            489,
+            175
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Yan Wang, Elizabeth S Chen, Serguei Pakhomov, Elliot Arsoniadis, Elizabeth W Carter, Elizabeth Lindemann, Indra Neil Sarkar, and Genevieve B Melton. 2015. Automated extraction of substance use information from clinical texts. In AMIA Annual Symposium Proceedings, volume 2015, page 2121. American Medical Informatics Association.",
+        "bbox": [
+            115,
+            184,
+            489,
+            275
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. 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 1112-1122.",
+        "bbox": [
+            115,
+            286,
+            489,
+            379
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Tamara J Winden, Elizabeth S Chen, Yan Wang, Elizabeth Lindemann, and Genevieve B Melton. 2017. Residence, living situation, and living conditions information documentation in clinical practice. In AMIA Annual Symposium Proceedings, volume 2017, page 1783. American Medical Informatics Association.",
+        "bbox": [
+            115,
+            388,
+            489,
+            479
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Meliha Yetisgen, Elena Pellicer, David R Crosslin, and Lucy Vanderwende. 2016. Automatic identification of lifestyle and environmental factors from social history in clinical text. In Proceedings of AMIA 2016 Joint Summits on Translational Science.",
+        "bbox": [
+            115,
+            489,
+            487,
+            555
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Meliha Yetisgen and Lucy Vanderwende. 2017. Automatic identification of substance abuse from social history in clinical text. In Artificial Intelligence in Medicine: 16th Conference on Artificial Intelligence in Medicine, AIME 2017, Vienna, Austria, June 21-24, 2017, Proceedings 16, pages 171-181. Springer.",
+        "bbox": [
+            115,
+            565,
+            489,
+            645
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Labeling Setup",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            680,
+            101
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Raters were given the following instructions:",
+        "bbox": [
+            509,
+            109,
+            843,
+            124
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "\"In this task, you will be given a list of snippets from transcribed medical records, describing a person's social history. Your job is to select the categories that are relevant to the snippet and, in each category, select statements about the person that are supported by the snippet.",
+        "bbox": [
+            507,
+            124,
+            882,
+            221
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Only select categories and statements relevant to the subject of the sentence, NOT if they apply to someone else (such as the subject's relatives).",
+        "bbox": [
+            507,
+            222,
+            880,
+            269
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The source of the medical transcript samples is a public dataset (MTSamples.com), and not your, other raters' or users health transcript data.\"",
+        "bbox": [
+            507,
+            269,
+            882,
+            317
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For each snippet, they were required select one or more of the following categories (including \"None of the above\" if none of the categories were relevant). If they selected a category, the statements within the category would be displayed, and they were required to select one of statements within that category, or \"None of the above\":",
+        "bbox": [
+            507,
+            319,
+            882,
+            430
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Smoking",
+        "bbox": [
+            531,
+            441,
+            618,
+            457
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person is currently a smoker.",
+            "- The person is currently not a smoker.",
+            "- The person was a smoker in the past.",
+            "- The person wasn't a smoker in the past.",
+            "- None of the above"
+        ],
+        "bbox": [
+            566,
+            464,
+            877,
+            551
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "- Alcohol use",
+        "bbox": [
+            531,
+            564,
+            638,
+            577
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person currently drinks alcohol.",
+            "- The person currently does not drink alcohol.",
+            "- The person drank alcohol in the past.",
+            "- The person did not drink alcohol in the past.",
+            "- None of the above"
+        ],
+        "bbox": [
+            566,
+            587,
+            882,
+            706
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "- Drug use",
+        "bbox": [
+            531,
+            719,
+            618,
+            733
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person is a drug user.",
+            "- The person is not a drug user.",
+            "- The person was a drug user in the past.",
+            "- The person wasn't a drug user in the past.",
+            "- The person uses cocaine.",
+            "- The person does not use cocaine.",
+            "- The person used cocaine in the past.",
+            "- The person did not use cocaine in the past.",
+            "- The person uses marijuana."
+        ],
+        "bbox": [
+            566,
+            741,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "4795",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person does not use marijuana.",
+            "- The person used marijuana in the past.",
+            "- The person did not use marijuana in the past.",
+            "- The person uses opioids (e.g., heroin, fentanyl, oxycodone).",
+            "- The person does not opioids (e.g., heroin, fentanyl, oxycodone).",
+            "- The person used opioids in the past (e.g., heroin, fentanyl, oxycodone).",
+            "- The person did not use opioids in the past (e.g., heroin, fentanyl, oxycodone).",
+            "- None of the above"
+        ],
+        "bbox": [
+            169,
+            85,
+            487,
+            305
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- Employment",
+        "text_level": 1,
+        "bbox": [
+            136,
+            318,
+            250,
+            332
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person is employed.",
+            "- The person is not employed.",
+            "- The person is employed part time.",
+            "- The person is a student.",
+            "- The person is a homemaker.",
+            "- The person is retired due to age or preference.",
+            "- The person is retired due to disability.",
+            "- The person is retired due to an unknown reason.",
+            "- None of the above"
+        ],
+        "bbox": [
+            169,
+            342,
+            487,
+            533
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Housing",
+        "text_level": 1,
+        "bbox": [
+            136,
+            545,
+            218,
+            560
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person lives in their own or their family's home.",
+            "- The person lives in a housing facility.",
+            "- The person is homeless.",
+            "- The person is stably housed.",
+            "- The person's housing is unsuited to their needs.",
+            "- None of the above"
+        ],
+        "bbox": [
+            169,
+            568,
+            485,
+            705
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Food",
+        "text_level": 1,
+        "bbox": [
+            136,
+            717,
+            193,
+            730
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person is able to obtain food on a consistent basis.",
+            "- The person is not able to obtain food on a consistent basis.",
+            "- The person has consistent fruit and vegetable intake.",
+            "- The person does not have consistent fruit and vegetable intake.",
+            "- None of the above"
+        ],
+        "bbox": [
+            169,
+            740,
+            487,
+            891
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- Transportation",
+        "text_level": 1,
+        "bbox": [
+            136,
+            903,
+            263,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person has access to transportation.",
+            "- The person does not have access to transportation.",
+            "- The person has access to a car.",
+            "- The person does not have access to a car.",
+            "- The person has access to public transit.",
+            "- The person does not have access to public transit.",
+            "- The person has issues with finding transportation.",
+            "- None of the above"
+        ],
+        "bbox": [
+            564,
+            84,
+            882,
+            275
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Health insurance",
+        "text_level": 1,
+        "bbox": [
+            532,
+            288,
+            675,
+            300
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person has private health insurance.",
+            "- The person is on Medicare.",
+            "- The person is on Medicaid.",
+            "- The person does not have health insurance.",
+            "- None of the above"
+        ],
+        "bbox": [
+            564,
+            311,
+            882,
+            414
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Social support",
+        "text_level": 1,
+        "bbox": [
+            532,
+            426,
+            655,
+            442
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person has social support.",
+            "- The person does not have social support.",
+            "- The person lives alone.",
+            "- The person does not live alone.",
+            "- None of the above"
+        ],
+        "bbox": [
+            564,
+            451,
+            882,
+            537
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Financial situation",
+        "text_level": 1,
+        "bbox": [
+            532,
+            550,
+            687,
+            564
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The person is below the poverty line.",
+            "- The person is above the poverty line.",
+            "- The person is able to afford medications.",
+            "- The person is not able to afford medications.",
+            "- None of the above"
+        ],
+        "bbox": [
+            564,
+            574,
+            882,
+            677
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- None of the above",
+        "text_level": 1,
+        "bbox": [
+            532,
+            690,
+            685,
+            703
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B Prompting Setup",
+        "text_level": 1,
+        "bbox": [
+            509,
+            717,
+            694,
+            734
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our goal with the design of our zero/few-shot experiments was to stay close to how the models were trained and evaluated on similar NLI tasks. Since the models we used were fine-tuned on the Flan collection (Longpre et al., 2023), which includes several NLI datasets, we reused some of the Flan collections's prompt templates.",
+        "bbox": [
+            507,
+            741,
+            882,
+            854
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In particular, we used the templates for the SNLI and MNLI dataset which were the most relevant to our data. (We excluded templates for other datasets such as RTE, ANLI, or WNLI because the wording",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "4796",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "of some of the prompts could be slightly misleading; e.g. by referring to the premise as a \"paragraph.\") For each example, we chose a prompt template uniformly at random from the 19 templates. For few-shot experiments, we first picked a random positive example from the training set, then picked the other few-shot examples uniformly at random from the training examples. (Without explicitly forcing at least one example to be positive, it would be very likely that all the few-shot examples would be negative, given the label imbalance in the dataset. We tried dropping that requirement and sampling all examples at random, which resulted in a slight drop in model performance, as we expected.)",
+        "bbox": [
+            115,
+            84,
+            489,
+            324
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Since SNLI and MNLI have three answer options (\"yes,\" \"it is not possible to tell,\" \"no\"), we kept all three of these options in the prompt template, even though our dataset is binary. We experimented with dropping either the \"no\" or the \"it is not possible to tell\" option; both of these deviations from the original prompts resulted in slight decreases in model performance.",
+        "bbox": [
+            115,
+            326,
+            489,
+            454
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We use rank classification to obtain binary labels (as is customary): i.e., instead of decoding a model prediction, we score the three answer options (for three copies of each input example) and consider the prediction positive if \"yes\" has the highest score. Similarly to our experiment with SENTLI, we also tried taking softmax over the three options and using a fixed threshold for \"yes\" which maximizes F1 score on the training set; this resulted in a small drop in test F1 score (e.g., .6501 to .6222 for Flan-UL2).",
+        "bbox": [
+            115,
+            456,
+            489,
+            630
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "C Fine-tuning Setup",
+        "text_level": 1,
+        "bbox": [
+            115,
+            645,
+            309,
+            662
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "All models were fine-tuned using the T5X framework (Roberts et al., 2022) on TPUv3 chips for  $10\\mathrm{k}$  steps with batch size 32 and learning rate 1e-4, with the exception of the T5 small size models which were finetuned for  $50\\mathrm{k}$  steps. For each model, we picked the checkpoint with the highest F1 score on the validation set.",
+        "bbox": [
+            115,
+            671,
+            489,
+            781
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "D Additional Model Fine-tuning Results",
+        "text_level": 1,
+        "bbox": [
+            115,
+            796,
+            478,
+            813
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "See Table 2 for the full set of results from our finetuning experiments. Here we also include metrics on the subset of the test set consisting on the new SDOH factors not included in the training and validation sets to highlight the differences in models' ability to generalize to unseen factors.",
+        "bbox": [
+            115,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Note that since Flan-T5 models between sizes large and XL fine-tuned on either ANLI or our SDOH dataset alone underperformed the same models fine-tuned on the combined dataset, we did not perform these dataset ablations on smaller Flan-T5 models or on T5 below the XXL size.",
+        "bbox": [
+            512,
+            84,
+            884,
+            179
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "4797",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/514aba77f9ea16175ddd817e9ee0dcca8bd69793c124ea2c004290a8d709c85b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Size</td><td>Dataset</td><td>Test P</td><td>Test R</td><td>Test F1</td><td>New factors P</td><td>New factors R</td><td>New factors F1</td></tr><tr><td>T5</td><td>small</td><td>ANLI + SDOH-NLI</td><td>.7987</td><td>.7069</td><td>.7500</td><td>.6667</td><td>.0833</td><td>.1481</td></tr><tr><td>T5</td><td>base</td><td>ANLI + SDOH-NLI</td><td>.7396</td><td>.8161</td><td>.7760</td><td>.75</td><td>.375</td><td>.5</td></tr><tr><td>T5</td><td>large</td><td>ANLI + SDOH-NLI</td><td>.8247</td><td>.7299</td><td>.7744</td><td>.75</td><td>.125</td><td>.2143</td></tr><tr><td>T5</td><td>XL</td><td>ANLI + SDOH-NLI</td><td>.8391</td><td>.8391</td><td>.8391</td><td>.9231</td><td>.5</td><td>.6486</td></tr><tr><td>T5</td><td>XXL</td><td>ANLI</td><td>.8391</td><td>.8391</td><td>.8391</td><td>.75</td><td>.1</td><td>.8571</td></tr><tr><td>T5</td><td>XXL</td><td>SDOH-NLI</td><td>.8859</td><td>.7586</td><td>.8173</td><td>.8571</td><td>.25</td><td>.3871</td></tr><tr><td>T5</td><td>XXL</td><td>ANLI + SDOH-NLI</td><td>.9024</td><td>.8506</td><td>.8757</td><td>.8889</td><td>1.</td><td>.9412</td></tr><tr><td>Flan-T5</td><td>small</td><td>ANLI + SDOH-NLI</td><td>.8075</td><td>.7471</td><td>.7761</td><td>.6667</td><td>.0833</td><td>.1481</td></tr><tr><td>Flan-T5</td><td>base</td><td>ANLI + SDOH-NLI</td><td>.7956</td><td>.8276</td><td>.8113</td><td>.6216</td><td>.9583</td><td>.7541</td></tr><tr><td>Flan-T5</td><td>large</td><td>ANLI</td><td>.1777</td><td>.8506</td><td>.2939</td><td>.2421</td><td>.9583</td><td>.3866</td></tr><tr><td>Flan-T5</td><td>large</td><td>SDOH-NLI</td><td>.8041</td><td>.6839</td><td>.7391</td><td>.2308</td><td>.125</td><td>.1622</td></tr><tr><td>Flan-T5</td><td>large</td><td>ANLI + SDOH-NLI</td><td>.8212</td><td>.8448</td><td>.8329</td><td>.8846</td><td>.9583</td><td>.92</td></tr><tr><td>Flan-T5</td><td>XL</td><td>ANLI</td><td>.2868</td><td>.8851</td><td>.4332</td><td>.7667</td><td>.9583</td><td>.8519</td></tr><tr><td>Flan-T5</td><td>XL</td><td>SDOH</td><td>.9071</td><td>.7299</td><td>.8089</td><td>1.</td><td>.125</td><td>.2222</td></tr><tr><td>Flan-T5</td><td>XL</td><td>ANLI + SDOH-NLI</td><td>.8647</td><td>.8448</td><td>.8547</td><td>.8846</td><td>.9583</td><td>.92</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>ANLI</td><td>.3560</td><td>.8736</td><td>.5058</td><td>.7931</td><td>.9583</td><td>.8679</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>SDOH-NLI</td><td>.9295</td><td>.8333</td><td>.8788</td><td>1.</td><td>.875</td><td>.9333</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>ANLI + SDOH-NLI</td><td>.8765</td><td>.8563</td><td>.8663</td><td>.8846</td><td>.9583</td><td>.92</td></tr></table>",
+        "bbox": [
+            117,
+            362,
+            878,
+            594
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Precision, recall, and F1 scores of various models on the full test set and on the subset of the test set consisting of the two held-out factors.",
+        "bbox": [
+            112,
+            604,
+            882,
+            633
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "4798",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    }
+]

2023/SDOH-NLI_ a Dataset for Inferring Social Determinants of Health from Clinical Notes/a4f22374-5104-49b6-a9f1-43a434ee9541_model.json

@@ -0,0 +1,2464 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.206,
+                0.09,
+                0.795,
+                0.131
+            ],
+            "angle": 0,
+            "content": "SDOH-NLI: a Dataset for Inferring Social Determinants of Health from Clinical Notes"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.139,
+                0.154,
+                0.866,
+                0.174
+            ],
+            "angle": 0,
+            "content": "Adam D. Lelkes<sup>1*</sup>, Eric Loreaux<sup>2†</sup>, Tal Schuster<sup>1</sup>, Ming-Jun Chen<sup>1‡</sup>, Alvin Rajkomar<sup>1‡</sup>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.365,
+                0.188,
+                0.64,
+                0.206
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>Google Research, <sup>2</sup>Curai Health"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.343,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.144,
+                0.274,
+                0.461,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Social and behavioral determinants of health (SDOH) play a significant role in shaping health outcomes, and extracting these determinants from clinical notes is a first step to help healthcare providers systematically identify opportunities to provide appropriate care and address disparities. Progress on using NLP methods for this task has been hindered by the lack of high-quality publicly available labeled data, largely due to the privacy and regulatory constraints on the use of real patients' information. This paper introduces a new dataset, SDOH-NLI, that is based on publicly available notes and which we release publicly. We formulate SDOH extraction as a natural language inference (NLI) task, and provide binary textual entailment labels obtained from human raters for a cross product of a set of social history snippets as premises and SDOH factors as hypotheses. Our dataset differs from standard NLI benchmarks in that our premises and hypotheses are obtained independently. We evaluate both \"off-the-shelf\" entailment models as well as models fine-tuned on our data, and highlight the ways in which our dataset appears more challenging than commonly used NLI datasets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.665,
+                0.262,
+                0.68
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.689,
+                0.491,
+                0.85
+            ],
+            "angle": 0,
+            "content": "There has been growing recognition that social and behavioral determinants of health (SDOH) play a significant role in shaping health outcomes for individuals and populations. The ability to accurately identify and extract social and behavioral determinants of health from clinical notes can provide valuable insights that can enable healthcare providers to better understand and address the underlying determinants of health that contribute to poor health outcomes and health disparities."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.254,
+                0.886,
+                0.445
+            ],
+            "angle": 0,
+            "content": "Social determinants of health are frequently recorded in clinical notes as unstructured text, so natural language processing (NLP) can be a valuable tool for extracting actionable insights for care teams. However, research in this area often uses patient records from private health systems' electronic health records (EHRs), which makes it difficult to compare results from other health systems or even replicate the studies. The development and release of high-quality publicly available datasets could enable more reproducible research in this area."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.449,
+                0.886,
+                0.706
+            ],
+            "angle": 0,
+            "content": "In this work, we introduce a new, public SDOH dataset based on MTSamples.com, an online collection of transcribed medical reports. Our setup is motivated by the use cases of slicing patient populations along social determinant dimensions for population analytics, and retrieving patients with certain social determinants of health to allow for more targeted outreach and intervention. Given a large set of social determinant factors, our goal is to make binary determinations for each patient about whether that patient's notes imply a particular SDOH factor. In other words, for example, we want to be able to find all patients who lack access to transportation, as opposed to just tagging transportation-related spans in their notes, as done in some previous work."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.708,
+                0.886,
+                0.901
+            ],
+            "angle": 0,
+            "content": "To achieve this goal, we formulate the task as a textual entailment problem, with patient note snippets as the premises, SDOH factors as the hypotheses, and binary entailment labels. We use human annotators to label 1,398 social history snippets according to a curated list of 60 SDOH statements, resulting in a dataset of 29,635 labeled premise-hypothesis examples after some filtering (see Section 3 for details). We release this dataset publicly. We also evaluate state-of-the-art publicly available large language models on our data in a range of different settings (see Section 4)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.528,
+                0.904,
+                0.882,
+                0.918
+            ],
+            "angle": 0,
+            "content": "A notable feature of our entailment dataset is"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.14,
+                0.855,
+                0.414,
+                0.867
+            ],
+            "angle": 0,
+            "content": "*Corresponding author. lelkes@google.com"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.143,
+                0.868,
+                0.32,
+                0.881
+            ],
+            "angle": 0,
+            "content": "\\(\\dagger\\) Work completed at Google."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.143,
+                0.882,
+                0.258,
+                0.894
+            ],
+            "angle": 0,
+            "content": "Equal leadership."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.117,
+                0.894,
+                0.479,
+                0.917
+            ],
+            "angle": 0,
+            "content": "<https://github.com/google-research-datasets/SDOH-NLI>"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.855,
+                0.479,
+                0.917
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4789"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.781,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4789-4798 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.293
+            ],
+            "angle": 0,
+            "content": "that unlike other entailment datasets, our premises and hypotheses were obtained independently, and we label the full cross product of premises and hypotheses. In traditional entailment datasets, the hypotheses are constructed to be tied to a particular premise; however, in our formulation, the hypotheses are drawn from a large set of SDOH factors that may or may not be discussed in a particular premise (drawn from a clinical note). Since all our text comes from the same domain, we still have a non-negligible fraction of positive entailment labels (albeit with a much larger label imbalance than in standard NLI benchmark datasets)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.295,
+                0.49,
+                0.486
+            ],
+            "angle": 0,
+            "content": "This requires NLI methods to understand both the premise and the hypothesis, and defeats typical shortcuts that have been observed to work for other entailment datasets, such as guessing the label based on the hypothesis alone, or relying on simple syntactic clues such as the presence of negations (see Section 2.2). Indeed, even though our task does not require domain-specific knowledge, we observe that state-of-the-art models struggle to generalize from common NLI benchmarks to our dataset, and highlight typical failure cases (see Section 4)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.488,
+                0.49,
+                0.649
+            ],
+            "angle": 0,
+            "content": "We evaluate both off-the-shelf and fine-tuned models in different setups: dual encoders, zero/few-shot prompting, and binary classification. We show that state-of-the-art off-the-shelf models, even if they were fine-tuned on various NLI datasets, do not reliably solve our problem; on the other hand, models fine-tuned on our training set robustly generalize both to unseen notes and to unseen factors, providing evidence for the usefulness of our dataset for model fine-tuning and evaluation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.662,
+                0.27,
+                0.677
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.689,
+                0.466,
+                0.719
+            ],
+            "angle": 0,
+            "content": "2.1 Social and Behavioral Determinants of Health"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.726,
+                0.49,
+                0.838
+            ],
+            "angle": 0,
+            "content": "There has been a lot of interest in using NLP techniques for SDOH extraction; see Patra et al. (2021) for a recent survey. The range and granularity of SDOH factors vary considerably across different papers. There has also been a range of methods used, from rule-based heuristics, to n-grams, to fine-tuning pretrained Transformer models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Many previous research studies on SDOH extraction were performed on EHR data from particular health systems and are not released publicly. Exceptions include the i2b2 NLP Smoking Challenge (Uzuner et al., 2008), which classified 502"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.262
+            ],
+            "angle": 0,
+            "content": "deidentified medical discharge records for smoking status only, and small number of datasets based on MIMIC-III (Johnson et al., 2016), a large publicly available database of deidentified health records of patients who stayed in critical care units of Beth Israel Deaconess Medical Center between 2001 and 2012. For example, Gehrmann et al. (2018) annotated MIMIC-III with binary labels for certain \"phenotypes\" including alcohol and substance abuse, and Lybarger et al. (2021) annotated note spans with SDOH information."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.268,
+                0.885,
+                0.365
+            ],
+            "angle": 0,
+            "content": "We are aware of four other previous SDOH-related papers which used MTSamples data (Wang et al., 2015; Winden et al., 2017; Yetisgen et al., 2016; Yetisgen and Vanderwende, 2017), all of which focused on extracting and tagging SDOH-related spans from social history sections."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.372,
+                0.885,
+                0.484
+            ],
+            "angle": 0,
+            "content": "Among previous papers, the one methodologically closest to ours is Lituiev et al. (2023) which also formulated SDOH extraction as an entailment problem and evaluated RoBERTa (Liu et al., 2019) fine-tuned on ANLI (Nie et al., 2020) on clinical notes from UCSF, without any in-domain fine-tuning experiments."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.516,
+                0.78,
+                0.532
+            ],
+            "angle": 0,
+            "content": "2.2 Natural Language Inference"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.55,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Natural language inference (NLI), also called recognizing textual entailment (RTE), has been a very well-studied NLP task; see e.g. Storks et al. (2019); Poliak (2020) for recent surveys. Many standard NLI datasets, such as SNLI (Bowman et al., 2015) or MultiNLI (Williams et al., 2018), are obtained by automatically collecting premises and, for each premise and target entailment label, having human annotators write a hypothesis with the specified entailment relation to the premise. (Even some of the datasets specifically designed to address these datasets' shortcomings, such as ANLI (Nie et al., 2020), follow a similar setup.) It has been observed that this leads to annotation artifacts that lets models do well on these tasks without requiring true understanding of entailment, including by using simple syntactic heuristics (McCoy et al., 2019) or by completely ignoring the premise, and considering only the hypothesis (Gururangan et al., 2018; Poliak et al., 2018). ContractNLI (Koreeda and Manning, 2021) is an example of a previous dataset which used the same fixed set of hypotheses for all the premises."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4790"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.333,
+                0.099
+            ],
+            "angle": 0,
+            "content": "3 Dataset Construction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.112,
+                0.49,
+                0.257
+            ],
+            "angle": 0,
+            "content": "We scraped all 5,003 medical reports from MTSamples. Within these reports, we obtained 1,030 note sections related to social history by searching for a collection of note section titles identified as synonyms of social history. We then split each note section into sentences, resulting in 3,281 sentences. Many sentences (such as \"He is married\") appear in multiple notes; after dedduplication, we have 1,398 unique sentences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.258,
+                0.49,
+                0.548
+            ],
+            "angle": 0,
+            "content": "We manually curated a collection of SDOH factors primarily from two sources, the AHRQ Social Determinants of Health Database (for Healthcare Research and Quality, 2020) and UCSF SIREN's Compendium of medical terminology codes for social risk factors (Arons et al., 2018). We rephrased each factor as a full English sentence stating a fact about a person; e.g. \"The person is employed.\" For binary factors, we included a statement and its negation; e.g. \"The person currently drinks alcohol\" and \"The person currently does not drink alcohol.\" For factors with multiple potential values (such as housing status), we aimed to list all the common options. We grouped these 60 statements into 10 categories: smoking, alcohol use, drug use, employment, housing, food, transportation, health insurance, social support, and financial situation. See the full list in Appendix A."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.549,
+                0.49,
+                0.71
+            ],
+            "angle": 0,
+            "content": "For each social history sentence, we asked human raters to select all the relevant categories and, within each category, all the statements that are supported by the social history snippet. Each snippet was rated by at least three raters. For each (snippet, statement) pair, we took the majority vote of raters to get binary entailment labels. Rater agreement was high, with a Krippendorff's alpha of 0.97 (computed over the binary entailment labels provided by different raters for (premise, hypothesis) pairs)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.711,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "We removed all SDOH statements which were not entailed by any of the social history snippets as well as all note sentences that were not relevant to any SDOH categories. Finally, after inspecting the dataset, we removed three SDOH factors (\"The person is stably housed.\", \"The person has social support.\", \"The person does not have social support.\") because raters weren't able to consistently give correct ratings. That left us with 787 unique sentences and 38 SDOH factors. Since each snippet would typically only entail one factor or a small number of factors, the resulting dataset is heavily imbalanced: only \\(4.6\\%\\) of the labels are positive."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.278
+            ],
+            "angle": 0,
+            "content": "We split the dataset along the snippets into training, validation, and test sets with a 70:15:15 ratio, with the following modification: we remove a single pair of factors, \"The person lives alone\" and \"The person does not live alone,\" from the training and validation sets, in order to evaluate fine-tuned models' ability to generalize to unseen factors. In other words, the training, validation, and test sets have disjoint note snippets but the same SDOH factors, except for the test set which also contains an additional two factors that are not present in the training or validation sets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.292,
+                0.698,
+                0.308
+            ],
+            "angle": 0,
+            "content": "4 Model Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.319,
+                0.884,
+                0.382
+            ],
+            "angle": 0,
+            "content": "Since the typical use case is retrieving patients with a particular SDOH factor, and because of the heavy label imbalance in our dataset, we use precision, recall, and F1 score as our evaluation metrics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.384,
+                0.884,
+                0.416
+            ],
+            "angle": 0,
+            "content": "We evaluate state-of-the-art public models in four different setups:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.43,
+                0.884,
+                0.542
+            ],
+            "angle": 0,
+            "content": "- Treating the problem as a retrieval task with SDOH factors as queries and note snippets as documents. We evaluate Sentence-T5 (Ni et al., 2022a) and GTR (Ni et al., 2022b), two state-of-the-art dual encoder models. We select the cosine similarity threshold which maximizes F1 score on the training set."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.556,
+                0.885,
+                0.732
+            ],
+            "angle": 0,
+            "content": "- A general-purpose NLI model. We evaluate the SENTLI model (Schuster et al., 2022), a state-of-the-art NLI model (T5 large fine-tuned on the SNLI (Bowman et al., 2015), MNLI (Williams et al., 2018), ANLI (Nie et al., 2020), Fever (Thorne et al., 2018), and VitaminC (Schuster et al., 2021) datasets). To help the model adapt to the label imbalance in our dataset, we also evaluate it with re-tuning the threshold for positive prediction to maximize F1 score on our training set."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.746,
+                0.884,
+                0.842
+            ],
+            "angle": 0,
+            "content": "- Zero/few-shot prompting. We evaluate FlanT5 XXL (Chung et al., 2022) and FlanUL2 (Tay et al., 2023) in both a zero- and a few-shot setting. These models instruction tuned for a large set of tasks, including NLI tasks. (See Appendix B for details.)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.855,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "- Fine-tuning experiments. We fine-tune T5 and Flan-T5 on our dataset (SDOH-NLI), on ANLI, and on a mixture of both at various model sizes. (See Appendix C for details.)"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.532,
+                0.43,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4791"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.195,
+                0.082,
+                0.803,
+                0.311
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td>Sentence-T5 11B</td><td>.2826</td><td>.523</td><td>.3669</td></tr><tr><td>GTR XXL</td><td>.3148</td><td>.4885</td><td>.3829</td></tr><tr><td>SENTLI</td><td>.4438</td><td>.8161</td><td>.5749</td></tr><tr><td>SENTLI, threshold tuned</td><td>.5721</td><td>.7299</td><td>.6414</td></tr><tr><td>Flan-T5 XXL 0-shot</td><td>.6867</td><td>.6552</td><td>.6706</td></tr><tr><td>Flan-T5 XXL 5-shot</td><td>.6297</td><td>.6839</td><td>.6556</td></tr><tr><td>Flan-UL2 0-shot</td><td>.5922</td><td>.7011</td><td>.6421</td></tr><tr><td>Flan-UL2 5-shot</td><td>.5536</td><td>.7126</td><td>.6231</td></tr><tr><td>Flan-T5 XXL finetuned on ANLI</td><td>.356</td><td>.8736</td><td>.5058</td></tr><tr><td>Flan-T5 XXL finetuned on SDOH-NLI</td><td>.9295</td><td>.8333</td><td>.8788</td></tr><tr><td>Flan-T5 XXL finetuned on ANLI + SDOH-NLI</td><td>.8765</td><td>.8563</td><td>.8663</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.321,
+                0.884,
+                0.351
+            ],
+            "angle": 0,
+            "content": "Table 1: Scores of retrieval-based, general-purpose NLI, in-context, and fine-tuned models on the SDOH test set. See Appendix C for additional fine-tuning experiments."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.376,
+                0.219,
+                0.39
+            ],
+            "angle": 0,
+            "content": "4.1 Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.397,
+                0.486,
+                0.509
+            ],
+            "angle": 0,
+            "content": "Table 1 shows selected results (see Appendix D for more model fine-tuning metrics). First, we observe that while our problem can naturally be framed as an information retrieval problem, even state-of-the-art retrieval models perform poorly on it; formulating the problem as natural language inference yields dramatically better results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.512,
+                0.489,
+                0.625
+            ],
+            "angle": 0,
+            "content": "However, even powerful models fine-tuned for NLI, either on NLI datasets alone or as part of the much larger Flan collection, do not reliably solve our problem, with an F1 score of at most .67 on the test set. Even T5 small fine-tuned on our dataset outperforms the largest off-the-shelf models, highlighting the added value of our dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.628,
+                0.489,
+                0.789
+            ],
+            "angle": 0,
+            "content": "When prompting instruction-tuned models, including few-shot examples does not appear to add any value compared to a zero-shot setup. We conjecture that this is because the models are already familiar with the NLI task and the prompt format from instruction tuning, and presenting them with a small number of additional examples is not sufficient for teaching them the difference between this task and the NLI benchmark datasets they have seen during fine-tuning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.791,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "For fine-tuned models below XXL size, when fine-tuning on the SDOH-NLI training set only, we observed poor generalization to the held-out SDOH factors in the test set (see Appendix D for detailed metrics). Because of this, why we experimented with fine-tuning on a combination of our data and ANLI R1, a challenging general-purpose NLI dataset of a similar size. For smaller models,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.376,
+                0.884,
+                0.424
+            ],
+            "angle": 0,
+            "content": "fine-tuning on this mixture enabled robust generalization to unseen factors without sacrificing overall test performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.454,
+                0.64,
+                0.468
+            ],
+            "angle": 0,
+            "content": "4.2 Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.485,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "What makes our dataset challenging for models fine-tuned on standard NLI datasets? We emphasize that it is not that it requires any specialized (e.g., medical) knowledge: most of our examples describe everyday situations in plain English (e.g., \"The patient is retired on disability due to her knee replacements\"); human raters without any domain-specific training had no difficulty understanding them. We conjecture that the main difference is that the hypotheses of typical NLI datasets are written to satisfy a given entailment label for a given premise, whereas ours are obtained independently from the premises, and therefore their entailment status can be more subtle and ambiguous. These models also struggled with distinguishing between statements about the present and the past; e.g., Flan-UL2 erroneously predicting that \"He is not a smoker\" entails \"The person wasn't a smoker in the past.\" Also, our dataset contains a lot of irrelevant hypotheses and requires the model to correctly classify all of those; we see off-the-shelf models occasionally giving them positive labels (e.g. predicting that \"She was living alone and is now living in assisted living.\" implies \"The person drank alcohol in the past.), hurting precision. As an example, Flan-T5 fine-tuned on ANLI had comparable recall to our best models, but much worse precision."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4792"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.248,
+                0.099
+            ],
+            "angle": 0,
+            "content": "5 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.11,
+                0.493,
+                0.288
+            ],
+            "angle": 0,
+            "content": "In this work, we introduced SDOH-NLI, a new entailment dataset containing social history snippets as premises and social and behavioral determinants of health as hypotheses. Our dataset was designed both to reflect realistic use cases of SDOH extraction in clinical settings as well as to provide a high quality entailment dataset to support broader NLI research. We evaluated baseline methods using state-of-the-art public models in a variety of setups, and highlighted novel and challenging features of our dataset."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.301,
+                0.279,
+                0.318
+            ],
+            "angle": 0,
+            "content": "Acknowledgments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.327,
+                0.49,
+                0.391
+            ],
+            "angle": 0,
+            "content": "The authors would like to thank Jonas B. Kemp, Von Nguyen, Birju Patel, Martin G. Seneviratne, Andy Strunk, and Vinh Q. Tran for their valuable feedback and discussions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.403,
+                0.22,
+                0.419
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.43,
+                0.492,
+                0.784
+            ],
+            "angle": 0,
+            "content": "Our dataset is English-only, and reflects the American healthcare system. While a lot of the social and behavioral determinants of health mentioned in the data could apply elsewhere, too, their distribution and the language used to describe them could reflect U.S. norms. Also, since the dataset is based on transcription samples, the text can be cleaner than in some other settings (such as notes in EHRs), where the task could be more challenging due to typos and abbreviations that do not appear in our dataset. In such settings, performance could be improved by first using the methods of Rajkumar et al. (2022) to decode abbreviations, but we have not included this in our evaluations. Finally, our dataset only contains short note snippets, and we have not evaluated the models' ability to reconcile contradictory statements or reason about the chronology of information in longer patient records. For longer contexts, especially if the social history sections don't fit in the Transformer model's context window, we recommend evaluating the methods of Schuster et al. (2022)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.797,
+                0.267,
+                0.812
+            ],
+            "angle": 0,
+            "content": "Ethics Statement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.823,
+                0.491,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Machine learning research in the clinical domain has many ethical considerations. Given the nature of this work is to identify opportunities for care teams to improve the health outcomes of patients due to factors commonly not addressed, we think it contributes to human well-being and avoids harms."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.343
+            ],
+            "angle": 0,
+            "content": "The use of public, non-identifiable data that is released for the NLP community helps balance the need to have reproducible (i.e., honest and trustworthy) data to enable technical advances while limiting the need for sensitive, private medical records for research. We acknowledge that the purpose of the work is to identify SDOH to provide additional help and services to patients, and we warn against any use to deny patients care or services based on any factor that can be identified. Although we picked a large number of SDOH factors to test our method, we acknowledge that there may be additional factors that are important for specific patients and populations, so we encourage researchers to reflect on those possible factors and create datasets to help others study them, as well."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.369,
+                0.61,
+                0.384
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.39,
+                0.884,
+                0.432
+            ],
+            "angle": 0,
+            "content": "Abigail Arons, Sarah DeSilvey, Caroline Fichtenberg, and Laura Gottlieb. 2018. Compendium of medical terminology codes for social risk factors."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.439,
+                0.885,
+                0.519
+            ],
+            "angle": 0,
+            "content": "Samuel Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.528,
+                0.887,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.603,
+                0.885,
+                0.643
+            ],
+            "angle": 0,
+            "content": "Agency for Healthcare Research and MD. Quality, Rockville. 2020. Social determinants of health database."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.651,
+                0.885,
+                0.743
+            ],
+            "angle": 0,
+            "content": "Sebastian Gehrmann, Franck Dernoncourt, Yeran Li, Eric T Carlson, Joy T Wu, Jonathan Welt, John Foote Jr, Edward T Moseley, David W Grant, Patrick D Tyler, et al. 2018. Comparing deep learning and concept extraction based methods for patient phenotyping from clinical narratives. *PloS one*, 13(2):e0192360."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.752,
+                0.886,
+                0.858
+            ],
+            "angle": 0,
+            "content": "Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. 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 107-112."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.866,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Alistair EW Johnson, Tom J Pollard, Lu Shen, Liwei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-iii,"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.51,
+                0.39,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4793"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.135,
+                0.086,
+                0.489,
+                0.113
+            ],
+            "angle": 0,
+            "content": "a freely accessible critical care database. Scientific data, 3(1):1-9."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.121,
+                0.49,
+                0.188
+            ],
+            "angle": 0,
+            "content": "Yuta Koreeda and Christopher D Manning. 2021. Contractnli: A dataset for document-level natural language inference for contracts. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1907-1919."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.196,
+                0.49,
+                0.275
+            ],
+            "angle": 0,
+            "content": "Dmytro S Lituiev, Benjamin Lacar, Sang Pak, Peter L Abramowitsch, Emilia H De Marchis, and Thomas A Peterson. 2023. Automatic extraction of social determinants of health from medical notes of chronic lower back pain patients. J. Am. Med. Inform. Assoc."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.282,
+                0.49,
+                0.349
+            ],
+            "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,
+                0.356,
+                0.49,
+                0.423
+            ],
+            "angle": 0,
+            "content": "Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. 2023. The flan collection: Designing data and methods for effective instruction tuning. arXiv preprint arXiv:2301.13688."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.43,
+                0.49,
+                0.496
+            ],
+            "angle": 0,
+            "content": "Kevin Lybarger, Mari Ostendorf, and Meliha Yetisgen. 2021. Annotating social determinants of health using active learning, and characterizing determinants using neural event extraction. Journal of Biomedical Informatics, 113:103631."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.504,
+                0.49,
+                0.571
+            ],
+            "angle": 0,
+            "content": "Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428-3448."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.578,
+                0.49,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Jianmo Ni, Gustavo Hernandez Abrego, Noah Constant, Ji Ma, Keith Hall, Daniel Cer, and Yinfei Yang. 2022a. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 1864–1874."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.666,
+                0.49,
+                0.771
+            ],
+            "angle": 0,
+            "content": "Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith Hall, Ming-Wei Chang, and Yinfei Yang. 2022b. Large dual encoders are generalizable retrievers. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9844–9855, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.779,
+                0.49,
+                0.858
+            ],
+            "angle": 0,
+            "content": "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial nli: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.866,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Braja G Patra, Mohit M Sharma, Veer Vekaria, Prakash Adekkanattu, Olga V Patterson, Benjamin Glicksberg, Lauren A Lepow, Euijung Ryu, Joanna M Biernacka, Al'ona Furmanchuk, et al. 2021. Extracting"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.885,
+                0.139
+            ],
+            "angle": 0,
+            "content": "social determinants of health from electronic health records using natural language processing: a systematic review. Journal of the American Medical Informatics Association, 28(12):2716-2727."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.15,
+                0.885,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Adam Poliak. 2020. A survey on recognizing textual entailment as an nlp evaluation. In Proceedings of the First Workshop on Evaluation and Comparison of NLP Systems, pages 92-109."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.213,
+                0.885,
+                0.292
+            ],
+            "angle": 0,
+            "content": "Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 180-191."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.303,
+                0.885,
+                0.381
+            ],
+            "angle": 0,
+            "content": "Alvin Rajkomar, Eric Loreaux, Yuchen Liu, Jonas Kemp, Benny Li, Ming-Jun Chen, Yi Zhang, Afroz Mohiuddin, and Juraj Gottweis. 2022. Deciphering clinical abbreviations with a privacy protecting machine learning system. Nature Communications, 13(1):7456."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.392,
+                0.885,
+                0.47
+            ],
+            "angle": 0,
+            "content": "Adam Roberts, Hyung Won Chung, Anselm Levskaya, Gaurav Mishra, James Bradbury, Daniel Andor, Sharan Narang, Brian Lester, Colin Gaffney, Afroz Mohiuddin, et al. 2022. Scaling up models and data with t5x and seqio. arXiv preprint arXiv:2203.17189."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.481,
+                0.885,
+                0.574
+            ],
+            "angle": 0,
+            "content": "Tal Schuster, Sihao Chen, Senaka Buthpitiya, Alex Fabrikant, and Donald Metzler. 2022. Stretching sentence-pair NLI models to reason over long documents and clusters. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 394-412, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.584,
+                0.885,
+                0.663
+            ],
+            "angle": 0,
+            "content": "Tal Schuster, Adam Fisch, and Regina Barzilay. 2021. Get your vitamin c! robust fact verification with contrastive evidence. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 624-643."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.673,
+                0.885,
+                0.727
+            ],
+            "angle": 0,
+            "content": "Shane Storks, Qiaozi Gao, and Joyce Y Chai. 2019. Recent advances in natural language inference: A survey of benchmarks, resources, and approaches. arXiv preprint arXiv:1904.01172."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.737,
+                0.885,
+                0.815
+            ],
+            "angle": 0,
+            "content": "Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Steven Zheng, et al. 2023. UI2: Unifying language learning paradigms. In The Eleventh International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.826,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "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"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4794"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.086,
+                0.489,
+                0.113
+            ],
+            "angle": 0,
+            "content": "Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.122,
+                0.49,
+                0.176
+            ],
+            "angle": 0,
+            "content": "Özlem Uzuner, Ira Goldstein, Yuan Luo, and Isaac Kohane. 2008. Identifying patient smoking status from medical discharge records. Journal of the American Medical Informatics Association, 15(1):14-24."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.185,
+                0.49,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Yan Wang, Elizabeth S Chen, Serguei Pakhomov, Elliot Arsoniadis, Elizabeth W Carter, Elizabeth Lindemann, Indra Neil Sarkar, and Genevieve B Melton. 2015. Automated extraction of substance use information from clinical texts. In AMIA Annual Symposium Proceedings, volume 2015, page 2121. American Medical Informatics Association."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.287,
+                0.49,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. 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 1112-1122."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.389,
+                0.49,
+                0.48
+            ],
+            "angle": 0,
+            "content": "Tamara J Winden, Elizabeth S Chen, Yan Wang, Elizabeth Lindemann, and Genevieve B Melton. 2017. Residence, living situation, and living conditions information documentation in clinical practice. In AMIA Annual Symposium Proceedings, volume 2017, page 1783. American Medical Informatics Association."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.491,
+                0.489,
+                0.556
+            ],
+            "angle": 0,
+            "content": "Meliha Yetisgen, Elena Pellicer, David R Crosslin, and Lucy Vanderwende. 2016. Automatic identification of lifestyle and environmental factors from social history in clinical text. In Proceedings of AMIA 2016 Joint Summits on Translational Science."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.566,
+                0.49,
+                0.646
+            ],
+            "angle": 0,
+            "content": "Meliha Yetisgen and Lucy Vanderwende. 2017. Automatic identification of substance abuse from social history in clinical text. In Artificial Intelligence in Medicine: 16th Conference on Artificial Intelligence in Medicine, AIME 2017, Vienna, Austria, June 21-24, 2017, Proceedings 16, pages 171-181. Springer."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.084,
+                0.681,
+                0.102
+            ],
+            "angle": 0,
+            "content": "A Labeling Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.11,
+                0.845,
+                0.125
+            ],
+            "angle": 0,
+            "content": "Raters were given the following instructions:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.126,
+                0.884,
+                0.222
+            ],
+            "angle": 0,
+            "content": "\"In this task, you will be given a list of snippets from transcribed medical records, describing a person's social history. Your job is to select the categories that are relevant to the snippet and, in each category, select statements about the person that are supported by the snippet."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.223,
+                0.882,
+                0.27
+            ],
+            "angle": 0,
+            "content": "Only select categories and statements relevant to the subject of the sentence, NOT if they apply to someone else (such as the subject's relatives)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.271,
+                0.884,
+                0.318
+            ],
+            "angle": 0,
+            "content": "The source of the medical transcript samples is a public dataset (MTSamples.com), and not your, other raters' or users health transcript data.\""
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.32,
+                0.884,
+                0.431
+            ],
+            "angle": 0,
+            "content": "For each snippet, they were required select one or more of the following categories (including \"None of the above\" if none of the categories were relevant). If they selected a category, the statements within the category would be displayed, and they were required to select one of statements within that category, or \"None of the above\":"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.442,
+                0.619,
+                0.458
+            ],
+            "angle": 0,
+            "content": "Smoking"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.466,
+                0.833,
+                0.481
+            ],
+            "angle": 0,
+            "content": "- The person is currently a smoker."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.485,
+                0.86,
+                0.499
+            ],
+            "angle": 0,
+            "content": "- The person is currently not a smoker."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.503,
+                0.858,
+                0.518
+            ],
+            "angle": 0,
+            "content": "- The person was a smoker in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.521,
+                0.878,
+                0.535
+            ],
+            "angle": 0,
+            "content": "- The person wasn't a smoker in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.539,
+                0.722,
+                0.552
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.567,
+                0.466,
+                0.878,
+                0.552
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.565,
+                0.64,
+                0.578
+            ],
+            "angle": 0,
+            "content": "- Alcohol use"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.588,
+                0.854,
+                0.603
+            ],
+            "angle": 0,
+            "content": "- The person currently drinks alcohol."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.606,
+                0.884,
+                0.636
+            ],
+            "angle": 0,
+            "content": "- The person currently does not drink alcohol."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.64,
+                0.86,
+                0.656
+            ],
+            "angle": 0,
+            "content": "- The person drank alcohol in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.659,
+                0.881,
+                0.69
+            ],
+            "angle": 0,
+            "content": "- The person did not drink alcohol in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.693,
+                0.722,
+                0.707
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.567,
+                0.588,
+                0.884,
+                0.707
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.72,
+                0.619,
+                0.734
+            ],
+            "angle": 0,
+            "content": "- Drug use"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.743,
+                0.776,
+                0.758
+            ],
+            "angle": 0,
+            "content": "- The person is a drug user."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.761,
+                0.804,
+                0.776
+            ],
+            "angle": 0,
+            "content": "- The person is not a drug user."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.78,
+                0.872,
+                0.794
+            ],
+            "angle": 0,
+            "content": "- The person was a drug user in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.798,
+                0.884,
+                0.813
+            ],
+            "angle": 0,
+            "content": "- The person wasn't a drug user in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.816,
+                0.77,
+                0.83
+            ],
+            "angle": 0,
+            "content": "- The person uses cocaine."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.834,
+                0.83,
+                0.848
+            ],
+            "angle": 0,
+            "content": "- The person does not use cocaine."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.852,
+                0.852,
+                0.866
+            ],
+            "angle": 0,
+            "content": "- The person used cocaine in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.87,
+                0.881,
+                0.902
+            ],
+            "angle": 0,
+            "content": "- The person did not use cocaine in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.904,
+                0.788,
+                0.919
+            ],
+            "angle": 0,
+            "content": "- The person uses marijuana."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.567,
+                0.743,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4795"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.086,
+                0.452,
+                0.101
+            ],
+            "angle": 0,
+            "content": "- The person does not use marijuana."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.104,
+                0.475,
+                0.119
+            ],
+            "angle": 0,
+            "content": "- The person used marijuana in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.122,
+                0.486,
+                0.154
+            ],
+            "angle": 0,
+            "content": "- The person did not use marijuana in the past."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.156,
+                0.488,
+                0.188
+            ],
+            "angle": 0,
+            "content": "- The person uses opioids (e.g., heroin, fentanyl, oxycodone)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.19,
+                0.488,
+                0.222
+            ],
+            "angle": 0,
+            "content": "- The person does not opioids (e.g., heroin, fentanyl, oxycodone)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.224,
+                0.488,
+                0.256
+            ],
+            "angle": 0,
+            "content": "- The person used opioids in the past (e.g., heroin, fentanyl, oxycodone)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.259,
+                0.486,
+                0.29
+            ],
+            "angle": 0,
+            "content": "- The person did not use opioids in the past (e.g., heroin, fentanyl, oxycodone)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.293,
+                0.328,
+                0.306
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.17,
+                0.086,
+                0.488,
+                0.306
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.137,
+                0.319,
+                0.252,
+                0.333
+            ],
+            "angle": 0,
+            "content": "- Employment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.343,
+                0.373,
+                0.358
+            ],
+            "angle": 0,
+            "content": "- The person is employed."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.36,
+                0.4,
+                0.376
+            ],
+            "angle": 0,
+            "content": "- The person is not employed."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.379,
+                0.442,
+                0.394
+            ],
+            "angle": 0,
+            "content": "- The person is employed part time."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.397,
+                0.366,
+                0.412
+            ],
+            "angle": 0,
+            "content": "- The person is a student."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.415,
+                0.398,
+                0.43
+            ],
+            "angle": 0,
+            "content": "- The person is a homemaker."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.433,
+                0.488,
+                0.463
+            ],
+            "angle": 0,
+            "content": "- The person is retired due to age or preference."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.467,
+                0.47,
+                0.483
+            ],
+            "angle": 0,
+            "content": "- The person is retired due to disability."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.485,
+                0.486,
+                0.515
+            ],
+            "angle": 0,
+            "content": "- The person is retired due to an unknown reason."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.52,
+                0.328,
+                0.534
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.17,
+                0.343,
+                0.488,
+                0.534
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.137,
+                0.546,
+                0.219,
+                0.561
+            ],
+            "angle": 0,
+            "content": "Housing"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.569,
+                0.486,
+                0.6
+            ],
+            "angle": 0,
+            "content": "- The person lives in their own or their family's home."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.603,
+                0.467,
+                0.619
+            ],
+            "angle": 0,
+            "content": "- The person lives in a housing facility."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.621,
+                0.369,
+                0.636
+            ],
+            "angle": 0,
+            "content": "- The person is homeless."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.64,
+                0.401,
+                0.655
+            ],
+            "angle": 0,
+            "content": "- The person is stably housed."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.657,
+                0.486,
+                0.687
+            ],
+            "angle": 0,
+            "content": "- The person's housing is unsuited to their needs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.691,
+                0.328,
+                0.706
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.17,
+                0.569,
+                0.486,
+                0.706
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.137,
+                0.718,
+                0.194,
+                0.731
+            ],
+            "angle": 0,
+            "content": "Food"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.741,
+                0.486,
+                0.771
+            ],
+            "angle": 0,
+            "content": "- The person is able to obtain food on a consistent basis."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.775,
+                0.486,
+                0.805
+            ],
+            "angle": 0,
+            "content": "- The person is not able to obtain food on a consistent basis."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.81,
+                0.488,
+                0.839
+            ],
+            "angle": 0,
+            "content": "- The person has consistent fruit and vegetable intake."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.844,
+                0.486,
+                0.875
+            ],
+            "angle": 0,
+            "content": "- The person does not have consistent fruit and vegetable intake."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.17,
+                0.878,
+                0.328,
+                0.892
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.17,
+                0.741,
+                0.488,
+                0.892
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.137,
+                0.904,
+                0.264,
+                0.919
+            ],
+            "angle": 0,
+            "content": "- Transportation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.085,
+                0.88,
+                0.101
+            ],
+            "angle": 0,
+            "content": "- The person has access to transportation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.104,
+                0.884,
+                0.135
+            ],
+            "angle": 0,
+            "content": "- The person does not have access to transportation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.139,
+                0.812,
+                0.153
+            ],
+            "angle": 0,
+            "content": "- The person has access to a car."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.157,
+                0.884,
+                0.171
+            ],
+            "angle": 0,
+            "content": "- The person does not have access to a car."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.175,
+                0.875,
+                0.19
+            ],
+            "angle": 0,
+            "content": "- The person has access to public transit."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.193,
+                0.884,
+                0.223
+            ],
+            "angle": 0,
+            "content": "- The person does not have access to public transit."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.228,
+                0.884,
+                0.259
+            ],
+            "angle": 0,
+            "content": "- The person has issues with finding transportation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.262,
+                0.723,
+                0.276
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.565,
+                0.085,
+                0.884,
+                0.276
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.533,
+                0.289,
+                0.676,
+                0.302
+            ],
+            "angle": 0,
+            "content": "Health insurance"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.312,
+                0.881,
+                0.327
+            ],
+            "angle": 0,
+            "content": "- The person has private health insurance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.33,
+                0.788,
+                0.345
+            ],
+            "angle": 0,
+            "content": "- The person is on Medicare."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.349,
+                0.788,
+                0.364
+            ],
+            "angle": 0,
+            "content": "- The person is on Medicaid."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.367,
+                0.884,
+                0.396
+            ],
+            "angle": 0,
+            "content": "- The person does not have health insurance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.401,
+                0.723,
+                0.415
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.565,
+                0.312,
+                0.884,
+                0.415
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.533,
+                0.428,
+                0.657,
+                0.443
+            ],
+            "angle": 0,
+            "content": "Social support"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.452,
+                0.81,
+                0.467
+            ],
+            "angle": 0,
+            "content": "- The person has social support."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.47,
+                0.884,
+                0.485
+            ],
+            "angle": 0,
+            "content": "- The person does not have social support."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.489,
+                0.757,
+                0.503
+            ],
+            "angle": 0,
+            "content": "- The person lives alone."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.507,
+                0.816,
+                0.521
+            ],
+            "angle": 0,
+            "content": "- The person does not live alone."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.525,
+                0.723,
+                0.538
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.565,
+                0.452,
+                0.884,
+                0.538
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.533,
+                0.551,
+                0.688,
+                0.565
+            ],
+            "angle": 0,
+            "content": "Financial situation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.575,
+                0.859,
+                0.59
+            ],
+            "angle": 0,
+            "content": "- The person is below the poverty line."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.594,
+                0.858,
+                0.609
+            ],
+            "angle": 0,
+            "content": "- The person is above the poverty line."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.612,
+                0.884,
+                0.626
+            ],
+            "angle": 0,
+            "content": "- The person is able to afford medications."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.63,
+                0.884,
+                0.659
+            ],
+            "angle": 0,
+            "content": "- The person is not able to afford medications."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.565,
+                0.665,
+                0.723,
+                0.678
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.565,
+                0.575,
+                0.884,
+                0.678
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.533,
+                0.691,
+                0.686,
+                0.705
+            ],
+            "angle": 0,
+            "content": "- None of the above"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.718,
+                0.695,
+                0.735
+            ],
+            "angle": 0,
+            "content": "B Prompting Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.743,
+                0.884,
+                0.855
+            ],
+            "angle": 0,
+            "content": "Our goal with the design of our zero/few-shot experiments was to stay close to how the models were trained and evaluated on similar NLI tasks. Since the models we used were fine-tuned on the Flan collection (Longpre et al., 2023), which includes several NLI datasets, we reused some of the Flan collections's prompt templates."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.856,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In particular, we used the templates for the SNLI and MNLI dataset which were the most relevant to our data. (We excluded templates for other datasets such as RTE, ANLI, or WNLI because the wording"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4796"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.085,
+                0.49,
+                0.325
+            ],
+            "angle": 0,
+            "content": "of some of the prompts could be slightly misleading; e.g. by referring to the premise as a \"paragraph.\") For each example, we chose a prompt template uniformly at random from the 19 templates. For few-shot experiments, we first picked a random positive example from the training set, then picked the other few-shot examples uniformly at random from the training examples. (Without explicitly forcing at least one example to be positive, it would be very likely that all the few-shot examples would be negative, given the label imbalance in the dataset. We tried dropping that requirement and sampling all examples at random, which resulted in a slight drop in model performance, as we expected.)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.327,
+                0.49,
+                0.455
+            ],
+            "angle": 0,
+            "content": "Since SNLI and MNLI have three answer options (\"yes,\" \"it is not possible to tell,\" \"no\"), we kept all three of these options in the prompt template, even though our dataset is binary. We experimented with dropping either the \"no\" or the \"it is not possible to tell\" option; both of these deviations from the original prompts resulted in slight decreases in model performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.457,
+                0.49,
+                0.631
+            ],
+            "angle": 0,
+            "content": "We use rank classification to obtain binary labels (as is customary): i.e., instead of decoding a model prediction, we score the three answer options (for three copies of each input example) and consider the prediction positive if \"yes\" has the highest score. Similarly to our experiment with SENTLI, we also tried taking softmax over the three options and using a fixed threshold for \"yes\" which maximizes F1 score on the training set; this resulted in a small drop in test F1 score (e.g., .6501 to .6222 for Flan-UL2)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.117,
+                0.646,
+                0.31,
+                0.663
+            ],
+            "angle": 0,
+            "content": "C Fine-tuning Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.672,
+                0.49,
+                0.782
+            ],
+            "angle": 0,
+            "content": "All models were fine-tuned using the T5X framework (Roberts et al., 2022) on TPUv3 chips for \\(10\\mathrm{k}\\) steps with batch size 32 and learning rate 1e-4, with the exception of the T5 small size models which were finetuned for \\(50\\mathrm{k}\\) steps. For each model, we picked the checkpoint with the highest F1 score on the validation set."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.117,
+                0.797,
+                0.48,
+                0.814
+            ],
+            "angle": 0,
+            "content": "D Additional Model Fine-tuning Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.823,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "See Table 2 for the full set of results from our finetuning experiments. Here we also include metrics on the subset of the test set consisting on the new SDOH factors not included in the training and validation sets to highlight the differences in models' ability to generalize to unseen factors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.085,
+                0.885,
+                0.18
+            ],
+            "angle": 0,
+            "content": "Note that since Flan-T5 models between sizes large and XL fine-tuned on either ANLI or our SDOH dataset alone underperformed the same models fine-tuned on the combined dataset, we did not perform these dataset ablations on smaller Flan-T5 models or on T5 below the XXL size."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4797"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.118,
+                0.363,
+                0.88,
+                0.595
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>Size</td><td>Dataset</td><td>Test P</td><td>Test R</td><td>Test F1</td><td>New factors P</td><td>New factors R</td><td>New factors F1</td></tr><tr><td>T5</td><td>small</td><td>ANLI + SDOH-NLI</td><td>.7987</td><td>.7069</td><td>.7500</td><td>.6667</td><td>.0833</td><td>.1481</td></tr><tr><td>T5</td><td>base</td><td>ANLI + SDOH-NLI</td><td>.7396</td><td>.8161</td><td>.7760</td><td>.75</td><td>.375</td><td>.5</td></tr><tr><td>T5</td><td>large</td><td>ANLI + SDOH-NLI</td><td>.8247</td><td>.7299</td><td>.7744</td><td>.75</td><td>.125</td><td>.2143</td></tr><tr><td>T5</td><td>XL</td><td>ANLI + SDOH-NLI</td><td>.8391</td><td>.8391</td><td>.8391</td><td>.9231</td><td>.5</td><td>.6486</td></tr><tr><td>T5</td><td>XXL</td><td>ANLI</td><td>.8391</td><td>.8391</td><td>.8391</td><td>.75</td><td>.1</td><td>.8571</td></tr><tr><td>T5</td><td>XXL</td><td>SDOH-NLI</td><td>.8859</td><td>.7586</td><td>.8173</td><td>.8571</td><td>.25</td><td>.3871</td></tr><tr><td>T5</td><td>XXL</td><td>ANLI + SDOH-NLI</td><td>.9024</td><td>.8506</td><td>.8757</td><td>.8889</td><td>1.</td><td>.9412</td></tr><tr><td>Flan-T5</td><td>small</td><td>ANLI + SDOH-NLI</td><td>.8075</td><td>.7471</td><td>.7761</td><td>.6667</td><td>.0833</td><td>.1481</td></tr><tr><td>Flan-T5</td><td>base</td><td>ANLI + SDOH-NLI</td><td>.7956</td><td>.8276</td><td>.8113</td><td>.6216</td><td>.9583</td><td>.7541</td></tr><tr><td>Flan-T5</td><td>large</td><td>ANLI</td><td>.1777</td><td>.8506</td><td>.2939</td><td>.2421</td><td>.9583</td><td>.3866</td></tr><tr><td>Flan-T5</td><td>large</td><td>SDOH-NLI</td><td>.8041</td><td>.6839</td><td>.7391</td><td>.2308</td><td>.125</td><td>.1622</td></tr><tr><td>Flan-T5</td><td>large</td><td>ANLI + SDOH-NLI</td><td>.8212</td><td>.8448</td><td>.8329</td><td>.8846</td><td>.9583</td><td>.92</td></tr><tr><td>Flan-T5</td><td>XL</td><td>ANLI</td><td>.2868</td><td>.8851</td><td>.4332</td><td>.7667</td><td>.9583</td><td>.8519</td></tr><tr><td>Flan-T5</td><td>XL</td><td>SDOH</td><td>.9071</td><td>.7299</td><td>.8089</td><td>1.</td><td>.125</td><td>.2222</td></tr><tr><td>Flan-T5</td><td>XL</td><td>ANLI + SDOH-NLI</td><td>.8647</td><td>.8448</td><td>.8547</td><td>.8846</td><td>.9583</td><td>.92</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>ANLI</td><td>.3560</td><td>.8736</td><td>.5058</td><td>.7931</td><td>.9583</td><td>.8679</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>SDOH-NLI</td><td>.9295</td><td>.8333</td><td>.8788</td><td>1.</td><td>.875</td><td>.9333</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>ANLI + SDOH-NLI</td><td>.8765</td><td>.8563</td><td>.8663</td><td>.8846</td><td>.9583</td><td>.92</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.605,
+                0.884,
+                0.634
+            ],
+            "angle": 0,
+            "content": "Table 2: Precision, recall, and F1 scores of various models on the full test set and on the subset of the test set consisting of the two held-out factors."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "4798"
+        }
+    ]
+]

2023/SDOH-NLI_ a Dataset for Inferring Social Determinants of Health from Clinical Notes/a4f22374-5104-49b6-a9f1-43a434ee9541_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17d95752c7da727a696486f4a06863b87d96648efc977b998349479c83bc42ad
+size 187655

2023/SDOH-NLI_ a Dataset for Inferring Social Determinants of Health from Clinical Notes/full.md

@@ -0,0 +1,294 @@
+# SDOH-NLI: a Dataset for Inferring Social Determinants of Health from Clinical Notes
+
+Adam D. Lelkes<sup>1*</sup>, Eric Loreaux<sup>2†</sup>, Tal Schuster<sup>1</sup>, Ming-Jun Chen<sup>1‡</sup>, Alvin Rajkomar<sup>1‡</sup>
+
+<sup>1</sup>Google Research, <sup>2</sup>Curai Health
+
+# Abstract
+
+Social and behavioral determinants of health (SDOH) play a significant role in shaping health outcomes, and extracting these determinants from clinical notes is a first step to help healthcare providers systematically identify opportunities to provide appropriate care and address disparities. Progress on using NLP methods for this task has been hindered by the lack of high-quality publicly available labeled data, largely due to the privacy and regulatory constraints on the use of real patients' information. This paper introduces a new dataset, SDOH-NLI, that is based on publicly available notes and which we release publicly. We formulate SDOH extraction as a natural language inference (NLI) task, and provide binary textual entailment labels obtained from human raters for a cross product of a set of social history snippets as premises and SDOH factors as hypotheses. Our dataset differs from standard NLI benchmarks in that our premises and hypotheses are obtained independently. We evaluate both "off-the-shelf" entailment models as well as models fine-tuned on our data, and highlight the ways in which our dataset appears more challenging than commonly used NLI datasets.
+
+# 1 Introduction
+
+There has been growing recognition that social and behavioral determinants of health (SDOH) play a significant role in shaping health outcomes for individuals and populations. The ability to accurately identify and extract social and behavioral determinants of health from clinical notes can provide valuable insights that can enable healthcare providers to better understand and address the underlying determinants of health that contribute to poor health outcomes and health disparities.
+
+Social determinants of health are frequently recorded in clinical notes as unstructured text, so natural language processing (NLP) can be a valuable tool for extracting actionable insights for care teams. However, research in this area often uses patient records from private health systems' electronic health records (EHRs), which makes it difficult to compare results from other health systems or even replicate the studies. The development and release of high-quality publicly available datasets could enable more reproducible research in this area.
+
+In this work, we introduce a new, public SDOH dataset based on MTSamples.com, an online collection of transcribed medical reports. Our setup is motivated by the use cases of slicing patient populations along social determinant dimensions for population analytics, and retrieving patients with certain social determinants of health to allow for more targeted outreach and intervention. Given a large set of social determinant factors, our goal is to make binary determinations for each patient about whether that patient's notes imply a particular SDOH factor. In other words, for example, we want to be able to find all patients who lack access to transportation, as opposed to just tagging transportation-related spans in their notes, as done in some previous work.
+
+To achieve this goal, we formulate the task as a textual entailment problem, with patient note snippets as the premises, SDOH factors as the hypotheses, and binary entailment labels. We use human annotators to label 1,398 social history snippets according to a curated list of 60 SDOH statements, resulting in a dataset of 29,635 labeled premise-hypothesis examples after some filtering (see Section 3 for details). We release this dataset publicly. We also evaluate state-of-the-art publicly available large language models on our data in a range of different settings (see Section 4).
+
+A notable feature of our entailment dataset is
+
+that unlike other entailment datasets, our premises and hypotheses were obtained independently, and we label the full cross product of premises and hypotheses. In traditional entailment datasets, the hypotheses are constructed to be tied to a particular premise; however, in our formulation, the hypotheses are drawn from a large set of SDOH factors that may or may not be discussed in a particular premise (drawn from a clinical note). Since all our text comes from the same domain, we still have a non-negligible fraction of positive entailment labels (albeit with a much larger label imbalance than in standard NLI benchmark datasets).
+
+This requires NLI methods to understand both the premise and the hypothesis, and defeats typical shortcuts that have been observed to work for other entailment datasets, such as guessing the label based on the hypothesis alone, or relying on simple syntactic clues such as the presence of negations (see Section 2.2). Indeed, even though our task does not require domain-specific knowledge, we observe that state-of-the-art models struggle to generalize from common NLI benchmarks to our dataset, and highlight typical failure cases (see Section 4).
+
+We evaluate both off-the-shelf and fine-tuned models in different setups: dual encoders, zero/few-shot prompting, and binary classification. We show that state-of-the-art off-the-shelf models, even if they were fine-tuned on various NLI datasets, do not reliably solve our problem; on the other hand, models fine-tuned on our training set robustly generalize both to unseen notes and to unseen factors, providing evidence for the usefulness of our dataset for model fine-tuning and evaluation.
+
+# 2 Related Work
+
+# 2.1 Social and Behavioral Determinants of Health
+
+There has been a lot of interest in using NLP techniques for SDOH extraction; see Patra et al. (2021) for a recent survey. The range and granularity of SDOH factors vary considerably across different papers. There has also been a range of methods used, from rule-based heuristics, to n-grams, to fine-tuning pretrained Transformer models.
+
+Many previous research studies on SDOH extraction were performed on EHR data from particular health systems and are not released publicly. Exceptions include the i2b2 NLP Smoking Challenge (Uzuner et al., 2008), which classified 502
+
+deidentified medical discharge records for smoking status only, and small number of datasets based on MIMIC-III (Johnson et al., 2016), a large publicly available database of deidentified health records of patients who stayed in critical care units of Beth Israel Deaconess Medical Center between 2001 and 2012. For example, Gehrmann et al. (2018) annotated MIMIC-III with binary labels for certain "phenotypes" including alcohol and substance abuse, and Lybarger et al. (2021) annotated note spans with SDOH information.
+
+We are aware of four other previous SDOH-related papers which used MTSamples data (Wang et al., 2015; Winden et al., 2017; Yetisgen et al., 2016; Yetisgen and Vanderwende, 2017), all of which focused on extracting and tagging SDOH-related spans from social history sections.
+
+Among previous papers, the one methodologically closest to ours is Lituiev et al. (2023) which also formulated SDOH extraction as an entailment problem and evaluated RoBERTa (Liu et al., 2019) fine-tuned on ANLI (Nie et al., 2020) on clinical notes from UCSF, without any in-domain fine-tuning experiments.
+
+# 2.2 Natural Language Inference
+
+Natural language inference (NLI), also called recognizing textual entailment (RTE), has been a very well-studied NLP task; see e.g. Storks et al. (2019); Poliak (2020) for recent surveys. Many standard NLI datasets, such as SNLI (Bowman et al., 2015) or MultiNLI (Williams et al., 2018), are obtained by automatically collecting premises and, for each premise and target entailment label, having human annotators write a hypothesis with the specified entailment relation to the premise. (Even some of the datasets specifically designed to address these datasets' shortcomings, such as ANLI (Nie et al., 2020), follow a similar setup.) It has been observed that this leads to annotation artifacts that lets models do well on these tasks without requiring true understanding of entailment, including by using simple syntactic heuristics (McCoy et al., 2019) or by completely ignoring the premise, and considering only the hypothesis (Gururangan et al., 2018; Poliak et al., 2018). ContractNLI (Koreeda and Manning, 2021) is an example of a previous dataset which used the same fixed set of hypotheses for all the premises.
+
+# 3 Dataset Construction
+
+We scraped all 5,003 medical reports from MTSamples. Within these reports, we obtained 1,030 note sections related to social history by searching for a collection of note section titles identified as synonyms of social history. We then split each note section into sentences, resulting in 3,281 sentences. Many sentences (such as "He is married") appear in multiple notes; after dedduplication, we have 1,398 unique sentences.
+
+We manually curated a collection of SDOH factors primarily from two sources, the AHRQ Social Determinants of Health Database (for Healthcare Research and Quality, 2020) and UCSF SIREN's Compendium of medical terminology codes for social risk factors (Arons et al., 2018). We rephrased each factor as a full English sentence stating a fact about a person; e.g. "The person is employed." For binary factors, we included a statement and its negation; e.g. "The person currently drinks alcohol" and "The person currently does not drink alcohol." For factors with multiple potential values (such as housing status), we aimed to list all the common options. We grouped these 60 statements into 10 categories: smoking, alcohol use, drug use, employment, housing, food, transportation, health insurance, social support, and financial situation. See the full list in Appendix A.
+
+For each social history sentence, we asked human raters to select all the relevant categories and, within each category, all the statements that are supported by the social history snippet. Each snippet was rated by at least three raters. For each (snippet, statement) pair, we took the majority vote of raters to get binary entailment labels. Rater agreement was high, with a Krippendorff's alpha of 0.97 (computed over the binary entailment labels provided by different raters for (premise, hypothesis) pairs).
+
+We removed all SDOH statements which were not entailed by any of the social history snippets as well as all note sentences that were not relevant to any SDOH categories. Finally, after inspecting the dataset, we removed three SDOH factors ("The person is stably housed.", "The person has social support.", "The person does not have social support.") because raters weren't able to consistently give correct ratings. That left us with 787 unique sentences and 38 SDOH factors. Since each snippet would typically only entail one factor or a small number of factors, the resulting dataset is heavily imbalanced: only  $4.6\%$  of the labels are positive.
+
+We split the dataset along the snippets into training, validation, and test sets with a 70:15:15 ratio, with the following modification: we remove a single pair of factors, "The person lives alone" and "The person does not live alone," from the training and validation sets, in order to evaluate fine-tuned models' ability to generalize to unseen factors. In other words, the training, validation, and test sets have disjoint note snippets but the same SDOH factors, except for the test set which also contains an additional two factors that are not present in the training or validation sets.
+
+# 4 Model Evaluation
+
+Since the typical use case is retrieving patients with a particular SDOH factor, and because of the heavy label imbalance in our dataset, we use precision, recall, and F1 score as our evaluation metrics.
+
+We evaluate state-of-the-art public models in four different setups:
+
+- Treating the problem as a retrieval task with SDOH factors as queries and note snippets as documents. We evaluate Sentence-T5 (Ni et al., 2022a) and GTR (Ni et al., 2022b), two state-of-the-art dual encoder models. We select the cosine similarity threshold which maximizes F1 score on the training set.  
+- A general-purpose NLI model. We evaluate the SENTLI model (Schuster et al., 2022), a state-of-the-art NLI model (T5 large fine-tuned on the SNLI (Bowman et al., 2015), MNLI (Williams et al., 2018), ANLI (Nie et al., 2020), Fever (Thorne et al., 2018), and VitaminC (Schuster et al., 2021) datasets). To help the model adapt to the label imbalance in our dataset, we also evaluate it with re-tuning the threshold for positive prediction to maximize F1 score on our training set.  
+- Zero/few-shot prompting. We evaluate FlanT5 XXL (Chung et al., 2022) and FlanUL2 (Tay et al., 2023) in both a zero- and a few-shot setting. These models instruction tuned for a large set of tasks, including NLI tasks. (See Appendix B for details.)  
+- Fine-tuning experiments. We fine-tune T5 and Flan-T5 on our dataset (SDOH-NLI), on ANLI, and on a mixture of both at various model sizes. (See Appendix C for details.)
+
+<table><tr><td>Model</td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td>Sentence-T5 11B</td><td>.2826</td><td>.523</td><td>.3669</td></tr><tr><td>GTR XXL</td><td>.3148</td><td>.4885</td><td>.3829</td></tr><tr><td>SENTLI</td><td>.4438</td><td>.8161</td><td>.5749</td></tr><tr><td>SENTLI, threshold tuned</td><td>.5721</td><td>.7299</td><td>.6414</td></tr><tr><td>Flan-T5 XXL 0-shot</td><td>.6867</td><td>.6552</td><td>.6706</td></tr><tr><td>Flan-T5 XXL 5-shot</td><td>.6297</td><td>.6839</td><td>.6556</td></tr><tr><td>Flan-UL2 0-shot</td><td>.5922</td><td>.7011</td><td>.6421</td></tr><tr><td>Flan-UL2 5-shot</td><td>.5536</td><td>.7126</td><td>.6231</td></tr><tr><td>Flan-T5 XXL finetuned on ANLI</td><td>.356</td><td>.8736</td><td>.5058</td></tr><tr><td>Flan-T5 XXL finetuned on SDOH-NLI</td><td>.9295</td><td>.8333</td><td>.8788</td></tr><tr><td>Flan-T5 XXL finetuned on ANLI + SDOH-NLI</td><td>.8765</td><td>.8563</td><td>.8663</td></tr></table>
+
+Table 1: Scores of retrieval-based, general-purpose NLI, in-context, and fine-tuned models on the SDOH test set. See Appendix C for additional fine-tuning experiments.
+
+# 4.1 Results
+
+Table 1 shows selected results (see Appendix D for more model fine-tuning metrics). First, we observe that while our problem can naturally be framed as an information retrieval problem, even state-of-the-art retrieval models perform poorly on it; formulating the problem as natural language inference yields dramatically better results.
+
+However, even powerful models fine-tuned for NLI, either on NLI datasets alone or as part of the much larger Flan collection, do not reliably solve our problem, with an F1 score of at most .67 on the test set. Even T5 small fine-tuned on our dataset outperforms the largest off-the-shelf models, highlighting the added value of our dataset.
+
+When prompting instruction-tuned models, including few-shot examples does not appear to add any value compared to a zero-shot setup. We conjecture that this is because the models are already familiar with the NLI task and the prompt format from instruction tuning, and presenting them with a small number of additional examples is not sufficient for teaching them the difference between this task and the NLI benchmark datasets they have seen during fine-tuning.
+
+For fine-tuned models below XXL size, when fine-tuning on the SDOH-NLI training set only, we observed poor generalization to the held-out SDOH factors in the test set (see Appendix D for detailed metrics). Because of this, why we experimented with fine-tuning on a combination of our data and ANLI R1, a challenging general-purpose NLI dataset of a similar size. For smaller models,
+
+fine-tuning on this mixture enabled robust generalization to unseen factors without sacrificing overall test performance.
+
+# 4.2 Discussion
+
+What makes our dataset challenging for models fine-tuned on standard NLI datasets? We emphasize that it is not that it requires any specialized (e.g., medical) knowledge: most of our examples describe everyday situations in plain English (e.g., "The patient is retired on disability due to her knee replacements"); human raters without any domain-specific training had no difficulty understanding them. We conjecture that the main difference is that the hypotheses of typical NLI datasets are written to satisfy a given entailment label for a given premise, whereas ours are obtained independently from the premises, and therefore their entailment status can be more subtle and ambiguous. These models also struggled with distinguishing between statements about the present and the past; e.g., Flan-UL2 erroneously predicting that "He is not a smoker" entails "The person wasn't a smoker in the past." Also, our dataset contains a lot of irrelevant hypotheses and requires the model to correctly classify all of those; we see off-the-shelf models occasionally giving them positive labels (e.g. predicting that "She was living alone and is now living in assisted living." implies "The person drank alcohol in the past.), hurting precision. As an example, Flan-T5 fine-tuned on ANLI had comparable recall to our best models, but much worse precision.
+
+# 5 Conclusion
+
+In this work, we introduced SDOH-NLI, a new entailment dataset containing social history snippets as premises and social and behavioral determinants of health as hypotheses. Our dataset was designed both to reflect realistic use cases of SDOH extraction in clinical settings as well as to provide a high quality entailment dataset to support broader NLI research. We evaluated baseline methods using state-of-the-art public models in a variety of setups, and highlighted novel and challenging features of our dataset.
+
+# Acknowledgments
+
+The authors would like to thank Jonas B. Kemp, Von Nguyen, Birju Patel, Martin G. Seneviratne, Andy Strunk, and Vinh Q. Tran for their valuable feedback and discussions.
+
+# Limitations
+
+Our dataset is English-only, and reflects the American healthcare system. While a lot of the social and behavioral determinants of health mentioned in the data could apply elsewhere, too, their distribution and the language used to describe them could reflect U.S. norms. Also, since the dataset is based on transcription samples, the text can be cleaner than in some other settings (such as notes in EHRs), where the task could be more challenging due to typos and abbreviations that do not appear in our dataset. In such settings, performance could be improved by first using the methods of Rajkumar et al. (2022) to decode abbreviations, but we have not included this in our evaluations. Finally, our dataset only contains short note snippets, and we have not evaluated the models' ability to reconcile contradictory statements or reason about the chronology of information in longer patient records. For longer contexts, especially if the social history sections don't fit in the Transformer model's context window, we recommend evaluating the methods of Schuster et al. (2022).
+
+# Ethics Statement
+
+Machine learning research in the clinical domain has many ethical considerations. Given the nature of this work is to identify opportunities for care teams to improve the health outcomes of patients due to factors commonly not addressed, we think it contributes to human well-being and avoids harms.
+
+The use of public, non-identifiable data that is released for the NLP community helps balance the need to have reproducible (i.e., honest and trustworthy) data to enable technical advances while limiting the need for sensitive, private medical records for research. We acknowledge that the purpose of the work is to identify SDOH to provide additional help and services to patients, and we warn against any use to deny patients care or services based on any factor that can be identified. Although we picked a large number of SDOH factors to test our method, we acknowledge that there may be additional factors that are important for specific patients and populations, so we encourage researchers to reflect on those possible factors and create datasets to help others study them, as well.
+
+# References
+
+Abigail Arons, Sarah DeSilvey, Caroline Fichtenberg, and Laura Gottlieb. 2018. Compendium of medical terminology codes for social risk factors.  
+Samuel Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416.  
+Agency for Healthcare Research and MD. Quality, Rockville. 2020. Social determinants of health database.  
+Sebastian Gehrmann, Franck Dernoncourt, Yeran Li, Eric T Carlson, Joy T Wu, Jonathan Welt, John Foote Jr, Edward T Moseley, David W Grant, Patrick D Tyler, et al. 2018. Comparing deep learning and concept extraction based methods for patient phenotyping from clinical narratives. *PloS one*, 13(2):e0192360.  
+Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. 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 107-112.  
+Alistair EW Johnson, Tom J Pollard, Lu Shen, Liwei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-iii,
+
+a freely accessible critical care database. Scientific data, 3(1):1-9.  
+Yuta Koreeda and Christopher D Manning. 2021. Contractnli: A dataset for document-level natural language inference for contracts. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1907-1919.  
+Dmytro S Lituiev, Benjamin Lacar, Sang Pak, Peter L Abramowitsch, Emilia H De Marchis, and Thomas A Peterson. 2023. Automatic extraction of social determinants of health from medical notes of chronic lower back pain patients. J. Am. Med. Inform. Assoc.  
+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.  
+Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. 2023. The flan collection: Designing data and methods for effective instruction tuning. arXiv preprint arXiv:2301.13688.  
+Kevin Lybarger, Mari Ostendorf, and Meliha Yetisgen. 2021. Annotating social determinants of health using active learning, and characterizing determinants using neural event extraction. Journal of Biomedical Informatics, 113:103631.  
+Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428-3448.  
+Jianmo Ni, Gustavo Hernandez Abrego, Noah Constant, Ji Ma, Keith Hall, Daniel Cer, and Yinfei Yang. 2022a. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 1864–1874.  
+Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith Hall, Ming-Wei Chang, and Yinfei Yang. 2022b. Large dual encoders are generalizable retrievers. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9844–9855, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial nli: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901.  
+Braja G Patra, Mohit M Sharma, Veer Vekaria, Prakash Adekkanattu, Olga V Patterson, Benjamin Glicksberg, Lauren A Lepow, Euijung Ryu, Joanna M Biernacka, Al'ona Furmanchuk, et al. 2021. Extracting
+
+social determinants of health from electronic health records using natural language processing: a systematic review. Journal of the American Medical Informatics Association, 28(12):2716-2727.  
+Adam Poliak. 2020. A survey on recognizing textual entailment as an nlp evaluation. In Proceedings of the First Workshop on Evaluation and Comparison of NLP Systems, pages 92-109.  
+Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 180-191.  
+Alvin Rajkomar, Eric Loreaux, Yuchen Liu, Jonas Kemp, Benny Li, Ming-Jun Chen, Yi Zhang, Afroz Mohiuddin, and Juraj Gottweis. 2022. Deciphering clinical abbreviations with a privacy protecting machine learning system. Nature Communications, 13(1):7456.  
+Adam Roberts, Hyung Won Chung, Anselm Levskaya, Gaurav Mishra, James Bradbury, Daniel Andor, Sharan Narang, Brian Lester, Colin Gaffney, Afroz Mohiuddin, et al. 2022. Scaling up models and data with t5x and seqio. arXiv preprint arXiv:2203.17189.  
+Tal Schuster, Sihao Chen, Senaka Buthpitiya, Alex Fabrikant, and Donald Metzler. 2022. Stretching sentence-pair NLI models to reason over long documents and clusters. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 394-412, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Tal Schuster, Adam Fisch, and Regina Barzilay. 2021. Get your vitamin c! robust fact verification with contrastive evidence. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 624-643.  
+Shane Storks, Qiaozi Gao, and Joyce Y Chai. 2019. Recent advances in natural language inference: A survey of benchmarks, resources, and approaches. arXiv preprint arXiv:1904.01172.  
+Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Steven Zheng, et al. 2023. UI2: Unifying language learning paradigms. In The Eleventh International Conference on Learning Representations.  
+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.
+
+Özlem Uzuner, Ira Goldstein, Yuan Luo, and Isaac Kohane. 2008. Identifying patient smoking status from medical discharge records. Journal of the American Medical Informatics Association, 15(1):14-24.
+
+Yan Wang, Elizabeth S Chen, Serguei Pakhomov, Elliot Arsoniadis, Elizabeth W Carter, Elizabeth Lindemann, Indra Neil Sarkar, and Genevieve B Melton. 2015. Automated extraction of substance use information from clinical texts. In AMIA Annual Symposium Proceedings, volume 2015, page 2121. American Medical Informatics Association.
+
+Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. 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 1112-1122.
+
+Tamara J Winden, Elizabeth S Chen, Yan Wang, Elizabeth Lindemann, and Genevieve B Melton. 2017. Residence, living situation, and living conditions information documentation in clinical practice. In AMIA Annual Symposium Proceedings, volume 2017, page 1783. American Medical Informatics Association.
+
+Meliha Yetisgen, Elena Pellicer, David R Crosslin, and Lucy Vanderwende. 2016. Automatic identification of lifestyle and environmental factors from social history in clinical text. In Proceedings of AMIA 2016 Joint Summits on Translational Science.
+
+Meliha Yetisgen and Lucy Vanderwende. 2017. Automatic identification of substance abuse from social history in clinical text. In Artificial Intelligence in Medicine: 16th Conference on Artificial Intelligence in Medicine, AIME 2017, Vienna, Austria, June 21-24, 2017, Proceedings 16, pages 171-181. Springer.
+
+# A Labeling Setup
+
+Raters were given the following instructions:
+
+"In this task, you will be given a list of snippets from transcribed medical records, describing a person's social history. Your job is to select the categories that are relevant to the snippet and, in each category, select statements about the person that are supported by the snippet.
+
+Only select categories and statements relevant to the subject of the sentence, NOT if they apply to someone else (such as the subject's relatives).
+
+The source of the medical transcript samples is a public dataset (MTSamples.com), and not your, other raters' or users health transcript data."
+
+For each snippet, they were required select one or more of the following categories (including "None of the above" if none of the categories were relevant). If they selected a category, the statements within the category would be displayed, and they were required to select one of statements within that category, or "None of the above":
+
+Smoking
+
+- The person is currently a smoker.  
+- The person is currently not a smoker.  
+- The person was a smoker in the past.  
+- The person wasn't a smoker in the past.  
+- None of the above
+
+- Alcohol use
+
+- The person currently drinks alcohol.  
+- The person currently does not drink alcohol.  
+- The person drank alcohol in the past.  
+- The person did not drink alcohol in the past.  
+- None of the above
+
+- Drug use
+
+- The person is a drug user.  
+- The person is not a drug user.  
+- The person was a drug user in the past.  
+- The person wasn't a drug user in the past.  
+- The person uses cocaine.  
+- The person does not use cocaine.  
+- The person used cocaine in the past.  
+- The person did not use cocaine in the past.  
+- The person uses marijuana.
+
+- The person does not use marijuana.  
+- The person used marijuana in the past.  
+- The person did not use marijuana in the past.  
+- The person uses opioids (e.g., heroin, fentanyl, oxycodone).  
+- The person does not opioids (e.g., heroin, fentanyl, oxycodone).  
+- The person used opioids in the past (e.g., heroin, fentanyl, oxycodone).  
+- The person did not use opioids in the past (e.g., heroin, fentanyl, oxycodone).  
+- None of the above
+
+# - Employment
+
+- The person is employed.  
+- The person is not employed.  
+- The person is employed part time.  
+- The person is a student.  
+- The person is a homemaker.  
+- The person is retired due to age or preference.  
+- The person is retired due to disability.  
+- The person is retired due to an unknown reason.  
+- None of the above
+
+# Housing
+
+- The person lives in their own or their family's home.  
+- The person lives in a housing facility.  
+- The person is homeless.  
+- The person is stably housed.  
+- The person's housing is unsuited to their needs.  
+- None of the above
+
+# Food
+
+- The person is able to obtain food on a consistent basis.  
+- The person is not able to obtain food on a consistent basis.  
+- The person has consistent fruit and vegetable intake.  
+- The person does not have consistent fruit and vegetable intake.  
+- None of the above
+
+# - Transportation
+
+- The person has access to transportation.  
+- The person does not have access to transportation.  
+- The person has access to a car.  
+- The person does not have access to a car.  
+- The person has access to public transit.  
+- The person does not have access to public transit.  
+- The person has issues with finding transportation.  
+- None of the above
+
+# Health insurance
+
+- The person has private health insurance.  
+- The person is on Medicare.  
+- The person is on Medicaid.  
+- The person does not have health insurance.  
+- None of the above
+
+# Social support
+
+- The person has social support.  
+- The person does not have social support.  
+- The person lives alone.  
+- The person does not live alone.  
+- None of the above
+
+# Financial situation
+
+- The person is below the poverty line.  
+- The person is above the poverty line.  
+- The person is able to afford medications.  
+- The person is not able to afford medications.  
+- None of the above
+
+# - None of the above
+
+# B Prompting Setup
+
+Our goal with the design of our zero/few-shot experiments was to stay close to how the models were trained and evaluated on similar NLI tasks. Since the models we used were fine-tuned on the Flan collection (Longpre et al., 2023), which includes several NLI datasets, we reused some of the Flan collections's prompt templates.
+
+In particular, we used the templates for the SNLI and MNLI dataset which were the most relevant to our data. (We excluded templates for other datasets such as RTE, ANLI, or WNLI because the wording
+
+of some of the prompts could be slightly misleading; e.g. by referring to the premise as a "paragraph.") For each example, we chose a prompt template uniformly at random from the 19 templates. For few-shot experiments, we first picked a random positive example from the training set, then picked the other few-shot examples uniformly at random from the training examples. (Without explicitly forcing at least one example to be positive, it would be very likely that all the few-shot examples would be negative, given the label imbalance in the dataset. We tried dropping that requirement and sampling all examples at random, which resulted in a slight drop in model performance, as we expected.)
+
+Since SNLI and MNLI have three answer options ("yes," "it is not possible to tell," "no"), we kept all three of these options in the prompt template, even though our dataset is binary. We experimented with dropping either the "no" or the "it is not possible to tell" option; both of these deviations from the original prompts resulted in slight decreases in model performance.
+
+We use rank classification to obtain binary labels (as is customary): i.e., instead of decoding a model prediction, we score the three answer options (for three copies of each input example) and consider the prediction positive if "yes" has the highest score. Similarly to our experiment with SENTLI, we also tried taking softmax over the three options and using a fixed threshold for "yes" which maximizes F1 score on the training set; this resulted in a small drop in test F1 score (e.g., .6501 to .6222 for Flan-UL2).
+
+# C Fine-tuning Setup
+
+All models were fine-tuned using the T5X framework (Roberts et al., 2022) on TPUv3 chips for  $10\mathrm{k}$  steps with batch size 32 and learning rate 1e-4, with the exception of the T5 small size models which were finetuned for  $50\mathrm{k}$  steps. For each model, we picked the checkpoint with the highest F1 score on the validation set.
+
+# D Additional Model Fine-tuning Results
+
+See Table 2 for the full set of results from our finetuning experiments. Here we also include metrics on the subset of the test set consisting on the new SDOH factors not included in the training and validation sets to highlight the differences in models' ability to generalize to unseen factors.
+
+Note that since Flan-T5 models between sizes large and XL fine-tuned on either ANLI or our SDOH dataset alone underperformed the same models fine-tuned on the combined dataset, we did not perform these dataset ablations on smaller Flan-T5 models or on T5 below the XXL size.
+
+<table><tr><td>Model</td><td>Size</td><td>Dataset</td><td>Test P</td><td>Test R</td><td>Test F1</td><td>New factors P</td><td>New factors R</td><td>New factors F1</td></tr><tr><td>T5</td><td>small</td><td>ANLI + SDOH-NLI</td><td>.7987</td><td>.7069</td><td>.7500</td><td>.6667</td><td>.0833</td><td>.1481</td></tr><tr><td>T5</td><td>base</td><td>ANLI + SDOH-NLI</td><td>.7396</td><td>.8161</td><td>.7760</td><td>.75</td><td>.375</td><td>.5</td></tr><tr><td>T5</td><td>large</td><td>ANLI + SDOH-NLI</td><td>.8247</td><td>.7299</td><td>.7744</td><td>.75</td><td>.125</td><td>.2143</td></tr><tr><td>T5</td><td>XL</td><td>ANLI + SDOH-NLI</td><td>.8391</td><td>.8391</td><td>.8391</td><td>.9231</td><td>.5</td><td>.6486</td></tr><tr><td>T5</td><td>XXL</td><td>ANLI</td><td>.8391</td><td>.8391</td><td>.8391</td><td>.75</td><td>.1</td><td>.8571</td></tr><tr><td>T5</td><td>XXL</td><td>SDOH-NLI</td><td>.8859</td><td>.7586</td><td>.8173</td><td>.8571</td><td>.25</td><td>.3871</td></tr><tr><td>T5</td><td>XXL</td><td>ANLI + SDOH-NLI</td><td>.9024</td><td>.8506</td><td>.8757</td><td>.8889</td><td>1.</td><td>.9412</td></tr><tr><td>Flan-T5</td><td>small</td><td>ANLI + SDOH-NLI</td><td>.8075</td><td>.7471</td><td>.7761</td><td>.6667</td><td>.0833</td><td>.1481</td></tr><tr><td>Flan-T5</td><td>base</td><td>ANLI + SDOH-NLI</td><td>.7956</td><td>.8276</td><td>.8113</td><td>.6216</td><td>.9583</td><td>.7541</td></tr><tr><td>Flan-T5</td><td>large</td><td>ANLI</td><td>.1777</td><td>.8506</td><td>.2939</td><td>.2421</td><td>.9583</td><td>.3866</td></tr><tr><td>Flan-T5</td><td>large</td><td>SDOH-NLI</td><td>.8041</td><td>.6839</td><td>.7391</td><td>.2308</td><td>.125</td><td>.1622</td></tr><tr><td>Flan-T5</td><td>large</td><td>ANLI + SDOH-NLI</td><td>.8212</td><td>.8448</td><td>.8329</td><td>.8846</td><td>.9583</td><td>.92</td></tr><tr><td>Flan-T5</td><td>XL</td><td>ANLI</td><td>.2868</td><td>.8851</td><td>.4332</td><td>.7667</td><td>.9583</td><td>.8519</td></tr><tr><td>Flan-T5</td><td>XL</td><td>SDOH</td><td>.9071</td><td>.7299</td><td>.8089</td><td>1.</td><td>.125</td><td>.2222</td></tr><tr><td>Flan-T5</td><td>XL</td><td>ANLI + SDOH-NLI</td><td>.8647</td><td>.8448</td><td>.8547</td><td>.8846</td><td>.9583</td><td>.92</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>ANLI</td><td>.3560</td><td>.8736</td><td>.5058</td><td>.7931</td><td>.9583</td><td>.8679</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>SDOH-NLI</td><td>.9295</td><td>.8333</td><td>.8788</td><td>1.</td><td>.875</td><td>.9333</td></tr><tr><td>Flan-T5</td><td>XXL</td><td>ANLI + SDOH-NLI</td><td>.8765</td><td>.8563</td><td>.8663</td><td>.8846</td><td>.9583</td><td>.92</td></tr></table>
+
+Table 2: Precision, recall, and F1 scores of various models on the full test set and on the subset of the test set consisting of the two held-out factors.

2023/SDOH-NLI_ a Dataset for Inferring Social Determinants of Health from Clinical Notes/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:723ee60bd823fc61f0e6ad7f1f5a1b8c67ad85563a3bcd685519542f9a8185c0
+size 167470

2023/SELFOOD_ Self-Supervised Out-Of-Distribution Detection via Learning to Rank/f2ac6bc4-d902-4552-a8de-3700ee905922_content_list.json

@@ -0,0 +1,1761 @@
+[
+    {
+        "type": "text",
+        "text": "SELFOOD: Self-Supervised Out-Of-Distribution Detection via Learning to Rank",
+        "text_level": 1,
+        "bbox": [
+            191,
+            89,
+            805,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dheeraj Mekala\\*, Adithya Samavedhi\\*, Chengyu Dong\\* Jingbo Shang\\*, University of California San Diego",
+        "bbox": [
+            127,
+            145,
+            875,
+            186
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{\\text{©}}$  Halicioglu Data Science Institute, University of California San Diego",
+        "bbox": [
+            211,
+            187,
+            789,
+            205
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{dmekala, asamavedhi, cdong, jshang}@ucsd.edu",
+        "bbox": [
+            329,
+            206,
+            672,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Deep neural classifiers trained with cross-entropy (CE) loss often suffer from poor calibration, necessitating the task of out-of-distribution (OOD) detection. Traditional supervised OOD detection methods require expensive manual annotation of in-distribution and OOD samples. To address the annotation bottleneck, we introduce SELFOOD, a self-supervised OOD detection method that requires only in-distribution samples as supervision. We cast OOD detection as an inter-document intralabel (IDIL) ranking problem and train the classifier with our pairwise ranking loss, referred to as IDIL loss. Specifically, given a set of in-distribution documents and their labels, for each label, we train the classifier to rank the softmax scores of documents belonging to that label to be higher than the scores of documents that belong to other labels. Unlike CE loss, our IDIL loss reaches zero when the desired confidence ranking is achieved and gradients are backpropagated to decrease probabilities associated with incorrect labels rather than continuously increasing the probability of the correct label. Extensive experiments with several classifiers on multiple classification datasets demonstrate the effectiveness of our method in both coarse-and fine-grained settings.",
+        "bbox": [
+            141,
+            282,
+            460,
+            680
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            695,
+            258,
+            709
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Deep neural networks (DNNs) are ubiquitously used for text classification (Liu et al., 2019; Devlin et al., 2019; Yang et al., 2019; Brown et al., 2020). However, they are generally poorly calibrated, resulting in erroneously high-confidence scores for both in-distribution and out-of-distribution (OOD) samples (Szegedy et al., 2013; Nguyen et al., 2015; Guo et al., 2017; Mekala et al., 2022a). Such poor calibration makes DNNs unreliable, and OOD detection task vital for the safe deployment of",
+        "bbox": [
+            112,
+            722,
+            489,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c911208f1877b4fdcb0be004bc715b8b1c27785693980a6d460f09e969f8b0ce.jpg",
+        "image_caption": [
+            "(a) CE Loss"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            250,
+            884,
+            329
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/47bfbf0f3c5f5ed60d454d099ba9821a6ea8a0bf8dadf2144d351de3798331a5.jpg",
+        "image_caption": [
+            "(b) IDIL Loss",
+            "Figure 1: CE Loss and IDIL loss optimization for documents  $D_{1} \\in \\text{Sports}$  and  $D_{2} \\in \\text{Arts}$ . CE loss increases scores corresponding to the Sports for  $D_{1}$  and Arts for  $D_{2}$ , implying an intra-document comparison. Instead, IDIL loss compares softmax scores in an inter-document intra-label fashion where it reduces scores corresponding to Sports for  $D_{2}$  to be less than that of  $D_{1}$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            359,
+            885,
+            435
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "deep learning models in safety-critical applications (Moon et al., 2020).",
+        "bbox": [
+            507,
+            576,
+            882,
+            606
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Traditional supervised OOD detection methods (Hendrycks et al., 2018; Larson et al., 2019; Kamath et al., 2020; Zeng et al., 2021b) assume access to high-quality manually annotated in-distribution and OOD samples. However, this requires extensive annotation of OOD samples belonging to diverse distributions, which is expensive to obtain. Moreover, text classifiers are ideally desired to be more confident on in-distribution samples than OOD samples. However, the poor calibration of DNN precludes this phenomenon.",
+        "bbox": [
+            507,
+            609,
+            884,
+            788
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address these problems, we propose SELFOOD, a self-supervised OOD detection framework that requires only in-distribution samples as supervision. To adhere to the aforementioned ranking constraint, we formulate OOD detection as an inter-document intra-label (IDIL) ranking problem and train the classifier using our pairwise ranking loss, referred to as IDIL loss. As",
+        "bbox": [
+            507,
+            790,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$\\spadesuit$  Equal Contribution",
+        "bbox": [
+            141,
+            892,
+            278,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Jingbo Shang is the corresponding author.",
+        "bbox": [
+            142,
+            904,
+            411,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10721",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10721-10734",
+        "bbox": [
+            208,
+            945,
+            786,
+            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/cae2b87cb7c1d319cdd7e34ae37f7b41387a297617da57a959bfdb29c3316ced.jpg",
+        "image_caption": [
+            "Figure 2: SELFOOD is a self-supervised framework that requires only annotated in-distribution data to train the OOD classifier. Firstly, we bucket documents based on their annotated label (in dotted lines). Then, we compare each document in a bucket with all documents in other buckets to compute IDIL loss (in solid lines). Finally, we backpropagate gradients to decrease scores associated with incorrect labels during the training of the OOD classifier."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            126,
+            93,
+            880,
+            237
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "shown in Figure 1(a), text classifiers are generally trained using cross-entropy (CE) loss (Good, 1952) in an intra-document fashion where for each document, the classifier is trained to distinguish between different labels by maximizing the score corresponding to the correct label. Instead, in our method, as shown in Figure 1(b), we propose to train in an inter-document, intra-label fashion where for each label, we train the model to rank the considered label probability score in documents belonging to the label higher compared to those not belonging to it. As OOD documents generally do not belong to any in-distribution label, we hypothesize such explicit learning to rank translates to accurately distinguishing OOD from in-distribution.",
+        "bbox": [
+            110,
+            338,
+            489,
+            580
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Moreover, minimizing CE loss involves continuous optimization to increase the probability of the correct label over the other labels, making the classifier overconfident (Wei et al., 2022). Instead, in our method, our IDIL loss function becomes zero once the desired ranking is achieved, and during training, we backpropagate gradients to decrease probabilities associated with incorrect labels rather than increasing the probability of the correct label. Theoretically, the perfect model trained using the CE loss with 0 training loss is a solution to our ranking problem, however, in Section 5.1, we empirically show that our ranking objective leads to a different solution, demonstrating the importance of the optimization procedure. Finally, it is important to note that our ranking formulation, loss function, and the self-supervised training strategy are specifically designed to improve the performance of OOD detection rather than classification accuracy.",
+        "bbox": [
+            110,
+            581,
+            489,
+            885
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We present our framework in Figure 2. Given a set of in-distribution documents and corresponding",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "labels as input, we bucket documents belonging to each label and train the classifier to rank the probabilities using our IDIL loss function. Specifically, for each document in a label bucket, we pair up with all documents in every other bucket and compute IDIL loss.",
+        "bbox": [
+            507,
+            338,
+            884,
+            434
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are summarized as follows:",
+        "bbox": [
+            526,
+            436,
+            870,
+            451
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose SELFFOOD, a novel self-supervised method to train an OOD detection model without any OOD samples as supervision.",
+            "- We formulate OOD detection as an inter-document intra-label ranking problem and optimize it using our IDIL ranking loss.",
+            "- We perform extensive experiments on multiple text classification datasets to demonstrate the effectiveness of our method in OOD detection.",
+            "We release the code on  $\\mathbf{G h i t b u^{1}}$"
+        ],
+        "bbox": [
+            507,
+            451,
+            882,
+            612
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            627,
+            665,
+            642
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional supervised methods cast the OOD detection as classification with binary labels (Kamath et al., 2020), with one additional label for unseen classes (Fei and Liu, 2016). Since manually annotating documents is expensive (Mekala and Shang, 2020; Mekala et al., 2022b), recent works have deviated from requiring such extensive annotations and leveraging various distance metrics to detect OOD samples. Leys et al. (2018); Xu et al. (2020) use Mahalanobis distance as a post-processing technique to identify OOD samples. These methods use distance-based scoring functions along with the intermediate model layer features to determine an OOD score. Lee et al. (2018); Hsu et al. (2020) use a similar distance metric called ODIN to detect",
+        "bbox": [
+            507,
+            653,
+            884,
+            894
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<https://github.com/dheeraj7596/SELF00D>",
+        "bbox": [
+            529,
+            903,
+            830,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10722",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "OOD images. An alternate approach to compensate for the lack of OOD training data involves generating pseudo-OOD data for training. Ouyang et al. (2021) propose a framework to generate high-quality OOD utterances and importance weights by selectively replacing phrases in in-domain samples. Zhan et al. (2021) generate pseudo-OOD samples for the task of intent detection using self-supervision. Zhou et al. (2021); Zeng et al. (2021a) introduce self-supervised approaches to OOD detection using a contrastive learning framework. They suggest fine-tuning transformers using a margin-based contrastive loss to learn text representations (Mekala et al., 2021) for OOD classification. Vyas et al. (2018); Li et al. (2021) treat a part of in-domain data as OOD samples as an alternate self-supervised approach. Further, Wu et al. (2022a) use a Reassigned Contrastive Loss (RCL) along with an adaptive class-dependent threshold mechanism to separate in-domain and OOD intents. Ren et al. (2019); Ggangal et al. (2020) leverage likelihood ratios crafted by generative models to classify OOD samples. Wei et al. (2022) observe that the norm of the logits keeps increasing during training, leading to overconfident outputs, and propose LogitNorm as a solution to decoupling the output norm during training optimization. Moon et al. (2020) introduce a novel Correctness Ranking Loss function in order to regularize output probabilities to produce well-ranked confidence estimates. Other calibration techniques include \"Top-label\" calibration which is used to regularize the reported probability for the predicted class (Gupta and Ramdas, 2021).",
+        "bbox": [
+            112,
+            84,
+            492,
+            632
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 SELFOOD: Self-Supervised OOD Detection",
+        "text_level": 1,
+        "bbox": [
+            112,
+            645,
+            443,
+            678
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We present the problem statement, the motivation for our ranking formulation, and our method including the loss function and its optimization strategy.",
+        "bbox": [
+            112,
+            690,
+            489,
+            740
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Problem Statement",
+        "text_level": 1,
+        "bbox": [
+            112,
+            752,
+            314,
+            766
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we work on the OOD detection task with only in-distribution samples and no OOD samples as supervision. Specifically, given a labeled dataset  $\\mathcal{D}^{InD} = \\{(x_1,y_1),(x_2,y_2),\\ldots (x_n,y_n)\\}$  sampled from a distribution space  $(\\mathcal{X},\\mathcal{C})$  where documents  $x_{i}\\in \\mathcal{X}$  and labels  $y_{i}\\in \\mathcal{C}$  as input, our goal is to train an OOD detector  $\\mathcal{M}$  that accurately distinguishes in-distribution documents  $\\mathcal{D}^{InD}$  and OOD documents  $\\mathcal{D}^{OOD}\\notin (\\mathcal{X},\\mathcal{C})$  without any",
+        "bbox": [
+            112,
+            774,
+            489,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "OOD documents required for training.",
+        "bbox": [
+            509,
+            84,
+            794,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Motivation",
+        "text_level": 1,
+        "bbox": [
+            509,
+            111,
+            643,
+            124
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Numerous neural text classifiers have been proposed, incorporating multiple hidden layers (Rosenblatt, 1957), convolutional layers (Kim, 2014), and various types of attention mechanisms (Devlin et al., 2019; Liu et al., 2019; Radford et al., 2019). All these models culminate in a softmax head, which produces probabilities corresponding to each class. These classifiers are generally trained with CE loss in an intra-document fashion i.e. each document is considered independently and the softmax score of the true label is maximized. Such training of neural text classifiers is known to increase the magnitude of logit vectors even when most training examples are correctly classified (Wei et al., 2022), making them poorly calibrated, producing unreasonably high probabilities even for incorrect predictions (Szegedy et al., 2013; Nguyen et al., 2015; Guo et al., 2017). This diminishes their ability to maintain the desired attribute of ordinal ranking for predictions based on confidence levels, wherein a prediction exhibiting a higher confidence value should be considered more likely to be accurate than one with a lower confidence value (Moon et al., 2020). Intuitively, a text classifier possessing such quality would be a perfect OOD detector.",
+        "bbox": [
+            507,
+            131,
+            885,
+            533
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3 OOD Detection as Inter-Document Intra-Label Ranking",
+        "text_level": 1,
+        "bbox": [
+            507,
+            544,
+            828,
+            576
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In order to align with the aforementioned characteristic, we propose formulating the OOD detection as an inter-document, intra-label ranking problem. Specifically, given a set of in-distribution documents, we compare across documents but within the same label and train our model to generate higher probability score for documents belonging to the label than for documents not belonging to the label. We consider the same model architecture as any text classifier with a softmax head that generates scores corresponding to each label, however, we train it using our IDIL loss instead of CE loss. Our assumption is that an OOD document does not fall under any specific label in the in-distribution space. Hence, we anticipate that the trained model would produce lower scores for OOD documents compared to in-distribution documents. This distinction in scores is expected to facilitate easy separation between OOD and in-distribution documents.",
+        "bbox": [
+            507,
+            581,
+            885,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "IDIL Loss is a pairwise-ranking loss that enforces",
+        "bbox": [
+            509,
+            903,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10723",
+        "bbox": [
+            477,
+            927,
+            524,
+            941
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "desired ordinal ranking of confidence estimates. This loss function reaches its minimum value for a particular label when the probability of that label being the annotated label is greater than its probability when it is not the annotated label. Specifically, for documents  $x_{1}, x_{2} \\in \\mathcal{D}^{InD}$  and their corresponding annotated labels  $y_{1}, y_{2}$  where  $y_{1} \\neq y_{2}$ , IDIL loss corresponding to label  $y_{1}$  is mathematically computed as follows:",
+        "bbox": [
+            112,
+            84,
+            489,
+            227
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {I D I L} \\left(y _ {1} \\mid x _ {1}, x _ {2}\\right) = S i L U \\left(p \\left(y _ {1} \\mid x _ {2}\\right) - p \\left(y _ {1} \\mid x _ {1}\\right)\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            240,
+            487,
+            271
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $SiLU(x) = x\\sigma(x)$  is the Sigmoid Linear Unit (SiLU) function (Elfwing et al., 2018). To ensure stable training and enhance performance, we incorporate the SiLU function, a continuous variant of the Rectified Linear Unit (ReLU) (Hahnloser et al., 2000), in conjunction with the ranking loss. The SiLU function introduces smooth gradients around zero, effectively mitigating potential instability issues during training. We observe that this inclusion contributes to the overall stability in training and improved performance of the model as shown in Section 4.5. Note that, in contrast to CE loss, IDIL loss becomes zero once desired ranking is achieved, addressing the overconfidence issue.",
+        "bbox": [
+            112,
+            273,
+            489,
+            499
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4 Implementation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            508,
+            287,
+            524
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Ideally, the loss has to be computed over all possible pairs of documents for each model update. However, it is computationally expensive. Therefore, following (Toneva et al., 2019; Moon et al., 2020), we approximate the computation by considering only documents in each mini-batch. Specifically, we bucket the documents in the mini-batch based on their annotated label and pair each document in a bucket with all documents in other buckets and compute the loss. Mathematically, the loss for a mini-batch  $b$  is computed as follows:",
+        "bbox": [
+            112,
+            529,
+            489,
+            706
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\sum_ {l \\in \\mathcal {C}} \\sum_ {x _ {1} \\in b _ {l}} \\sum_ {x _ {2} \\in b _ {- l}} \\mathcal {L} _ {I D I L} (l | x _ {1}, x _ {2})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            164,
+            713,
+            436,
+            749
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $b_{l}$  denotes the set of training data points  $x$  in this batch  $b$  whose label are  $l$ , and  $b_{-l}$  denotes the set of training data points  $x$  in this batch  $b$  whose label is not  $l$ .",
+        "bbox": [
+            112,
+            758,
+            487,
+            820
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In contrast to CE loss, where the optimization involves increasing the score corresponding to the correct label, we backpropagate gradients to decrease scores associated with incorrect labels. Specifically, during the backpropagation of gradients, we detach the gradients for the subrahend",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/9bdb0da6b112ee56df8749e7fe60111befda3fc0ae57684ccb0265fa78f44c11.jpg",
+        "table_caption": [
+            "Table 1: Dataset statistics."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Domain</td><td>Criteria</td><td>#Docs</td><td>#labels</td></tr><tr><td>NYT</td><td>News</td><td>Topic</td><td>13081</td><td>26</td></tr><tr><td>Yelp</td><td>Reviews</td><td>Sentiment</td><td>70000</td><td>5</td></tr><tr><td>Clima</td><td>Climate</td><td>Question Type</td><td>17175</td><td>8</td></tr><tr><td>TREC</td><td>General</td><td>Question Type</td><td>5952</td><td>6</td></tr></table>",
+        "bbox": [
+            514,
+            98,
+            878,
+            180
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "of the difference and exclusively propagate the gradients through the minuend. In Equation 1, for instance, we detach the gradients for  $p(y_{1}|x_{1})$  and solely backpropagate the gradients through  $p(y_{1}|x_{2})$ . This detachment allows for a more controlled and selective gradient flow, aiding in the optimization process, and improvement in performance as shown in Section 4.5.",
+        "bbox": [
+            507,
+            203,
+            884,
+            331
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It is important to note that our optimization focuses solely on the inter-document ranking loss. Consequently, while the trained model would serve as a reliable OOD detector, it may not perform as effectively as a classifier.",
+        "bbox": [
+            507,
+            332,
+            884,
+            412
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            424,
+            655,
+            439
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our OOD detection method against state-of-the-art baselines with two classifiers on multiple datasets belonging to different domains.",
+        "bbox": [
+            507,
+            449,
+            882,
+            498
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            508,
+            623,
+            521
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our method and baselines on four publicly available English text classification datasets belonging to different domains. In particular, we consider the news topic classification dataset New York Times (NYT), restaurant review sentiment classification dataset Yelp, and question-type classification datasets related to climate: ClimaInsurance+ (Clima) (Laud et al., 2023), and a general domain: TREC (Li and Roth, 2002; Hovy et al., 2001). The documents within the New York Times dataset are labeled with both coarse and fine-grained labels. For our training and testing process, we utilize fine-grained labels. The dataset statistics are provided in Table 1.",
+        "bbox": [
+            507,
+            527,
+            884,
+            753
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Compared Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            764,
+            714,
+            780
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Cross Entropy Loss (CE Loss) trains a classifier using CE loss on in-distribution documents. The predicted probabilities from the classifier are used as confidence estimates for OOD detection.",
+            "- Correctness Ranking Loss (CRL) (Moon et al., 2020) is a regularization term added to the CE"
+        ],
+        "bbox": [
+            507,
+            785,
+            884,
+            881
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>2</sup>http://developer.nytimes.com/",
+        "bbox": [
+            529,
+            889,
+            759,
+            903
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://www.yelp.com/dataset/",
+        "bbox": [
+            529,
+            903,
+            759,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10724",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/9eb3baddd9b6c501202816898c5d5b370ba35d03efd9b8f3375f45595aa2fe12.jpg",
+        "table_caption": [
+            "Table 2: OOD detection results with BERT & RoBERTa classifiers. Each experiment is repeated with three random seeds and the mean scores are reported. The false-positive-rate at  $95\\%$  true-positive-rate (FPR95), minimum detection error over all thresholds (ERR), the area under the risk-coverage curve (AURC), and the area under the precision-recall curve (AUPR) using in-distribution samples as the positives are used as evaluation metrics."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">In-dist</td><td rowspan=\"2\">OOD</td><td rowspan=\"2\">Method</td><td colspan=\"4\">BERT</td><td colspan=\"4\">RoBERTa</td></tr><tr><td>FPR95(↓)</td><td>ERR(↓)</td><td>AUROC(↑)</td><td>AUPR(↑)</td><td>FPR95(↓)</td><td>ERR(↓)</td><td>AUROC(↑)</td><td>AUPR(↑)</td></tr><tr><td rowspan=\"21\">Yelp</td><td rowspan=\"7\">NYT</td><td>LogitNorm</td><td>82.0</td><td>33.2</td><td>69.4</td><td>57.9</td><td>84.4</td><td>33.9</td><td>66.6</td><td>55.0</td></tr><tr><td>OneVsRest</td><td>85.1</td><td>35.8</td><td>62.4</td><td>49.3</td><td>88.5</td><td>29.6</td><td>69.8</td><td>56.5</td></tr><tr><td>RCL</td><td>56.9</td><td>31.2</td><td>73.2</td><td>55.2</td><td>87.7</td><td>34.5</td><td>56.7</td><td>45.0</td></tr><tr><td>Bayes Approx</td><td>64.3</td><td>35.9</td><td>66.5</td><td>50.3</td><td>79.6</td><td>36.4</td><td>54.1</td><td>40.8</td></tr><tr><td>CE Loss</td><td>82.1</td><td>34.9</td><td>63.4</td><td>50.7</td><td>79.8</td><td>32.9</td><td>66.8</td><td>58.0</td></tr><tr><td>CRL</td><td>99.9</td><td>36.9</td><td>43.9</td><td>51.1</td><td>99.8</td><td>28.1</td><td>64.8</td><td>60.5</td></tr><tr><td>SELFOOD</td><td>63.2</td><td>19.6</td><td>79.4</td><td>82.7</td><td>69.5</td><td>21.6</td><td>72.6</td><td>78.8</td></tr><tr><td rowspan=\"7\">Clima</td><td>LogitNorm</td><td>79.2</td><td>25.5</td><td>72.5</td><td>54.8</td><td>82.6</td><td>25.4</td><td>70.7</td><td>54.0</td></tr><tr><td>OneVsRest</td><td>83.0</td><td>25.1</td><td>69.5</td><td>53.2</td><td>86.7</td><td>28.0</td><td>53.8</td><td>33.7</td></tr><tr><td>RCL</td><td>59.9</td><td>27.8</td><td>67.4</td><td>42.1</td><td>73.9</td><td>22.5</td><td>70.7</td><td>56.3</td></tr><tr><td>Bayes Approx</td><td>74.7</td><td>28.7</td><td>56.7</td><td>33.6</td><td>63.0</td><td>28.0</td><td>66.8</td><td>40.9</td></tr><tr><td>CE Loss</td><td>83.2</td><td>27.8</td><td>48.6</td><td>37.0</td><td>76.4</td><td>22.0</td><td>78.2</td><td>68.4</td></tr><tr><td>CRL</td><td>99.6</td><td>29.0</td><td>49.1</td><td>38.8</td><td>99.9</td><td>20.6</td><td>63.0</td><td>56.9</td></tr><tr><td>SELFOOD</td><td>17.6</td><td>4.8</td><td>97.9</td><td>96.4</td><td>65.6</td><td>16.6</td><td>73.2</td><td>77.2</td></tr><tr><td rowspan=\"7\">TREC</td><td>LogitNorm</td><td>68.2</td><td>27.0</td><td>79.8</td><td>82.1</td><td>60.8</td><td>23.0</td><td>83.9</td><td>84.4</td></tr><tr><td>OneVsRest</td><td>84.8</td><td>40.0</td><td>63.9</td><td>69.7</td><td>75.5</td><td>33.0</td><td>57.9</td><td>61.2</td></tr><tr><td>RCL</td><td>89.8</td><td>43.2</td><td>40.1</td><td>51.5</td><td>77.4</td><td>32.4</td><td>65.0</td><td>68.6</td></tr><tr><td>Bayes Approx</td><td>92.5</td><td>44.7</td><td>32.1</td><td>44.7</td><td>51.8</td><td>25.9</td><td>72.7</td><td>69.4</td></tr><tr><td>CE Loss</td><td>74.2</td><td>33.7</td><td>52.8</td><td>62.7</td><td>58.5</td><td>27.5</td><td>75.3</td><td>78.3</td></tr><tr><td>CRL</td><td>33.3</td><td>15.5</td><td>66.7</td><td>77.2</td><td>100.0</td><td>37.3</td><td>56.7</td><td>73.9</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>55.1</td><td>20.6</td><td>80.2</td><td>90.3</td></tr><tr><td rowspan=\"21\">TREC</td><td rowspan=\"7\">Yelp</td><td>LogitNorm</td><td>32.3</td><td>0.3</td><td>95.5</td><td>78.1</td><td>15.2</td><td>0.5</td><td>95.7</td><td>44.7</td></tr><tr><td>OneVsRest</td><td>4.2</td><td>0.3</td><td>97.9</td><td>83.4</td><td>39.3</td><td>0.4</td><td>91.6</td><td>63.0</td></tr><tr><td>RCL</td><td>78.5</td><td>0.6</td><td>80.2</td><td>20.2</td><td>67.7</td><td>0.6</td><td>83.0</td><td>24.2</td></tr><tr><td>Bayes Approx</td><td>46.6</td><td>0.4</td><td>93.0</td><td>61.6</td><td>63.7</td><td>0.7</td><td>87.7</td><td>21.3</td></tr><tr><td>CE Loss</td><td>27.2</td><td>0.3</td><td>94.9</td><td>71.2</td><td>11.3</td><td>0.4</td><td>96.9</td><td>62.3</td></tr><tr><td>CRL</td><td>96.6</td><td>0.7</td><td>70.9</td><td>16.5</td><td>99.8</td><td>0.7</td><td>24.3</td><td>9.7</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan=\"7\">NYT</td><td>LogitNorm</td><td>9.3</td><td>1.0</td><td>97.7</td><td>91.2</td><td>10.9</td><td>2.3</td><td>97.2</td><td>77.3</td></tr><tr><td>OneVsRest</td><td>1.8</td><td>0.5</td><td>99.0</td><td>95.8</td><td>32.5</td><td>1.8</td><td>93.2</td><td>74.8</td></tr><tr><td>RCL</td><td>63.6</td><td>3.3</td><td>84.1</td><td>45.7</td><td>74.2</td><td>3.0</td><td>80.4</td><td>45.1</td></tr><tr><td>Bayes Approx</td><td>39.1</td><td>2.5</td><td>93.5</td><td>68.1</td><td>64.9</td><td>3.0</td><td>88.4</td><td>54.4</td></tr><tr><td>CE Loss</td><td>8.4</td><td>1.3</td><td>97.6</td><td>87.5</td><td>6.8</td><td>1.4</td><td>98.1</td><td>87.1</td></tr><tr><td>CRL</td><td>97.2</td><td>3.3</td><td>73.0</td><td>35.2</td><td>99.9</td><td>3.8</td><td>22.9</td><td>12.2</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan=\"7\">Clima</td><td>LogitNorm</td><td>11.3</td><td>0.8</td><td>97.6</td><td>88.5</td><td>10.4</td><td>1.6</td><td>97.3</td><td>75.9</td></tr><tr><td>OneVsRest</td><td>3.4</td><td>0.6</td><td>98.7</td><td>93.2</td><td>36.3</td><td>1.8</td><td>91.7</td><td>60.1</td></tr><tr><td>RCL</td><td>73.5</td><td>2.4</td><td>81.2</td><td>37.1</td><td>60.7</td><td>2.0</td><td>86.4</td><td>48.2</td></tr><tr><td>Bayes Approx</td><td>51.7</td><td>2.0</td><td>90.5</td><td>54.4</td><td>61.1</td><td>2.4</td><td>87.6</td><td>40.4</td></tr><tr><td>CE Loss</td><td>15.1</td><td>1.7</td><td>96.2</td><td>69.5</td><td>12.2</td><td>1.8</td><td>96.4</td><td>67.8</td></tr><tr><td>CRL</td><td>86.3</td><td>2.2</td><td>80.7</td><td>39.5</td><td>98.9</td><td>2.7</td><td>30.1</td><td>13.0</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td></tr></table>",
+        "bbox": [
+            117,
+            148,
+            878,
+            659
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "loss to make class probabilities better confidence estimates. It estimates the true class probability to be proportional to the number of times a sample is classified correctly during training.",
+        "bbox": [
+            126,
+            684,
+            487,
+            747
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- LogitNorm (Wei et al., 2022) is a variant of CE loss that normalizes the logit vector to have a constant norm during training.",
+            "- Bayesian Approximation (Bayes Approx) (Wu et al., 2022b) calibrates distribution uncertainty for OOD detection using Monte-Carlo dropout.",
+            "- Reassigned Contrastive Learning (RCL) (Wu et al., 2022a) discriminates over-confident OOD samples using adaptive class-dependent local threshold mechanism."
+        ],
+        "bbox": [
+            114,
+            749,
+            485,
+            908
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- OneVsRest trains a one vs rest binary classifier per label. The probability scores of each label from its corresponding classifier are normalized and used as their respective confidence estimates.",
+        "bbox": [
+            509,
+            684,
+            882,
+            747
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            765,
+            729,
+            780
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We experiment with BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019) as text classifiers. For SELFOOD, we train the classifier for 5 epochs with a batch size of 16 using an AdamW optimizer. We use a learning rate of 5e-5 using a linear scheduler with no warm-up steps. For all baselines, we train the classifier for the same number of steps.",
+        "bbox": [
+            507,
+            788,
+            882,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In our evaluation, for each dataset as in-",
+        "bbox": [
+            526,
+            903,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10725",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "distribution, we treat all other datasets as OOD and compute the performance. Our evaluation follows a standard approach for each in-distribution dataset where we begin by splitting the in-distribution dataset into three subsets:  $80\\%$  for train,  $10\\%$  for val, and  $10\\%$  for test. The model is trained using the train split, and its performance is evaluated on both the test split of the in-distribution dataset and the entire OOD dataset.",
+        "bbox": [
+            112,
+            84,
+            487,
+            227
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics. We utilize evaluation metrics from (Hendrycks and Gimpel, 2017; DeVries and Taylor, 2018; Moon et al., 2020) such as the false positive rate at  $95\\%$  true positive rate (FPR95), minimum detection error over all thresholds (ERR), the area under the risk-coverage curve (AURC), and the area under the precision-recall curve (AUPR) using in-distribution samples as the positives.",
+        "bbox": [
+            112,
+            229,
+            489,
+            357
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            368,
+            218,
+            382
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We summarize the evaluation results with BERT and RoBERTa as classifiers on Yahoo, TREC as in-distribution in Table 2 and NYT, Clima as in-distribution in Appendix A.1. All experiments are run on three random seeds and the mean performance scores are reported. As shown in Table 2, we observe that SELFOOD performs better than the baselines on most of the in-distribution, OOD dataset pairs for both classifiers. A low FPR95 value indicates that the top-  $95\\%$  confident in-distribution samples, selected based on their probability scores, predominantly rank higher than a majority of OOD samples and SELFOOD achieves improvements of up to 82 points in FPR95 with Yelp as in-distribution and Clima as OOD datasets when compared to CRL with BERT classifier. SELFOOD also exhibits substantial improvements of up to 33 points in Detection Error, 48 points in AUROC, and 58 points in AUPR when compared to CE-Loss with BERT classifier. SELFOOD achieves a perfect OOD detection score for some settings such as TREC as in-distribution dataset for both BERT and RoBERTa classifiers. These results highlight the effectiveness of our ranking formulation with self-supervised training using IDIL loss.",
+        "bbox": [
+            112,
+            388,
+            489,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.5 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            818,
+            280,
+            834
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To understand the impact of each component in our IDIL loss design and implementation, we compare our method with four ablated versions with BERT classifier in Table 3: (1) SELFOOD + Grad Sub represents our method with backpropagating gradi",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/6b9b333df54869d365493aa3bec62ad48b66a87f24bb57ed4495eb585ff77d27.jpg",
+        "table_caption": [
+            "Table 3: Ablation Study."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>FPR95</td><td>ERR\nIn Dist: Clima\nOOD: NYT</td><td>AUROC</td><td>AUPR</td></tr><tr><td>SELFOOD</td><td>17.2</td><td>2.6</td><td>97.8</td><td>94.5</td></tr><tr><td>+ Grad Sub</td><td>67.9</td><td>8.9</td><td>82.4</td><td>61.1</td></tr><tr><td>+ Grad Sub &amp; Min</td><td>98.4</td><td>10.7</td><td>49.2</td><td>32.3</td></tr><tr><td>+ Intra-Doc</td><td>92.7</td><td>10.6</td><td>67.3</td><td>41.6</td></tr><tr><td>- SILu</td><td>14.6</td><td>3.9</td><td>96.7</td><td>89.3</td></tr><tr><td></td><td></td><td>In Dist: NYT\nOOD: Clima</td><td></td><td></td></tr><tr><td>SELFOOD</td><td>10.5</td><td>1.7</td><td>98.4</td><td>90.5</td></tr><tr><td>+ Grad Sub only</td><td>83.5</td><td>5.1</td><td>77.4</td><td>44.8</td></tr><tr><td>+ Grad Sub &amp; Min</td><td>96.5</td><td>5.2</td><td>60.3</td><td>36.0</td></tr><tr><td>+ Intra-Doc</td><td>48.3</td><td>2.6</td><td>93.2</td><td>80.2</td></tr><tr><td>- SiLU</td><td>19.0</td><td>2.3</td><td>97.4</td><td>86.0</td></tr></table>",
+        "bbox": [
+            515,
+            97,
+            877,
+            269
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ents through the subtrahend instead of the minuend, (2) SELFOOD  $^+$  Grad Sub & Min represents our method with gradients backpropagating through both minuend and subtrahend, (3) SELFOOD  $^+$  Intra-Doc considers intra-document comparison similar to CE loss, in addition to inter-document intra-label comparison for loss computation, and (4) SELFOOD-SiLU excludes SiLU function from the IDIL loss formulation. We also present the performance of SELFOOD for reference. SELFOOD performs better than SELFOOD  $^+$  Grad Sub & Min demonstrating that backpropagating through one part of the difference is more beneficial than both, and the comparison between SELFOOD and SELFOOD  $^+$  Grad Sub indicates that backpropagating through minuend is better than subtrahend. We observe that incorporating intra-document comparison into loss formulation leads to a decrement in the ranking ability of the model. Finally, we observe that removing the SiLU function from the IDIL loss leads to a decrease in most of the metrics.",
+        "bbox": [
+            507,
+            294,
+            884,
+            631
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Analysis & Case Studies",
+        "text_level": 1,
+        "bbox": [
+            507,
+            646,
+            754,
+            663
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present a comprehensive analysis of our proposed method from different perspectives to understand its effectiveness.",
+        "bbox": [
+            507,
+            674,
+            884,
+            722
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 CE Loss vs SELFOOD: Training Trajectory Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            734,
+            815,
+            766
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Mathematically, the model trained using CE loss is a solution to our IDIL loss optimization. However, our empirical results reveal a significant disparity in OOD detection performance between SELFOOD and model trained on CE loss. To gain a deeper understanding, we plot the CE Loss and IDIL losses for models optimized with IDIL and CE loss, as shown in Figure 3(a). We consider Clima as the in-distribution dataset and BERT as the classifier.",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10726",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8a5752b037c51fb605f668b88e60675e422e15801264977d92dd949eca9fafaf.jpg",
+        "image_caption": [
+            "(a) Losses vs Training Steps"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            99,
+            357,
+            217
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d26a489c9693b1aebcf29e79ac3e9702c1e60c1eee6c0ab51b27a036ea0909dd.jpg",
+        "image_caption": [
+            "(b) FPR vs Training Steps"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            369,
+            98,
+            610,
+            216
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/50ebefbb99df794abc58210ece2bed50f8d107877b01fecfafbde4ad5b7e3c9b.jpg",
+        "image_caption": [
+            "(c) Learned weights"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            631,
+            99,
+            810,
+            216
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/843480b4fbf688fbab1bcf7a6f8714956a78b49a5e9b82558dde471fbc6e5b1a.jpg",
+        "image_caption": [
+            "Figure 3: To understand training trajectories of CE Loss and SELFOOD, we plot their respective losses, validation FPR, and learned weights while training. We observe IDIL loss minimization is more correlated with performance and the learned weights of CE-Loss and SELFOOD show divergent paths since the beginning.",
+            "(a) CE Loss",
+            "Figure 4: Maximum softmax score vs percentile in-distribution & OOD data associated with that score. By using the least in-distribution maximum softmax score as the threshold (dotted line) for OOD classification, CE loss considers more than  $50\\%$  and CRL considers almost  $100\\%$  of the OOD data as in-distribution (red region). However, in the case of SELFOOD, we observe a clear margin in maximum softmax scores that separate OOD and in-distribution."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            288,
+            359,
+            412
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/245c4617d1f5667dd2b7a75204cdbd36e881e23daa2410db3b77879900551b3d.jpg",
+        "image_caption": [
+            "(b) CRL"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            371,
+            288,
+            608,
+            412
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/21ead48aa7a316cff58b0ae8920390495bcb30935f403e1fa7bef841be453af7.jpg",
+        "image_caption": [
+            "(c) SELFOOD"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            621,
+            292,
+            867,
+            412
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Additionally, we plot the validation FPR95 using Yelp as the OOD dataset to analyze the correlation between loss minimization and performance in Figure 3(b). Figure 3(a) illustrates that while minimizing CE loss leads to fluctuations in the IDIL loss, ultimately reaching zero, Figure 3(b) shows that the performance is more closely associated with explicit IDIL minimization. Specifically, minimizing the IDIL loss results in improved FPR95. The findings indicate that the minimization of CE loss and IDIL loss follows distinct trajectories. To validate this hypothesis, we conduct an experiment using a toy 3-dimensional dataset comprising five isometric Gaussian blobs and train a logistic regression classifier initialized with same weights, minimizing IDIL and CE-loss. We flatten the weight matrix and plot the top-3 principal components. As shown in Figure 3(c), the learned weights of CE-Loss and SELFOOD progress in different directions, thereby confirming our hypothesis.",
+        "bbox": [
+            112,
+            502,
+            489,
+            825
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Maximum softmax score Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            845,
+            423,
+            860
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We plot the maximum softmax score with the percentile in-distribution and OOD data associated with that score for CE loss, CRL, and SELFOOD",
+        "bbox": [
+            112,
+            871,
+            487,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "with BERT classifier on Clima as in-distribution and NYT as OOD datasets in Figure 4. When comparing the maximum softmax score with the percentile data associated with that score, interesting observations can be made regarding its distribution. Specifically, when using the least in-distribution maximum softmax score as the threshold for OOD classification, we find that the CE loss considers over  $50\\%$  of the OOD data and CRL considers almost  $100\\%$  of the OOD data as in-distribution. However, in the case of SELFOOD, we observe a clear margin in the maximum softmax scores that effectively separates OOD and in-distribution data. This suggests that the threshold needs to be carefully tuned for CE loss and CRL, requiring more effort and annotated data, whereas in the case of SELFOOD, we do not require such tuning. This demonstrates that SELFOOD is capable of accurately classifying between OOD and in-distribution samples based on their respective maximum softmax scores, resulting in superior performance.",
+        "bbox": [
+            507,
+            502,
+            884,
+            841
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Batch size Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            860,
+            705,
+            876
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compare documents within the mini-batch when computing the loss. Consequently, the num",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10727",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6ecab13b9bdcccdfea4e0a94db63b28458700569f0881a722f44a8c11c04c977.jpg",
+        "image_caption": [
+            "(a) NYT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            87,
+            297,
+            171
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7e66588085baee7ee42017a07ad781cf15e66b75644e44d1c10cd37b0d5c9140.jpg",
+        "image_caption": [
+            "(b) Yelp",
+            "Figure 5: Performance of SELFOOD w.r.t. batch size. We consider BERT classifier, Clima as in-distribution, and NYT, Yelp as OOD datasets. We observe an increase in performance with an increase in batch size."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            302,
+            87,
+            482,
+            171
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ber of document pairs over which the loss is computed depends on the batch size used during training. To investigate the impact of batch size on performance, we vary the batch size and evaluate the corresponding performance in Figure 5 with BERT classifier. We consider Clima as in-distribution, and NYT, Yelp as OOD datasets. As the batch size increases, the performance of the model also improves, until a batch size of 16, where the performance reaches a plateau. This observation aligns with our intuition that a larger batch size allows for more in-batch document pairs to be compared, leading to more accurate loss computation. Based on these findings, we recommend utilizing a batch size of 16 or higher to achieve optimal performance.",
+        "bbox": [
+            112,
+            263,
+            489,
+            505
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4 Fine-grained OOD Detection",
+        "text_level": 1,
+        "bbox": [
+            112,
+            517,
+            386,
+            532
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To explore the limits of SELFOOD, we consider in-distribution and OOD datasets from the same domain. The objective is to assess its ability to differentiate between samples that are closely related but still OOD, which is a challenging task. Specifically, we consider the news domain and choose NYT as the in-distribution, and AGNews (Zhang et al., 2015),  $20\\mathrm{News}^4$  as OOD datasets and train BERT classifier. As shown in Table 4, SELFOOD performs significantly better than the baselines on most of the metrics. Moreover, it also achieves near-perfect scores for the AGNews, highlighting its ability to accurately identify OOD samples even when they belong to the same domain as the in-distribution dataset.",
+        "bbox": [
+            112,
+            538,
+            489,
+            778
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.5 Classification performance of SELFOOD",
+        "text_level": 1,
+        "bbox": [
+            112,
+            791,
+            482,
+            807
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The BERT classifier trained with CE-Loss demonstrates exceptional classification performance, achieving accuracy scores of  $97\\%$ ,  $85\\%$ , and  $97\\%$  on NYT, Clima, and TREC datasets respectively. In contrast, SELFOOD achieves only  $0.3\\%$ ,  $9.5\\%$ ,",
+        "bbox": [
+            112,
+            813,
+            489,
+            894
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e3a607b499b2b3c4490b3557709e1549b681d25ef9645cbd6ef42da801b89b9d.jpg",
+        "table_caption": [
+            "Table 4: Fine-grained OOD detection results with NYT as the in-distribution dataset with BERT classifier. We choose two datasets AGNews, 20News from the news domain, the same as NYT, and consider them as OOD datasets. The results show that SELFOOD can accurately detect OOD samples within the same domain."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>OOD</td><td>Method</td><td>FPR95</td><td>ERR</td><td>AUROC</td><td>AUPR</td></tr><tr><td rowspan=\"5\">AGNews</td><td>LogitNorm</td><td>45.9</td><td>0.9</td><td>86.7</td><td>15.8</td></tr><tr><td>Bayes Approx</td><td>57.5</td><td>0.9</td><td>87.7</td><td>23.9</td></tr><tr><td>CE-Loss</td><td>45.5</td><td>0.9</td><td>86.5</td><td>16.6</td></tr><tr><td>CRL</td><td>77.2</td><td>0.9</td><td>80.6</td><td>9.5</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.1</td><td>100.0</td><td>99.4</td></tr><tr><td rowspan=\"5\">20News</td><td>LogitNorm</td><td>24.5</td><td>6.2</td><td>91.7</td><td>39.2</td></tr><tr><td>Bayes Approx</td><td>35.0</td><td>4.3</td><td>93.7</td><td>66.6</td></tr><tr><td>CE-Loss</td><td>25.4</td><td>6.2</td><td>91.9</td><td>39.6</td></tr><tr><td>CRL</td><td>59.9</td><td>6.4</td><td>87.4</td><td>36.9</td></tr><tr><td>SELFOOD</td><td>28.2</td><td>4.9</td><td>94.2</td><td>61.5</td></tr></table>",
+        "bbox": [
+            515,
+            168,
+            877,
+            297
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and  $1.8\\%$ . This highlights that while SELFOOD serves as a reliable OOD detector, its performance as a text classifier is subpar. This observation can be attributed to our IDIL loss formulation, which focuses on comparing confidence levels across documents for each label rather than across labels for each document. As a result, the IDIL loss primarily promotes the ordinal ranking of confidence levels across documents, which enhances the model's OOD detection capabilities. However, this emphasis on inter-document ranking comes at the expense of inter-label ranking, resulting in limited classification capabilities. Moreover, when we introduce the intra-document comparison to the IDIL loss, as discussed in Section 4.5, we observe a decline in the model's ranking ability. Podolskiy et al. (2021) use mahalanobis distance for OOD detection and observe that although mahalanobis distance alongside transformer-based models is effective, such methods are sensitive to the geometric features in the embedding space and can be spoilt if the embedder is used for classification and is overfit. This matches with our observation, upon adding intra-document comparison to the loss, the OOD detection performance drops. This further supports the notion that balancing the inter-document and intra-document comparisons is crucial for achieving optimal performance in both OOD detection and text classification tasks.",
+        "bbox": [
+            505,
+            321,
+            884,
+            788
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.6 SELFOOD + Mahalanobis distance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            800,
+            835,
+            815
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Mahalanobis distance-based estimation of OOD scores is an effective post-processing method used on trained OOD classifiers (Leys et al., 2018; Xu et al., 2020). We investigate whether this postprocessing further improves the performance of our method. OOD scores are estimated as the dis",
+        "bbox": [
+            507,
+            822,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>4</sup>http://qwone.com/~jason/20Newsgroups/",
+        "bbox": [
+            134,
+            902,
+            425,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10728",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/8b4a65c2c4a5883714a7842663631d53db01b27e2773fddcb48717633d8ff359.jpg",
+        "table_caption": [
+            "Table 5: OOD detection results with Mahalanobis Distance post-processing technique. We choose two in-distribution, OOD pairs with BERT classifier. The results show that SELFOOD's OOD detection capabilities are enhanced with the post-processing technique."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>InD</td><td>OOD</td><td>Method</td><td>FPR95</td><td>ERR</td><td>AUROC</td><td>AUPR</td></tr><tr><td rowspan=\"2\">Clima</td><td rowspan=\"2\">Yelp</td><td>SELFOOD</td><td>14.7</td><td>0.6</td><td>98.0</td><td>90.8</td></tr><tr><td>SELFOOD + Maha</td><td>0.8</td><td>0.2</td><td>99.2</td><td>97.2</td></tr><tr><td rowspan=\"2\">Yelp</td><td rowspan=\"2\">NYT</td><td>SELFOOD</td><td>63.2</td><td>19.6</td><td>79.4</td><td>82.7</td></tr><tr><td>SELFOOD + Maha</td><td>12.4</td><td>4.4</td><td>97.4</td><td>97.5</td></tr></table>",
+        "bbox": [
+            121,
+            164,
+            480,
+            230
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "tance between a test data point and the distribution of in-distribution samples using the Mahalanobis distance metric. Following (Xu et al., 2020), we consider the intermediate layer encodings of the OOD classifier trained on in-distribution samples as its representative distribution. We apply this post-processing on top of SELFOOD and experiment with two in-distribution, OOD pairs with BERT classifier. As shown in Table 5, we observe further improvement in performance, demonstrating the quality of learned representative distribution using SELFOOD.",
+        "bbox": [
+            112,
+            255,
+            487,
+            448
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.7 Qualitative Analysis of SELFOOD",
+        "text_level": 1,
+        "bbox": [
+            112,
+            458,
+            431,
+            474
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We present a few examples with Yelp as in-distribution and TREC as OOD where SELFOOD accurately distinguishes and CE-loss fails to distinguish between in-distribution and OOD samples in Table 6. Although the first example has food-related words like lunch, cork and sentiment-indicative words such as contemptible scoundrel, SELFOOD accurately identifies it as OOD and CE loss model fails to do so. Examples 3 and 4 have questions that revolve around eateries. CE-Loss model confuses and predicts them as OOD as TREC dataset contains only questions. However, SELFOOD was able to rank them higher and accurately identify them as in-distribution.",
+        "bbox": [
+            112,
+            480,
+            489,
+            705
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            717,
+            247,
+            732
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present SELFOOD, a novel framework for OOD detection that leverages only in-distribution samples as supervision. Building upon the insight that OOD samples typically do not belong to any in-distribution labels, we formulate the OOD detection problem as an inter-document intra-label ranking task. To address this challenge, we propose IDIL loss which guides the training process. Through extensive experiments on multiple datasets, we demonstrate the effectiveness of SELFOOD on OOD detection task. However, we",
+        "bbox": [
+            112,
+            741,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/8fdfe9fd323f511cb13b9e3f9b76b8870373430a1fdfcce1f4727f34c868ed28.jpg",
+        "table_caption": [
+            "Table 6: Qualitative analysis of CE-Loss vs SELFOOD with Yelp as in-distribution and TREC as OOD datasets. Wrong predictions are in Red and correct predictions are in Green."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Text</td><td>CE-Loss</td><td>SELFOOD</td></tr><tr><td>Who said “What contemptible scoundrel stole the cork from my lunch?”</td><td>InD</td><td>OOD</td></tr><tr><td>Who said “Give me liberty or give me death”?</td><td>InD</td><td>OOD</td></tr><tr><td>Its a tequila house - wouldn’t you want your staff to be knowledgeable about tequila? Call me silly. So Gregg and I are wondering Bourbon Street, looking for the next &quot;place&quot;...</td><td>OOD</td><td>InD</td></tr><tr><td>the pitch reads: &quot;what’s worse than mean girl cheerleaders? how about resurrected mean-girl cheerleaders with supernatural powers?&quot; And if this place could be farther away from my house, it’d be in Camden... Oh, the Vomitorium, every time i grace the hipster-stained ...</td><td>OOD</td><td>InD</td></tr></table>",
+        "bbox": [
+            515,
+            140,
+            873,
+            338
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "also acknowledge that it comes at the expense of text classification performance. Future research can focus on developing techniques that effectively balance inter-document and intra-document comparisons, enabling improved performance in both OOD detection and text classification tasks.",
+        "bbox": [
+            507,
+            353,
+            884,
+            449
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            462,
+            645,
+            476
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in the previous sections, the OOD detection performance of our method comes at the cost of classification performance. This limitation needs to be addressed.",
+        "bbox": [
+            507,
+            488,
+            882,
+            552
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            507,
+            565,
+            690,
+            579
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper proposes a self-supervised OOD detection method. The aim of the paper is to minimize the human effort in annotating OOD documents and using only in-distribution documents for OOD detection. OOD detection is helpful in detection potential harmful content. Hence, we do not anticipate any major ethical concerns.",
+        "bbox": [
+            507,
+            590,
+            882,
+            703
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "9 Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            715,
+            702,
+            732
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We thank the anonymous reviewers and our colleagues from Shang Data Lab, especially Bill Hogan and Zihan Wang for their helpful feedback. Our work is sponsored in part by NSF CAREER Award 2239440, NSF Proto-OKN Award 2333790, NIH Bridge2AI Center Program under award 1U54HG012510-01, Cisco-UCSD Sponsored Research Project, as well as generous gifts from Google, Adobe, and Teradata. Any opinions, findings, and conclusions or recommendations expressed herein are those of the authors and should",
+        "bbox": [
+            507,
+            741,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10729",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "not be interpreted as necessarily representing the views, either expressed or implied, of the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for government purposes notwithstanding any copyright annotation hereon.",
+        "bbox": [
+            112,
+            84,
+            489,
+            180
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            206,
+            213,
+            223
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D 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 Ziegler, Jeffrey Wu, Clemens Winter, Chris 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 Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805.",
+            "Terrance DeVries and Graham W Taylor. 2018. Learning confidence for out-of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865.",
+            "Stefan Elfwing, Eiji Uchibe, and Kenji Doya. 2018. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural Networks, 107:3-11.",
+            "Geli Fei and Bing Liu. 2016. Breaking the closed world assumption in text classification. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 506-514.",
+            "Varun Gangal, Abhinav Arora, Arash Einolghozati, and Sonal Gupta. 2020. Likelihood ratios and generative classifiers for unsupervised out-of-domain detection in task-oriented dialog. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7764-7771.",
+            "Irving John Good. 1952. Rational decisions. Journal of the Royal Statistical Society: Series B (Methodological), 14(1):107-114.",
+            "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR.",
+            "Chirag Gupta and Aaditya Ramdas. 2021. Top-label calibration and multiclass-to-binary reductions. arXiv preprint arXiv:2107.08353."
+        ],
+        "bbox": [
+            115,
+            231,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Richard HR Hahnloser, Rahul Sarpeshkar, Misha A Mahowald, Rodney J Douglas, and H Sebastian Seung. 2000. Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit. nature, 405(6789):947-951.",
+            "Dan Hendrycks and Kevin Gimpel. 2017. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations.",
+            "Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. 2018. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606.",
+            "Eduard Hovy, Laurie Gerber, Ulf Hermjakob, ChinYew Lin, and Deepak Ravichandran. 2001. Toward semantics-based answer pinpointing. In Proceedings of the First International Conference on Human Language Technology Research.",
+            "Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. 2020. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10951-10960.",
+            "Amita Kamath, Robin Jia, and Percy Liang. 2020. Selective question answering under domain shift. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684-5696.",
+            "Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746-1751, Doha, Qatar. Association for Computational Linguistics.",
+            "Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang, and et al. 2019. An evaluation dataset for intent classification and out-of-scope prediction. 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 1311-1316.",
+            "Tanmay Laud, Daniel Spokoyny, Tom Corringham, and Taylor Berg-Kirkpatrick. 2023. Climabench: A benchmark dataset for climate change text understanding in english. arXiv preprint arXiv:2301.04253.",
+            "Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. 2018. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems, 31."
+        ],
+        "bbox": [
+            510,
+            84,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10730",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Christophe Leys, Olivier Klein, Yves Dominicy, and Christophe Ley. 2018. Detecting multivariate outliers: Use a robust variant of the mahalanobis distance. Journal of experimental social psychology, 74:150-156.",
+            "Xiaoya Li, Jiwei Li, Xiaofei Sun, Chun Fan, Tianwei Zhang, Fei Wu, Yuxian Meng, and Jun Zhang. 2021.  $k$  folden:  $k$ -fold ensemble for out-of-distribution detection. arXiv preprint arXiv:2108.12731.",
+            "Xin Li and Dan Roth. 2002. Learning question classifiers. In COLING 2002: The 19th International Conference on Computational Linguistics.",
+            "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.",
+            "Dheeraj Mekala, Chengyu Dong, and Jingbo Shang. 2022a. LOPS: Learning order inspired pseudo-label selection for weakly supervised text classification. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4894-4908, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Dheeraj Mekala, Varun Gangal, and Jingbo Shang. 2021. Coarse2Fine: Fine-grained text classification on coarsely-grained annotated data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 583-594, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Dheeraj Mekala and Jingbo Shang. 2020. Contextualized weak supervision for text classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 323-333.",
+            "Dheeraj Mekala, Tu Vu, Timo Schick, and Jingbo Shang. 2022b. Leveraging qa datasets to improve generative data augmentation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing.",
+            "Jooyoung Moon, Jihyo Kim, Younghak Shin, and Sangheum Hwang. 2020. Confidence-aware learning for deep neural networks. In international conference on machine learning, pages 7034-7044. PMLR.",
+            "Anh Nguyen, Jason Yosinski, and Jeff Clune. 2015. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 427-436.",
+            "Yawen Ouyang, Jiasheng Ye, Yu Chen, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2021. Energy-based unknown intent detection with data manipulation. arXiv preprint arXiv:2107.12542."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Alexander Podolskiy, Dmitry Lipin, Andrey Bout, Ekaterina Artemova, and Irina Pionkovskaya. 2021. Revisiting mahalanobis distance for transformer-based out-of-domain detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13675-13682.",
+            "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.",
+            "Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. 2019. Likelihood ratios for out-of-distribution detection. Advances in neural information processing systems, 32.",
+            "Frank Rosenblatt. 1957. The perceptron, a perceiving and recognizing automaton Project Para. Cornell Aeronautical Laboratory.",
+            "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199.",
+            "Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J. Gordon. 2019. An empirical study of example forgetting during deep neural network learning. In International Conference on Learning Representations.",
+            "Apoory Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L Willke. 2018. Out-of-distribution detection using an ensemble of self supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), pages 550-564.",
+            "Hongxin Wei, Renchunzi Xie, Hao Cheng, Lei Feng, Bo An, and Yixuan Li. 2022. Mitigating neural network overconfidence with logit normalization. In International Conference on Machine Learning, pages 23631-23644. PMLR.",
+            "Yanan Wu, Keqing He, Yuanmeng Yan, QiXiang Gao, Zhiyuan Zeng, Fujia Zheng, Lulu Zhao, Huixing Jiang, Wei Wu, and Weiran Xu. 2022a. Revisit overconfidence for ood detection: Reassigned contrastive learning with adaptive class-dependent threshold. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4165-4179.",
+            "Yanan Wu, Zhiyuan Zeng, Keqing He, Yutao Mou, Pei Wang, and Weiran Xu. 2022b. Distribution calibration for out-of-domain detection with Bayesian approximation. In Proceedings of the 29th International Conference on Computational Linguistics, pages 608-615, Gyeongju, Republic of Korea. International Committee on Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "10731",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Hong Xu, Keqing He, Yuanmeng Yan, Sihong Liu, Zijun Liu, and Weiran Xu. 2020. A deep generative distance-based classifier for out-of-domain detection with mahalanobis space. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1452-1460.",
+            "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Neural Information Processing Systems.",
+            "Zhiyuan Zeng, Keqing He, Yuanmeng Yan, Zijun Liu, Yanan Wu, Hong Xu, Huixing Jiang, and Weiran Xu. 2021a. Modeling discriminative representations for out-of-domain detection with supervised contrastive learning. arXiv preprint arXiv:2105.14289.",
+            "Zhiyuan Zeng, Hong Xu, Keqing He, Yuanmeng Yan, Sihong Liu, Zijun Liu, and Weiran Xu. 2021b. Adversarial generative distance-based classifier for robust out-of-domain detection. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7658-7662. IEEE.",
+            "Li-Ming Zhan, Haowen Liang, Bo Liu, Lu Fan, Xiaoming Wu, and Albert Lam. 2021. Out-of-scope intent detection with self-supervision and discriminative training. arXiv preprint arXiv:2106.08616.",
+            "Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 649-657. Curran Associates, Inc.",
+            "Wenxuan Zhou, Fangyu Liu, and Muhao Chen. 2021. Contrastive out-of-distribution detection for pretrained transformers. arXiv preprint arXiv:2104.08812."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            630
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "10732",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            238,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A.1 Results with NYT & Clima as in-distribution datasets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            109,
+            396,
+            139
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We present experimental results on Clima and NYT as in-distribution datasets with BERT and RoBERTa classifiers in Table 7.",
+        "bbox": [
+            112,
+            146,
+            487,
+            193
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A.2 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            114,
+            205,
+            339,
+            219
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Training trajectory analysis experiment. We compare two experiments showcasing empirical evidence on how the trajectory of CE-loss models and IDI loss models differ. In 3(b) we plot the change in validation FPR on BERT-classifiers optimized on CE-loss and IDIL loss. Here, we calculate the FPR across training steps on a validation dataset consisting of 500 in-distribution and OOD data points each to compare the training trajectories of the two models.",
+        "bbox": [
+            112,
+            225,
+            489,
+            384
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We generate the toy dataset by creating five isometric Gaussian blobs of 3-dimensional data points, with each cluster containing 1000 samples. The centers of all the clusters are bounded by the range  $[-10.0, 10.0]$  and the cluster standard deviation is set to 2.0 to account for limited overlap between cluster datapoints. We consider four clusters as being in-distribution and the remaining cluster as OOD. We train a logistic regression classifier on the in-domain datapoints initialized with the same weights and minimizing IDIL and CE-loss. The model is optimized using Stochastic Gradient Descent with a learning rate of 0.01 for 300 epochs. The weights are visualized at the beginning of each epoch. In order to facilitate visualization, the weight matrices of both approaches of dimension are flattened and Principal Component Analysis (PCA) is applied to reduce the dimensionality of the weight matrix. We plot the top-3 principal components after each epoch to conclude that the learned weights of CE-loss and SELFOOD progress in a different manner.",
+        "bbox": [
+            112,
+            387,
+            489,
+            740
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "10733",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/28bd720c516d0ea3088961387b884e4957a96758ee5d5e06303c61dbd0ff18d3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">In-dist</td><td rowspan=\"2\">OOD</td><td rowspan=\"2\">Method</td><td colspan=\"4\">BERT</td><td colspan=\"4\">RoBERTa</td></tr><tr><td>FPR95(↓)</td><td>ERR(↓)</td><td>AUROC(↑)</td><td>AUPR(↑)</td><td>FPR95(↓)</td><td>ERR(↓)</td><td>AUROC(↑)</td><td>AUPR(↑)</td></tr><tr><td rowspan=\"21\">NYT</td><td rowspan=\"7\">Clima</td><td>LogitNorm</td><td>20.6</td><td>1.0</td><td>96.8</td><td>92.3</td><td>18.4</td><td>1.0</td><td>97.3</td><td>93.6</td></tr><tr><td>OneVsRest</td><td>84.7</td><td>1.7</td><td>87.3</td><td>82.6</td><td>92.7</td><td>1.9</td><td>91.1</td><td>85.1</td></tr><tr><td>RCL</td><td>46.9</td><td>1.5</td><td>94.4</td><td>87.5</td><td>86.7</td><td>2.3</td><td>86.7</td><td>76.6</td></tr><tr><td>Bayes Approx</td><td>40.6</td><td>1.2</td><td>95.4</td><td>90.0</td><td>80.0</td><td>2.4</td><td>88.4</td><td>76.5</td></tr><tr><td>CE Loss</td><td>22.1</td><td>1.0</td><td>96.9</td><td>92.5</td><td>15.7</td><td>1.0</td><td>97.3</td><td>92.8</td></tr><tr><td>CRL</td><td>32.3</td><td>1.7</td><td>95.4</td><td>86.5</td><td>79.3</td><td>3.2</td><td>79.8</td><td>67.0</td></tr><tr><td>SELFOOD</td><td>10.5</td><td>1.6</td><td>98.4</td><td>90.5</td><td>4.6</td><td>1.0</td><td>99.3</td><td>95.4</td></tr><tr><td rowspan=\"7\">TREC</td><td>LogitNorm</td><td>13.3</td><td>5.5</td><td>97.2</td><td>87.9</td><td>17.5</td><td>6.4</td><td>96.1</td><td>84.1</td></tr><tr><td>OneVsRest</td><td>96.0</td><td>6.2</td><td>84.0</td><td>79.4</td><td>97.2</td><td>6.3</td><td>90.5</td><td>82.8</td></tr><tr><td>RCL</td><td>62.9</td><td>6.1</td><td>92.2</td><td>85.8</td><td>73.9</td><td>8.8</td><td>88.4</td><td>73.1</td></tr><tr><td>Bayes Approx</td><td>38.8</td><td>4.5</td><td>94.8</td><td>89.9</td><td>67.5</td><td>8.3</td><td>88.7</td><td>74.7</td></tr><tr><td>CE Loss</td><td>14.5</td><td>5.8</td><td>97.0</td><td>87.1</td><td>21.9</td><td>6.7</td><td>95.5</td><td>83.5</td></tr><tr><td>CRL</td><td>67.9</td><td>6.5</td><td>89.9</td><td>79.1</td><td>85.6</td><td>9.6</td><td>71.4</td><td>61.4</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan=\"7\">Yelp</td><td>LogitNorm</td><td>1.7</td><td>0.6</td><td>99.3</td><td>87.9</td><td>1.2</td><td>0.6</td><td>98.5</td><td>87.9</td></tr><tr><td>OneVsRest</td><td>15.1</td><td>0.7</td><td>97.4</td><td>78.1</td><td>8.7</td><td>0.6</td><td>97.6</td><td>78.8</td></tr><tr><td>RCL</td><td>45.1</td><td>0.5</td><td>94.9</td><td>84.7</td><td>76.7</td><td>1.0</td><td>90.2</td><td>60.0</td></tr><tr><td>Bayes Approx</td><td>28.7</td><td>0.4</td><td>96.4</td><td>86.8</td><td>83.0</td><td>1.1</td><td>87.7</td><td>56.6</td></tr><tr><td>CE Loss</td><td>3.4</td><td>0.7</td><td>99.2</td><td>82.3</td><td>1.8</td><td>0.6</td><td>99.3</td><td>85.4</td></tr><tr><td>CRL</td><td>69.5</td><td>0.5</td><td>89.3</td><td>77.8</td><td>73.3</td><td>1.0</td><td>83.8</td><td>58.3</td></tr><tr><td>SELFOOD</td><td>22.5</td><td>1.1</td><td>95.7</td><td>57.8</td><td>0.2</td><td>0.2</td><td>99.9</td><td>97.5</td></tr><tr><td rowspan=\"21\">Clima</td><td rowspan=\"7\">TREC</td><td>LogitNorm</td><td>37.8</td><td>5.5</td><td>94.6</td><td>91.9</td><td>60.5</td><td>7.0</td><td>91.6</td><td>88.0</td></tr><tr><td>OneVsRest</td><td>25.5</td><td>6.8</td><td>95.5</td><td>91.8</td><td>54.3</td><td>15.5</td><td>84.5</td><td>67.7</td></tr><tr><td>RCL</td><td>80.1</td><td>18.4</td><td>75.7</td><td>55.3</td><td>91.2</td><td>18.4</td><td>66.6</td><td>47.7</td></tr><tr><td>Bayes Approx</td><td>83.7</td><td>15.8</td><td>77.5</td><td>63.8</td><td>65.7</td><td>13.9</td><td>83.6</td><td>70.7</td></tr><tr><td>CE Loss</td><td>34.9</td><td>5.6</td><td>94.9</td><td>92.2</td><td>60.2</td><td>9.1</td><td>90.4</td><td>83.8</td></tr><tr><td>CRL</td><td>83.0</td><td>12.7</td><td>82.0</td><td>72.2</td><td>60.9</td><td>12.6</td><td>79.2</td><td>63.2</td></tr><tr><td>SELFOOD</td><td>0.3</td><td>1.1</td><td>99.4</td><td>99.4</td><td>38.9</td><td>8.8</td><td>90.1</td><td>84.7</td></tr><tr><td rowspan=\"7\">Yelp</td><td>LogitNorm</td><td>57.7</td><td>0.8</td><td>93.0</td><td>81.3</td><td>54.7</td><td>1.2</td><td>92.6</td><td>72.3</td></tr><tr><td>OneVsRest</td><td>43.7</td><td>1.0</td><td>92.7</td><td>72.5</td><td>58.8</td><td>1.8</td><td>73.5</td><td>34.7</td></tr><tr><td>RCL</td><td>91.0</td><td>2.3</td><td>60.7</td><td>12.2</td><td>98.0</td><td>2.3</td><td>56.4</td><td>12.7</td></tr><tr><td>Bayes Approx</td><td>81.4</td><td>2.2</td><td>76.5</td><td>22.7</td><td>96.8</td><td>2.2</td><td>52.3</td><td>16.3</td></tr><tr><td>CE Loss</td><td>46.3</td><td>0.9</td><td>93.8</td><td>80.1</td><td>79.5</td><td>1.9</td><td>86.5</td><td>47.2</td></tr><tr><td>CRL</td><td>98.7</td><td>1.4</td><td>67.3</td><td>53.3</td><td>72.9</td><td>2.1</td><td>50.9</td><td>19.0</td></tr><tr><td>SELFOOD</td><td>14.7</td><td>0.6</td><td>98.0</td><td>90.8</td><td>37.6</td><td>1.5</td><td>72.3</td><td>54.9</td></tr><tr><td rowspan=\"7\">NYT</td><td>LogitNorm</td><td>37.6</td><td>4.2</td><td>94.4</td><td>87.3</td><td>46.1</td><td>4.5</td><td>93.1</td><td>85.4</td></tr><tr><td>OneVsRest</td><td>34.6</td><td>4.8</td><td>94.3</td><td>85.3</td><td>40.6</td><td>9.4</td><td>83.1</td><td>51.3</td></tr><tr><td>RCL</td><td>90.3</td><td>12.1</td><td>64.6</td><td>28.8</td><td>97.9</td><td>12.4</td><td>51.5</td><td>21.2</td></tr><tr><td>Bayes Approx</td><td>87.4</td><td>11.2</td><td>70.4</td><td>38.3</td><td>96.8</td><td>11.3</td><td>55.8</td><td>29.2</td></tr><tr><td>CE Loss</td><td>44.9</td><td>4.8</td><td>93.2</td><td>84.4</td><td>62.5</td><td>6.7</td><td>89.6</td><td>75.2</td></tr><tr><td>CRL</td><td>85.6</td><td>7.6</td><td>78.0</td><td>62.5</td><td>61.3</td><td>9.4</td><td>61.8</td><td>39.9</td></tr><tr><td>SELFOOD</td><td>17.2</td><td>2.6</td><td>97.8</td><td>94.5</td><td>55.3</td><td>6.7</td><td>68.6</td><td>61.5</td></tr></table>",
+        "bbox": [
+            117,
+            208,
+            878,
+            720
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 7: OOD detection results with BERT & RoBERTa classifiers. Each experiment is repeated with three random seeds and the mean scores are reported. The false-positive-rate at  $95\\%$  true-positive-rate (FPR95), minimum detection error over all thresholds (ERR), the area under the risk-coverage curve (AURC), and the area under the precision-recall curve (AUPR) using in-distribution samples as the positives are used as evaluation metrics.",
+        "bbox": [
+            112,
+            728,
+            882,
+            787
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "10734",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/SELFOOD_ Self-Supervised Out-Of-Distribution Detection via Learning to Rank/f2ac6bc4-d902-4552-a8de-3700ee905922_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85df53e816e2b2aab9a02c63a3fb7feff836ef055d499ae0e47ce34c0f7c2584
+size 655530

2023/SELFOOD_ Self-Supervised Out-Of-Distribution Detection via Learning to Rank/full.md

@@ -0,0 +1,323 @@
+# SELFOOD: Self-Supervised Out-Of-Distribution Detection via Learning to Rank
+
+Dheeraj Mekala\*, Adithya Samavedhi\*, Chengyu Dong\* Jingbo Shang\*, University of California San Diego
+
+$^{\text{©}}$  Halicioglu Data Science Institute, University of California San Diego
+
+{dmekala, asamavedhi, cdong, jshang}@ucsd.edu
+
+# Abstract
+
+Deep neural classifiers trained with cross-entropy (CE) loss often suffer from poor calibration, necessitating the task of out-of-distribution (OOD) detection. Traditional supervised OOD detection methods require expensive manual annotation of in-distribution and OOD samples. To address the annotation bottleneck, we introduce SELFOOD, a self-supervised OOD detection method that requires only in-distribution samples as supervision. We cast OOD detection as an inter-document intralabel (IDIL) ranking problem and train the classifier with our pairwise ranking loss, referred to as IDIL loss. Specifically, given a set of in-distribution documents and their labels, for each label, we train the classifier to rank the softmax scores of documents belonging to that label to be higher than the scores of documents that belong to other labels. Unlike CE loss, our IDIL loss reaches zero when the desired confidence ranking is achieved and gradients are backpropagated to decrease probabilities associated with incorrect labels rather than continuously increasing the probability of the correct label. Extensive experiments with several classifiers on multiple classification datasets demonstrate the effectiveness of our method in both coarse-and fine-grained settings.
+
+# 1 Introduction
+
+Deep neural networks (DNNs) are ubiquitously used for text classification (Liu et al., 2019; Devlin et al., 2019; Yang et al., 2019; Brown et al., 2020). However, they are generally poorly calibrated, resulting in erroneously high-confidence scores for both in-distribution and out-of-distribution (OOD) samples (Szegedy et al., 2013; Nguyen et al., 2015; Guo et al., 2017; Mekala et al., 2022a). Such poor calibration makes DNNs unreliable, and OOD detection task vital for the safe deployment of
+
+![](images/c911208f1877b4fdcb0be004bc715b8b1c27785693980a6d460f09e969f8b0ce.jpg)  
+(a) CE Loss
+
+![](images/47bfbf0f3c5f5ed60d454d099ba9821a6ea8a0bf8dadf2144d351de3798331a5.jpg)  
+(b) IDIL Loss  
+Figure 1: CE Loss and IDIL loss optimization for documents  $D_{1} \in \text{Sports}$  and  $D_{2} \in \text{Arts}$ . CE loss increases scores corresponding to the Sports for  $D_{1}$  and Arts for  $D_{2}$ , implying an intra-document comparison. Instead, IDIL loss compares softmax scores in an inter-document intra-label fashion where it reduces scores corresponding to Sports for  $D_{2}$  to be less than that of  $D_{1}$ .
+
+deep learning models in safety-critical applications (Moon et al., 2020).
+
+Traditional supervised OOD detection methods (Hendrycks et al., 2018; Larson et al., 2019; Kamath et al., 2020; Zeng et al., 2021b) assume access to high-quality manually annotated in-distribution and OOD samples. However, this requires extensive annotation of OOD samples belonging to diverse distributions, which is expensive to obtain. Moreover, text classifiers are ideally desired to be more confident on in-distribution samples than OOD samples. However, the poor calibration of DNN precludes this phenomenon.
+
+To address these problems, we propose SELFOOD, a self-supervised OOD detection framework that requires only in-distribution samples as supervision. To adhere to the aforementioned ranking constraint, we formulate OOD detection as an inter-document intra-label (IDIL) ranking problem and train the classifier using our pairwise ranking loss, referred to as IDIL loss. As
+
+![](images/cae2b87cb7c1d319cdd7e34ae37f7b41387a297617da57a959bfdb29c3316ced.jpg)  
+Figure 2: SELFOOD is a self-supervised framework that requires only annotated in-distribution data to train the OOD classifier. Firstly, we bucket documents based on their annotated label (in dotted lines). Then, we compare each document in a bucket with all documents in other buckets to compute IDIL loss (in solid lines). Finally, we backpropagate gradients to decrease scores associated with incorrect labels during the training of the OOD classifier.
+
+shown in Figure 1(a), text classifiers are generally trained using cross-entropy (CE) loss (Good, 1952) in an intra-document fashion where for each document, the classifier is trained to distinguish between different labels by maximizing the score corresponding to the correct label. Instead, in our method, as shown in Figure 1(b), we propose to train in an inter-document, intra-label fashion where for each label, we train the model to rank the considered label probability score in documents belonging to the label higher compared to those not belonging to it. As OOD documents generally do not belong to any in-distribution label, we hypothesize such explicit learning to rank translates to accurately distinguishing OOD from in-distribution.
+
+Moreover, minimizing CE loss involves continuous optimization to increase the probability of the correct label over the other labels, making the classifier overconfident (Wei et al., 2022). Instead, in our method, our IDIL loss function becomes zero once the desired ranking is achieved, and during training, we backpropagate gradients to decrease probabilities associated with incorrect labels rather than increasing the probability of the correct label. Theoretically, the perfect model trained using the CE loss with 0 training loss is a solution to our ranking problem, however, in Section 5.1, we empirically show that our ranking objective leads to a different solution, demonstrating the importance of the optimization procedure. Finally, it is important to note that our ranking formulation, loss function, and the self-supervised training strategy are specifically designed to improve the performance of OOD detection rather than classification accuracy.
+
+We present our framework in Figure 2. Given a set of in-distribution documents and corresponding
+
+labels as input, we bucket documents belonging to each label and train the classifier to rank the probabilities using our IDIL loss function. Specifically, for each document in a label bucket, we pair up with all documents in every other bucket and compute IDIL loss.
+
+Our contributions are summarized as follows:
+
+- We propose SELFFOOD, a novel self-supervised method to train an OOD detection model without any OOD samples as supervision.  
+- We formulate OOD detection as an inter-document intra-label ranking problem and optimize it using our IDIL ranking loss.  
+- We perform extensive experiments on multiple text classification datasets to demonstrate the effectiveness of our method in OOD detection.  
+We release the code on  $\mathbf{G h i t b u^{1}}$
+
+# 2 Related Work
+
+Traditional supervised methods cast the OOD detection as classification with binary labels (Kamath et al., 2020), with one additional label for unseen classes (Fei and Liu, 2016). Since manually annotating documents is expensive (Mekala and Shang, 2020; Mekala et al., 2022b), recent works have deviated from requiring such extensive annotations and leveraging various distance metrics to detect OOD samples. Leys et al. (2018); Xu et al. (2020) use Mahalanobis distance as a post-processing technique to identify OOD samples. These methods use distance-based scoring functions along with the intermediate model layer features to determine an OOD score. Lee et al. (2018); Hsu et al. (2020) use a similar distance metric called ODIN to detect
+
+OOD images. An alternate approach to compensate for the lack of OOD training data involves generating pseudo-OOD data for training. Ouyang et al. (2021) propose a framework to generate high-quality OOD utterances and importance weights by selectively replacing phrases in in-domain samples. Zhan et al. (2021) generate pseudo-OOD samples for the task of intent detection using self-supervision. Zhou et al. (2021); Zeng et al. (2021a) introduce self-supervised approaches to OOD detection using a contrastive learning framework. They suggest fine-tuning transformers using a margin-based contrastive loss to learn text representations (Mekala et al., 2021) for OOD classification. Vyas et al. (2018); Li et al. (2021) treat a part of in-domain data as OOD samples as an alternate self-supervised approach. Further, Wu et al. (2022a) use a Reassigned Contrastive Loss (RCL) along with an adaptive class-dependent threshold mechanism to separate in-domain and OOD intents. Ren et al. (2019); Ggangal et al. (2020) leverage likelihood ratios crafted by generative models to classify OOD samples. Wei et al. (2022) observe that the norm of the logits keeps increasing during training, leading to overconfident outputs, and propose LogitNorm as a solution to decoupling the output norm during training optimization. Moon et al. (2020) introduce a novel Correctness Ranking Loss function in order to regularize output probabilities to produce well-ranked confidence estimates. Other calibration techniques include "Top-label" calibration which is used to regularize the reported probability for the predicted class (Gupta and Ramdas, 2021).
+
+# 3 SELFOOD: Self-Supervised OOD Detection
+
+We present the problem statement, the motivation for our ranking formulation, and our method including the loss function and its optimization strategy.
+
+# 3.1 Problem Statement
+
+In this paper, we work on the OOD detection task with only in-distribution samples and no OOD samples as supervision. Specifically, given a labeled dataset  $\mathcal{D}^{InD} = \{(x_1,y_1),(x_2,y_2),\ldots (x_n,y_n)\}$  sampled from a distribution space  $(\mathcal{X},\mathcal{C})$  where documents  $x_{i}\in \mathcal{X}$  and labels  $y_{i}\in \mathcal{C}$  as input, our goal is to train an OOD detector  $\mathcal{M}$  that accurately distinguishes in-distribution documents  $\mathcal{D}^{InD}$  and OOD documents  $\mathcal{D}^{OOD}\notin (\mathcal{X},\mathcal{C})$  without any
+
+OOD documents required for training.
+
+# 3.2 Motivation
+
+Numerous neural text classifiers have been proposed, incorporating multiple hidden layers (Rosenblatt, 1957), convolutional layers (Kim, 2014), and various types of attention mechanisms (Devlin et al., 2019; Liu et al., 2019; Radford et al., 2019). All these models culminate in a softmax head, which produces probabilities corresponding to each class. These classifiers are generally trained with CE loss in an intra-document fashion i.e. each document is considered independently and the softmax score of the true label is maximized. Such training of neural text classifiers is known to increase the magnitude of logit vectors even when most training examples are correctly classified (Wei et al., 2022), making them poorly calibrated, producing unreasonably high probabilities even for incorrect predictions (Szegedy et al., 2013; Nguyen et al., 2015; Guo et al., 2017). This diminishes their ability to maintain the desired attribute of ordinal ranking for predictions based on confidence levels, wherein a prediction exhibiting a higher confidence value should be considered more likely to be accurate than one with a lower confidence value (Moon et al., 2020). Intuitively, a text classifier possessing such quality would be a perfect OOD detector.
+
+# 3.3 OOD Detection as Inter-Document Intra-Label Ranking
+
+In order to align with the aforementioned characteristic, we propose formulating the OOD detection as an inter-document, intra-label ranking problem. Specifically, given a set of in-distribution documents, we compare across documents but within the same label and train our model to generate higher probability score for documents belonging to the label than for documents not belonging to the label. We consider the same model architecture as any text classifier with a softmax head that generates scores corresponding to each label, however, we train it using our IDIL loss instead of CE loss. Our assumption is that an OOD document does not fall under any specific label in the in-distribution space. Hence, we anticipate that the trained model would produce lower scores for OOD documents compared to in-distribution documents. This distinction in scores is expected to facilitate easy separation between OOD and in-distribution documents.
+
+IDIL Loss is a pairwise-ranking loss that enforces
+
+desired ordinal ranking of confidence estimates. This loss function reaches its minimum value for a particular label when the probability of that label being the annotated label is greater than its probability when it is not the annotated label. Specifically, for documents  $x_{1}, x_{2} \in \mathcal{D}^{InD}$  and their corresponding annotated labels  $y_{1}, y_{2}$  where  $y_{1} \neq y_{2}$ , IDIL loss corresponding to label  $y_{1}$  is mathematically computed as follows:
+
+$$
+\mathcal {L} _ {I D I L} \left(y _ {1} \mid x _ {1}, x _ {2}\right) = S i L U \left(p \left(y _ {1} \mid x _ {2}\right) - p \left(y _ {1} \mid x _ {1}\right)\right) \tag {1}
+$$
+
+where  $SiLU(x) = x\sigma(x)$  is the Sigmoid Linear Unit (SiLU) function (Elfwing et al., 2018). To ensure stable training and enhance performance, we incorporate the SiLU function, a continuous variant of the Rectified Linear Unit (ReLU) (Hahnloser et al., 2000), in conjunction with the ranking loss. The SiLU function introduces smooth gradients around zero, effectively mitigating potential instability issues during training. We observe that this inclusion contributes to the overall stability in training and improved performance of the model as shown in Section 4.5. Note that, in contrast to CE loss, IDIL loss becomes zero once desired ranking is achieved, addressing the overconfidence issue.
+
+# 3.4 Implementation
+
+Ideally, the loss has to be computed over all possible pairs of documents for each model update. However, it is computationally expensive. Therefore, following (Toneva et al., 2019; Moon et al., 2020), we approximate the computation by considering only documents in each mini-batch. Specifically, we bucket the documents in the mini-batch based on their annotated label and pair each document in a bucket with all documents in other buckets and compute the loss. Mathematically, the loss for a mini-batch  $b$  is computed as follows:
+
+$$
+\mathcal {L} = \sum_ {l \in \mathcal {C}} \sum_ {x _ {1} \in b _ {l}} \sum_ {x _ {2} \in b _ {- l}} \mathcal {L} _ {I D I L} (l | x _ {1}, x _ {2})
+$$
+
+where  $b_{l}$  denotes the set of training data points  $x$  in this batch  $b$  whose label are  $l$ , and  $b_{-l}$  denotes the set of training data points  $x$  in this batch  $b$  whose label is not  $l$ .
+
+In contrast to CE loss, where the optimization involves increasing the score corresponding to the correct label, we backpropagate gradients to decrease scores associated with incorrect labels. Specifically, during the backpropagation of gradients, we detach the gradients for the subrahend
+
+Table 1: Dataset statistics.  
+
+<table><tr><td>Dataset</td><td>Domain</td><td>Criteria</td><td>#Docs</td><td>#labels</td></tr><tr><td>NYT</td><td>News</td><td>Topic</td><td>13081</td><td>26</td></tr><tr><td>Yelp</td><td>Reviews</td><td>Sentiment</td><td>70000</td><td>5</td></tr><tr><td>Clima</td><td>Climate</td><td>Question Type</td><td>17175</td><td>8</td></tr><tr><td>TREC</td><td>General</td><td>Question Type</td><td>5952</td><td>6</td></tr></table>
+
+of the difference and exclusively propagate the gradients through the minuend. In Equation 1, for instance, we detach the gradients for  $p(y_{1}|x_{1})$  and solely backpropagate the gradients through  $p(y_{1}|x_{2})$ . This detachment allows for a more controlled and selective gradient flow, aiding in the optimization process, and improvement in performance as shown in Section 4.5.
+
+It is important to note that our optimization focuses solely on the inter-document ranking loss. Consequently, while the trained model would serve as a reliable OOD detector, it may not perform as effectively as a classifier.
+
+# 4 Experiments
+
+We evaluate our OOD detection method against state-of-the-art baselines with two classifiers on multiple datasets belonging to different domains.
+
+# 4.1 Datasets
+
+We evaluate our method and baselines on four publicly available English text classification datasets belonging to different domains. In particular, we consider the news topic classification dataset New York Times (NYT), restaurant review sentiment classification dataset Yelp, and question-type classification datasets related to climate: ClimaInsurance+ (Clima) (Laud et al., 2023), and a general domain: TREC (Li and Roth, 2002; Hovy et al., 2001). The documents within the New York Times dataset are labeled with both coarse and fine-grained labels. For our training and testing process, we utilize fine-grained labels. The dataset statistics are provided in Table 1.
+
+# 4.2 Compared Methods
+
+- Cross Entropy Loss (CE Loss) trains a classifier using CE loss on in-distribution documents. The predicted probabilities from the classifier are used as confidence estimates for OOD detection.  
+- Correctness Ranking Loss (CRL) (Moon et al., 2020) is a regularization term added to the CE
+
+Table 2: OOD detection results with BERT & RoBERTa classifiers. Each experiment is repeated with three random seeds and the mean scores are reported. The false-positive-rate at  $95\%$  true-positive-rate (FPR95), minimum detection error over all thresholds (ERR), the area under the risk-coverage curve (AURC), and the area under the precision-recall curve (AUPR) using in-distribution samples as the positives are used as evaluation metrics.  
+
+<table><tr><td rowspan="2">In-dist</td><td rowspan="2">OOD</td><td rowspan="2">Method</td><td colspan="4">BERT</td><td colspan="4">RoBERTa</td></tr><tr><td>FPR95(↓)</td><td>ERR(↓)</td><td>AUROC(↑)</td><td>AUPR(↑)</td><td>FPR95(↓)</td><td>ERR(↓)</td><td>AUROC(↑)</td><td>AUPR(↑)</td></tr><tr><td rowspan="21">Yelp</td><td rowspan="7">NYT</td><td>LogitNorm</td><td>82.0</td><td>33.2</td><td>69.4</td><td>57.9</td><td>84.4</td><td>33.9</td><td>66.6</td><td>55.0</td></tr><tr><td>OneVsRest</td><td>85.1</td><td>35.8</td><td>62.4</td><td>49.3</td><td>88.5</td><td>29.6</td><td>69.8</td><td>56.5</td></tr><tr><td>RCL</td><td>56.9</td><td>31.2</td><td>73.2</td><td>55.2</td><td>87.7</td><td>34.5</td><td>56.7</td><td>45.0</td></tr><tr><td>Bayes Approx</td><td>64.3</td><td>35.9</td><td>66.5</td><td>50.3</td><td>79.6</td><td>36.4</td><td>54.1</td><td>40.8</td></tr><tr><td>CE Loss</td><td>82.1</td><td>34.9</td><td>63.4</td><td>50.7</td><td>79.8</td><td>32.9</td><td>66.8</td><td>58.0</td></tr><tr><td>CRL</td><td>99.9</td><td>36.9</td><td>43.9</td><td>51.1</td><td>99.8</td><td>28.1</td><td>64.8</td><td>60.5</td></tr><tr><td>SELFOOD</td><td>63.2</td><td>19.6</td><td>79.4</td><td>82.7</td><td>69.5</td><td>21.6</td><td>72.6</td><td>78.8</td></tr><tr><td rowspan="7">Clima</td><td>LogitNorm</td><td>79.2</td><td>25.5</td><td>72.5</td><td>54.8</td><td>82.6</td><td>25.4</td><td>70.7</td><td>54.0</td></tr><tr><td>OneVsRest</td><td>83.0</td><td>25.1</td><td>69.5</td><td>53.2</td><td>86.7</td><td>28.0</td><td>53.8</td><td>33.7</td></tr><tr><td>RCL</td><td>59.9</td><td>27.8</td><td>67.4</td><td>42.1</td><td>73.9</td><td>22.5</td><td>70.7</td><td>56.3</td></tr><tr><td>Bayes Approx</td><td>74.7</td><td>28.7</td><td>56.7</td><td>33.6</td><td>63.0</td><td>28.0</td><td>66.8</td><td>40.9</td></tr><tr><td>CE Loss</td><td>83.2</td><td>27.8</td><td>48.6</td><td>37.0</td><td>76.4</td><td>22.0</td><td>78.2</td><td>68.4</td></tr><tr><td>CRL</td><td>99.6</td><td>29.0</td><td>49.1</td><td>38.8</td><td>99.9</td><td>20.6</td><td>63.0</td><td>56.9</td></tr><tr><td>SELFOOD</td><td>17.6</td><td>4.8</td><td>97.9</td><td>96.4</td><td>65.6</td><td>16.6</td><td>73.2</td><td>77.2</td></tr><tr><td rowspan="7">TREC</td><td>LogitNorm</td><td>68.2</td><td>27.0</td><td>79.8</td><td>82.1</td><td>60.8</td><td>23.0</td><td>83.9</td><td>84.4</td></tr><tr><td>OneVsRest</td><td>84.8</td><td>40.0</td><td>63.9</td><td>69.7</td><td>75.5</td><td>33.0</td><td>57.9</td><td>61.2</td></tr><tr><td>RCL</td><td>89.8</td><td>43.2</td><td>40.1</td><td>51.5</td><td>77.4</td><td>32.4</td><td>65.0</td><td>68.6</td></tr><tr><td>Bayes Approx</td><td>92.5</td><td>44.7</td><td>32.1</td><td>44.7</td><td>51.8</td><td>25.9</td><td>72.7</td><td>69.4</td></tr><tr><td>CE Loss</td><td>74.2</td><td>33.7</td><td>52.8</td><td>62.7</td><td>58.5</td><td>27.5</td><td>75.3</td><td>78.3</td></tr><tr><td>CRL</td><td>33.3</td><td>15.5</td><td>66.7</td><td>77.2</td><td>100.0</td><td>37.3</td><td>56.7</td><td>73.9</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>55.1</td><td>20.6</td><td>80.2</td><td>90.3</td></tr><tr><td rowspan="21">TREC</td><td rowspan="7">Yelp</td><td>LogitNorm</td><td>32.3</td><td>0.3</td><td>95.5</td><td>78.1</td><td>15.2</td><td>0.5</td><td>95.7</td><td>44.7</td></tr><tr><td>OneVsRest</td><td>4.2</td><td>0.3</td><td>97.9</td><td>83.4</td><td>39.3</td><td>0.4</td><td>91.6</td><td>63.0</td></tr><tr><td>RCL</td><td>78.5</td><td>0.6</td><td>80.2</td><td>20.2</td><td>67.7</td><td>0.6</td><td>83.0</td><td>24.2</td></tr><tr><td>Bayes Approx</td><td>46.6</td><td>0.4</td><td>93.0</td><td>61.6</td><td>63.7</td><td>0.7</td><td>87.7</td><td>21.3</td></tr><tr><td>CE Loss</td><td>27.2</td><td>0.3</td><td>94.9</td><td>71.2</td><td>11.3</td><td>0.4</td><td>96.9</td><td>62.3</td></tr><tr><td>CRL</td><td>96.6</td><td>0.7</td><td>70.9</td><td>16.5</td><td>99.8</td><td>0.7</td><td>24.3</td><td>9.7</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan="7">NYT</td><td>LogitNorm</td><td>9.3</td><td>1.0</td><td>97.7</td><td>91.2</td><td>10.9</td><td>2.3</td><td>97.2</td><td>77.3</td></tr><tr><td>OneVsRest</td><td>1.8</td><td>0.5</td><td>99.0</td><td>95.8</td><td>32.5</td><td>1.8</td><td>93.2</td><td>74.8</td></tr><tr><td>RCL</td><td>63.6</td><td>3.3</td><td>84.1</td><td>45.7</td><td>74.2</td><td>3.0</td><td>80.4</td><td>45.1</td></tr><tr><td>Bayes Approx</td><td>39.1</td><td>2.5</td><td>93.5</td><td>68.1</td><td>64.9</td><td>3.0</td><td>88.4</td><td>54.4</td></tr><tr><td>CE Loss</td><td>8.4</td><td>1.3</td><td>97.6</td><td>87.5</td><td>6.8</td><td>1.4</td><td>98.1</td><td>87.1</td></tr><tr><td>CRL</td><td>97.2</td><td>3.3</td><td>73.0</td><td>35.2</td><td>99.9</td><td>3.8</td><td>22.9</td><td>12.2</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan="7">Clima</td><td>LogitNorm</td><td>11.3</td><td>0.8</td><td>97.6</td><td>88.5</td><td>10.4</td><td>1.6</td><td>97.3</td><td>75.9</td></tr><tr><td>OneVsRest</td><td>3.4</td><td>0.6</td><td>98.7</td><td>93.2</td><td>36.3</td><td>1.8</td><td>91.7</td><td>60.1</td></tr><tr><td>RCL</td><td>73.5</td><td>2.4</td><td>81.2</td><td>37.1</td><td>60.7</td><td>2.0</td><td>86.4</td><td>48.2</td></tr><tr><td>Bayes Approx</td><td>51.7</td><td>2.0</td><td>90.5</td><td>54.4</td><td>61.1</td><td>2.4</td><td>87.6</td><td>40.4</td></tr><tr><td>CE Loss</td><td>15.1</td><td>1.7</td><td>96.2</td><td>69.5</td><td>12.2</td><td>1.8</td><td>96.4</td><td>67.8</td></tr><tr><td>CRL</td><td>86.3</td><td>2.2</td><td>80.7</td><td>39.5</td><td>98.9</td><td>2.7</td><td>30.1</td><td>13.0</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td></tr></table>
+
+loss to make class probabilities better confidence estimates. It estimates the true class probability to be proportional to the number of times a sample is classified correctly during training.
+
+- LogitNorm (Wei et al., 2022) is a variant of CE loss that normalizes the logit vector to have a constant norm during training.  
+- Bayesian Approximation (Bayes Approx) (Wu et al., 2022b) calibrates distribution uncertainty for OOD detection using Monte-Carlo dropout.  
+- Reassigned Contrastive Learning (RCL) (Wu et al., 2022a) discriminates over-confident OOD samples using adaptive class-dependent local threshold mechanism.
+
+- OneVsRest trains a one vs rest binary classifier per label. The probability scores of each label from its corresponding classifier are normalized and used as their respective confidence estimates.
+
+# 4.3 Experimental Settings
+
+We experiment with BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019) as text classifiers. For SELFOOD, we train the classifier for 5 epochs with a batch size of 16 using an AdamW optimizer. We use a learning rate of 5e-5 using a linear scheduler with no warm-up steps. For all baselines, we train the classifier for the same number of steps.
+
+In our evaluation, for each dataset as in-
+
+distribution, we treat all other datasets as OOD and compute the performance. Our evaluation follows a standard approach for each in-distribution dataset where we begin by splitting the in-distribution dataset into three subsets:  $80\%$  for train,  $10\%$  for val, and  $10\%$  for test. The model is trained using the train split, and its performance is evaluated on both the test split of the in-distribution dataset and the entire OOD dataset.
+
+Evaluation Metrics. We utilize evaluation metrics from (Hendrycks and Gimpel, 2017; DeVries and Taylor, 2018; Moon et al., 2020) such as the false positive rate at  $95\%$  true positive rate (FPR95), minimum detection error over all thresholds (ERR), the area under the risk-coverage curve (AURC), and the area under the precision-recall curve (AUPR) using in-distribution samples as the positives.
+
+# 4.4 Results
+
+We summarize the evaluation results with BERT and RoBERTa as classifiers on Yahoo, TREC as in-distribution in Table 2 and NYT, Clima as in-distribution in Appendix A.1. All experiments are run on three random seeds and the mean performance scores are reported. As shown in Table 2, we observe that SELFOOD performs better than the baselines on most of the in-distribution, OOD dataset pairs for both classifiers. A low FPR95 value indicates that the top-  $95\%$  confident in-distribution samples, selected based on their probability scores, predominantly rank higher than a majority of OOD samples and SELFOOD achieves improvements of up to 82 points in FPR95 with Yelp as in-distribution and Clima as OOD datasets when compared to CRL with BERT classifier. SELFOOD also exhibits substantial improvements of up to 33 points in Detection Error, 48 points in AUROC, and 58 points in AUPR when compared to CE-Loss with BERT classifier. SELFOOD achieves a perfect OOD detection score for some settings such as TREC as in-distribution dataset for both BERT and RoBERTa classifiers. These results highlight the effectiveness of our ranking formulation with self-supervised training using IDIL loss.
+
+# 4.5 Ablation Study
+
+To understand the impact of each component in our IDIL loss design and implementation, we compare our method with four ablated versions with BERT classifier in Table 3: (1) SELFOOD + Grad Sub represents our method with backpropagating gradi
+
+Table 3: Ablation Study.  
+
+<table><tr><td>Method</td><td>FPR95</td><td>ERR
+In Dist: Clima
+OOD: NYT</td><td>AUROC</td><td>AUPR</td></tr><tr><td>SELFOOD</td><td>17.2</td><td>2.6</td><td>97.8</td><td>94.5</td></tr><tr><td>+ Grad Sub</td><td>67.9</td><td>8.9</td><td>82.4</td><td>61.1</td></tr><tr><td>+ Grad Sub &amp; Min</td><td>98.4</td><td>10.7</td><td>49.2</td><td>32.3</td></tr><tr><td>+ Intra-Doc</td><td>92.7</td><td>10.6</td><td>67.3</td><td>41.6</td></tr><tr><td>- SILu</td><td>14.6</td><td>3.9</td><td>96.7</td><td>89.3</td></tr><tr><td></td><td></td><td>In Dist: NYT
+OOD: Clima</td><td></td><td></td></tr><tr><td>SELFOOD</td><td>10.5</td><td>1.7</td><td>98.4</td><td>90.5</td></tr><tr><td>+ Grad Sub only</td><td>83.5</td><td>5.1</td><td>77.4</td><td>44.8</td></tr><tr><td>+ Grad Sub &amp; Min</td><td>96.5</td><td>5.2</td><td>60.3</td><td>36.0</td></tr><tr><td>+ Intra-Doc</td><td>48.3</td><td>2.6</td><td>93.2</td><td>80.2</td></tr><tr><td>- SiLU</td><td>19.0</td><td>2.3</td><td>97.4</td><td>86.0</td></tr></table>
+
+ents through the subtrahend instead of the minuend, (2) SELFOOD  $^+$  Grad Sub & Min represents our method with gradients backpropagating through both minuend and subtrahend, (3) SELFOOD  $^+$  Intra-Doc considers intra-document comparison similar to CE loss, in addition to inter-document intra-label comparison for loss computation, and (4) SELFOOD-SiLU excludes SiLU function from the IDIL loss formulation. We also present the performance of SELFOOD for reference. SELFOOD performs better than SELFOOD  $^+$  Grad Sub & Min demonstrating that backpropagating through one part of the difference is more beneficial than both, and the comparison between SELFOOD and SELFOOD  $^+$  Grad Sub indicates that backpropagating through minuend is better than subtrahend. We observe that incorporating intra-document comparison into loss formulation leads to a decrement in the ranking ability of the model. Finally, we observe that removing the SiLU function from the IDIL loss leads to a decrease in most of the metrics.
+
+# 5 Analysis & Case Studies
+
+In this section, we present a comprehensive analysis of our proposed method from different perspectives to understand its effectiveness.
+
+# 5.1 CE Loss vs SELFOOD: Training Trajectory Analysis
+
+Mathematically, the model trained using CE loss is a solution to our IDIL loss optimization. However, our empirical results reveal a significant disparity in OOD detection performance between SELFOOD and model trained on CE loss. To gain a deeper understanding, we plot the CE Loss and IDIL losses for models optimized with IDIL and CE loss, as shown in Figure 3(a). We consider Clima as the in-distribution dataset and BERT as the classifier.
+
+![](images/8a5752b037c51fb605f668b88e60675e422e15801264977d92dd949eca9fafaf.jpg)  
+(a) Losses vs Training Steps
+
+![](images/d26a489c9693b1aebcf29e79ac3e9702c1e60c1eee6c0ab51b27a036ea0909dd.jpg)  
+(b) FPR vs Training Steps
+
+![](images/50ebefbb99df794abc58210ece2bed50f8d107877b01fecfafbde4ad5b7e3c9b.jpg)  
+(c) Learned weights
+
+![](images/843480b4fbf688fbab1bcf7a6f8714956a78b49a5e9b82558dde471fbc6e5b1a.jpg)  
+Figure 3: To understand training trajectories of CE Loss and SELFOOD, we plot their respective losses, validation FPR, and learned weights while training. We observe IDIL loss minimization is more correlated with performance and the learned weights of CE-Loss and SELFOOD show divergent paths since the beginning.  
+(a) CE Loss  
+Figure 4: Maximum softmax score vs percentile in-distribution & OOD data associated with that score. By using the least in-distribution maximum softmax score as the threshold (dotted line) for OOD classification, CE loss considers more than  $50\%$  and CRL considers almost  $100\%$  of the OOD data as in-distribution (red region). However, in the case of SELFOOD, we observe a clear margin in maximum softmax scores that separate OOD and in-distribution.
+
+![](images/245c4617d1f5667dd2b7a75204cdbd36e881e23daa2410db3b77879900551b3d.jpg)  
+(b) CRL
+
+![](images/21ead48aa7a316cff58b0ae8920390495bcb30935f403e1fa7bef841be453af7.jpg)  
+(c) SELFOOD
+
+Additionally, we plot the validation FPR95 using Yelp as the OOD dataset to analyze the correlation between loss minimization and performance in Figure 3(b). Figure 3(a) illustrates that while minimizing CE loss leads to fluctuations in the IDIL loss, ultimately reaching zero, Figure 3(b) shows that the performance is more closely associated with explicit IDIL minimization. Specifically, minimizing the IDIL loss results in improved FPR95. The findings indicate that the minimization of CE loss and IDIL loss follows distinct trajectories. To validate this hypothesis, we conduct an experiment using a toy 3-dimensional dataset comprising five isometric Gaussian blobs and train a logistic regression classifier initialized with same weights, minimizing IDIL and CE-loss. We flatten the weight matrix and plot the top-3 principal components. As shown in Figure 3(c), the learned weights of CE-Loss and SELFOOD progress in different directions, thereby confirming our hypothesis.
+
+# 5.2 Maximum softmax score Analysis
+
+We plot the maximum softmax score with the percentile in-distribution and OOD data associated with that score for CE loss, CRL, and SELFOOD
+
+with BERT classifier on Clima as in-distribution and NYT as OOD datasets in Figure 4. When comparing the maximum softmax score with the percentile data associated with that score, interesting observations can be made regarding its distribution. Specifically, when using the least in-distribution maximum softmax score as the threshold for OOD classification, we find that the CE loss considers over  $50\%$  of the OOD data and CRL considers almost  $100\%$  of the OOD data as in-distribution. However, in the case of SELFOOD, we observe a clear margin in the maximum softmax scores that effectively separates OOD and in-distribution data. This suggests that the threshold needs to be carefully tuned for CE loss and CRL, requiring more effort and annotated data, whereas in the case of SELFOOD, we do not require such tuning. This demonstrates that SELFOOD is capable of accurately classifying between OOD and in-distribution samples based on their respective maximum softmax scores, resulting in superior performance.
+
+# 5.3 Batch size Analysis
+
+We compare documents within the mini-batch when computing the loss. Consequently, the num
+
+![](images/6ecab13b9bdcccdfea4e0a94db63b28458700569f0881a722f44a8c11c04c977.jpg)  
+(a) NYT
+
+![](images/7e66588085baee7ee42017a07ad781cf15e66b75644e44d1c10cd37b0d5c9140.jpg)  
+(b) Yelp  
+Figure 5: Performance of SELFOOD w.r.t. batch size. We consider BERT classifier, Clima as in-distribution, and NYT, Yelp as OOD datasets. We observe an increase in performance with an increase in batch size.
+
+ber of document pairs over which the loss is computed depends on the batch size used during training. To investigate the impact of batch size on performance, we vary the batch size and evaluate the corresponding performance in Figure 5 with BERT classifier. We consider Clima as in-distribution, and NYT, Yelp as OOD datasets. As the batch size increases, the performance of the model also improves, until a batch size of 16, where the performance reaches a plateau. This observation aligns with our intuition that a larger batch size allows for more in-batch document pairs to be compared, leading to more accurate loss computation. Based on these findings, we recommend utilizing a batch size of 16 or higher to achieve optimal performance.
+
+# 5.4 Fine-grained OOD Detection
+
+To explore the limits of SELFOOD, we consider in-distribution and OOD datasets from the same domain. The objective is to assess its ability to differentiate between samples that are closely related but still OOD, which is a challenging task. Specifically, we consider the news domain and choose NYT as the in-distribution, and AGNews (Zhang et al., 2015),  $20\mathrm{News}^4$  as OOD datasets and train BERT classifier. As shown in Table 4, SELFOOD performs significantly better than the baselines on most of the metrics. Moreover, it also achieves near-perfect scores for the AGNews, highlighting its ability to accurately identify OOD samples even when they belong to the same domain as the in-distribution dataset.
+
+# 5.5 Classification performance of SELFOOD
+
+The BERT classifier trained with CE-Loss demonstrates exceptional classification performance, achieving accuracy scores of  $97\%$ ,  $85\%$ , and  $97\%$  on NYT, Clima, and TREC datasets respectively. In contrast, SELFOOD achieves only  $0.3\%$ ,  $9.5\%$ ,
+
+Table 4: Fine-grained OOD detection results with NYT as the in-distribution dataset with BERT classifier. We choose two datasets AGNews, 20News from the news domain, the same as NYT, and consider them as OOD datasets. The results show that SELFOOD can accurately detect OOD samples within the same domain.  
+
+<table><tr><td>OOD</td><td>Method</td><td>FPR95</td><td>ERR</td><td>AUROC</td><td>AUPR</td></tr><tr><td rowspan="5">AGNews</td><td>LogitNorm</td><td>45.9</td><td>0.9</td><td>86.7</td><td>15.8</td></tr><tr><td>Bayes Approx</td><td>57.5</td><td>0.9</td><td>87.7</td><td>23.9</td></tr><tr><td>CE-Loss</td><td>45.5</td><td>0.9</td><td>86.5</td><td>16.6</td></tr><tr><td>CRL</td><td>77.2</td><td>0.9</td><td>80.6</td><td>9.5</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.1</td><td>100.0</td><td>99.4</td></tr><tr><td rowspan="5">20News</td><td>LogitNorm</td><td>24.5</td><td>6.2</td><td>91.7</td><td>39.2</td></tr><tr><td>Bayes Approx</td><td>35.0</td><td>4.3</td><td>93.7</td><td>66.6</td></tr><tr><td>CE-Loss</td><td>25.4</td><td>6.2</td><td>91.9</td><td>39.6</td></tr><tr><td>CRL</td><td>59.9</td><td>6.4</td><td>87.4</td><td>36.9</td></tr><tr><td>SELFOOD</td><td>28.2</td><td>4.9</td><td>94.2</td><td>61.5</td></tr></table>
+
+and  $1.8\%$ . This highlights that while SELFOOD serves as a reliable OOD detector, its performance as a text classifier is subpar. This observation can be attributed to our IDIL loss formulation, which focuses on comparing confidence levels across documents for each label rather than across labels for each document. As a result, the IDIL loss primarily promotes the ordinal ranking of confidence levels across documents, which enhances the model's OOD detection capabilities. However, this emphasis on inter-document ranking comes at the expense of inter-label ranking, resulting in limited classification capabilities. Moreover, when we introduce the intra-document comparison to the IDIL loss, as discussed in Section 4.5, we observe a decline in the model's ranking ability. Podolskiy et al. (2021) use mahalanobis distance for OOD detection and observe that although mahalanobis distance alongside transformer-based models is effective, such methods are sensitive to the geometric features in the embedding space and can be spoilt if the embedder is used for classification and is overfit. This matches with our observation, upon adding intra-document comparison to the loss, the OOD detection performance drops. This further supports the notion that balancing the inter-document and intra-document comparisons is crucial for achieving optimal performance in both OOD detection and text classification tasks.
+
+# 5.6 SELFOOD + Mahalanobis distance
+
+Mahalanobis distance-based estimation of OOD scores is an effective post-processing method used on trained OOD classifiers (Leys et al., 2018; Xu et al., 2020). We investigate whether this postprocessing further improves the performance of our method. OOD scores are estimated as the dis
+
+Table 5: OOD detection results with Mahalanobis Distance post-processing technique. We choose two in-distribution, OOD pairs with BERT classifier. The results show that SELFOOD's OOD detection capabilities are enhanced with the post-processing technique.  
+
+<table><tr><td>InD</td><td>OOD</td><td>Method</td><td>FPR95</td><td>ERR</td><td>AUROC</td><td>AUPR</td></tr><tr><td rowspan="2">Clima</td><td rowspan="2">Yelp</td><td>SELFOOD</td><td>14.7</td><td>0.6</td><td>98.0</td><td>90.8</td></tr><tr><td>SELFOOD + Maha</td><td>0.8</td><td>0.2</td><td>99.2</td><td>97.2</td></tr><tr><td rowspan="2">Yelp</td><td rowspan="2">NYT</td><td>SELFOOD</td><td>63.2</td><td>19.6</td><td>79.4</td><td>82.7</td></tr><tr><td>SELFOOD + Maha</td><td>12.4</td><td>4.4</td><td>97.4</td><td>97.5</td></tr></table>
+
+tance between a test data point and the distribution of in-distribution samples using the Mahalanobis distance metric. Following (Xu et al., 2020), we consider the intermediate layer encodings of the OOD classifier trained on in-distribution samples as its representative distribution. We apply this post-processing on top of SELFOOD and experiment with two in-distribution, OOD pairs with BERT classifier. As shown in Table 5, we observe further improvement in performance, demonstrating the quality of learned representative distribution using SELFOOD.
+
+# 5.7 Qualitative Analysis of SELFOOD
+
+We present a few examples with Yelp as in-distribution and TREC as OOD where SELFOOD accurately distinguishes and CE-loss fails to distinguish between in-distribution and OOD samples in Table 6. Although the first example has food-related words like lunch, cork and sentiment-indicative words such as contemptible scoundrel, SELFOOD accurately identifies it as OOD and CE loss model fails to do so. Examples 3 and 4 have questions that revolve around eateries. CE-Loss model confuses and predicts them as OOD as TREC dataset contains only questions. However, SELFOOD was able to rank them higher and accurately identify them as in-distribution.
+
+# 6 Conclusion
+
+In this paper, we present SELFOOD, a novel framework for OOD detection that leverages only in-distribution samples as supervision. Building upon the insight that OOD samples typically do not belong to any in-distribution labels, we formulate the OOD detection problem as an inter-document intra-label ranking task. To address this challenge, we propose IDIL loss which guides the training process. Through extensive experiments on multiple datasets, we demonstrate the effectiveness of SELFOOD on OOD detection task. However, we
+
+Table 6: Qualitative analysis of CE-Loss vs SELFOOD with Yelp as in-distribution and TREC as OOD datasets. Wrong predictions are in Red and correct predictions are in Green.  
+
+<table><tr><td>Text</td><td>CE-Loss</td><td>SELFOOD</td></tr><tr><td>Who said “What contemptible scoundrel stole the cork from my lunch?”</td><td>InD</td><td>OOD</td></tr><tr><td>Who said “Give me liberty or give me death”?</td><td>InD</td><td>OOD</td></tr><tr><td>Its a tequila house - wouldn’t you want your staff to be knowledgeable about tequila? Call me silly. So Gregg and I are wondering Bourbon Street, looking for the next &quot;place&quot;...</td><td>OOD</td><td>InD</td></tr><tr><td>the pitch reads: &quot;what’s worse than mean girl cheerleaders? how about resurrected mean-girl cheerleaders with supernatural powers?&quot; And if this place could be farther away from my house, it’d be in Camden... Oh, the Vomitorium, every time i grace the hipster-stained ...</td><td>OOD</td><td>InD</td></tr></table>
+
+also acknowledge that it comes at the expense of text classification performance. Future research can focus on developing techniques that effectively balance inter-document and intra-document comparisons, enabling improved performance in both OOD detection and text classification tasks.
+
+# 7 Limitations
+
+As mentioned in the previous sections, the OOD detection performance of our method comes at the cost of classification performance. This limitation needs to be addressed.
+
+# 8 Ethics Statement
+
+This paper proposes a self-supervised OOD detection method. The aim of the paper is to minimize the human effort in annotating OOD documents and using only in-distribution documents for OOD detection. OOD detection is helpful in detection potential harmful content. Hence, we do not anticipate any major ethical concerns.
+
+# 9 Acknowledgments
+
+We thank the anonymous reviewers and our colleagues from Shang Data Lab, especially Bill Hogan and Zihan Wang for their helpful feedback. Our work is sponsored in part by NSF CAREER Award 2239440, NSF Proto-OKN Award 2333790, NIH Bridge2AI Center Program under award 1U54HG012510-01, Cisco-UCSD Sponsored Research Project, as well as generous gifts from Google, Adobe, and Teradata. Any opinions, findings, and conclusions or recommendations expressed herein are those of the authors and should
+
+not be interpreted as necessarily representing the views, either expressed or implied, of the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for government purposes notwithstanding any copyright annotation hereon.
+
+# References
+
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D 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 Ziegler, Jeffrey Wu, Clemens Winter, Chris 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 Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805.  
+Terrance DeVries and Graham W Taylor. 2018. Learning confidence for out-of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865.  
+Stefan Elfwing, Eiji Uchibe, and Kenji Doya. 2018. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural Networks, 107:3-11.  
+Geli Fei and Bing Liu. 2016. Breaking the closed world assumption in text classification. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 506-514.  
+Varun Gangal, Abhinav Arora, Arash Einolghozati, and Sonal Gupta. 2020. Likelihood ratios and generative classifiers for unsupervised out-of-domain detection in task-oriented dialog. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7764-7771.  
+Irving John Good. 1952. Rational decisions. Journal of the Royal Statistical Society: Series B (Methodological), 14(1):107-114.  
+Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR.  
+Chirag Gupta and Aaditya Ramdas. 2021. Top-label calibration and multiclass-to-binary reductions. arXiv preprint arXiv:2107.08353.
+
+Richard HR Hahnloser, Rahul Sarpeshkar, Misha A Mahowald, Rodney J Douglas, and H Sebastian Seung. 2000. Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit. nature, 405(6789):947-951.  
+Dan Hendrycks and Kevin Gimpel. 2017. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations.  
+Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. 2018. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606.  
+Eduard Hovy, Laurie Gerber, Ulf Hermjakob, ChinYew Lin, and Deepak Ravichandran. 2001. Toward semantics-based answer pinpointing. In Proceedings of the First International Conference on Human Language Technology Research.  
+Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. 2020. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10951-10960.  
+Amita Kamath, Robin Jia, and Percy Liang. 2020. Selective question answering under domain shift. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684-5696.  
+Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746-1751, Doha, Qatar. Association for Computational Linguistics.  
+Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang, and et al. 2019. An evaluation dataset for intent classification and out-of-scope prediction. 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 1311-1316.  
+Tanmay Laud, Daniel Spokoyny, Tom Corringham, and Taylor Berg-Kirkpatrick. 2023. Climabench: A benchmark dataset for climate change text understanding in english. arXiv preprint arXiv:2301.04253.  
+Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. 2018. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems, 31.
+
+Christophe Leys, Olivier Klein, Yves Dominicy, and Christophe Ley. 2018. Detecting multivariate outliers: Use a robust variant of the mahalanobis distance. Journal of experimental social psychology, 74:150-156.  
+Xiaoya Li, Jiwei Li, Xiaofei Sun, Chun Fan, Tianwei Zhang, Fei Wu, Yuxian Meng, and Jun Zhang. 2021.  $k$  folden:  $k$ -fold ensemble for out-of-distribution detection. arXiv preprint arXiv:2108.12731.  
+Xin Li and Dan Roth. 2002. Learning question classifiers. In COLING 2002: The 19th International Conference on Computational Linguistics.  
+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.  
+Dheeraj Mekala, Chengyu Dong, and Jingbo Shang. 2022a. LOPS: Learning order inspired pseudo-label selection for weakly supervised text classification. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4894-4908, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Dheeraj Mekala, Varun Gangal, and Jingbo Shang. 2021. Coarse2Fine: Fine-grained text classification on coarsely-grained annotated data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 583-594, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Dheeraj Mekala and Jingbo Shang. 2020. Contextualized weak supervision for text classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 323-333.  
+Dheeraj Mekala, Tu Vu, Timo Schick, and Jingbo Shang. 2022b. Leveraging qa datasets to improve generative data augmentation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing.  
+Jooyoung Moon, Jihyo Kim, Younghak Shin, and Sangheum Hwang. 2020. Confidence-aware learning for deep neural networks. In international conference on machine learning, pages 7034-7044. PMLR.  
+Anh Nguyen, Jason Yosinski, and Jeff Clune. 2015. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 427-436.  
+Yawen Ouyang, Jiasheng Ye, Yu Chen, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2021. Energy-based unknown intent detection with data manipulation. arXiv preprint arXiv:2107.12542.
+
+Alexander Podolskiy, Dmitry Lipin, Andrey Bout, Ekaterina Artemova, and Irina Pionkovskaya. 2021. Revisiting mahalanobis distance for transformer-based out-of-domain detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13675-13682.  
+Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.  
+Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. 2019. Likelihood ratios for out-of-distribution detection. Advances in neural information processing systems, 32.  
+Frank Rosenblatt. 1957. The perceptron, a perceiving and recognizing automaton Project Para. Cornell Aeronautical Laboratory.  
+Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199.  
+Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J. Gordon. 2019. An empirical study of example forgetting during deep neural network learning. In International Conference on Learning Representations.  
+Apoory Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L Willke. 2018. Out-of-distribution detection using an ensemble of self supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), pages 550-564.  
+Hongxin Wei, Renchunzi Xie, Hao Cheng, Lei Feng, Bo An, and Yixuan Li. 2022. Mitigating neural network overconfidence with logit normalization. In International Conference on Machine Learning, pages 23631-23644. PMLR.  
+Yanan Wu, Keqing He, Yuanmeng Yan, QiXiang Gao, Zhiyuan Zeng, Fujia Zheng, Lulu Zhao, Huixing Jiang, Wei Wu, and Weiran Xu. 2022a. Revisit overconfidence for ood detection: Reassigned contrastive learning with adaptive class-dependent threshold. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4165-4179.  
+Yanan Wu, Zhiyuan Zeng, Keqing He, Yutao Mou, Pei Wang, and Weiran Xu. 2022b. Distribution calibration for out-of-domain detection with Bayesian approximation. In Proceedings of the 29th International Conference on Computational Linguistics, pages 608-615, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.
+
+Hong Xu, Keqing He, Yuanmeng Yan, Sihong Liu, Zijun Liu, and Weiran Xu. 2020. A deep generative distance-based classifier for out-of-domain detection with mahalanobis space. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1452-1460.  
+Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Neural Information Processing Systems.  
+Zhiyuan Zeng, Keqing He, Yuanmeng Yan, Zijun Liu, Yanan Wu, Hong Xu, Huixing Jiang, and Weiran Xu. 2021a. Modeling discriminative representations for out-of-domain detection with supervised contrastive learning. arXiv preprint arXiv:2105.14289.  
+Zhiyuan Zeng, Hong Xu, Keqing He, Yuanmeng Yan, Sihong Liu, Zijun Liu, and Weiran Xu. 2021b. Adversarial generative distance-based classifier for robust out-of-domain detection. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7658-7662. IEEE.  
+Li-Ming Zhan, Haowen Liang, Bo Liu, Lu Fan, Xiaoming Wu, and Albert Lam. 2021. Out-of-scope intent detection with self-supervision and discriminative training. arXiv preprint arXiv:2106.08616.  
+Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 649-657. Curran Associates, Inc.  
+Wenxuan Zhou, Fangyu Liu, and Muhao Chen. 2021. Contrastive out-of-distribution detection for pretrained transformers. arXiv preprint arXiv:2104.08812.
+
+# A Appendix
+
+# A.1 Results with NYT & Clima as in-distribution datasets
+
+We present experimental results on Clima and NYT as in-distribution datasets with BERT and RoBERTa classifiers in Table 7.
+
+# A.2 Experimental Settings
+
+Training trajectory analysis experiment. We compare two experiments showcasing empirical evidence on how the trajectory of CE-loss models and IDI loss models differ. In 3(b) we plot the change in validation FPR on BERT-classifiers optimized on CE-loss and IDIL loss. Here, we calculate the FPR across training steps on a validation dataset consisting of 500 in-distribution and OOD data points each to compare the training trajectories of the two models.
+
+We generate the toy dataset by creating five isometric Gaussian blobs of 3-dimensional data points, with each cluster containing 1000 samples. The centers of all the clusters are bounded by the range  $[-10.0, 10.0]$  and the cluster standard deviation is set to 2.0 to account for limited overlap between cluster datapoints. We consider four clusters as being in-distribution and the remaining cluster as OOD. We train a logistic regression classifier on the in-domain datapoints initialized with the same weights and minimizing IDIL and CE-loss. The model is optimized using Stochastic Gradient Descent with a learning rate of 0.01 for 300 epochs. The weights are visualized at the beginning of each epoch. In order to facilitate visualization, the weight matrices of both approaches of dimension are flattened and Principal Component Analysis (PCA) is applied to reduce the dimensionality of the weight matrix. We plot the top-3 principal components after each epoch to conclude that the learned weights of CE-loss and SELFOOD progress in a different manner.
+
+<table><tr><td rowspan="2">In-dist</td><td rowspan="2">OOD</td><td rowspan="2">Method</td><td colspan="4">BERT</td><td colspan="4">RoBERTa</td></tr><tr><td>FPR95(↓)</td><td>ERR(↓)</td><td>AUROC(↑)</td><td>AUPR(↑)</td><td>FPR95(↓)</td><td>ERR(↓)</td><td>AUROC(↑)</td><td>AUPR(↑)</td></tr><tr><td rowspan="21">NYT</td><td rowspan="7">Clima</td><td>LogitNorm</td><td>20.6</td><td>1.0</td><td>96.8</td><td>92.3</td><td>18.4</td><td>1.0</td><td>97.3</td><td>93.6</td></tr><tr><td>OneVsRest</td><td>84.7</td><td>1.7</td><td>87.3</td><td>82.6</td><td>92.7</td><td>1.9</td><td>91.1</td><td>85.1</td></tr><tr><td>RCL</td><td>46.9</td><td>1.5</td><td>94.4</td><td>87.5</td><td>86.7</td><td>2.3</td><td>86.7</td><td>76.6</td></tr><tr><td>Bayes Approx</td><td>40.6</td><td>1.2</td><td>95.4</td><td>90.0</td><td>80.0</td><td>2.4</td><td>88.4</td><td>76.5</td></tr><tr><td>CE Loss</td><td>22.1</td><td>1.0</td><td>96.9</td><td>92.5</td><td>15.7</td><td>1.0</td><td>97.3</td><td>92.8</td></tr><tr><td>CRL</td><td>32.3</td><td>1.7</td><td>95.4</td><td>86.5</td><td>79.3</td><td>3.2</td><td>79.8</td><td>67.0</td></tr><tr><td>SELFOOD</td><td>10.5</td><td>1.6</td><td>98.4</td><td>90.5</td><td>4.6</td><td>1.0</td><td>99.3</td><td>95.4</td></tr><tr><td rowspan="7">TREC</td><td>LogitNorm</td><td>13.3</td><td>5.5</td><td>97.2</td><td>87.9</td><td>17.5</td><td>6.4</td><td>96.1</td><td>84.1</td></tr><tr><td>OneVsRest</td><td>96.0</td><td>6.2</td><td>84.0</td><td>79.4</td><td>97.2</td><td>6.3</td><td>90.5</td><td>82.8</td></tr><tr><td>RCL</td><td>62.9</td><td>6.1</td><td>92.2</td><td>85.8</td><td>73.9</td><td>8.8</td><td>88.4</td><td>73.1</td></tr><tr><td>Bayes Approx</td><td>38.8</td><td>4.5</td><td>94.8</td><td>89.9</td><td>67.5</td><td>8.3</td><td>88.7</td><td>74.7</td></tr><tr><td>CE Loss</td><td>14.5</td><td>5.8</td><td>97.0</td><td>87.1</td><td>21.9</td><td>6.7</td><td>95.5</td><td>83.5</td></tr><tr><td>CRL</td><td>67.9</td><td>6.5</td><td>89.9</td><td>79.1</td><td>85.6</td><td>9.6</td><td>71.4</td><td>61.4</td></tr><tr><td>SELFOOD</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td><td>0.0</td><td>0.0</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan="7">Yelp</td><td>LogitNorm</td><td>1.7</td><td>0.6</td><td>99.3</td><td>87.9</td><td>1.2</td><td>0.6</td><td>98.5</td><td>87.9</td></tr><tr><td>OneVsRest</td><td>15.1</td><td>0.7</td><td>97.4</td><td>78.1</td><td>8.7</td><td>0.6</td><td>97.6</td><td>78.8</td></tr><tr><td>RCL</td><td>45.1</td><td>0.5</td><td>94.9</td><td>84.7</td><td>76.7</td><td>1.0</td><td>90.2</td><td>60.0</td></tr><tr><td>Bayes Approx</td><td>28.7</td><td>0.4</td><td>96.4</td><td>86.8</td><td>83.0</td><td>1.1</td><td>87.7</td><td>56.6</td></tr><tr><td>CE Loss</td><td>3.4</td><td>0.7</td><td>99.2</td><td>82.3</td><td>1.8</td><td>0.6</td><td>99.3</td><td>85.4</td></tr><tr><td>CRL</td><td>69.5</td><td>0.5</td><td>89.3</td><td>77.8</td><td>73.3</td><td>1.0</td><td>83.8</td><td>58.3</td></tr><tr><td>SELFOOD</td><td>22.5</td><td>1.1</td><td>95.7</td><td>57.8</td><td>0.2</td><td>0.2</td><td>99.9</td><td>97.5</td></tr><tr><td rowspan="21">Clima</td><td rowspan="7">TREC</td><td>LogitNorm</td><td>37.8</td><td>5.5</td><td>94.6</td><td>91.9</td><td>60.5</td><td>7.0</td><td>91.6</td><td>88.0</td></tr><tr><td>OneVsRest</td><td>25.5</td><td>6.8</td><td>95.5</td><td>91.8</td><td>54.3</td><td>15.5</td><td>84.5</td><td>67.7</td></tr><tr><td>RCL</td><td>80.1</td><td>18.4</td><td>75.7</td><td>55.3</td><td>91.2</td><td>18.4</td><td>66.6</td><td>47.7</td></tr><tr><td>Bayes Approx</td><td>83.7</td><td>15.8</td><td>77.5</td><td>63.8</td><td>65.7</td><td>13.9</td><td>83.6</td><td>70.7</td></tr><tr><td>CE Loss</td><td>34.9</td><td>5.6</td><td>94.9</td><td>92.2</td><td>60.2</td><td>9.1</td><td>90.4</td><td>83.8</td></tr><tr><td>CRL</td><td>83.0</td><td>12.7</td><td>82.0</td><td>72.2</td><td>60.9</td><td>12.6</td><td>79.2</td><td>63.2</td></tr><tr><td>SELFOOD</td><td>0.3</td><td>1.1</td><td>99.4</td><td>99.4</td><td>38.9</td><td>8.8</td><td>90.1</td><td>84.7</td></tr><tr><td rowspan="7">Yelp</td><td>LogitNorm</td><td>57.7</td><td>0.8</td><td>93.0</td><td>81.3</td><td>54.7</td><td>1.2</td><td>92.6</td><td>72.3</td></tr><tr><td>OneVsRest</td><td>43.7</td><td>1.0</td><td>92.7</td><td>72.5</td><td>58.8</td><td>1.8</td><td>73.5</td><td>34.7</td></tr><tr><td>RCL</td><td>91.0</td><td>2.3</td><td>60.7</td><td>12.2</td><td>98.0</td><td>2.3</td><td>56.4</td><td>12.7</td></tr><tr><td>Bayes Approx</td><td>81.4</td><td>2.2</td><td>76.5</td><td>22.7</td><td>96.8</td><td>2.2</td><td>52.3</td><td>16.3</td></tr><tr><td>CE Loss</td><td>46.3</td><td>0.9</td><td>93.8</td><td>80.1</td><td>79.5</td><td>1.9</td><td>86.5</td><td>47.2</td></tr><tr><td>CRL</td><td>98.7</td><td>1.4</td><td>67.3</td><td>53.3</td><td>72.9</td><td>2.1</td><td>50.9</td><td>19.0</td></tr><tr><td>SELFOOD</td><td>14.7</td><td>0.6</td><td>98.0</td><td>90.8</td><td>37.6</td><td>1.5</td><td>72.3</td><td>54.9</td></tr><tr><td rowspan="7">NYT</td><td>LogitNorm</td><td>37.6</td><td>4.2</td><td>94.4</td><td>87.3</td><td>46.1</td><td>4.5</td><td>93.1</td><td>85.4</td></tr><tr><td>OneVsRest</td><td>34.6</td><td>4.8</td><td>94.3</td><td>85.3</td><td>40.6</td><td>9.4</td><td>83.1</td><td>51.3</td></tr><tr><td>RCL</td><td>90.3</td><td>12.1</td><td>64.6</td><td>28.8</td><td>97.9</td><td>12.4</td><td>51.5</td><td>21.2</td></tr><tr><td>Bayes Approx</td><td>87.4</td><td>11.2</td><td>70.4</td><td>38.3</td><td>96.8</td><td>11.3</td><td>55.8</td><td>29.2</td></tr><tr><td>CE Loss</td><td>44.9</td><td>4.8</td><td>93.2</td><td>84.4</td><td>62.5</td><td>6.7</td><td>89.6</td><td>75.2</td></tr><tr><td>CRL</td><td>85.6</td><td>7.6</td><td>78.0</td><td>62.5</td><td>61.3</td><td>9.4</td><td>61.8</td><td>39.9</td></tr><tr><td>SELFOOD</td><td>17.2</td><td>2.6</td><td>97.8</td><td>94.5</td><td>55.3</td><td>6.7</td><td>68.6</td><td>61.5</td></tr></table>
+
+Table 7: OOD detection results with BERT & RoBERTa classifiers. Each experiment is repeated with three random seeds and the mean scores are reported. The false-positive-rate at  $95\%$  true-positive-rate (FPR95), minimum detection error over all thresholds (ERR), the area under the risk-coverage curve (AURC), and the area under the precision-recall curve (AUPR) using in-distribution samples as the positives are used as evaluation metrics.

2023/SELFOOD_ Self-Supervised Out-Of-Distribution Detection via Learning to Rank/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d85ad0c614444455a7c806ea28693d1e1e0d9104bb0f1c2216a36db2169d2d3
+size 738329

2023/SGP-TOD_ Building Task Bots Effortlessly via Schema-Guided LLM Prompting/586b129b-aeb9-4225-8c30-4a6d80f38793_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5566bfa6cf159e5a02203048d0ecec14f87cf05164e87293af7ead24acd384f8
+size 1498454

2023/SGP-TOD_ Building Task Bots Effortlessly via Schema-Guided LLM Prompting/full.md

@@ -0,0 +1,644 @@
+# SGP-TOD: Building Task Bots Effortlessly via Schema-Guided LLM Prompting
+
+Xiaoying Zhang $^{1}$ , Baolin Peng $^{2}$ , Kun Li $^{1}$ , Jingyan Zhou $^{1}$ , Helen Meng $^{1,3}$
+
+<sup>1</sup>The Chinese University of Hong Kong, Hong Kong
+
+2Tencent AI Lab, Bellevue
+
+<sup>3</sup>Centre for Perceptual and Interactive Intelligence, Hong Kong
+
+{zhangxy, kunli, jyzhou, hmmeng} @se.cuhk.edu.hk
+
+{baolinpeng} @ global.tencent.com
+
+# Abstract
+
+Building and maintaining end-to-end task bots using minimal human effort is a longstanding challenge in dialog research. In this work, we introduce SGP-TOD, Schema-Guided Prompting for building Task-Oriented Dialog systems effortlessly based on large language models (LLMs). Utilizing the predefined task schema, i.e., belief instruction and dialog policy, we instruct fixed LLMs to generate appropriate responses on novel tasks, without the need for training data. Specifically, SGP-TOD comprises three components: an LLM for interacting with users, a Dialog State Tracking (DST) Prompter to aid the LLM in tracking dialog states with the given belief instruction, and a Policy Prompter to direct the LLM to generate proper responses adhering to the provided dialog policy. Experimental results on Multiwoz, RADDLE, and STAR datasets show that our training-free strategy, SGP-TOD, yields state-of-the-art (SOTA) zero-shot performance, significantly surpassing the few-shot approaches. In a domain-extension setting, SGP-TOD aptly adapts to new functionalities by merely adding supplementary schema rules. We make our code and data publicly available.
+
+# 1 Introduction
+
+Building task-oriented dialog (TOD) systems has been a long-standing challenge in artificial intelligence. The prevailing approach for creating task bots (Hosseini-Asl et al., 2020; Peng et al., 2021a; Sun et al., 2022) is to fine-tune pre-trained language models (PLMs), such as T5 (Raffel et al., 2020) and GPT-2 (Radford et al., 2019). Despite their great success, developing and maintaining such task bots generally requires adequate annotated data and extensive fine-tuning/re-training.
+
+Recently, large Language Models (LLMs), such as ChatGPT (OpenAI, 2022) and GPT-4 (OpenAI,
+
+![](images/e0a3353d60c7384ec98e6f3102ab0499d5863922fc30ac74f2f445653774882e.jpg)  
+Figure 1: The proposed SGP-TOD is depicted with a dialog example, where the prompters integrate the task schema (right) to assist the frozen LLM in generating an appropriate response (left).
+
+2023), have revolutionized natural language processing (NLP) applications (Wei et al., 2022; Wang et al., 2023), owing to their remarkable conversational skills (Qin et al., 2023), instruction-following abilities (Ouyang et al., 2022) and zero-shot generalization capabilities (Chowdhery et al., 2022a; Hu et al., 2022). This raises a research question: can LLMs be effectively utilized for building task bots with minimum human effort?
+
+A contemporary study (Hudecek and Dusek, 2023) explores the potential of LLMs for rapidly building task bots via few-shot prompting, a.k.a. in-context learning (ICL) paradigm (Brown et al., 2020; Madotto et al., 2021). Though demonstrably effective, the ICL performance is highly influenced by the quality of the in-context exemplars (Zhao et al., 2021; Liu et al., 2022; Dong et al., 2023), as they struggle to provide comprehensive information for dialog task completion.
+
+In this work, we introduce symbolic knowledge (Nye et al., 2021; Cheng et al., 2023), i.e., the task schema into LLMs, for creating task bots. Task schema (Mosig et al., 2020; Mehri and Eskenazi, 2021) encompasses a concise symbolic representation of a task, supplying LLMs with a comprehensive blueprint. It comprises  $(i)$  task-specific ontology containing all slots and their appropriate values (Budzianowski et al., 2018); and  $(ii)$  a dialog flow explicitly outlining fundamental interaction patterns (Peng et al., 2021b). Specifically, we propose SGP-TOD (as depicted in Figure 1), a schema-guided prompting method for rapidly building task bots. We integrate the predefined task schema and dialog context into prompts through the use of two specifically-designed prompters, namely a DST Prompter and a Policy Prompter. Utilizing these prompters, we adeptly guide fixed LLMs to track dialog states, retrieve database entries, and generate appropriate responses for novel tasks in a zero-shot manner, without the need for additional training or fine-tuning. By incorporating task-specific symbolic knowledge into LLMs, SGP-TOD provides knowledge-based, coherent and human-like responses. Moreover, this training-free design empowers developers to flexibly prototype dialog systems on new tasks, while seamlessly extending system functionalities through modifying the task schema.
+
+We perform empirical automatic evaluations on two multi-domain datasets, namely, Multiwoz 2.0 and 2.2 (Budzianowski et al., 2018; Zang et al., 2020), as well as two single-domain/task datasets, RADDLE (Peng et al., 2021a) and STAR (Mosig et al., 2020), within zero-shot scenarios. Additionally, we complement these assessments with interactive human evaluations. The results indicate that SGP-TOD, employing merely task schema devoid of any training or fine-tuning, substantially boosts the SOTA zero-shot results, markedly outperforming few-shot prompting/fine-tuning methods, and even attaining competitive results cf. full-shot finetuning approaches. In a domain-extension context, SGP-TOD proficiently adapts to new functionalities by simply adding a handful of schema rules without necessitating further data collection, significantly exceeding the few-shot prompting/fine-tuning methods reinforced by machine teaching (Williams and Liden, 2017).
+
+In summary, our contributions are three-fold:
+
+- We propose SGP-TOD, a schema-guided
+
+LLM prompting strategy that facilitates in effortlessly creating task bots, eliminating the necessity for task-specific data or fine-tuning.
+
+- We integrate symbolic knowledge - task schema into LLMs, allowing them to generate schema-compliant responses and adaptively expand their functionalities to tackle new tasks by solely modifying the task schema.
+
+- We demonstrate the effectiveness of SGP-TOD on Multiwoz, RADDLE, STAR datasets in zero-shot settings using both automatic and human evaluations. SGP-TOD notably elevates the SOTA zero-shot performance.
+
+# 2 Related work
+
+Zero-Shot Task-Oriented Dialog Modeling. Zero-shot generalization is an essential yet challenging task in TOD research. A comprehensive study is shown in Appendix A. In this paper, we focus on zero-shot end-to-end dialog modeling, including policy management and dialog generation.
+
+The works by Zhao and Eskenazi (2018) and Qian and Yu (2019) utilize ontology and response templates to train dialog models, enabling the discovery of shared dialog policies between the source and target domains. To enable broader adaptation to diverse dialog policies, Mosig et al. (2020); Mehri and Eskenazi (2021) implement task-specific policy skeletons, training dialog models to adhere to novel policies. Furthermore, Zhao et al. (2022) employs a neural language model (LM) for tracking dialog states and user actions using slot and action descriptions; subsequently, a policy program is deployed to facilitate an LM in generating system actions and responses. Despite the effectiveness of previous approaches, they still require ample fine-tuning and copious annotated dialog corpora on source or heterogeneous domains/tasks.
+
+A concurrent study to ours is Hudecek and Dusek (2023), which employs a prompting strategy - IG-TOD (instruction-guided TOD) to guide frozen LLMs in generating suitable responses. Specifically, IG-TOD first tracks belief states by utilizing slot descriptions as prompts, then retrieves database entries, and generates responses. Our SGP-TOD differs in that:  $(i)$  we employ slot names and value examples, rather than slot descriptions, as prompts to facilitate frozen LLMs in generating belief states, thereby reducing human effort;  $(ii)$  we offer a policy skeleton to guide
+
+LLMs in producing appropriate responses. In addition, experimental results indicate that SGP-TOD substantially outperforms IG-TOD.
+
+Leveraging LLMs for Dialog Tasks. LLMs (Chowdhery et al., 2022b; OpenAI, 2023) have exhibited unparalleled mastery of natural language understanding, reasoning and generation (Wei et al., 2022; Bubeck et al., 2023). Three primary research directions have obtained substantial success in numerous dialog tasks by utilizing LLMs.  $(i)$  Few-shot prompting (Brown et al., 2020) has showcased remarkable performance in intent classification (Yu et al., 2021), semantic parsing (Shin and Van Durme, 2022), dialog state tracking (Hu et al., 2022; Xie et al., 2022), and response generation (Madotto et al., 2021).  $(ii)$  Li et al. (2022); Mehri et al. (2022); Dai et al. (2023) employ LLMs for data augmentation, i.e., generating synthetic task-oriented dialogs to train smaller models for inference.  $(iii)$  Recently, several studies endeavor to support LLMs in specialized tasks by incorporating external knowledge. Peng et al. (2023) advocates for enhancing LLMs' responses with external knowledge and automated feedback to reduce hallucination. Liang et al. (2023) suggests connecting LLMs with millions of APIs to accomplish diverse tasks. Different from the aforementioned works, we aim to employ LLMs in building task bots in a zero-shot manner using pre-defined task schema.
+
+# 3 SGP-TOD
+
+# 3.1 Overview
+
+The overall architecture of the proposed SGP-TOD (Figure 1) consists of three key components:  $(i)$  an LLM, responsible for adhering to instructions, comprehending user queries, and generating coherent responses for user interaction;  $(ii)$  a DST Prompter, tasked with supporting the LLM in tracking dialogue states using the belief instruction;  $(iii)$  a Policy Prompter, guiding the LLM to adhere to the predefined task policy for providing suitable system actions and responses.
+
+At each dialog turn  $t$ , the end-to-end generation task is systematically divided into three subsequent sub-tasks: (i) Belief State Prediction - given the dialog history up to current dialog turn  $\pmb{h}_t$ , which is a sequence of utterances alternating between the user and the system  $\pmb{h}_t = [u_1, r_1, u_2, r_2, \dots, u_t]$  (where  $u$  and  $r$  denote user and system utterances, respectively), the DST Prompter embeds the belief instruction  $\pmb{BI}$  to direct the frozen LLM (pa
+
+rameterized by  $\theta$  ) in generating a belief state  $b_{t}$  (Equation 1). The belief state is then used to query a database and obtain the database (DB) state  $c_{t}$  (Equation 2). (ii) System Action Determination - the Policy Prompt incorporates a policy skeleton  $PS$ , assisting the LLM in generating a system action  $a_{t}$ , based on  $h_t, b_t$ , and  $c_{t}$  (Equation 3). (iii) Dialog Response Generation - grounded in the dialog history  $h_t$ , belief state  $b_{t}$ , DB state  $c_{t}$ , system action  $a_{t}$ , the Policy Prompt aids the LLM in generating a delexicalized response by providing the policy skeleton  $PS$  (Equation 4). Ultimately, the delexicalized response is automatically postprocessed to generate system response in natural language. Detailed illustration with a dialog example is shown in Appendix L.
+
+$$
+\boldsymbol {b} _ {t} = \boldsymbol {L L M} _ {\theta} \left(\boldsymbol {h} _ {t}, \boldsymbol {B I}\right) \tag {1}
+$$
+
+$$
+\boldsymbol {c} _ {t} = \boldsymbol {D} \boldsymbol {B} \left(\boldsymbol {b} _ {t}\right) \tag {2}
+$$
+
+$$
+\boldsymbol {a} _ {t} = \boldsymbol {L L M} _ {\boldsymbol {\theta}} \left(\boldsymbol {h} _ {t}, \boldsymbol {b} _ {t}, \boldsymbol {c} _ {t}, \boldsymbol {P S}\right) \tag {3}
+$$
+
+$$
+\boldsymbol {r} _ {t} = \boldsymbol {L L M} _ {\theta} \left(\boldsymbol {h} _ {t}, \boldsymbol {b} _ {t}, \boldsymbol {c} _ {t}, \boldsymbol {a} _ {t}, \boldsymbol {P S}\right) \tag {4}
+$$
+
+# 3.2 LLM
+
+An LLM is responsible for following task-specific instructions and generating appropriate responses.
+
+Many off-the-shelf LLMs, e.g., ChatGPT, Codex (Chen et al., 2021), are pre-trained on massive corpora of text data and/or code data. In addition, they are trained to follow instructions in the prompts (Ouyang et al., 2022) and provide pertinent responses. Exhibiting remarkable proficiencies in natural language processing, instruction compliance, and zero-shot generalization across diverse downstream dialog tasks, these LLMs serve as valuable foundation models for our approach.
+
+# 3.3 DST Prompter
+
+Given the dialog history  $h_t$ , the DST prompter aims to guide the LLM in predicting the belief state  $\mathbf{b}_t$  at each turn  $t$ , using the belief instruction  $\mathbf{BI}$ . The belief state  $\mathbf{b}_t$  is defined as the concatenation of the domain/task (i.e., user intent)  $\mathbf{d}_t$  and a set of slot-value pairs  $\{(s_t^i, v_t^i); i = 1, \dots, n_t\}$ , where  $n_t$  is the total number of pairs in the set.
+
+As shown in Figure 2, the proposed DST prompter contains four parts: (i) a task instruction that offers general guidance on belief state prediction; (ii) belief instructions  $\mathbf{BI}$  of all domains/tasks; (iii) a formatting example illustrating
+
+![](images/244328e92372df955ccf4910dc61c3eef7286cbddb07beb437c4a63f89fc5422.jpg)  
+Figure 2: Illustration of belief state prediction utilizing DST Prompter. The predicted belief state is highlighted.
+
+the anticipated output format to direct the LLM, in addition, we follow Hu et al. (2022) and adopt SQL state to represent the dialog state  $\boldsymbol{b}_t^3$ ; and  $(iv)$  the test input, i.e., the given dialog history  $\boldsymbol{h}_t$ . Since the prompt is fixed and no labeled data from the target task or domain is used, we refer to this setting as "zero-shot", following Wang et al. (2022b).
+
+Belief Instruction. For each task/domain, the belief instruction contains the task/domain name, all potential slot names, and their possible values (Figure 2). Regarding categorical slots, such as the "price range" in the restaurant domain, all plausible values are included, i.e., "don't care", "cheap", "moderate", and "expensive"; whereas, for non-categorical slots, such as "name", only a few value examples are injected, e.g., Pizza Hut City, Golden Wok, etc. Detailed belief instructions for all tasks/domains can be found in Appendix B.
+
+![](images/5f0112cb022cbc5ef025613602d020369fbe24c11c4c7db8c06a26c37538b4e9.jpg)  
+Figure 3: Illustration of system action determination and response generation employing the Policy Prompter. The pertinent template turns, previously predicted belief state, retrieved DB state within the input, alongside the generated system action and generated response in the output are accentuated.
+
+# 3.4 Policy Prompt
+
+Dialog policy, governing the behavior of task bots, plays a crucial role in task-oriented dialogs. To represent the dialog policy for a given task, we utilize a policy skeleton, which delineates interaction patterns and encompasses business logic in the form of template dialog flows (Peng et al., 2021b). The Policy Prompter is devised to guide the static LLM in adhering to the policy skeleton  $PS$ , enabling the sequential generation of appropriate system actions  $\mathbf{a}_t$  and responses  $\mathbf{r}_t$ .
+
+Analogous to the DST Prompt, the Policy Prompt (Figure 3) comprises four components: (i) a task instruction; (ii) a formatting example derived from another task/domain, consisting of a par
+
+tial policy skeleton and its associated dialogue turn exemplar (in Appendix C); (iii) a policy skeleton for the previously predicted domain/task; and (iv) the test input, i.e., the dialog history  $\pmb{h}_t$ , generated belief state  $\pmb{b}_t$ , and obtained DB state  $\pmb{c}_t$ .
+
+Policy Skeleton. Given that user behaviors and DB results jointly determine system actions and responses, policy skeleton is designed to cover all fundamental user behaviors and characteristic DB results, along with their corresponding system actions and responses.<sup>5</sup> Considering the infeasibility of developing a multi-task/domain policy skeleton for every possible combination of tasks and domains, we opt to develop a distinct policy skeleton tailored to each specific task and domain.
+
+Following Mehri and Eskenazi (2021), our strategy converts the established dialog policy into a series of template dialog turns  $\mathcal{X}$  that are logically arranged and concentrate on task completion:
+
+$$
+\begin{array}{l} \mathcal {X} = \left\{\boldsymbol {x} _ {i} \right\} _ {i = 1} ^ {N}, \tag {5} \\ \boldsymbol {x} _ {i} = \left(u ^ {i}, a ^ {i}, r ^ {i}\right) o r \left(c ^ {i}, a ^ {i}, r ^ {i}\right) \\ \end{array}
+$$
+
+where  $\pmb{x}_i$  is a template dialog turn, which contains a user utterance  $u^i$  or a DB state  $c^i$ , matching system action  $a^i$ , and system response  $r^i$ .  $N$  denotes the total number of template turns within the policy skeleton (around 10-20 template turns depending on the task complexity). In order to equip the frozen LLM with new capabilities or modify current ones, we only need insert, amend, or eliminate a few template turns within the policy skeleton.
+
+# 4 Experiments
+
+# 4.1 Experimental Setup
+
+Datasets. (i) Two multi-domain dialog datasets: Multiwoz 2.0 (Budzianowski et al., 2018) and Multiwoz 2.2 (Zang et al., 2020). (ii) Two single-domain/task datasets: RADDLE (Peng et al., 2021a,c) and STAR (Mosig et al., 2020) (single-task dialogs from the corpus, following the "happy path"). Details are elaborated in Appendix D.
+
+Automatic Evaluation Metrics. We evaluate the end-to-end dialog generation performance using the same metrics as those listed in Budzianowski et al. (2018): Inform  $(\%)$ , Success  $(\%)$ , BLEU  $(\%)$  (Papineni et al., 2002) and Combined  $(\%)$  judges the overall quality, defined as Combined =
+
+(Inform + Success) × 0.5 + BLEU. Additionally, we utilize BERTScore(%) (Zhang* et al., 2020).
+
+Following Mehri and Eskenazi (2021), we perform the next action prediction task on STAR (wherein the system actions and response templates are mapped one to one), which predicts next system action given the dialog history. We report the results using F1score  $(\%)$  and accuracy  $(\%)$ .
+
+Human Evaluation Metrics. We conduct interactive human evaluations (by five student helpers), following the evaluation protocol in the DSTC9 Track 1 challenge (Gunasekara et al., 2020). For each dialog session, students are mandated to interact with a dialog agent via natural language and assess the overall dialog quality employing these five metrics: (i) Success w/o g(%), (ii) Success w/ g(%), (iii) Understanding(1-5), (iv) Appropriateness(1-5) and (v) Turns. Full details are elaborated in Appendix D.
+
+Compared Methods. We compare the proposed SGP-TOD with SOTA zero-shot transfer methods and zero-shot/few-shot prompting strategies. (We report the mean results of three different runs.)
+
+# Zero-shot transfer methods:
+
+- BERT+S (Mosig et al., 2020) augments a BERT-base classifier (Devlin et al., 2019) with a system-side schema to predict system action.  
+- SAM (Mehri and Eskenazi, 2021) is based on BERT-base, which uses a user-aware schema to predict the next system action.  
+- ANYTOD-XXL (Zhao et al., 2022) adopts T5-XXL (Roberts et al., 2022) to generate system actions and responses utilizing slot/action descriptions and a policy program. It is pretrained on SGD dataset (Rastogi et al., 2020a).
+
+# Prompting methods:
+
+- IG-TOD-CHATGPT (Hudecek and Dusek, 2023) is a prompting approach based on ChatGPT, exploiting slot descriptions for tracking dialog states, fetching DB entries, and generating responses. IG-TOD-CHATGPT-ZS and IG-TOD-CHATGPT-FS are in the zero-shot and few-shot settings, respectively.  
+- FEW-SHOT-CHATGPT is a few-shot prompting approach applied to ChatGPT, utilizing a few (i.e.,  $\pmb{k}$ ) training dialog turns as prompts. Optimal results are achieved with  $k = 15$  on Multiwoz and  $k = 10$  on RADDLE.  
+- SGP-TOD (Ours) is compatible with any off-the-shelf LLMs. In this paper, we employ
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">Multiwoz 2.0</td><td colspan="4">Multiwoz 2.2</td></tr><tr><td>Inform</td><td>Success</td><td>BLEU</td><td>Combined</td><td>Inform</td><td>Success</td><td>BLEU</td><td>Combined</td></tr><tr><td colspan="9">Full-shot fine-tuning (with 8.4k+ training dialogs):</td></tr><tr><td>DAMD (Zhang et al., 2020)</td><td>76.33</td><td>60.40</td><td>16.60</td><td>84.97</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SIMPLETOD (Hosseini-Asl et al., 2020)</td><td>84.40</td><td>70.10</td><td>15.01</td><td>92.26</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SOLOIST (Peng et al., 2021a)</td><td>85.50</td><td>72.90</td><td>16.54</td><td>95.74</td><td>81.70</td><td>67.10</td><td>13.60</td><td>88.00</td></tr><tr><td>PPTOD (Su et al., 2022)</td><td>89.20</td><td>79.40</td><td>18.62</td><td>102.92</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MARS (Sun et al., 2022)</td><td>88.90</td><td>78.00</td><td>19.90</td><td>103.35</td><td>88.90</td><td>78.00</td><td>19.60</td><td>103.05</td></tr><tr><td colspan="9">Zero-shot transfer method (pre-trained on SGD):</td></tr><tr><td>ANYTOD-XXL</td><td>-</td><td>-</td><td>-</td><td>-</td><td>73.90</td><td>24.40</td><td>3.40</td><td>52.55</td></tr><tr><td colspan="9">Few-shot prompting:</td></tr><tr><td>IG-TOD-CHATGPT-FS</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>20.00</td><td>7.17</td><td>-</td></tr><tr><td>FEW-SHOT-CHATGPT</td><td>44.74</td><td>24.32</td><td>7.88</td><td>42.41</td><td>45.40</td><td>24.50</td><td>7.72</td><td>42.67</td></tr><tr><td colspan="9">Zero-shot prompting:</td></tr><tr><td>IG-TOD-CHATGPT-ZS</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>15.00</td><td>3.58</td><td>-</td></tr><tr><td>SGP-TOD-CHATGPT (Ours)</td><td>64.56</td><td>54.05</td><td>7.17</td><td>66.48</td><td>64.70</td><td>54.70</td><td>6.96</td><td>66.66</td></tr><tr><td>SGP-TOD-CODEX (Ours)</td><td>71.67</td><td>52.55</td><td>7.91</td><td>70.02</td><td>75.50</td><td>52.30</td><td>6.62</td><td>70.53</td></tr><tr><td>SGP-TOD-GPT3.5 (Ours)</td><td>83.88</td><td>69.87</td><td>9.09</td><td>85.97</td><td>82.00</td><td>72.50</td><td>9.22</td><td>86.47</td></tr></table>
+
+Table 1: End-to-end dialog generation evaluation results on Multiwoz. Results of SOLOIST, MARS, ANYTOD-XXL on Multiwoz 2.2 are cited from Zhao et al. (2022). Results of IG-TOD-CHATGPT are cited from Hudecek and Dusek (2023). Other results of the full-shot fine-tuning methods are cited from He et al. (2022) and Sun et al. (2022). (Difference in mean is significant with  $p < 0.01$ .)  
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">Attraction</td><td colspan="4">Train</td><td colspan="4">Hotel</td><td colspan="4">Restaurant</td></tr><tr><td>Info.</td><td>Succ.</td><td>BLEU</td><td>Comb.</td><td>Info.</td><td>Succ.</td><td>BLEU</td><td>Comb.</td><td>Info.</td><td>Succ.</td><td>BLEU</td><td>Comb.</td><td>Info.</td><td>Succ.</td><td>BLEU</td><td>Comb.</td></tr><tr><td colspan="17">Few-shot fine-tuning (with 50 training dialogs):</td></tr><tr><td>SIMPLETOD</td><td>65.66</td><td>46.97</td><td>5.85</td><td>62.17</td><td>59.00</td><td>44.00</td><td>7.07</td><td>58.57</td><td>62.50</td><td>40.00</td><td>7.70</td><td>58.95</td><td>75.50</td><td>44.50</td><td>11.00</td><td>71.00</td></tr><tr><td>SOLOIST</td><td>86.00</td><td>65.00</td><td>12.90</td><td>88.40</td><td>80.81</td><td>64.65</td><td>9.96</td><td>82.69</td><td>74.50</td><td>43.50</td><td>8.12</td><td>67.12</td><td>81.00</td><td>55.50</td><td>12.80</td><td>81.50</td></tr><tr><td colspan="17">Few-shot prompting:</td></tr><tr><td>Few-SHOT-CHATGPT</td><td>75.00</td><td>67.00</td><td>8.22</td><td>79.23</td><td>79.80</td><td>65.66</td><td>8.12</td><td>80.85</td><td>51.00</td><td>26.50</td><td>5.80</td><td>44.55</td><td>80.00</td><td>55.50</td><td>7.71</td><td>75.46</td></tr><tr><td colspan="17">Zero-shot prompting:</td></tr><tr><td>SGP-TOD-CHATGPT</td><td>95.00</td><td>94.00</td><td>7.13</td><td>101.63</td><td>76.77</td><td>74.24</td><td>6.75</td><td>82.26</td><td>76.50</td><td>57.00</td><td>5.16</td><td>71.91</td><td>90.00</td><td>82.50</td><td>6.72</td><td>92.97</td></tr><tr><td>SGP-TOD-CODEX</td><td>98.00</td><td>93.00</td><td>10.45</td><td>105.95</td><td>78.79</td><td>70.20</td><td>8.56</td><td>83.06</td><td>83.50</td><td>69.50</td><td>7.86</td><td>84.36</td><td>91.00</td><td>85.00</td><td>10.50</td><td>98.50</td></tr><tr><td>SGP-TOD-GPT3.5</td><td>96.00</td><td>93.00</td><td>9.53</td><td>104.03</td><td>82.83</td><td>77.27</td><td>8.72</td><td>88.77</td><td>82.50</td><td>71.50</td><td>7.05</td><td>84.05</td><td>91.50</td><td>84.00</td><td>12.90</td><td>100.65</td></tr></table>
+
+Table 2: End-to-end dialog generation evaluation results on RADDLE. The few-shot fine-tuning results are cited from Peng et al. (2021a). (Difference in mean is significant with  $p < 0.01$ .)
+
+ChatGPT, GPT-3.5 and Codex. Implementation details are provided in Appendix E.
+
+# 4.2 End-to-End Evaluation on Multiwoz
+
+Results. We present the evaluation results in multi-domain contexts on Multiwoz in Table 1. In addition to the aforementioned methods, we include the results of SOTA full-shot fine-tuning approaches to facilitate a more comprehensive comparison. SGP-TOD obtains SOTA zero-shot performance, substantially outperforming few-shot prompting approaches across all metrics, while even exhibiting competitive results in comparison to full-shot fine-tuning methods concerning Success and Inform. This confirms the effectiveness of integrating the task schema with the LLMs' proficient language processing capabilities.
+
+Comparison with Prompting Methods. SGP-TOD-CHATGPT distinctly surpasses the zero-shot prompting approach IG-TOD-CHATGPT-zs with respect to Success (surpassing by  $40\%$ ) and BLEU (exceeding by  $3\%$ ). Moreover, SGP-TOD-CHATGPT, without requiring task-specific data,
+
+considerably outperforms the few-shot prompting methods, i.e., IG-TOD-CHATGPT-FS and FEWSHOT-CHATGPT (e.g., about 30 points improvement over Success). This suggests that providing explicit and concise task instructions via task schema is preferable to imparting implicit task guidance through the selected dialog turns.
+
+Comparison with Zero-Shot Transfer Methods. Our SGP-TOD demonstrates a substantial advantage over ANYTOD-XXL, which necessitates task-specific pre-training and additional annotations, e.g., slot and action descriptions, over all the metrics. This exemplifies the potency of SGP-TOD, which markedly reduces the necessity for human labor and computational resources.
+
+Comparison with Full-Shot Fine-Tuning Methods. SGP-TOD exhibits competitive performance over Inform and Success. The lower BLEU is due to a lack of linguistic variations of the template utterances, which is acceptable considering the trade-off between human effort and efficacy.
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Task transfer</td><td colspan="2">Domain transfer</td></tr><tr><td>F1</td><td>Accuracy</td><td>F1</td><td>Accuracy</td></tr><tr><td colspan="5">Zero-shot transfer (leave-one fine-tuning with 2.5k training dialogs):</td></tr><tr><td>BERT+S</td><td>24.25</td><td>24.89</td><td>25.70</td><td>28.56</td></tr><tr><td>SAM</td><td>49.82</td><td>51.30</td><td>55.91</td><td>57.92</td></tr><tr><td colspan="5">Zero-shot prompting:</td></tr><tr><td>SGP-TOD-CODEX-INI</td><td>45.18</td><td>47.99</td><td>47.21</td><td>49.97</td></tr><tr><td>SGP-TOD-GPT3.5</td><td>47.67</td><td>48.27</td><td>49.76</td><td>50.39</td></tr><tr><td>SGP-TOD-CODEX</td><td>49.78</td><td>51.01</td><td>52.72</td><td>53.66</td></tr><tr><td>SGP-TOD-GPT3.5-E2E</td><td>50.84</td><td>50.74</td><td>53.50</td><td>53.21</td></tr></table>
+
+# 4.3 End-to-End Evaluation on RADDLE
+
+Results. Table 2 reports the results in single-domain settings on RADDLE. On all four dialog tasks, SGP-TOD demonstrates remarkable zero-shot performance that consistently surpasses both few-shot prompting and fine-tuning approaches. This results in substantial improvements of up to  $12\%$  in Inform,  $45\%$  in Success, and  $19\%$  in Combined metrics, while maintaining competitive BLEU scores. This evidence further substantiates the efficacy of SGP-TOD.
+
+# 4.4 End-to-End Evaluation on STAR
+
+Setup. BERT+S, SAM are fine-tuned on source tasks/domains then zero-shot on the held-out task/domain. SGP-TOD is presented with two formatting turns from the source tasks/domains.
+
+Results. Following Mehri and Eskenazi (2021), we report the zero-shot evaluation results in two settings, i.e., task transfer and domain transfer in Table 3. SGP-TOD, merely with two formatting sample turns, demonstrates exceptional performance, surpassing or rivaling SOTA zero-shot transfer methods in both settings. This outcome signifies that, even when faced with complicated business logic and system actions in dialog policies, the proposed SGP-TOD continues to exhibit commendable performance. Additionally, we investigate the impact of changing the number of training dialogs and formatting example turns in Appendix F.
+
+Impact of Different LLMs and Prompt Formats. SGP-TOD-CODEX surpasses SGP-TOD-GPT3.5 while rivaling SGP-TOD-GPT3.5-E2E (with template responses affixed to action labels in the policy prompt, demonstrated in Figure 10 in Appendix M). We conjecture that Codex, benefiting from extensive pre-training on copious code
+
+Table 3: Zero-shot end-to-end next action prediction evaluation results on STAR. (Difference in mean is significant with  $p < 0.01$ .)  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">FT/FS/ZS</td><td colspan="4">Restaurant-Ext</td></tr><tr><td>Info.</td><td>Succ.</td><td>BLEU</td><td>BERTS.</td></tr><tr><td colspan="6">Without domain-relevant knowledge</td></tr><tr><td>ChatGPT</td><td>ZS</td><td>44.00</td><td>6.00</td><td>4.31</td><td>85.96</td></tr><tr><td>GPT-3.5</td><td>ZS</td><td>34.00</td><td>16.00</td><td>8.70</td><td>84.31</td></tr><tr><td colspan="6">With prior knowledge on Restaurant</td></tr><tr><td>SOLOIST</td><td>FT</td><td>78.00</td><td>0.00</td><td>10.62</td><td>87.24</td></tr><tr><td>SGP-TOD-CHATGPT</td><td>ZS</td><td>88.00</td><td>34.00</td><td>5.45</td><td>86.11</td></tr><tr><td>SGP-TOD-GPT3.5</td><td>ZS</td><td>94.00</td><td>30.00</td><td>10.68</td><td>87.30</td></tr><tr><td colspan="6">With knowledge on Restaurant-Ext</td></tr><tr><td>SOLOIST+TEACH</td><td>FT</td><td>82.00</td><td>38.00</td><td>10.99</td><td>87.66</td></tr><tr><td>FEW-SHOT-GPT3.5+TEACH</td><td>FS</td><td>88.00</td><td>54.00</td><td>12.95</td><td>88.90</td></tr><tr><td>SGP-TOD-CHATGPT-EXT</td><td>ZS</td><td>88.00</td><td>78.00</td><td>6.25</td><td>86.15</td></tr><tr><td>SGP-TOD-GPT3.5-EXT</td><td>ZS</td><td>96.00</td><td>86.00</td><td>14.57</td><td>89.01</td></tr></table>
+
+Table 4: End-to-end evaluation results on domain extension. FT: fine-tuning, FS: few-shot prompting, ZS: zero-shot prompting, Info.: Inform, Succ.: Success, BERTS.: BERTScore. (Difference in mean is significant with  $p < 0.01$ .)
+
+data, demonstrates enhanced proficiency compared to GPT-3.5 in interpreting action labels. In addition, appending template responses is presumed to facilitate the explication of action labels for GPT-3.5.
+
+Impact of Different Task Schemas. SGP-TOD-CODEX-INI, utilizing an identical task schema as employed in training SAM, manifests commendable performance. This result highlights that SGP-TOD as a flexible prompting strategy, compatible with any manually-crafted task schema.
+
+# 4.5 End-to-End Evaluation on Domain Extension
+
+Setup. We conduct experiments in a domain extension setting (Gasic et al., 2014; Lipton et al., 2018) to assess the efficacy of SGP-TOD in adapting deployed task bots to incorporate novel functionalities. Following Zhang et al. (2022), we construct the Restaurant-ext corpus by extending the Restaurant in RADDLE with four new slots: [restaurant_dish], [value_price], [start_time], and [end_time]. Details are shown in Appendix J.
+
+# Compared Methods.
+
+- ChatGPT, GPT-3.5 denote zero-shot prompting that receive two formatting examples.  
+- SGP-TOD-CHATGPT, SGP-TOD-GPT3.5 represent our SGP-TOD implementation, with the Restaurant policy skeleton.  
+- SOLOIST is trained with 50 training dialogs in Restaurant domain (reported in Table 2).  
+- SOLOIST+TEACH is fine-tuning method enhanced with machine teaching (Simard et al., 2017). We deploy SOLOIST to converse with
+
+real users, then implement machine teaching to obtain 10/50/50 annotated dialogs in Restaurant-ext for training, validating, and testing. We fine-tune SOLOIST with the gathered 10 training dialogs covering new slots.
+
+- FEW-SHOT-GPT3.5+TEACH is the few-shot prompting strategy augmented with machine teaching. We use 10 randomly selected dialog turns from the collected 10 training dialogs as prompts (with peak performance at 10).  
+- SGP-TOD-CHATGPT-EXT, SGP-TOD-GP3.5-EXT refer to SGP-TOD with Restaurant-Ext policy skeleton, where we only add four template turns about four new slots to the policy skeleton of Restaurant.
+
+Results. In Table 4, SGP-TOD-CHATGPT-EXT, and notably SGP-TOD-GPT3.5-EXT surpasses all other evaluated approaches by a substantial margin over all the metrics. This demonstrates the strong adaptability of our SGP-TOD in accommodating novel functionalities, revealing its immense potential for lifelong learning. Two interactive dialog examples are supplied in Appendix K.
+
+Comparison with Approaches Augmented by Machine Teaching. SOLOIST yields zero Success, a predictable result given its lack of awareness regarding the new features. Augmented by machine teaching, SOLOIST+TEACH substantially improves SOLOIST in terms of Inform and Success. Nevertheless, relying solely on prior Restaurant knowledge, both SGP-TOD-CHATGPT and SGP-TOD-GP3.5 exhibit performance on par with SOLOIST+TEACH, demonstrating that SGP-TOD provides enhanced robustness in zero-shot generalization. Moreover, SGP-TOD-GP3.5-Ext obtains substantially higher Success rates than SOLOIST+TEACH (a rise of  $48\%$ ) and FEW-SHOTGPT3.5+TEACH (an increase of  $32\%$ ). Compared to fine-tuning/prompting strategies utilizing additional dialogs corrected through machine teaching, SGP-TOD facilitates a more agile adaptation to novel functionalities by merely modifying template turns within the task schema.
+
+# 4.6 Interactive Human Evaluation
+
+Setup. We conduct interactive human evaluations on Restaurant domain to evaluate the performance of SOLOIST, FEW-SHOT-CHATGPT, SGP-TOD-CHATGPT (reported in Table 2), with 50 dialogs gathered for analysis, respectively. Details can be found in Appendix H.
+
+<table><tr><td rowspan="2">Model</td><td colspan="5">Restaurant</td></tr><tr><td>Sw/o g ↑</td><td>Sw/ g ↑</td><td>Und. ↑</td><td>App. ↑</td><td>T. ↓</td></tr><tr><td>SOLOIST</td><td>34.00</td><td>30.00</td><td>2.18</td><td>2.10</td><td>10.64</td></tr><tr><td>FEW-SHOT-CHATGPT</td><td>94.00</td><td>74.00</td><td>4.58</td><td>4.72</td><td>8.32</td></tr><tr><td>SGP-TOD-CHATGPT</td><td>100.00</td><td>92.00</td><td>4.86</td><td>4.88</td><td>7.28</td></tr></table>
+
+Table 5: Human evaluation results. S w/o g, S w/ g: Success without / with grounding; Und.: Understanding; App.: Appropriateness; T.: Turns.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">Multiwoz 2.2</td></tr><tr><td>Inform</td><td>Success</td><td>BLEU</td><td>Combined</td></tr><tr><td>SP-TOD-GPT3.5</td><td>82.00</td><td>72.50</td><td>9.22</td><td>86.47</td></tr><tr><td>-policy</td><td>81.80</td><td>56.20</td><td>6.63</td><td>75.63</td></tr><tr><td>-policy -DB</td><td>81.40</td><td>52.30</td><td>6.57</td><td>73.42</td></tr><tr><td>-policy -DB -belief</td><td>38.60</td><td>33.90</td><td>6.29</td><td>42.54</td></tr></table>
+
+Table 6: Ablation study results on the impact of the three components in the proposed SGP-TOD and the database expertise on Multiwoz 2.2 using GPT-3.5. -policy: removing Policy Prompt, -DB: removing database expertise, -belief: removing DST Prompt.
+
+Results. In Table 5, our proposed SGP-TOD-CHATGPT attains a remarkably high performance in a zero-shot context, consistently outpacing SOLOIST and FEW-SHOT-CHATGPT across all metrics. Particularly, regarding Success w/ g, SGP-TOD-CHATGPT significantly surpasses FEW-SHOT-CHATGPT (by  $18\%$ ) and SOLOIST (by  $62\%$ ), illustrating its proficiency in accomplishing tasks within real-world scenarios. Furthermore, SGP-TOD-CHATGPT exhibits a more stable performance (demonstrated in Appendix I). A detailed analysis is provided in Appendix I.
+
+# 4.7 Ablation Study
+
+In Table 6, we study the impact of the three components of SGP-TOD (namely, Policy Prompter, DST Prompter, and LLM) as well as the database expertise, on Multiwoz 2.2 utilizing GPT-3.5. Combining the three elements in SGP-TOD with the database expertise produces the optimal result, underscoring the value of enhancing the LLM with the task schema and external database information. Detailed analyses are provided in Appendix G.
+
+# 5 Conclusion
+
+We present SGP-TOD, a schema-guided prompting strategy aimed at the expeditious construction of end-to-end task bots, relying exclusively on LLMs and the corresponding task schema. Employing the symbolic knowledge - task schema, SGP-TOD guides fixed LLMs to generate suitable responses for novel tasks in a zero-shot fashion. Empirical findings on four well-studied datasets
+
+reveal that SGP-TOD attains remarkable SOTA zero-shot performance, using both automatic and human evaluations. For future work, we plan to explore the use of SGP-TOD to develop personalized chatbots by utilizing pertinent task schema.
+
+# Limitations
+
+This work is accompanied by two primary limitations. (i) Data contamination (Brown et al., 2020; Madotto et al., 2021) in prompt-based zero-shot learning pertains to the potential presence of test samples during the LLM pre-training phase. Given that data utilized for pre-training LLMs, such as ChatGPT and GPT-4, remains undisclosed and continuously expands, verifying data contamination presents a formidable challenge. Consequently, our research cannot preclude data contamination in the experimental process, deferring a more comprehensive investigation to future endeavors. Nevertheless, we undertake domain-extension experiments (Table 4 in Section 4.5), subjecting our proposed SGP-TOD to evaluation on a novel test set (currently not publicly available), encompassing recently obtained and annotated human-bot dialogs. The remarkable zero-shot performance of SGP-TOD demonstrates its substantial potential for adeptly adapting to innovative functionalities, without reliance on task-specific data.
+
+(ii) We employ the manually-crafted task schema as prompts to steer the LLMs towards generating suitable responses on novel tasks. As illustrated in Table 3 of Section 4.4, SGP-TOD exhibits minor performance discrepancies when implementing disparate task schema formulated by various authors. Notwithstanding such variations, our objective is to offer a foundational basis for schemaguided LLM prompting; future research may investigate approaches to designing more efficient task schema, i.e., diverse formats and coverage.
+
+# Ethics Statement
+
+Throughout the interactive human evaluations and domain-extension experiments, all participating student helpers were informed of the research objectives prior to the collection and annotation of human-bot dialog logs. Their privacy was ensured to remain protected and undisclosed during the research period. Each participant received equitable remuneration.
+
+The Promoters utilized in this research incorporate no language that discriminates against specific
+
+individuals or groups (Zhou et al., 2022) and avoid any negative impact on users' well-being (Bergman et al., 2022). Instances of these Prompts are provided in Appendix M. Furthermore, subsequent research endeavors may consider utilizing the OpenAI moderation API<sup>7</sup> in conjunction with other related APIs to systematically filter out unsuitable user inputs and system responses.
+
+# Acknowledgements
+
+This Project is partially supported by the HKSARG General Research Fund (Ref No. 14207619). We would like to express our gratitude to Xiaohan Feng and Haohan Guo for their valuable comments.
+
+# References
+
+A. Stevie Bergman, Gavin Abercrombie, Shannon Spruit, Dirk Hovy, Emily Dinan, Y-Lan Boureau, and Verena Rieser. 2022. Guiding the release of safer E2E conversational AI through value sensitive design. In Proceedings of the 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 39-52, Edinburgh, UK. Association for Computational Linguistics.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D 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 Ziegler, Jeffrey Wu, Clemens Winter, Chris 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 Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.  
+Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4.  
+Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Inigo Casanueva, Ultes Stefan, Ramadan Osman, and Milica Gašić. 2018. Multiwoz - a large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Giovanni Campagna, Agata Foryciarz, Mehrad Moradshahi, and Monica Lam. 2020. Zero-shot transfer learning with synthesized data for multi-domain dialogue state tracking. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 122-132, Online. Association for Computational Linguistics.  
+Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.  
+Zhoujun Cheng, Tianbao Xie, Peng Shi, Chengzhu Li, Rahul Nadkarni, Yushi Hu, Caiming Xiong, Dragomir Radev, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Tao Yu. 2023. Binding language models in symbolic languages. *ICLR*, abs/2210.02875.  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts,
+
+Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022a. Palm: Scaling language modeling with pathways.  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022b. Palm: Scaling language modeling with pathways. CoRR, abs/2204.02311.  
+Haixing Dai, Zhengliang Liu, Wenxiong Liao, Xiaoke Huang, Yihan Cao, Zihao Wu, Lin Zhao, Shaochen Xu, Wei Liu, Ninghao Liu, Sheng Li, Dajiang Zhu, Hongmin Cai, Lichao Sun, Quanzheng Li, Dinggang Shen, Tianming Liu, and Xiang Li. 2023. Auggpt: Leveraging chatgpt for text data augmentation.  
+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, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.  
+Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong
+
+Wu, Baobao Chang, Xu Sun, Jingjing Xu, Lei Li, and Zhifang Sui. 2023. A survey on in-context learning.  
+Milica Gasic, Dongho Kim, Pirros Tsiakoulis, Catherine Breslin, Matthew Henderson, Martin Szummer, Blaise Thomson, and Steve J. Young. 2014. Incremental on-line adaptation of pomdp-based dialogue managers to extended domains. In INTER-SPEECH 2014, 15th Annual Conference of the International Speech Communication Association, Singapore, September 14-18, 2014, pages 140-144. ISCA.  
+Chulaka Gunasekara, Seokhwan Kim, Luis Fernando D'Haro, Abhinav Rastogi, Yun-Nung Chen, Mihail Eric, Behnam Hedayatnia, Karthik Gopalakrishnan, Yang Liu, Chao-Wei Huang, et al. 2020. Overview of the ninth dialog system technology challenge: Dstc9. arXiv preprint arXiv:2011.06486.  
+Wanwei He, Yinpei Dai, Yinhe Zheng, Yuchuan Wu, Zheng Cao, Dermot Liu, Peng Jiang, Min Yang, Fei Huang, Luo Si, et al. 2022. Galaxy: A generative pre-trained model for task-oriented dialog with semi-supervised learning and explicit policy injection. Proceedings of the AAAI Conference on Artificial Intelligence.  
+Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. 2020. A simple language model for task-oriented dialogue. Advances in Neural Information Processing Systems, 33:20179-20191.  
+Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A. Smith, and Mari Ostendorf. 2022. Incontext learning for few-shot dialogue state tracking. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2627-2643, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Vojtech Hudecek and Ondrej Dusek. 2023. Are llms all you need for task-oriented dialogue? CoRR, abs/2304.06556.  
+Mihir Kale and Abhinav Rastogi. 2020. Template guided text generation for task-oriented dialogue. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 6505-6520. Association for Computational Linguistics.  
+Zekun Li, Wenhu Chen, Shiyang Li, Hong Wang, Jing Qian, and Xifeng Yan. 2022. Controllable dialogue simulation with in-context learning. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4330-4347, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Yaobo Liang, Chenfei Wu, Ting Song, Wenshan Wu, Yan Xia, Yu Liu, Yang Ou, Shuai Lu, Lei Ji, Shaoguang Mao, Yun Wang, Linjun Shou, Ming Gong, and Nan Duan. 2023. Taskmatrix.ai: Completing tasks by connecting foundation models with millions of apis.
+
+Zhaojiang Lin, Bing Liu, Andrea Madotto, Seungwhan Moon, Zhenpeng Zhou, Paul A Crook, Zhiguang Wang, Zhou Yu, Eunjoon Cho, Rajen Subba, et al. 2021a. Zero-shot dialogue state tracking via cross-task transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7890-7900.  
+Zhaojiang Lin, Bing Liu, Seungwhan Moon, Paul A Crook, Zhenpeng Zhou, Zhiguang Wang, Zhou Yu, Andrea Madotto, Eunjoon Cho, and Rajen Subba. 2021b. Leveraging slot descriptions for zero-shot cross-domain dialogue statetracking. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5640-5648.  
+Zachary Lipton, Xiujun Li, Jianfeng Gao, Lihong Li, Faisal Ahmed, and Li Deng. 2018. BBQ-networks: Efficient exploration in deep reinforcement learning for task-oriented dialogue systems. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32.  
+Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pages 100–114, Dublin, Ireland and Online. Association for Computational Linguistics.  
+Andrea Madotto, Zhaojiang Lin, Genta Indra Winata, and Pascale Fung. 2021. Few-shot bot: Prompt-based learning for dialogue systems. arXiv preprint arXiv:2110.08118.  
+Shikib Mehri, Yasemin Altun, and Maxine Eskenazi. 2022. LAD: Language models as data for zero-shot dialog. In Proceedings of the 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 595-604, Edinburgh, UK. Association for Computational Linguistics.  
+Shikib Mehri and Maxine Eskenazi. 2021. Schemaguided paradigm for zero-shot dialog. arXiv preprint arXiv:2106.07056.  
+Johannes EM Mosig, Shikib Mehri, and Thomas Kober. 2020. Star: A schema-guided dialog dataset for transfer learning. arXiv preprint arXiv:2010.11853.  
+Maxwell I. Nye, Michael Henry Tessler, Joshua B. Tenenbaum, and Brenden M. Lake. 2021. Improving coherence and consistency in neural sequence models with dual-system, neuro-symbolic reasoning. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 25192-25204.  
+OpenAI. 2022. large-scale generative pre-training model for conversation. OpenAI blog.  
+OpenAI. 2023. Gpt-4 technical report.
+
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, volume 35, pages 27730-27744. Curran Associates, Inc.  
+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, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.  
+Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. 2023. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813.  
+Baolin Peng, Chunyuan Li, Jinchao Li, Shahin Shayan-deh, Lars Liden, and Jianfeng Gao. 2021a. Soloist: Building task bots at scale with transfer learning and machine teaching. Transactions of the Association for Computational Linguistics, 9:807-824.  
+Baolin Peng, Chunyuan Li, Zhu Zhang, Jinchao Li, Chenguang Zhu, and Jianfeng Gao. 2021b. Synergy: Building task bots at scale using symbolic knowledge and machine teaching. arXiv preprint arXiv:2110.11514.  
+Baolin Peng, Chunyuan Li, Zhu Zhang, Chenguang Zhu, Jinchao Li, and Jianfeng Gao. 2021c. RADDLE: an evaluation benchmark and analysis platform for robust task-oriented dialog systems. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 4418-4429. Association for Computational Linguistics.  
+Kun Qian and Zhou Yu. 2019. Domain adaptive dialog generation via meta learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2639-2649, Florence, Italy. Association for Computational Linguistics.  
+Chengwei Qin, Aston Zhang, Zhuosheng Zhang, Jiaao Chen, Michihiro Yasunaga, and Diyi Yang. 2023. Is chatgpt a general-purpose natural language processing task solver?  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.
+
+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. Journal of Machine Learning Research, 21(140):1-67.  
+Abhinav Rastogi, Xiaoxue Zang, Srinivas Sunkara, Raghav Gupta, and Pranav Khaitan. 2020a. Towards scalable multi-domain conversational agents: The schema-guided dialogue dataset. 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 8689-8696. AAAI Press.  
+Abhinav Rastogi, Xiaoxue Zang, Srinivas Sunkara, Raghav Gupta, and Pranav Khaitan. 2020b. Towards scalable multi-domain conversational agents: The schema-guided dialogue dataset. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8689-8696.  
+Adam Roberts, Hyung Won Chung, Anselm Levskaya, Gaurav Mishra, James Bradbury, Daniel Andor, Sharan Narang, Brian Lester, Colin Gaffney, Afroz Mohiuddin, Curtis Hawthorne, Aitor Lewkowycz, Alex Salcianu, Marc van Zee, Jacob Austin, Sebastian Goodman, Livio Baldini Soares, Haitang Hu, Sasha Tsvyashchenko, Aakanksha Chowdhery, Jasmina Bastings, Jannis Bulian, Xavier Garcia, Jianmo Ni, Andrew Chen, Kathleen Kenealy, Jonathan H. Clark, Stephan Lee, Dan Garrette, James Lee-Thorp, Colin Raffel, Noam Shazeer, Marvin Ritter, Maarten Bosma, Alexandre Passos, Jeremy Maitin-Shepard, Noah Fiedel, Mark Omernick, Brennan Saeta, Ryan Sepassi, Alexander Spiridonov, Joshua Newlan, and Andrea Gesmundo. 2022. Scaling up models and data with t5x and seqio.  
+Darsh J Shah, Raghav Gupta, Amir A Fayazi, and Dilek Hakkani-Tur. 2019. Robust zero-shot cross-domain slot filling with example values. arXiv preprint arXiv:1906.06870.  
+Richard Shin and Benjamin Van Durme. 2022. Few-shot semantic parsing with language models trained on code. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5417-5425, Seattle, United States. Association for Computational Linguistics.  
+Swadheen Shukla, Lars Liden, Shahin Shayandeh, Eslam Kamal, Jinchao Li, Matt Mazzola, Thomas Park, Baolin Peng, and Jianfeng Gao. 2020. Conversation Learner - a machine teaching tool for building dialog managers for task-oriented dialog systems. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 343-349, Online. Association for Computational Linguistics.
+
+Patrice Y. Simard, Saleema Amershi, David Maxwell Chickering, Alicia Edelman Pelton, Soroush Ghorashi, Christopher Meek, Gonzalo A. Ramos, Jina Suh, Johan Verwey, Mo Wang, and John Werns-ing. 2017. Machine teaching: A new paradigm for building machine learning systems. CoRR, abs/1707.06742.  
+Yixuan Su, Lei Shu, Elman Mansimov, Arshit Gupta, Deng Cai, Yi-An Lai, and Yi Zhang. 2022. Multi-task pre-training for plug-and-play task-oriented dialogue system. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL).  
+Haipeng Sun, Junwei Bao, Youzheng Wu, and Xiaodong He. 2022. Mars: Semantic-aware contrastive learning for end-to-end task-oriented dialog. arXiv preprint arXiv:2210.08917.  
+Jindong Wang, Xixu Hu, Wenxin Hou, Hao Chen, Runkai Zheng, Yidong Wang, Linyi Yang, Haojun Huang, Wei Ye, Xiubo Geng, Binxin Jiao, Yue Zhang, and Xing Xie. 2023. On the robustness of chatgpt: An adversarial and out-of-distribution perspective.  
+Qingyue Wang, Yanan Cao, Piji Li, Yanhe Fu, Zheng Lin, and Li Guo. 2022a. Slot dependency modeling for zero-shot cross-domain dialogue state tracking. In Proceedings of the 29th International Conference on Computational Linguistics, pages 510-520, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, et al. 2022b. Supernaturalinstructions: Generalization via declarative instructions on  $1600+$  nlp tasks. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5085-5109.  
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent abilities of large language models.  
+Jason D. Williams and Lars Liden. 2017. Demonstration of interactive teaching for end-to-end dialog control with hybrid code networks. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue, Saarbrücken, Germany, August 15-17, 2017, pages 82-85. Association for Computational Linguistics.  
+Chien-Sheng Wu, Andrea Madotto, Ehsan Hosseini-Asl, Caiming Xiong, Richard Socher, and Pascale Fung. 2019. Transferable multi-domain state generator for task-oriented dialogue systems. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 808-819, Florence, Italy. Association for Computational Linguistics.
+
+Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I. Wang, Victor Zhong, Bailin Wang, Chengzu Li, Connor Boyle, Ansong Ni, Ziyu Yao, Dragomir Radev, Caiming Xiong, Lingpeng Kong, Rui Zhang, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2022. UnifiedSKG: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 602-631, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Dian Yu, Luheng He, Yuan Zhang, Xinya Du, Panupong Pasupat, and Qi Li. 2021. Few-shot intent classification and slot filling with retrieved examples. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021, pages 734-749. Association for Computational Linguistics.  
+Xiaoxue Zang, Abhinav Rastogi, Srinivas Sunkara, Raghav Gupta, Jianguo Zhang, and Jindong Chen. 2020. MultiWOZ 2.2: A dialogue dataset with additional annotation corrections and state tracking baselines. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 109-117, Online. Association for Computational Linguistics.  
+Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. *Bertscore: Evaluating text generation with bert.* In International Conference on Learning Representations.  
+Xiaoying Zhang, Baolin Peng, Jianfeng Gao, and Helen Meng. 2022. Toward self-learning end-to-end task-oriented dialog systems. In Proceedings of the 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 516-530.  
+Yichi Zhang, Zhijian Ou, and Zhou Yu. 2020. Task-oriented dialog systems that consider multiple appropriate responses under the same context. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 9604-9611.  
+Jeffrey Zhao, Yuan Cao, Raghav Gupta, Harrison Lee, Abhinav Rastogi, Mingqiu Wang, Hagen Soltau, Izhak Shafran, and Yonghui Wu. 2022. Anytod: A programmable task-oriented dialog system. arXiv preprint arXiv:2212.09939.  
+Tiancheng Zhao and Maxine Eskenazi. 2018. Zero-shot dialog generation with cross-domain latent actions. In Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue, pages 1-10, Melbourne, Australia. Association for Computational Linguistics.  
+Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the 38th International Conference
+
+on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12697-12706. PMLR.
+
+Jingyan Zhou, Jiawen Deng, Fei Mi, Yitong Li, Yasheng Wang, Minlie Huang, Xin Jiang, Qun Liu, and Helen Meng. 2022. Towards identifying social bias in dialog systems: Framework, dataset, and benchmark. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 3576-3591, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+# DST Prompt
+
+# Task instruction
+
+Following the instructions, predict the belief state based on the history.
+
+# Belief instructions
+
+restaurant name = pizza hut city, golden wok, etc.; food = dont care, chinese, mediterranean, british, etc.; pricerange = dontcare, cheap, moderate, expensive; area = dont care, centre, east, north, south, west; booking_day = monday, tuesday, wednesday, thursday, friday, saturday, sunday; booking_people = 1,2,3,4,5,6,7; booking_time = 12:15, 13:30, etc.
+
+attraction attraction type = swimmingpool, theatre, architecture, museum, nightclub, boat, park, college, concerthall, entertainment, multiple sports, cinema; name = the cherry hinton village centre, soul tree nightclub, etc.; area = dont care, centre, east, north, south, west
+
+```python
+hotel name = huntingdon marriott hotel, a and b guest house, etc.; pricerange = dont care, cheap, moderate, expensive; area = dont care, centre, east, north, south, west; stars = dont care, 0,1,2,3,4,5; parking = dont care, yes, no; internet = dont care, yes, no; type = hotel, guest house; booking_day = monday, tuesday, etc.; booking_people = 1,2,3,4,5,6,etc.; booking Stay = 1,2,3,4,etc.
+
+train leaveat = 10:45, 12:06, etc.; destination = norwich, cambridge, etc.; day = monday, tuesday, wednesday, thursday, friday, saturday, sunday; arriveby = 21:00, 09:45, etc.; departure = cambridge, stansted airport, etc.; booking_people = 1,2,3,4,5,6, etc.
+
+taxi leaveat  $= 08:45$  16:15,etc.;destination  $=$  saint john's college,kettle's yard,galleria,etc.;departure  $=$  huntingdon marriot taxi,cineworld cinema,bridge guest house,etc.;arriveby  $= 17:15$  17:30,etc.
+
+···
+
+Figure 4: Detailed belief instructions in DST Prompter.
+
+# A Zero-Shot Task-Oriented Dialog Modeling.
+
+Table 7 summarizes four main research directions in zero-shot task-oriented dialog modeling: slot filling (SF), dialog state tracking (DST), end-to-end policy management (E2E policy) and end-to-end dialog generation (E2E dialog).
+
+# B Detailed Belief Instructions in DST Prompter
+
+Figure 4 shows the detailed belief instructions in DST Prompter.
+
+# C A Formatting Example in Policy Prompt
+
+Figure 5 presents a formatting example in Policy Prompter.
+
+# D Experimental Setup
+
+# Datasets.
+
+- Multiwoz 2.0 (Budzianowski et al., 2018) is a multi-domain task-oriented dataset, which
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Task</td><td rowspan="2">Schema types</td><td colspan="3">Training strategy</td></tr><tr><td>Fine-tuning</td><td>Pre-training</td><td>Prompting</td></tr><tr><td>ROBUSTSF (Shah et al., 2019)</td><td>SF</td><td>slot names/value examples</td><td>✓</td><td></td><td></td></tr><tr><td>TRADE (Wu et al., 2019)</td><td>DST</td><td>slot names/value examples</td><td>✓</td><td></td><td></td></tr><tr><td>ZSTL-SD (Campaigna et al., 2020)</td><td>DST</td><td>ontology, dialog templates</td><td>✓(+synthesized data)</td><td></td><td></td></tr><tr><td>S-DST (Rastogi et al., 2020b)</td><td>DST</td><td>slot names/descriptionss+service, intent names/descriptionss</td><td>✓</td><td></td><td></td></tr><tr><td>T5DST (Lin et al., 2021b)</td><td>DST</td><td>slot names/descriptionss</td><td>✓</td><td></td><td></td></tr><tr><td>TRANSFERQA (Lin et al., 2021a)</td><td>DST</td><td>slot names/value examples</td><td></td><td>✓(QA tasks)</td><td></td></tr><tr><td>IC-DST (Hu et al., 2022)</td><td>DST</td><td>slot names/value examples</td><td></td><td></td><td>✓</td></tr><tr><td>SDM-DST (Wang et al., 2022a)</td><td>DST</td><td>slot names/value examples</td><td>✓</td><td></td><td></td></tr><tr><td>BERT+S (Mosig et al., 2020)</td><td>E2E policy</td><td>system-side policy skeletons</td><td>✓</td><td></td><td></td></tr><tr><td>SAM (Mehri and Eskenazi, 2021)</td><td>E2E policy</td><td>user-aware policy skeletons</td><td>✓</td><td></td><td></td></tr><tr><td>ZSDG (Zhao and Eskenazi, 2018)</td><td>E2E dialog</td><td>ontology, response templates</td><td>✓</td><td></td><td></td></tr><tr><td>DAML (Qian and Yu, 2019)</td><td>E2E dialog</td><td>ontology, response templates</td><td>✓</td><td></td><td></td></tr><tr><td>ANYTOD (Zhao et al., 2022)</td><td>E2E dialog</td><td>policy programs+slot names/value examples+slot descriptions+user action names/states/descriptionss</td><td>✓</td><td>✓(heterogeneous tasks)</td><td></td></tr><tr><td>IG-TOD (Hudecek and Dusek, 2023)</td><td>E2E dialog</td><td>slot names+slot descriptions</td><td></td><td></td><td>✓</td></tr><tr><td>SGP-TOD (ours)</td><td>E2E dialog</td><td>user-aware policy skeletons(+slot names/value examples)</td><td></td><td></td><td>✓</td></tr></table>
+
+Table 7: Zero-shot task-oriented dialog modeling. (Schema items enclosed in parentheses are required only when accessible.)
+
+contains 8,438/1,000/1,000 dialogs for training-validating/testing, spanning seven domains: restaurant, attraction, train, hotel, taxi, police, and hospital. Multiwoz 2.0 is annotated with belief states and system actions.
+
+- Multiwoz 2.2 (Zang et al., 2020) is a improved version of Multiwoz 2.0, encompassing refined belief state annotations, slot descriptions, user action annotations, etc.  
+- RADDLE (Peng et al., 2021a,c) consists of four single-domain dialog datasets derived from Multiwoz 2.0 (i.e., restaurant, train, hotel, attraction), reorganized by Peng et al. (2021a). Each corpus contains 50/50/200 dialogs for training-validating/testing, expect for 100 testing dialogs in attraction domain.  
+- STAR (Mosig et al., 2020) includes 24 tasks in 13 domains (e.g., "apartment" domain comprises "apartment-search" and "apartment-schedule"), requiring the dialog model to conform to the provided task schema. We use 2,688 single-task dialogs from the corpus, which follow a "happy path", i.e., the user is not instructed to execute any action exceeding the schema's expectations. Without additional annotations, STAR only provides a flow chart diagram that outlines the dialog policy for each task. The flow chart outlines the task, including the sequence in which at
+
+tributes should be asked (for example, ask for the user's name before asking for the hotel name), how to query a database, etc.
+
+Automatic Evaluation Metrics. We evaluate the end-to-end dialog generation performance using the same metrics as those listed in Budzianowski et al. (2018): (i) Inform(\%) assesses whether the agent returns an acceptable entity. (ii) Success(\%) determines if the agent appropriately responds to each attribute request. (iii) BLEU(\%) (Papineni et al., 2002) measures the word overlap of the generated response against the human response in the corpus. (iv) Combined(\%) judges the overall quality, which is defined as Combined = (Inform + Success) × 0.5 + BLEU. Additionally, we utilize BERTScore(\%) (Zhang* et al., 2020), which focuses on computing semantic similarity between the generated responses and the ground truth, and correlates better with human judgments.
+
+Following Mehri and Eskenazi (2021), we perform the next action prediction task on STAR, which predicts next system action based on the dialog history. Since the system actions and deterministic response templates are mapped one to one in STAR corpus, we believe the end-to-end next action prediction task falls within end-to-end dialog modeling, following Mosig et al. (2020); Mehri and Eskenazi (2021). In addition, we report the results using weighted F1score(%) and mean accuracy(%)
+
+# Policy Prompt
+
+# Task instruction
+
+Following the instructions, generate appropriate response based on the history.
+
+#  Formatting example (from other task/domain)
+
+# Policy skeleton
+
+(1) user: I'm looking for information on [attraction_name]. action: attraction (inform (name, address, area, entrance fee)) system: [attraction_name] is in the [value_area] and their address is [attraction_address], the entrance fee is [value_count] pounds. [eos]  
+(2) user: What is the phone number? action: attraction (inform phone), require (more)) system: Their phone number is [attraction_phone]. is there anything else i can do for you? [eos]  
+(3) user: I need the area/location and the postcode. action: attraction (inform (area, postcode)) system: The [attraction_name] is in the [value_area], their postal code is [attraction_postcode]. [eos]  
+(4) user: That is all i need today. thank you! action: attraction (general (bye)) system: You are welcome , thank you for calling the [value_place] town info [value_area]. have a great day. [eos]  
+(5) DB: attraction zero match. action: attraction (inform (none)) system: There are no results matching your query. Can i try a different area or type? [eos]  
+(6) DB: attraction one match. action: attraction (inform (name)) system: We have [attraction_name]. [eos]  
+(7) DB: attraction five match. action: attraction (inform (choices), request (area)) system: There are [value_count], which area were you looking at? [eos]
+
+# history.
+
+user: Hello, i'm looking for the museum of classical archaeology, could you tell me more about it? system: Certainly, the [attraction_name] is located at [attraction_address] and has free entrance, phone [attraction_phone]. do you also need the postcode? user: Yes, i'd like that. thank you so much!  
+SQL: select * from attraction where name = museum of classical archaeology DB: Attraction one match  
+action:attraction (inform (postcode),require (more)) system:The postcode for the [attribution_name] is [attribution_postcode].Is there anything else i can help you with?
+
+Figure 5: A formatting example in Policy Prompt.
+
+Human Evaluation Metrics. We employ interactive human evaluations to assess the quality of dialog agents, following the evaluation protocol in the DSTC9 Track 1 challenge (Gunasekara et al., 2020). We recruit student helpers to help with evaluations. For each dialog session, student helpers are provided with a goal and accompanying instructions, subsequently necessitating a discourse with the agent to achieve the goal via natural language. Upon the conclusion of each dialog session, students are mandated to assess the overall dialog quality employing these five metrics: (i) Success w/o g(%) evaluates whether the agent accomplishes the task. (ii) Success w/ g(%) judges whether the agent accomplishes the task and of
+
+fers matched slot values compared to the database record. (iii) Understanding(1-5) quantifies the accuracy with which the agent comprehends user utterances. (iv) Appropriateness(1-5) signifies the naturalness, appropriateness and fluency of an agent response. (v) Turns denotes the average number of dialog turns within successful dialog sessions.
+
+# E Implementation Details
+
+# Regarding SGP-TOD:
+
+- LLMs: We employ ChatGPT ("gpt-3.5-turbo"), GPT-3.5 ("text-davinci-003") and Codex ("code-davinci-002") as the fixed LLMs to implement the proposed SGP-TOD. Throughout the evaluation, we set temperature to 0.5.  
+- DST Prompt - belief instruction: In the context of multi-domain scenarios, the belief instructions encompassing all domains are incorporated, while solely the target domain's belief instruction is introduced in single-domain settings.  
+- Policy Prompt - policy skeleton: For the Multiwoz datasets, we manually construct the policy skeleton through observing a few dialogs in the training corpus, following Mosig et al. (2020); Mehri and Eskenazi (2021). In the case of the STAR corpus, we employ flow chart diagrams and several dialogs to develop the policy skeleton, following the guidelines set forth by Mehri and Eskenazi (2021). We integrate the relevant user template utterance and the system action into the policy skeleton, thereby augmenting the LLM's understanding of directives, in the absence of belief annotations. The prompt examples for the STAR dataset are shown in Appendix M.  
+- Formatting example: Following the zero-shot scenario in Wang et al. (2022b), we insert one formatting example from different tasks (fixed through the experimental procedure) into the prompt. The formatting example employed within DST Prompter/Policy Prompter is randomly chosen from the training corpus of different tasks/domains, conforming to zero-shot scenario proposed by Wang et al. (2022b). We appraise multiple randomly selected formatting examples, the evaluation results reveal minor deviations. In the experiments on domain
+
+extension (Section 4.5) and ablation analysis (Section 4.7), we employ the same (two) formatting exemplar turns originating from other domains within the RADDLE corpus for all prompting techniques.
+
+# Regarding compared methods:
+
+# (i) Zero-shot transfer methods:
+
+- BERT+S (Mosig et al., 2020) is a schemaguided method that augments a BERT-base classifier (Devlin et al., 2019) with a provided system-side schema to predict the next system action.  
+- SAM (Mehri and Eskenazi, 2021) represents a schema-guided model based on BERT-base, which aligns the dialog context to a user-aware schema to predict the next system action.  
+- ANYTOD-XXL (Zhao et al., 2022) adopts a neural LM to track dialog states and user actions utilizing slot and action descriptions. Then a program that outlines a predefined task policy is executed to recommend appropriate system actions. Upon considering these system actions, an LM generates the ultimate system action and formulates the corresponding template response using the approach proposed by Kale and Rastogi (2020). ANYTOD-XXL is implemented on T5-XXL (Roberts et al., 2022) and pre-trained on SGD dataset (Rastogi et al., 2020a) $^8$
+
+# (ii) Prompting methods:
+
+- IG-TOD-CHATGPT (Hudecek and Dusek, 2023) is a prompting approach based on ChatGPT that leverages the dialog context and manually-crafted slot descriptions as the prompt, to track dialog states, fetch DB entries, and produce responses. IG-TOD-CHATGPT-ZS and IG-TOD-CHATGPT-FS are in the zero-shot and few-shot settings, respectively.  
+- FEW-SHOT-CHATGPT is a few-shot prompting strategy implemented on ChatGPT, where we use a few (i.e.,  $\mathbf{k}$ ) dialog turns, randomly sampled from the training corpus to instruct ChatGPT on task execution. Upon evaluating
+
+various configurations of  $\pmb{k}$ , the optimal results manifest with 15 on Multiwoz (2.0 and 2.2), and 10 on RADDLE, exhibiting no further substantial enhancements.
+
+- SGP-TOD (Ours) is a schema-guided prompting strategy, which is compatible with any off-the-shelf LLMs. In this paper, we employ ChatGPT ("gpt-3.5-turbo"), GPT-3.5 ("text-davinci-003") and Codex ("codedavinci-002") as the fixed LLMs. Following the zero-shot scenario in Wang et al. (2022b), we insert one formatting example from different tasks (fixed through the experimental procedure) into the prompt. More implementation details are provided in Appendix E.
+
+# F Zero-Shot End-to-End Evaluation Results on STAR
+
+Figure 6 exhibits the zero-shot evaluation results on STAR, utilizing varying amounts of training dialogs (ranging from 1 to 1,000) and formatting example turns (spanning from 1 to 10) from source domains/tasks. SGP-TOD, merely with two formatting sample turns, achieves superior or comparable performance compared to BERT+S, SAM, which are fine-tuned on adequate source data.
+
+We observe that SGP-TOD, employing only two formatting sample turns, attains superior or commensurate performance in terms of both F1score and Accuracy, when compared to SAM trained with 1,000 dialogs. Given that a single dialog contains more than 10 dialog turns, this result suggests that SGP-TOD diminishes labeling expenses by a minimum factor of 1,000. Furthermore, it is noteworthy that augmenting the quantity of formatting exemplar turns exerts a negligible influence on the performance of SGP-TOD.
+
+# G Ablation Study
+
+Table 8 exhibits the findings from an ablation investigation, addressing the effects of the three integral aspects of SGP-TOD in conjunction with the database expertise, implemented on Multiwoz 2.0 and 2.2, employing GPT-3.5. Combining the three elements in SGP-TOD with the database expertise produces optimal results across both datasets. The removal of the Policy Prompter, database knowledge, and DST Prompter leads to consistent declines in all evaluation metrics, underscoring the
+
+![](images/7e137f02489fb56bd43286ce987898256a44d649264a8c4a3c925e81209049e0.jpg)  
+(a) Task transfer
+
+![](images/978d9deab0e162fec860e455ae43b219c23bf463aaa67c817c8e566fa84435af.jpg)  
+(b) Domain transfer
+
+![](images/d73b187c1733c199dbfc9325cfce1b1836e5d7ceb3d24bb707bf2bfcd4005aa0.jpg)  
+(c) Task transfer
+
+![](images/117c9f062cc70de1892b6c16965b16e39c67e2a7884543f7b297bf57a96b5794.jpg)  
+(d) Domain transfer  
+Figure 6: Zero-shot end-to-end evaluation results on STAR with different numbers of training dialogs (1, 10, 100, 1,000) / demonstration example turns (1, 10) from source domain/tasks.
+
+value of enhancing the fixed LLM with the task schema and external database information.
+
+Specifically, GPT-3.5 (in the final row) exhibits commendable zero-shot performance, highlighting the need of exploiting its superior zero-shot generalization capabilities in dialog generation tasks. Additionally, Disabling the Policy Prompt incurs a discernible decline in performance regarding Success (approximately  $16\%$ ) and BLEU (roughly  $3\%$ ), as the Policy Prompt's primary function is to provide task completion guidelines and interaction patterns. Eliminating the database expertise primarily reduces Success (by approximately  $4\%$ ), implying that incorporating database information contributes to task completion. Lastly, excising the DST Prompt engenders a considerable diminution in performance concerning Inform (around  $43\%$ ) and Success (nearly  $18\%$ ), due to the DST Prompt's intended purpose of assisting the frozen LLM in apprehending the dialog context.
+
+# H Human Evaluation Details
+
+We enlisted 5 student helpers (i.e., undergraduate students possessing basic proficiency in English communication) to participate in the evaluations. For each dialog agent, we collected 50 dialogs for analysis. Followed the methodology proposed by Li et al. (2022), we generated user goals through the subsequent techniques: (i) Randomly selecting slots and slot values within the Restaurant domain from RADDLE corpus to construct a user goal; (ii) Replacing the slot values of the user goals in randomly chosen dialogs from the Restaurant corpus with corresponding new values from randomly sampled database entries, thus forming a new user goal; (iii) Merging the user goals of several randomly selected dialogs from the Restaurant corpus to create a composite user goal. Lastly, we randomly chose 50 distinct user goals from these newly generated goals.
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">Multiwoz 2.0</td><td colspan="4">Multiwoz 2.2</td></tr><tr><td>Inform</td><td>Success</td><td>BLEU</td><td>Combined</td><td>Inform</td><td>Success</td><td>BLEU</td><td>Combined</td></tr><tr><td>SP-TOD-GPT3.5</td><td>83.88</td><td>69.87</td><td>9.09</td><td>85.97</td><td>82.00</td><td>72.50</td><td>9.22</td><td>86.47</td></tr><tr><td>-policy</td><td>82.28</td><td>55.65</td><td>6.51</td><td>75.48</td><td>81.80</td><td>56.20</td><td>6.63</td><td>75.63</td></tr><tr><td>-policy -DB</td><td>81.20</td><td>50.95</td><td>6.48</td><td>72.56</td><td>81.40</td><td>52.30</td><td>6.57</td><td>73.42</td></tr><tr><td>-policy -DB -belief</td><td>38.74</td><td>33.13</td><td>6.18</td><td>42.12</td><td>38.60</td><td>33.90</td><td>6.29</td><td>42.54</td></tr></table>
+
+Table 8: Ablation study on the impact of the three components in the proposed SGP-TOD and the database expertise on Multiwoz using GPT-3.5. -policy: removing Policy Prompter, -DB: removing database information, -belief: removing DST Prompter.
+
+# I Human Evaluation Results
+
+Figure 7 shows the interactive human evaluation results. SGP-TOD-CHATGPT exhibits a more stable performance. In contrast to the automated evaluation results shown in Table 2, FEW-SHOT-CHATGPT significantly outperforms SOLOIST over all metrics. This indicates that corpus-based evaluations might be biased, given that real user inputs tend to be more dynamic, complex, even with noise. Notably, SGP-TOD-CHATGPT consistently excels compared to the other methods in both evaluations, implying its robustness in handling diverse user inputs.
+
+# J More Details and Results on Domain Extension
+
+Setup. Following Zhang et al. (2022), we construct the Restaurant-ext corpus by extending the pre-existing Restaurant in RADDLE (Peng et al., 2021c) with additional functions. Specifically, we introduce four new slots: [restaurant_dish], [value_price], [start_time], and [end_time]. The initial slot pertains to recommendations for signature restaurant meals, while the final three concern delivery service details. All database entries are updated with corresponding values. Table 10 exhibits a dialog example on domain extension. The associated Restaurant-Ext database entry is illustrated in Table 9.
+
+# Compared Methods.
+
+- ChatGPT, GPT-3.5 denote zero-shot prompting with base LLMs that receive merely two formatting example turns from other domains in RADDLE. $^{10}$  
+- SGP-TOD-CHATGPT, SGP-TOD-GPT3.5 represent our SGP-TOD implementation, with the Restaurant policy skeleton.
+
+- SOLOIST is trained with 50 training dialogs in the Restaurant domain (previously reported in Table 2).  
+- SOLOIST+TEACH is fine-tuning method enhanced with machine teaching (Simard et al., 2017). Machine teaching is an efficient approach to equip deployed task bots with the ability to handle new functions by correcting representative failed human-bot dialogs. We deploy SOLOIST to converse with real users, then implement machine teaching via Conversational learner (Shukla et al., 2020), an effective machine teaching tool, to obtain 10/50/50 examples in Restaurant-ext for training, validating, and testing. Finally, we fine-tune SOLOIST with gathered 10 training dialogs covering four new slots, resulting in dialog agent SOLOIST+TEACH.  
+- FEW-SHOT-GPT3.5+TEACH is the few-shot prompting strategy augmented with machine teaching. Based on GPT-3.5, we utilize 10 randomly selected dialog turns from the collected 10 training dialogs as the prompt (with peak performance at 10), resulting in FEWSHOT-GPT3.5+TEACH.  
+- SGP-TOD-CHATGPT-EXT, SGP-TOD-GP3.5-Ext refer to SGP-TOD with Restaurant-Ext policy skeleton, where we only add four template turns about four new slots to the policy skeleton of Restaurant.
+
+Results. Comparison with Base LLMs. The substantial improvement of SGP-TOD-CHATGPT-EXT and SGP-TOD-GPT3.5-EXT over ChatGPT and GPT-3.5 illustrates SGP-TOD's efficiency in supplying task-specific knowledge in a zero-shot way.
+
+Impact of Different LLMs. SGP-TODCHATGPT-EXT attains a lower BLEU yet a comparable BERTScore, suggesting that ChatGPT generates more diverse responses.
+
+![](images/3cbdf3b1513ee9f96fecc4c728499a947c0b85bf3068aa2ef178542059a9c198.jpg)  
+(a) Success w/o g ↑
+
+![](images/1b326e76977705fe2590979ae990c490fdccf84470d4f4cb1f6df8f16d164f7c.jpg)  
+(b) Success w/ g ↑
+
+![](images/f6f85da5f3f15a5499c0d3af27c71554fb8667e4caf4bbd88c8f7db8dcbfccdd.jpg)  
+(c) Understanding  
+Figure 7: Interactive human evaluation results. SOLO.: SOLOIST, FS.: FEW-SHOT-CHATGPT, SGP.: SGP-TOD-CHATGPT.
+
+![](images/89175ae70b826217c987becd91e980100eb09425e9bd43786d2f220a7ab72e2c.jpg)  
+(d) Appropriateness  $\uparrow$
+
+![](images/bf9a90d6dab578f2777f2864788e94c5b4a92a8b02d016faf87888216ac4a2ce.jpg)  
+(e)Turns  $\downarrow$
+
+<table><tr><td>Slot</td><td>Value</td></tr><tr><td>&quot;address&quot;</td><td>&quot;21 - 24 Northampton Street&quot;</td></tr><tr><td>&quot;area&quot;</td><td>&quot;west&quot;</td></tr><tr><td>&quot;food&quot;</td><td>&quot;british&quot;</td></tr><tr><td>&quot;id&quot;</td><td>&quot;14810&quot;</td></tr><tr><td>&quot;location&quot;</td><td>[52.21031, 0.11381]</td></tr><tr><td>&quot;name&quot;</td><td>&quot;saint johns chop house&quot;</td></tr><tr><td>&quot;phone&quot;</td><td>&quot;01223353110&quot;</td></tr><tr><td>&quot;postcode&quot;</td><td>&quot;cb30ad&quot;</td></tr><tr><td>&quot;pricerange&quot;</td><td>&quot;moderate&quot;</td></tr><tr><td>&quot;type&quot;</td><td>&quot;restaurant&quot;</td></tr><tr><td>&quot;delivery&quot;</td><td>&quot;Yes&quot;</td></tr><tr><td>&quot;delivery fee&quot;</td><td>&quot;6 pounds&quot;</td></tr><tr><td>&quot;dish&quot;</td><td>&quot;Beef Wellington&quot;</td></tr><tr><td>&quot;start_time&quot;</td><td>&quot;10:30 am&quot;</td></tr><tr><td>&quot;end_time&quot;</td><td>&quot;22:40 pm&quot;</td></tr></table>
+
+Table 9: An example of Restaurant-Ext DB entry. The newly introduced slot-value pairs relevant to the extended functionality are highlighted.
+
+# K Case Study
+
+Despite the superior performance of the proposed SGP-TOD on GPT-3.5, we showcase interactive examples utilizing ChatGPT, a renowned and potent chatbot. In Table 11, a user engages with ChatGPT (left) and SGP-TOD-CHATGPT-EXT (right) to complete the identical task on Restaurant-Ext. $^{11}$  The user initiates the conversation by seeking recommendations for a Tuscan restaurant with no price range preference. Lacking external database information, ChatGPT conveys inaccurate details (Turn 2), whereas SGP-TOD-CHATGPT-EXT informs users of the absence of matching restaurants and proposes alternatives (Turn 2). This exemplifies the benefits of integrating real-world expertise into the fixed LLM. Furthermore, ChatGPT persistently inquires about
+
+the desired price range despite the user's indifference. We argue that SGP-TOD assists the frozen LLM in discerning user intentions. In Turn 4, ChatGPT continues to furnish fabricated details (i.e., the restaurant name and phone number) concerning the nonexistent eatery, while SGP-TOD-CHATGPT-EXT identifies a suitable Korean restaurant and the corresponding factual information. In contrast with ChatGPT, SGP-TOD-CHATGPT-EXT adeptly addresses inquiries about the delivery service (Turn 6), indicating that SGP-TOD is capable of endowing the frozen LLM with novel functionalities.
+
+# L Our Proposed SGP-TOD with a Dialog Example
+
+Figure 8 illustrates the proposed SGP-TOD with a dialog example.
+
+# M Prompt Examples for STAR Dataset
+
+Two exemplars of prompts for STAR dataset are presented in Figure 9 and Figure 10.
+
+<table><tr><td>Turn</td><td>Utterance</td></tr><tr><td>1</td><td>User: I want a restaurant on the west end of town.</td></tr><tr><td>2</td><td>System: Ok. Are you looking for any particular type of food?</td></tr><tr><td>3</td><td>User: I don&#x27;t care.</td></tr><tr><td>4</td><td>System: Are you looking for a particular price range?</td></tr><tr><td>5</td><td>User: A moderately priced restaurant would be good.</td></tr><tr><td>6</td><td>System: How about Saint Johns Chop House? They serve moderately priced British food on the west side of town.</td></tr><tr><td>7</td><td>User: Can i please have their address and phone number?</td></tr><tr><td>8</td><td>System: Saint Johns Chop House&#x27;s address is 21-24 Northampton Street, their phone number is 01223353110.</td></tr><tr><td>9</td><td>User: Does the restaurant offers delivery service? If it does, what&#x27;s the delivery fee?</td></tr><tr><td>10</td><td>System: Yes, it does, and the delivery charges 6 pounds.</td></tr><tr><td>11</td><td>User: Thank you, goodbye.</td></tr><tr><td>12</td><td>System: Thank you for using our system. Goodbye.</td></tr></table>
+
+Table 10: A dialog example on domain extension. Task bots need learn to provide pertinent responses concerning the expanded delivery service in subsequent interaction turns, adapting to the continually evolving user/enterprise demands.  
+
+<table><tr><td>Turn</td><td>Interaction between a user and ChatGPT</td><td>Turn</td><td>Interaction between a user and SP-TOD-ChatGPT-Ext</td></tr><tr><td>1</td><td>User: Hi, i&#x27;m looking for a place with Tuscan food in any price range.</td><td>1</td><td>User: Hi, i&#x27;m looking for a place with Tuscan food in any price range.</td></tr><tr><td>2</td><td>System: Certainly, there are 3 Tuscan restaurants in town. What is your preferred price range?</td><td>2</td><td>System: I am sorry, there are no Tuscan restaurants. Here are some other options: Korean, Italian, Chinese. Which one would you like more information on?</td></tr><tr><td>3</td><td>User: I don&#x27;t care about the price. Please recommend one for me, i also need the phone number please.</td><td>3</td><td>User: How about any Korean restaurants? I also need the phone number please.</td></tr><tr><td>4</td><td>System: The phone number for Pizza Hut City Centre is 01223323737. Is there anything else i can help you with today?</td><td>4</td><td>System: Little Seoul is a Korean restaurant. Their phone number is 01223308681.</td></tr><tr><td>5</td><td>User: Does the restaurant offer delivery service? How much does the delivery charge?</td><td>5</td><td>User: Does the restaurant offer delivery service? How much does the delivery charge?</td></tr><tr><td>6</td><td>System: I&#x27;m sorry, but I could not find any information on delivery fee in our database. You may want to contact the restaurant directly to inquire about this. Is there anything else I can assist you with today?</td><td>6</td><td>System: Yes, they offer delivery service and the delivery charge is 4 pounds. Can I help you with anything else?</td></tr><tr><td>7</td><td>User: No. Thank you, goodbye.</td><td>7</td><td>User: No. Thank you, goodbye.</td></tr><tr><td>8</td><td>System: Glad i could assist you. Have a good day.</td><td>8</td><td>System: You&#x27;re welcome, have a great meal! Goodbye!</td></tr></table>
+
+Table 11: Illustration of two interactive examples between a user and CHATGPT (left), SGP-TOD-CHATGPT-EXT (right). Appropriate responses and inappropriate responses are highlighted.
+
+![](images/711ee537a0ff2b321768070906a272a68796a2bff8b062acf12b09b943f48d74.jpg)  
+Figure 8: Illustration of the proposed SGP-TOD with a dialog example. Note that the belief state in the represented in the SQL format, the details of which are described in Section 3.3.
+
+# Policy Prompt
+
+# Task instruction
+
+Generate appropriate system actions based on the history, following the most relevant task rule.
+
+#  Formatting example (from other task/domain)
+
+# Task schema
+
+(1) user: Hello [sep] system action: hello [eos]  
+(2) user: I'd like to find out the status of my ride [sep] system action: ask_name [eos]  
+(3) user: My name is [NAME] [sep] system action: ride Asking Booking number [eos]
+
+# History.
+
+user: Hi, i need to check my ride status.
+
+# Answer
+
+(2) user: i'd like to find out the status of my ride [sep] system action: ask_name [eos]
+
+# Test (on target task/domain)
+
+# Task schema
+
+(1) user: hello [sep] system action: hello [eos]  
+(2) user: what is the weather like? [sep] system action: weather Asking_day [eos]  
+(3) user: on [DAY_OF_WEEK] [sep] system action: weather Asking_location [eos]
+
+··
+
+# History.
+
+user: i want to know the weather for Tuesday.
+
+# LLM
+
+Answer
+
+(3) user: on [DAY_OF_WEEK] [sep] system action: weather Asking_location [eos]
+
+# Policy Prompt
+
+# Task instruction
+
+Generate appropriate system actions based on the history, following the most relevant task rule.
+
+#  Formatting example (from other task/domain)
+
+# Task schema
+
+(1) user: Hello [sep] system action: hello system: Hello, how can i help? [eos]  
+(2) user: I'd like to find out the status of my ride [sep] system action: ask_name system: Could you give me your name, please? [eos]  
+(3) user: My name is [NAME] [sep] system action:  
+ride Asking Booking number system: Can I get your booking ID, please? [eos]
+
+# History.
+
+user: Hi, i need to check my ride status.
+
+# Answer
+
+(2) user: I'd like to find out the status of my ride [sep] system action: ask_name system: Could you give me your name, please? [eos]
+
+# Test (on target task/domain)
+
+# Task schema
+
+(1) user: hello [sep] system action: hello system: Hello, how can i help? [eos]  
+(2) user: what is the weather like? [sep] system action: weather Asking day system: For what day would you like the weather forecast? [eos]  
+(3) user: on [DAY_OF_WEEK] [sep] system action: weather Asking_location system: For what location would you like the weather forecast? [eos]
+
+# History.
+
+user: i want to know the weather for Tuesday.
+
+# LLM
+
+![](images/652d6fb90c49a942fc38da923756dd7b2ec6f3e3feb60f63e855f637d2ca9bb5.jpg)  
+Figure 9: Policy Prompt of SGP-TOD on STAR. The relevant template turn within the input, the generated user template utterance, and the system action in the output are accentuated.  
+Figure 10: Policy Prompt of SGP-TOD-E2E on STAR. The relevant template turn in the input, the generated user template utterance, the system action and the system response in the output are highlighted.
+
+Answer
+
+(3) user: on [DAY_OF_WEEK] [sep] system action: weather Asking_location system: For what location would you like the weather forecast? [eos]

2023/SGP-TOD_ Building Task Bots Effortlessly via Schema-Guided LLM Prompting/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5628ac7221398003e289f9246334d27a8b8aa5f77d0a6b4b3c7a49a82a66f4e6
+size 1124609

2023/SHARCS_ Efficient Transformers Through Routing with Dynamic Width Sub-networks/29c91769-471e-4e25-9798-c851bcedbfe6_content_list.json

@@ -0,0 +1,1830 @@
+[
+    {
+        "type": "text",
+        "text": "SHARCS: Efficient Transformers through Routing with Dynamic Width Sub-networks",
+        "text_level": 1,
+        "bbox": [
+            247,
+            89,
+            749,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Mohammadreza Salehi $^{\\dagger}$  Sachin Mehta $^{\\dagger}$  Aditya Kusupati $^{\\dagger}$  Ali Farhadi $^{\\dagger\\diamond}$  Hannaneh Hajishirzi $^{\\dagger\\diamond}$",
+        "bbox": [
+            242,
+            147,
+            757,
+            181
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "†University of Washington †Allen Institute for Artificial Intelligence {mrsalehi, sacmehta, kusupati, ali, hannaneh}@cs.washington.edu",
+        "bbox": [
+            201,
+            184,
+            800,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We introduce SHARCS for adaptive inference that takes into account the hardness of input samples. SHARCS can train a router on any transformer network, enabling the model to direct different samples to sub-networks with varying widths. Our experiments demonstrate that: (1) SHARCS outperforms or complements existing per-sample adaptive inference methods across various classification tasks in terms of accuracy vs. FLOPs; (2) SHARCS generalizes across different architectures and can be even applied to compressed and efficient transformer encoders to further improve their efficiency; (3) SHARCS can provide a  $2 \\times$  inference speed up at an insignificant drop in accuracy.",
+        "bbox": [
+            141,
+            280,
+            460,
+            494
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            506,
+            260,
+            521
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Web-scale pretrained models, including Large Language Models (LLMs), are widely used in various applications (Devlin et al., 2018; Liu et al., 2019a; Brown et al., 2020). However, their computational resource requirements can be problematic, especially in environments with limited resources. To address this issue, more efficient methods are needed, particularly those that can run on-device with efficient inference (Sanh et al., 2019).",
+        "bbox": [
+            112,
+            532,
+            489,
+            676
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Several methods (e.g., knowledge distillation (Hinton et al., 2015), pruning (Lagunas et al., 2021; Xia et al., 2022), and quantization (Shen et al., 2019)) have been proposed to improve the inference efficiency of transformer-based models. While these methods are promising, one drawback is that the resulting model is static. This raises concerns about whether the model is too complex for simple samples and not complex enough for more challenging ones. To tackle this problem, previous work have investigated sample-adaptive inference methods that use varying amount of compute to process different input samples (Kaya and Dumitras, 2018). Two predominant approaches exist in the field of sample adaptive inference: early-joining",
+        "bbox": [
+            112,
+            678,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "and token dropping. The former incorporates internal classifiers into intermediate layers of the model. Various techniques have been explored for early exiting, including using confidence scores or entropy of internal classifier predictions (Liu et al., 2020; Xin et al., 2020), using a module that predicts whether a layer should exit early (Xin et al., 2021), or implementing a patience-based method to adjust these internal predictions (Zhou et al., 2020). The other category of sample-adaptive inference methods, token dropping, enhances efficiency by progressively decreasing sequence length as the forward pass proceeds layer by layer (Goyal et al., 2020; Guan et al., 2022).",
+        "bbox": [
+            507,
+            253,
+            884,
+            476
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose Sample Hardness Aware Routing based on Confidence Scores  $(\\mathsf{SHARCS}^1)$ , which is a novel category within the efficient sample adaptive inference domain. Our approach introduces training a light-weight router. The router dynamically assigns input samples, based on their hardness (Ethayarajh et al., 2021), to one of the sub-networks with varying widths. Due to the lack of ground truth notion of hardness, we estimate sample hardness heuristically based on the network's prediction history during training for each sample. These estimates are then used as labels to train the router.",
+        "bbox": [
+            507,
+            478,
+            884,
+            686
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We make the following contributions:",
+        "bbox": [
+            527,
+            694,
+            808,
+            709
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. SHARCS introduces a router that predicts the hardness of a sample and can be trained on any transformer network. It enables dynamic input processing using sub-networks of varying widths, determined by the predicted hardness.",
+            "2. SHARCS delivers substantial efficiency improvements in terms of accuracy vs. FLOPs trade-off across different datasets & transformer encoders. For example, on QQP (Wang et al., 2017), it reduces the FLOPs of RoBERTa $_{\\text{base}}$  by  $2.75 \\times$  with only  $1\\%$  accuracy drop. Compared to other"
+        ],
+        "bbox": [
+            509,
+            713,
+            884,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "1 pronounced sharks.",
+        "bbox": [
+            529,
+            904,
+            658,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10519",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10519-10532 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            210,
+            945,
+            786,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e3f00069023724b960735641b0a5045627027c182b951d8bcb48a1f51a385768.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            151,
+            82,
+            443,
+            287
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/88fed47bcd69f387bcfef750639c6fbd11e3d631c95fe0d133027754b44fc572.jpg",
+        "image_caption": [
+            "(a) Inference.",
+            "(b) Generating hardness labels to train the router.",
+            "Figure 1: (a) At inference the router selects the reduction factor. Red parts denote the changes enforced by the router. (b) Training the router. The confidence scores of the last  $W$  epochs for each input sample is recorded. Then they are compared with confidence thresholds and labels for training the router are assigned accordingly."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            142,
+            309,
+            450,
+            463
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "sample adaptive inference techniques, SHARCS either outperforms them or can be paired with them to achieve even greater efficiency.",
+        "bbox": [
+            134,
+            607,
+            487,
+            655
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. The gains from SHARCS can be realized in real-world deployment with significant latency reduction in CPU-based inference. On QQP, SHARCS can speed up  $\\mathrm{BERT}_{\\mathrm{base}}$  more than  $2\\times$  with less than  $1\\%$  drop in accuracy.",
+        "bbox": [
+            114,
+            663,
+            489,
+            744
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            776,
+            218,
+            791
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We introduce SHARCS for adaptive inference that takes into account the hardness of input samples. Our approach has three main steps: (1) Obtaining labels that represent sample hardness ( $\\S 2.1$ ); (2) Training a router that can predict sample hardness ( $\\S 2.2$ ); (3) Adjusting the network's inference capacity according to the predicted hardness ( $\\S 2.3$ ).",
+        "bbox": [
+            112,
+            806,
+            490,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Estimating Hardness of a Sample",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            816,
+            99
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our objective is to learn a sample hardness aware router, enabling us to dynamically direct the input sample to one of the sub-networks with varying capacities (including the main network) for efficient inference. As there are no ground-truth labels that represent hardness, we leverage network's prediction history during training. We assume that there are  $M$  possible hardness levels for a sample, with level 1 being the easiest and  $M$  being the hardest. Our goal is to assign a hardness label  $\\hat{y} \\in \\{0,1\\}^{M}$  to samples in the training set and train the router with them. Toward this end, we employ a heuristic procedure which is illustrated in Figure 1b: If the model predicts the ground truth class for a training sample with a probability within a confidence interval  $[T_{low}^{i}, T_{high}^{i}]$ , then the  $i$ th entry in the label  $\\hat{y}$  would be 1; otherwise it would be zero. Here,  $i$  ranges from 1 to  $M$ , and  $T^{i}$ 's are hyperparameters associated with hardness level  $i$ .",
+        "bbox": [
+            507,
+            107,
+            884,
+            412
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Because of the stochastic nature of training, it is possible that the samples denoted as easy earlier will potentially be denoted as hard after that, posing instabilities while training the router. To mitigate such randomness, the  $i$ th entry is 1 only if the predicted ground truth probability is within the interval for a moving window of last  $W$  epochs. The assigned hardness labels will be used as labels to train the router in the next epoch. We do not train the router during the first  $W$  epochs and just train the network and record the network's predictions. Please see appendix E.0.5 for more details on the role of window size.",
+        "bbox": [
+            507,
+            414,
+            884,
+            621
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Training Sample Hardness Aware Router",
+        "text_level": 1,
+        "bbox": [
+            507,
+            637,
+            880,
+            653
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We split the main transformer network with  $L$  layers into non-adaptive and adaptive parts, and incorporate the router between these networks (see Figure 1a). The non-adaptive network is comprised of the first  $0 < K < L - 1$  layers while the adaptive part consists of the remaining  $L - K$  layers. The adaptive component is a shared network that consists of sub-networks with varying widths, where the width of each sub-network is a fraction of the width of the network.",
+        "bbox": [
+            507,
+            659,
+            882,
+            819
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "More formally, sub-networks are associated with a given a set of reduction factors  $\\{r_i\\}_{i = 1}^M$ $0 < r_{i}\\leq$  1, where the width of  $i$  -th sub-network is  $1 / r_{i}\\times$  smaller than that of main network. We map the hardness level  $i$  to the width reduction factor  $r_i$ . During training, for each input, we sample one of the re",
+        "bbox": [
+            507,
+            822,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10520",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/864fef6028196209d80370d23ebff5194ff362b25e1b363694949254727100de.jpg",
+        "image_caption": [
+            "Figure 2: Results on the three of the sub-tasks in the GLUE benchmark. For the full set of plots on 8 tasks please refer to the appendix D.0.1. Best viewed in color."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            164,
+            80,
+            833,
+            236
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "duction factors with entry 1 in the router label for that input and do the forward and backward pass with just the network associated with that reduction factor. During inference, given the output of non-adaptive network for input sample  $x$ , the objective of the router is to determine the width of sub-network in the adaptive module to process  $x$ . We train the sub-networks and the router with the following objective:",
+        "bbox": [
+            112,
+            299,
+            490,
+            444
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {t a s k}} \\cdot \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {\\text {r o u t e r}} \\cdot \\mathcal {L} _ {\\text {r o u t e r}} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            161,
+            478,
+            487,
+            495
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_{\\text{router}}$  is a binary cross-entropy loss between predicted and ground-truth hardness label,  $\\mathcal{L}_{\\text{task}}$  is a task-specific loss, and  $\\lambda_{\\text{task}}$  and  $\\lambda_{\\text{router}}$  are loss weights which are hyper-parameters.",
+        "bbox": [
+            112,
+            511,
+            489,
+            576
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Reducing Width of Adaptive Module",
+        "text_level": 1,
+        "bbox": [
+            112,
+            590,
+            450,
+            606
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The basic building block in both multi-head attention (MHA) module and feed forward network (FFN) in transformers is the linear layer (Vaswani et al., 2017). Given the fully-connected nature of linear layers, to reduce their width, we retain the leftmost  $r \\cdot d_{\\mathrm{model}}$  neurons in both input and output (Kusupati et al., 2022). It is worth noting that as we reduce the input and output dimensions of matrix multiplications by a factor of  $1 / r$ , the flops will be reduced by a factor of  $1 / r^2$ . We follow a similar procedure for reducing the width of affine parameters in LayerNorm (Ba et al., 2016). Also in our setup, we do not change the head dimensions in MHA and instead decrease the number of heads by a factor of  $1 / r$ . As Figure 1a illustrates, we down-project the input hidden states to the adaptive layers using a pooled module and up-project it back before feeding to the single classifier for all of the sub-networks. Please see Appendix B for a",
+        "bbox": [
+            112,
+            613,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "detailed description of width reduction of different components in transformer layer.",
+        "bbox": [
+            507,
+            299,
+            882,
+            332
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Experimental setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            344,
+            714,
+            361
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Datasets. We evaluate SHARCS on 8 classification tasks in the standard GLUE benchmark (Wang et al., 2018): MNLI-m (Williams et al., 2018), QNLI (Wang et al., 2018; Rajpurkar et al., 2016), QQP (Wang et al., 2017), SST-2 (Socher et al., 2013), MRPC (Dolan and Brockett, 2005), RTE (Dagan et al., 2005), CoLA (Warstadt et al., 2019), WNLI (Levesque et al., 2012).",
+        "bbox": [
+            507,
+            370,
+            884,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics. Following previous work, we report the accuracy on the validation set, with an exception to CoLA for which we report Matthews correlation. We measure the total FLOPs on the validation set as it is invariant to the run time environment (Liu et al., 2020).",
+        "bbox": [
+            507,
+            510,
+            885,
+            606
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Training details. We train network with AdamW optimizer (Loshchilov and Hutter, 2019). We choose the number of epochs in  $\\{5,10,15\\}$  and use learning rate in  $\\{2e - 5,5e - 5\\}$  in our experiments. Please see Appendix C for more details.",
+        "bbox": [
+            507,
+            615,
+            884,
+            697
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            709,
+            608,
+            725
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 SHARCS is Better",
+        "text_level": 1,
+        "bbox": [
+            507,
+            737,
+            682,
+            751
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We compare SHARCS with existing sample adaptive inference methods: RTJ (Schwartz et al., 2020), DeeBERT (Xin et al., 2020), FastBERT (Liu et al., 2020), BERxiT (Xin et al., 2021), and PABEE (Zhou et al., 2020). Figure 2 shows the accuracy-FLOPs trade-off plots for different sample adaptive approaches with RoBERTa $_{\\text{base}}$  model on three GLUE subtasks (See appendix D.0.1 for the full set of plots). Our results show that SHARCS significantly outperforms other methods, especially in",
+        "bbox": [
+            507,
+            758,
+            884,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10521",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/06545fefbe9f7da7780bfef39175e19aa031064fee03529bba7f2b2d61d24295.jpg",
+        "table_caption": [
+            "Table 1: Comparison of different adaptive inference methods on MNLI-m dataset for different FLOP ranges. For the full table, please see Appendix D.0.2"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">FLOPs range (Tera FLOPs)</td><td colspan=\"2\">Best Baseline</td><td colspan=\"2\">SHARCS</td></tr><tr><td>Acc.(%) ↑</td><td>FLOPs ↓</td><td>Acc.</td><td>FLOPs</td></tr><tr><td colspan=\"5\">Roberta (Acc: 87.6, FLOPs: 33.5)</td></tr><tr><td>0-10</td><td>61.83</td><td>9.56</td><td>76.37</td><td>4.91</td></tr><tr><td>10-20</td><td>84.19</td><td>18</td><td>85.93</td><td>17.92</td></tr><tr><td>20-30</td><td>87.53</td><td>29.86</td><td>87.38</td><td>28.35</td></tr><tr><td colspan=\"5\">BERT (Acc: 84.8, FLOPs: 33)</td></tr><tr><td>0-10</td><td>48.87</td><td>8.76</td><td>72.7</td><td>8.94</td></tr><tr><td>10-20</td><td>72.50</td><td>19.99</td><td>81.61</td><td>16.73</td></tr><tr><td>20-30</td><td>83.27</td><td>27.99</td><td>83.04</td><td>22.76</td></tr><tr><td colspan=\"5\">DistilBERT (Acc: 82.2, FLOPs: 16.57)</td></tr><tr><td>0-5</td><td>48.87</td><td>4.7</td><td>64.72</td><td>3.80</td></tr><tr><td>5-10</td><td>64.44</td><td>9.95</td><td>76.02</td><td>8.87</td></tr><tr><td>10-15</td><td>80.38</td><td>14.91</td><td>81.61</td><td>14.39</td></tr><tr><td colspan=\"5\">DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26)</td></tr><tr><td>0-4</td><td>65.61</td><td>3.78</td><td>78.92</td><td>3.99</td></tr><tr><td>4-6</td><td>76.01</td><td>5.42</td><td>81.48</td><td>5.21</td></tr><tr><td>6-8</td><td>83.86</td><td>7.92</td><td>83.40</td><td>7.33</td></tr></table>",
+        "bbox": [
+            117,
+            128,
+            485,
+            370
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "the low-FLOPs regime. This can suggest that by substantially reducing the width of deeper layers in the model, SHARCS can achieve a significant reduction in FLOPs. Furthermore, we can maintain the accuracy better compared to fully skipping deeper layers as commonly done in early exiting methods.",
+        "bbox": [
+            112,
+            387,
+            489,
+            486
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 SHARCS is Model Agnostic",
+        "text_level": 1,
+        "bbox": [
+            112,
+            495,
+            361,
+            511
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "SHARCS can be seamlessly integrated with any state-of-the-art non-efficient and efficient transformer-based encoders to further improve their efficiency. We use SHARCS and the baseline sample adaptive inference methods in (§4.1) with two standard nonefficient (RoBERTa<sub>base</sub> and BERT-Base<sub>base</sub>) and efficient (DistilBERT (Sanh et al., 2019) and DynaBERT (Hou et al., 2020)) models. Table 1 shows that SHARCS consistently improves the efficiency of different models while maintaining the accuracy better than other baseline sample adaptive methods. Note that, for brevity, we have selected the highest accuracy among all the sample adaptive baselines for each FLOPs range; please see Table 5 in Appendix D.0.2 for the full set of results. As an example, SHARCS can reduce the FLOPs of BERT<sub>base</sub> to half with near 3% reduction in accuracy whereas in other methods the drop is more than 12%. Interestingly, SHARCS can further improve the efficiency of already optimized networks. For instance, inference with SHARCS on DynaBERT 0.25 width takes 10-15% less FLOPs with less than 1% drop in accuracy. More accuracy vs. FLOPs trade-off results for DynaBERT 0.25 and DistilBERT can be found in Figure 10 and 11.",
+        "bbox": [
+            112,
+            517,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3 SHARCS is Fast",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            665,
+            98
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We compare the latency of SHARCS applied to  $\\mathrm{BERT}_{\\mathrm{base}}$  with the baseline adaptive approaches on the QQP (Wang et al., 2017) dataset. The latency measurements are conducted on an Intel Xeon Gold 6230 CPU with a batch size of 1. We use two reduction factors  $\\{0.25, 1.0\\}$  and place the router after layer 2 or 4 to get different accuracy-speed up trade offs. We keep the speed up between 2 to 3 and report the best accuracy for each method in this range. Table 2 shows that our method obtains higher or comparable accuracy in this speed up range. Interestingly, SHARCS achieves a performance on par with DistilBERT, which is trained with distillation on a much larger dataset, but with higher speed up.",
+        "bbox": [
+            507,
+            111,
+            885,
+            336
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.4 SHARCS Paired with Token Dropping",
+        "text_level": 1,
+        "bbox": [
+            507,
+            357,
+            838,
+            375
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As discussed in Section 1, in addition to early-exiting, token dropping is another well-known approach to sample-adaptive inference efficiency (Goyal et al., 2020; Guan et al., 2022). Token dropping enhances efficiency by progressively dropping tokens layer by layer, thereby decreasing the hidden states sequence length as the forward pass proceeds. In contrast, SHARCS improves efficiency via reducing the transformer block's width. Therefore token dropping and SHARCS should not interfere with each other and in fact, they can be paired together to bring more efficiency. To showcase this, we combine SHARCS with a RoBERTa $_{\\text{base}}$  network that has already been made more efficient using Transkimmer token dropping. Table 3 shows that on QNLI dataset, SHARCS reduces the FLOPs of RoBERTa + Transkimmer by  $40\\%$  with a negligible drop in accuracy.",
+        "bbox": [
+            507,
+            385,
+            884,
+            675
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/e85acf074ab31f13a812d74247211abde92d359527c510010c3f7b8325804ef3.jpg",
+        "table_caption": [
+            "Table 2: Inference speed up results on the QQP dataset on CPU for various adaptive inference techniques applied to BERT. For SHARCS we place the router after layers 2 and 4 and report two numbers."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Accuracy (%)↑</td><td>Speed up (×)↑</td></tr><tr><td>BERT</td><td>90.90</td><td>1</td></tr><tr><td>DistilBERT</td><td>88.50</td><td>2</td></tr><tr><td>BERxiT</td><td>85.60</td><td>2.04</td></tr><tr><td>FastBERT</td><td>83.99</td><td>2.17</td></tr><tr><td>DeeBERT</td><td>83.82</td><td>2.14</td></tr><tr><td>PABEE</td><td>89.09</td><td>2.06</td></tr><tr><td>RTJ</td><td>84.44</td><td>2</td></tr><tr><td>SHARCS (layer 2)</td><td>88.43</td><td>2.57</td></tr><tr><td>SHARCS (layer 4)</td><td>90.05</td><td>2.05</td></tr></table>",
+        "bbox": [
+            512,
+            753,
+            884,
+            908
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10522",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/3b9ccc4895901e6e9b76274d6803771948027439f22fd4e0ab14b3501752ce5b.jpg",
+        "table_caption": [
+            "Table 3: SHARCS can be applied to RoBERTa in addition to Transkimmer to bring more efficiency. The decrease in FLOPs and Accuracy in parentheses are with respect to RoBERTa + Transkimmer."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>QNLI Acc.</td><td>TeraFLOPs</td></tr><tr><td>RoBERTa Base</td><td>92.8%</td><td>24.5</td></tr><tr><td>+Transkimmer</td><td>89.45%</td><td>14.08</td></tr><tr><td>+Transkimmer + SHARCS</td><td>88.83% (-0.62%)</td><td>8.56 (~40%↓)</td></tr></table>",
+        "bbox": [
+            117,
+            149,
+            485,
+            221
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.5 Ablating the Router",
+        "text_level": 1,
+        "bbox": [
+            112,
+            246,
+            317,
+            262
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To show the efficacy of the router and its training strategy in SHARCS, we replace it with the routing strategy used in BERxiT (Xin et al., 2021): while training two sub-networks with reduction factors  $\\{0.25, 1.0\\}$ , we feed the training sample to both of them; if a sub-network correctly classifies the sample, the router label for that sub-network in that iteration would be 1. Otherwise, it would be zero. The backward pass is then done with all the losses of sub-networks and the router loss. Table 4 shows the area under the normalized accuracy FLOPs curve (AUC) for both methods on MNLIm. Please find more detailed ablation results and experiments in Appendix E.",
+        "bbox": [
+            112,
+            268,
+            489,
+            493
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/1947568018e9a0eb3f58b4b46156c2f310f193bd1dad3c82f61795d68a938e7e.jpg",
+        "table_caption": [
+            "Table 4: While using adaptive width with RoBERTa  ${}_{\\text{base }}$  on MNLI-m, SHARCS router outperforms BERxiT router."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Router on RoBERTabase</td><td>AUC ↑</td></tr><tr><td>SHARCS</td><td>0.78</td></tr><tr><td>BERxiT</td><td>0.73</td></tr></table>",
+        "bbox": [
+            176,
+            533,
+            418,
+            590
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            612,
+            247,
+            627
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We presented SHARCS as a new sample adaptive inference approach that can improve any network's inference efficiency. SHARCS incorporated a lightweight router which is trained with a novel approach using the confidence of the network predictions during training. Our experiments showed the superiority or complementary role of SHARCS compared to other sample adaptive inference methods across various datasets and backbones.",
+        "bbox": [
+            112,
+            638,
+            489,
+            782
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            796,
+            220,
+            810
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "While the router and training approach in SHARCS are general purpose, a limitation of this paper is its focus solely on studying the impact on transformer encoders. Nonetheless, decoder-only (Radford et al., 2018, 2019) and encoder-decoder (Raffel et al., 2019; Lewis et al., 2020) models are",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "widely used classes of models for which we plan to integrate SHARCS in the future work. It should be noted that although our approach can be applied to regression tasks with minor adjustments, this work does not include any results for regression tasks.",
+        "bbox": [
+            507,
+            84,
+            882,
+            165
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            175,
+            682,
+            193
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We thank members of RAIVN and H2Lab for helpful discussion and feedback, and Hyak cluster team at the University of Washington for infrastructure support. This research was supported by NSF IIS-2044660, ONR MURI N00014- 18-1-2670, and gifts from AI2, Google and Apple.",
+        "bbox": [
+            507,
+            200,
+            884,
+            298
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            325,
+            608,
+            338
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. Layer normalization.",
+            "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. CoRR, abs/2005.14165.",
+            "Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In Proceedings of the PASCAL Challenges Workshop on Recognising Textual Entailment.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805.",
+            "Devvrit, Sneha Kudugunta, Aditya Kusupati, Tim Dettmers, Kaifeng Chen, Inderjit Dhillon, Yulia Tsvetkov, Hajishirzi Hannaneh, Sham Kakade, Ali Farhadi, and Prateek Jain. 2023. Matformer: Nested transformer for elastic inference. arXiv preprint arxiv:2310.07707.",
+            "William B. Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In Proceedings of the Third International Workshop on Paraphrasing (IWP2005).",
+            "Kawin Ethayarajh, Yejin Choi, and Swabha Swayamdipta. 2021. Information-theoretic measures of dataset difficulty. CoRR, abs/2110.08420.",
+            "Saurabh Goyal, Anamitra R. Choudhury, Saurabh M. Raje, Venkatesan T. Chakaravarthy, Yogish Sabharwal, and Ashish Verma. 2020. Power-bert: Accel"
+        ],
+        "bbox": [
+            507,
+            347,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10523",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "erating bert inference via progressive word-vector elimination.",
+            "Yue Guan, Zhengyi Li, Jingwen Leng, Zhouhan Lin, and Minyi Guo. 2022. Transkimmer: Transformer learns to layer-wise skim. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7275-7286, Dublin, Ireland. Association for Computational Linguistics.",
+            "Dan Hendrycks and Kevin Gimpel. 2016. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, abs/1606.08415.",
+            "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network.",
+            "Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, and Qun Liu. 2020. Dynabert: Dynamic bert with adaptive width and depth. In Advances in Neural Information Processing Systems, volume 33, pages 9782-9793. Curran Associates, Inc.",
+            "Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. 2019. Tinybert: Distilling BERT for natural language understanding. CoRR, abs/1909.10351.",
+            "Yigitcan Kaya and Tudor Dumitras. 2018. How to stop off-the-shelf deep neural networks from overthinking. CoRR, abs/1810.07052.",
+            "Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W Mahoney, and Kurt Keutzer. 2021. I-bert: Integer-only bert quantization. International Conference on Machine Learning (Accepted).",
+            "Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. 2022. Matryoshka representation learning. Advances in Neural Information Processing Systems, 35:30233-30249.",
+            "François Lagunas, Ella Charlaix, Victor Sanh, and Alexander M. Rush. 2021. Block pruning for faster transformers. CoRR, abs/2109.04838.",
+            "Hector J. Levesque, Ernest Davis, and Leora Morgenstern. 2012. The Winograd Schema Challenge. In Proceedings of the Thirteenth International Conference on Principles of Knowledge Representation and Reasoning, KR'12, pages 552-561. AAAI Press, Rome, Italy.",
+            "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, pages 7871-7880, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Weijie Liu, Peng Zhou, Zhiruo Wang, Zhe Zhao, Haotang Deng, and Qi Ju. 2020. FastBERT: a self-distilling BERT with adaptive inference time. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6035-6044, Online. Association for Computational Linguistics.",
+            "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019a. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692.",
+            "Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting (Tim) Cheng, and Jian Sun. 2019b. Metapruning: Meta learning for automatic neural network channel pruning. CoRR, abs/1903.10258.",
+            "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization.",
+            "Paul Michel, Omer Levy, and Graham Neubig. 2019. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.",
+            "Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training.",
+            "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. CoRR, abs/1910.10683.",
+            "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, Austin, Texas. Association for Computational Linguistics.",
+            "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter. CoRR, abs/1910.01108.",
+            "Victor Sanh, Thomas Wolf, and Alexander M. Rush. 2020. Movement pruning: Adaptive sparsity by fin-tuning. CoRR, abs/2005.07683.",
+            "Roy Schwartz, Gabriel Stanovsky, Swabha Swayamdipta, Jesse Dodge, and Noah A. Smith. 2020. The right tool for the job: Matching model and instance complexities. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6640-6651, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10524",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Sheng Shen, Zhen Dong, Jiayu Ye, Linjian Ma, Zhewei Yao, Amir Gholami, Michael W. Mahoney, and Kurt Keutzer. 2019. Q-BERT: hessian based ultra low precision quantization of BERT. CoRR, abs/1909.05840.",
+            "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics.",
+            "Tianxiang Sun, Xiangyang Liu, Wei Zhu, Zhichao Geng, Lingling Wu, Yilong He, Yuan Ni, Guotong Xie, Xuanjing Huang, and Xipeng Qiu. 2022. A simple hash-based early exiting approach for language understanding and generation.",
+            "Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. 2020. MobileBERT: a compact task-agnostic BERT for resource-limited devices. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2158-2170, Online. Association for Computational Linguistics.",
+            "Surat Teerapittayanon, Bradley McDanel, and H. T. Kung. 2017. Branchynet: Fast inference via early exiting from deep neural networks. CoRR, abs/1709.01686.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.",
+            "Elena Voita, David Talbot, Fedor Moiseev, Rico Senrich, and Ivan Titov. 2019. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned.",
+            "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. CoRR, abs/1804.07461.",
+            "Zhiguo Wang, Wael Hamza, and Radu Florian. 2017. Bilateral multi-perspective matching for natural language sentences. CoRR, abs/1702.03814.",
+            "Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625-641.",
+            "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume"
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics.",
+            "Mengzhou Xia, Zexuan Zhong, and Danqi Chen. 2022. Structured pruning learns compact and accurate models.",
+            "Ji Xin, Raphael Tang, Jaejun Lee, Yaoliang Yu, and Jimmy Lin. 2020. DeeBERT: Dynamic early exiting for accelerating BERT inference. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2246-2251, Online. Association for Computational Linguistics.",
+            "Ji Xin, Raphael Tang, Yaoliang Yu, and Jimmy Lin. 2021. BERxiT: Early exiting for BERT with better fine-tuning and extension to regression. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 91-104, Online. Association for Computational Linguistics.",
+            "Wangchunshu Zhou, Canwen Xu, Tao Ge, Julian McAuley, Ke Xu, and Furu Wei. 2020. Bert loses patience: Fast and robust inference with early exit. In Advances in Neural Information Processing Systems, volume 33, pages 18330-18341. Curran Associates, Inc."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            453
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10525",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Background and Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            421,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Efficient non-adaptive inference In the context of non-adaptive methods, numerous studies in the literature have improved the inference efficiency of transformers through various techniques such as knowledge distillation (Hinton et al., 2015) into smaller models (Sanh et al., 2019; Sun et al., 2020; Hou et al., 2020; Jiao et al., 2019), pruning unimportant weights of the model (Xia et al., 2022; Lagunas et al., 2021; Sanh et al., 2020; Liu et al., 2019b), and weight quantization (Shen et al., 2019; Kim et al., 2021) to store weights of a network with lower precision values. The aforementioned approaches result in smaller and more efficient albeit fixed and static models.",
+        "bbox": [
+            112,
+            110,
+            489,
+            335
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Efficient adaptive inference. Another line of work have proposed adaptive inference methods that allow the network to allocate varying amounts of compute for each sample. The predominant technique in this area is early exiting via adding internal classifiers to intermediate layers (Schwartz et al., 2020; Kaya and Dumitras, 2018; Xin et al., 2020; Teerapittayanon et al., 2017; Zhou et al., 2020; Liu et al., 2020): To early exit, prior work either use the confidence score of internal classifiers' predictions (Schwartz et al., 2020; Xin et al., 2020); the entropy of these predictions (Liu et al., 2020); a module that predicts whether a layer should early exit or not (Xin et al., 2021); a patience based change in these internal prediction (Zhou et al., 2020); or a hash based mechanism to do token level early exiting (Sun et al., 2022).",
+        "bbox": [
+            115,
+            347,
+            489,
+            619
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Devvrit et al. (2023) recently proposed Mat-Former that can enable adaptive compute based on the resource constraints but does not utilize dynamic token-based routing making SHARCS complementary to it.",
+        "bbox": [
+            112,
+            621,
+            489,
+            702
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            715,
+            356,
+            732
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B.0.1 Transformer Networks",
+        "text_level": 1,
+        "bbox": [
+            114,
+            743,
+            357,
+            757
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Transformer networks (Vaswani et al., 2017) are composed of a stack of  $L$  layers and each layer has two main components: Multi-head attention (MHA) and Feed-forward network (FFN).",
+        "bbox": [
+            112,
+            763,
+            487,
+            827
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "MHA consists of  $n_h$  heads where each head computes the attention operation (Vaswani et al., 2017) on the projection of a sequence of  $l$  tokens  $x = (x_1, x_2, \\dots, x_l)$  into key, query, and values:  $o_{\\mathrm{head}}^i = \\mathrm{Attention}(\\mathbf{W}_Kx, \\mathbf{W}_{\\mathbf{Q}}x, \\mathbf{W}_{\\mathbf{V}}x)$ , wherein",
+        "bbox": [
+            112,
+            838,
+            489,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "$1\\leq i\\leq H$ $o_{\\mathrm{head}}^i\\in \\mathbb{R}^{d_{\\mathrm{head}}}$  and  $\\mathbf{W}_K,\\mathbf{W}_Q,\\mathbf{W}_V\\in$ $\\mathbb{R}^{d_{\\mathrm{model}}\\times d_{\\mathrm{head}}}$  are the key, query, and value projection matrices, and  $d_{\\mathrm{head}} = d_{\\mathrm{model}} / n_h$  . The outputs from different heads are concatenated into  $o_{\\mathrm{heads}}\\in \\mathbb{R}^{d_{\\mathrm{model}}}$  and projected with another matrix  $\\mathbf{W}_O\\in \\mathbb{R}^{d_{\\mathrm{model}}\\times d_{\\mathrm{model}}}$  and passed through a layer norm (Ba et al., 2016) to get the output:  $o_{\\mathrm{MHA}} = \\mathrm{LayerNorm}(x + \\mathbf{W}_o o_{\\mathrm{heads}})$",
+        "bbox": [
+            507,
+            83,
+            885,
+            214
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "FFN consists of two feed-forward layers  $W_{1}$  and  $W_{2}$  with GeLU (Hendrycks and Gimpel, 2016) non-linearity and takes the output of MHA module and computes LayerNorm  $(o_{\\mathrm{MHA}} + \\mathrm{GeLU}(\\mathbf{W}_1 o_{\\mathrm{MHA}}) \\mathbf{W}_2)$ .",
+        "bbox": [
+            507,
+            224,
+            885,
+            305
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B.0.2 Reducing Width of a Transformer Layer",
+        "text_level": 1,
+        "bbox": [
+            507,
+            317,
+            840,
+            349
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 illustrates our approach in reducing the capacity of a transformer network based on the given reduction factor. We leave the first  $K$  layers (aka non-adaptive module) of the model intact, where  $K$  is a hyperparameter. Given an input sequence  $x$ , after computing the intermediate representation  $h \\in \\mathbb{R}^{l \\times d_{\\mathrm{model}}}$  by the non-adaptive module, and the reduction factor  $r$  by the router, we reduce the width of different components which we will describe in detail next.",
+        "bbox": [
+            507,
+            353,
+            885,
+            513
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Before passing  $h$  to adaptive layers, we pass it through a pooler module that reduces its dimensionality to  $d_{\\mathrm{model}} \\times r$ . For example, if  $d_{\\mathrm{model}} = 768$  and  $r = 0.25$ , the input to the adaptive module will be 192 dimensional. Although we experimented with affine transformations for the pooler, in practice we observed that a simpler pooling operation such as selecting just the first  $d_{\\mathrm{model}} \\times r$  dimensions works as good. Therefore, as this pooler is simpler and importantly adds no extra FLOPs to inference, we opted for that. For the unpooler, we used an affine transformation that transforms the hidden state with reduced dimensionality (i.e.  $r \\cdot d_{\\mathrm{model}}$ ) to  $d_{\\mathrm{model}}$ . Note that for the unpooler layer, the dimensionality of the input can change while the output dimensionality is always  $d_{\\mathrm{model}}$ .",
+        "bbox": [
+            507,
+            514,
+            885,
+            772
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Throughout this paper, whenever we decrease a dimension of a component (e.g. input vectors or weights) from  $d$  to  $d'$ , we exclusively use the first  $d'$  dimensions and disregards the remainder in the computations. This is illustrated schematically in Figure 4 for FFN and MHA components in transformer models.",
+        "bbox": [
+            507,
+            772,
+            885,
+            884
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In what follows we describe how we reduce the width of different components by the reduction",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10526",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9a44d48560b2ec04290012b686de013f6a6be101ecc6da65272120d543326732.jpg",
+        "image_caption": [
+            "Figure 3: Detailed schematic of reducing with of different components in a transformer model based on the prediction of the router."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            80,
+            485,
+            506
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "factor  $r$  which is also depicted in detail in Figure 3.",
+        "bbox": [
+            112,
+            583,
+            487,
+            600
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Reducing MHA Width. We reduce the width of MHA using the following steps:",
+        "bbox": [
+            112,
+            611,
+            487,
+            643
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. The input dimension  $d_{\\mathrm{model}}$  of self-attention projection matrices  $\\mathbf{W}_K$ ,  $\\mathbf{W}_Q$ , and  $\\mathbf{W}_v$  is decreased to  $d_{\\mathrm{model}} \\cdot r$ . We do not change the output dimension  $d_{\\mathrm{head}}$  of the linear projections.",
+            "2. The adaptive module only computes the output of the first  $n_h \\cdot r$  heads and disregards the other heads. Therefore the dimensionality of  $o_{\\mathrm{heads}}$  is decreased to  $d_{\\mathrm{model}} \\cdot r$ . Note that we could also reduce the dimensionality of each head instead, however, we built on previous findings in the literature that many of the heads in transformers are redundant and can be pruned without hurting the performance (Michel et al., 2019; Voita et al., 2019).",
+            "3. For the output projection  $\\mathbf{W}_o$  of MHA, the input and output dimensions, which are both equal to  $d_{\\mathrm{model}}$ , are reduced to  $d_{\\mathrm{model}}\\cdot r$ ."
+        ],
+        "bbox": [
+            112,
+            644,
+            487,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Throughout all the changes above, we do not",
+        "bbox": [
+            131,
+            903,
+            487,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "alter the sequence length of the input (i.e.  $l$ ) or any of the hidden states.",
+        "bbox": [
+            507,
+            84,
+            882,
+            115
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Reducing FFN width. Similar to the output projection  $\\mathbf{W}_o$  of MHA, the input and output dimensions of the feed-forward layers W1 and W2 in the FFN are reduced by a factor of  $1 / r$ . Therefore, the output dimension of all the adaptive layers are reduced by a factor of  $r$  to  $D_{\\mathrm{model}}\\times r$ .",
+        "bbox": [
+            507,
+            124,
+            884,
+            219
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "It is important to note that majority of operations in FFN and MHA components are comprised of matrix multiplications and reducing the input and output dimensions of these multiplications by a factor of  $r_j$  will reduce their flops by a factor of  $r_j^2$ .",
+        "bbox": [
+            507,
+            221,
+            882,
+            305
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Reducing layernorms width. We also reduce the width of the layer norm parameters by a factor of  $r$ . In our experiments, we find it beneficial to initialize the layernorm parameters with the corresponding first dimensions of original layernorm parameters instead of training them from scratch.",
+        "bbox": [
+            507,
+            309,
+            882,
+            405
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B.0.3 Inference",
+        "text_level": 1,
+        "bbox": [
+            509,
+            414,
+            648,
+            428
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In contrast to training where the width of the adaptive layers doing forward pass is enforced by the router labels, at inference, the router predicts the reduction factor. More formally, given the input sample  $x$  and router logits  $\\hat{W} \\in \\mathbb{R}^{M}$ , we select the reduction factor  $r_j$  to do the forward pass where  $j$  is  $\\arg \\max (\\hat{W})$ .",
+        "bbox": [
+            507,
+            432,
+            882,
+            546
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "C Training Details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            556,
+            690,
+            573
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Similar to (Hou et al., 2020), we reorder the heads based on their importance before fine-tuning on the down-stream task. We set  $\\lambda_{task}$  to 1 and  $\\lambda_{routert}$  to 0.5 in our approach and use a set of two reduction factors  $\\{0.25,1.0\\}$ . We choose the batch size in  $\\{16,32\\}$  depending on the model and dataset. We do a grid search over the lower and upper bound on the confidence thresholds and do the forward pass of each reduction factor on a separate GPU. We choose the window size values in  $\\{3,5\\}$ . We set  $\\lambda_{task}$  to 1 and  $\\lambda_{routert}$  to 0.5 in our approach and use a set of two reduction factors  $\\{0.25,1.0\\}$ . We do a grid search over the lower and upper bound on the confidence thresholds and the hyperparameters of the baselines and do the forward pass of each reduction factor on a separate GPU. Similar to previous work (Hou et al., 2020), we pad the input sequences in a batch to 128 tokens while training. To train SHARCS with Transkimer on RoBERTa, we first train RoBERTa + Transkimmer with skim coefficient of 0.3. We then start from this model as",
+        "bbox": [
+            507,
+            581,
+            882,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10527",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/05e9e082825ff27a58464e01a82590a852dd5999f580000d2f9bf1503f0add7a.jpg",
+        "image_caption": [
+            "(a) FFN"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            201,
+            99,
+            478,
+            288
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "image",
+        "img_path": "images/c57f8a98286552dc41e1d74aac8529252e487827100efada3cf742fd610c7c07.jpg",
+        "image_caption": [
+            "(b) MHA"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            99,
+            796,
+            290
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "an initial checkpoint and train for 10 epochs using both our loss and Transkimmer loss. Here, training the router is done with with 0.25 and 1.0 as reduction factors.",
+        "bbox": [
+            112,
+            368,
+            487,
+            431
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "D Results Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            450,
+            280,
+            464
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "D.0.1 GLUE Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            481,
+            292,
+            494
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We do a thorough hyperparameter search with different sample adaptive inference methods to get the accuracy of the RoBERTa $_{\\text{base}}$  model across different FLOPs values. We report the total number of FLOPs on the validation set and do not pad the input sequences before feeding to the model. Figure 12 shows the plots for 8 of the sub-tasks in the GLUE benchmark.",
+        "bbox": [
+            112,
+            504,
+            489,
+            632
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "D.0.2 Different Backbone Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            648,
+            393,
+            662
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 5 shows the comparison of different sample adaptive inference methods applied to different backbones.",
+        "bbox": [
+            112,
+            671,
+            487,
+            717
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "E Ablations and Discussion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            737,
+            369,
+            752
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "E.0.1 Router Ablation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            766,
+            305,
+            781
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Figure 5 shows the accuracy FLOPs plot for two methods: 1) The router introduced in SHARCS with adaptive width sub-networks. 2) A router similar to the early-exiting router used in BERxiT (Xin et al., 2021) with adaptive-width sub-networks. We report the AUC for a similar range of FLOPs for both methods and scale the accuracy and FLOPs values to [0, 1].",
+        "bbox": [
+            112,
+            790,
+            487,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a34ba4ad53d34380a5d75e19b2fb8c1eeb4b0f2187b01e7bc54be9532d3a9e5.jpg",
+        "image_caption": [
+            "Figure 4: Reducing the width of FFN and MHA in adaptive layers.",
+            "Figure 5: Accuracy vs. FLOPs plot for Adaptive width sub-networks trained with SHARCS router (orange) and BERxiT (Xin et al., 2021) router (blue)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            394,
+            848,
+            614
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "E.0.2 Number of Reduction Factors",
+        "text_level": 1,
+        "bbox": [
+            507,
+            714,
+            806,
+            728
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We change the number of reduction factors (or hardness levels)  $M$  to three and four and use the reduction factors  $\\{0.25, 0.5, 1.0\\}$  and  $\\{0.25, 0.5, 0.75, 1.0\\}$  respectively. Figure 6 shows the results of changing the number of reduction factors with  $\\mathrm{RoBERTa_{base}}$  model on MNLI-m dataset. The model trained with two reduction factors outperform the other cases for FLOPs values above 15 teraFLOPs significantly and four reduction factors can get better results in the range of 10-15 Tera FLOPs.",
+        "bbox": [
+            507,
+            741,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10528",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/c31d869e3f80b7ce895fbb7bb614b6f941e09dfd5c96fae6f5e571b76d205c18.jpg",
+        "table_caption": [
+            "Table 5: Comparison of different adaptive inference methods on MNLI-m dataset for different FLOP ranges."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">FLOPs range (Tera Flops)</td><td colspan=\"2\">RTJ</td><td colspan=\"2\">DeeBERT</td><td colspan=\"2\">PABEE</td><td colspan=\"2\">FastBERT</td><td colspan=\"2\">BERxiT</td><td colspan=\"2\">SHARCS</td></tr><tr><td>Acc. ↑</td><td>FLOPs ↓</td><td>Acc.</td><td>FLOPs</td><td>Acc.</td><td>FLOPs</td><td>Acc.</td><td>FLOPs</td><td>Acc.</td><td>FLOPs</td><td>Acc.</td><td>FLOPs</td></tr><tr><td colspan=\"13\">Roberta (Acc: 87.6, FLOPs: 33.5)</td></tr><tr><td>0-10</td><td>54.27</td><td>9.59</td><td>61.83</td><td>9.56</td><td>53.98</td><td>8.53</td><td>45.39</td><td>2.84</td><td>53.69</td><td>8.41</td><td>76.37</td><td>4.91</td></tr><tr><td>10-20</td><td>82.29</td><td>19.77</td><td>77.18</td><td>19.80</td><td>84.19</td><td>18</td><td>59.79</td><td>15.06</td><td>83.39</td><td>19.91</td><td>85.93</td><td>17.92</td></tr><tr><td>20-30</td><td>86.92</td><td>25.76</td><td>81.35</td><td>23.67</td><td>87.53</td><td>29.86</td><td>77.24</td><td>29.95</td><td>87.23</td><td>27.92</td><td>87.38</td><td>28.35</td></tr><tr><td colspan=\"13\">BERT (Acc: 84.8, FLOPs: 33)</td></tr><tr><td>0-10</td><td>47.32</td><td>9.46</td><td>42.32</td><td>9.97</td><td>42.97</td><td>7.51</td><td>48.87</td><td>8.76</td><td>46.86</td><td>9.88</td><td>72.7</td><td>8.94</td></tr><tr><td>10-20</td><td>58.12</td><td>15.61</td><td>61.57</td><td>19.89</td><td>66.33</td><td>17.29</td><td>70.25</td><td>19.18</td><td>72.50</td><td>19.99</td><td>81.61</td><td>16.73</td></tr><tr><td>20-30</td><td>78.96</td><td>28.53</td><td>80.57</td><td>29.34</td><td>81.13</td><td>27.66</td><td>83.27</td><td>27.99</td><td>80.82</td><td>25.73</td><td>83.04</td><td>22.76</td></tr><tr><td colspan=\"13\">DistilBERT (Acc: 82.2, FLOPs: 16.57)</td></tr><tr><td>0-5</td><td>41.92</td><td>2.76</td><td>43.98</td><td>2.76</td><td>41.45</td><td>2.76</td><td>48.87</td><td>4.7</td><td>41.47</td><td>4.07</td><td>64.72</td><td>3.80</td></tr><tr><td>5-10</td><td>52.72</td><td>9.56</td><td>59.87</td><td>9.45</td><td>64.44</td><td>9.95</td><td>64.24</td><td>9.76</td><td>55.10</td><td>9.76</td><td>76.02</td><td>8.87</td></tr><tr><td>10-15</td><td>72.93</td><td>13.90</td><td>80.38</td><td>14.91</td><td>79.29</td><td>14.12</td><td>79.72</td><td>14.84</td><td>60.25</td><td>11.68</td><td>81.61</td><td>14.39</td></tr><tr><td colspan=\"13\">DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26)</td></tr><tr><td>0-4</td><td>60.95</td><td>3.98</td><td>56.13</td><td>3.38</td><td>56.09</td><td>3.97</td><td>65.61</td><td>3.78</td><td>56.19</td><td>3.58</td><td>78.92</td><td>3.99</td></tr><tr><td>4-6</td><td>74.55</td><td>5.82</td><td>75.26</td><td>5.62</td><td>69.81</td><td>5.93</td><td>76.01</td><td>5.42</td><td>-</td><td>-</td><td>81.48</td><td>5.21</td></tr><tr><td>6-8</td><td>83.43</td><td>7.70</td><td>83.36</td><td>7.70</td><td>81.31</td><td>7.87</td><td>83.86</td><td>7.92</td><td>83.08</td><td>7.60</td><td>83.40</td><td>7.33</td></tr></table>",
+        "bbox": [
+            114,
+            107,
+            895,
+            331
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/a16e1f8928a4e2cf5c603b9ce28b863736461b63322287c25996ec6bb1f1284b.jpg",
+        "image_caption": [
+            "Figure 6: Results of different number of sub-networks (or reduction factors) with RoBERTa $_{\\text{base}}$  model on MNLI-m."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            164,
+            374,
+            423,
+            549
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "E.0.3 Confidence Thresholds",
+        "text_level": 1,
+        "bbox": [
+            112,
+            648,
+            357,
+            662
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We study the effect of changing confidence thresholds on our results with a simple experiment: With two reduction factors  $\\{0.25, 1.0\\}$  we use the following confidence score lower and upper bounds:  $[0.0, x]$  for full network and  $[x, 1.0]$  for 0.25 network, where  $x \\in \\{0.5, 0.7, 0.9\\}$ . We place the router after layer 1 (i.e.  $K = 1$ ). Figure 7 shows the results of RoBERTa<sub>base</sub> model on MNLI-m dataset. According to the figure, higher value of  $x$  leads to better accuracies for a fixed number of FLOPs. Furthermore, as we decrease  $x$ , we can reach lower values of FLOPs. This is intuitive as decreasing the lower bound of smallest reduction factor makes more samples be routed to that which reduces the overall FLOPs.",
+        "bbox": [
+            112,
+            677,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/1b8379324c27338c0d17ca4a3b8eb69621907d6b1a92828f5fd967191e081341.jpg",
+        "image_caption": [
+            "Figure 7: Results of changing confidence thresholds with RoBERTabase model on MNLI-m dataset. We use reduction factors  $\\{0.25,1.0\\}$  and confidence thresholds  $[0,x]$  for reduction factor 0.25 and confidence thresholds  $[x,1]$  for reduction factor 1, where  $x\\in$ $\\{0.5,0.7,0.9\\}$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            559,
+            374,
+            820,
+            549
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "E.0.4 Using Entropy or Confidence-based Hardness Labels to Train Router",
+        "text_level": 1,
+        "bbox": [
+            507,
+            683,
+            853,
+            713
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Similar to sample adaptive inference in early exiting methods (Xin et al., 2020; Liu et al., 2020), one can get hardness labels by defining thresholds on entropy of the network predictions instead of the confidence. We use the following formula to compute the entropy of network's prediction:",
+        "bbox": [
+            507,
+            719,
+            884,
+            816
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH = - \\sum_ {i = 1} ^ {C} p _ {i} \\cdot \\log p _ {i},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            611,
+            831,
+            779,
+            873
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "wherein  $C$  denotes the number of classes of the classifier, and  $p_i$  is the softmax probability that the",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "10529",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/89adfb1b376e0587f7131ba35925fe1396b504148361c8dfd383df6566cde0b3.jpg",
+        "image_caption": [
+            "Figure 8: Comparing entropy and confidence based hardness labels for training the router."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            111,
+            453,
+            331
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/e4bebf2b1a74074efe09d7bde4d017fd318196e8db7e209572267cdfb3fa386a.jpg",
+        "image_caption": [
+            "Figure 9: Effect of changing history window size in SHARCS on MNLI-m Accuracy FLOPs trade-off."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            169,
+            417,
+            408,
+            575
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "classifier assigns to class  $i$ . Figure 8 shows the accuracy FLOPs trade-off for both of these metrics on MNLI-m dataset with RoBERTa $_{\\text{base}}$  model. Given that the confidence score is a simpler approach and does not require any additional computation, we opted for using that in our method.",
+        "bbox": [
+            112,
+            646,
+            489,
+            743
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "E.0.5 Changing History Window Size in Training the Router",
+        "text_level": 1,
+        "bbox": [
+            112,
+            753,
+            443,
+            785
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in (§2.1), having a larger window size helps stabilizing training the router as the hardness label for a training sample might change throughout training. To illustrate this effect, we train SHARCS on RoBERTa $_{\\text{base}}$  model with three different window sizes  $\\{1, 3, 5\\}$  for 10 epochs on MNLI-m dataset. We place the router at layer 2, use two reduction factors  $\\{0.25, 1.0\\}$ , and set the",
+        "bbox": [
+            112,
+            790,
+            489,
+            919
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/7034b2357ef10d7e0c04c21de42a8d34dd162fbdcadf2917c533f66412323c6e.jpg",
+        "image_caption": [
+            "Figure 10: Accuracy FLOPs plot for SHARCS applied to DynaBERT 0.25 W (Hou et al., 2020) model. We set the reduction factors to  $\\{2/3, 1.0\\}$ . By placing router at different layers and changing the confidence thresholds, we can get different points in the plot. Note that the FLOPs for each sub-network in each point is also reported."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            86,
+            857,
+            318
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "confidence thresholds to [0.0, 0.8] for reduction factor 1 and [0.8, 1.0] for reduction factor 0.25. Figure 9 shows the accuracy FLOPs trade offs that different checkpoints with different window sizes get throughout training on the validation set. According to the figure, the network can reach higher accuracy when trained with window size 3 or 5. We did not see any significant improvements by using a window size larger than 5.",
+        "bbox": [
+            507,
+            451,
+            884,
+            598
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "10530",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/9ac553ab79c6eb7e321a17530ed927d1e7052f20e380716a6a5af33a9313a541.jpg",
+        "image_caption": [
+            "Figure 11: Accuracy FLOPs plot for SHARCS applied to DistilBERT(Sanh et al., 2019) model. We set the reduction factors to  $\\{2/3, 1.0\\}$  or  $\\{0.5, 1.0\\}$ . By placing router at different layers and changing the confidence thresholds, we can get different points in the plot. Note that the FLOPs for each sub-network in each point is also reported."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            326,
+            463,
+            558
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "10531",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/8885e9a306c7751bd99d564a50c13203ddf30d79b10b5a2a1672a99d4bd39e52.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            144,
+            221,
+            853,
+            237
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/a05d1a2c52e7c555d27612f907bc8403f58597bded762763024c5d94ca9845ba.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            265,
+            369,
+            428
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/73c428675368f761f3b3447bb34de1d7f213c883a3c276f68e7c2c8631db35bf.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            371,
+            263,
+            623,
+            428
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/c36876da88cfe6c1416e8cc096fcb24ca31d27180affbd2dd838a5b10264db50.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            628,
+            263,
+            884,
+            428
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/ae23cc1bb912f2366b10057d7c74677bca03533e2ec22be447969fec214816fe.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            429,
+            366,
+            590
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/1a890f8a994dd58a55c3bcf847c04e33d5dd36a87811147f9d07396c0fe39831.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            371,
+            429,
+            625,
+            590
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/50547702ea5ade889da2b1660f068d54ee20c72eac5846a42fd0ea8861003b92.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            628,
+            429,
+            882,
+            590
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf92c78996ddf323036bd017f7e88195e3b48fa578efbc65b44d6061b608a50a.jpg",
+        "image_caption": [
+            "Figure 12: GLUE benchmark results. Best viewed in color."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            242,
+            592,
+            495,
+            755
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/2b92e068d931f35bb484a6e4f930c92fa4798b723686e4c4b61dcf3332f4399b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            495,
+            592,
+            754,
+            755
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "10532",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    }
+]