Add Batch 01142821-d0da-454d-94d9-0c8e7f69d204 data

.gitattributes

@@ -5475,3 +5475,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/Imagination[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need![[:space:]]Curved[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Abstract[[:space:]]Sequence[[:space:]]Modeling[[:space:]]Utilized[[:space:]]on[[:space:]]Long[[:space:]]Short-Term[[:space:]]Dialogue[[:space:]]Planning/86e59497-dae5-4baa-bcfc-ac1d5a96027f_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Impact[[:space:]]of[[:space:]]Adversarial[[:space:]]Training[[:space:]]on[[:space:]]Robustness[[:space:]]and[[:space:]]Generalizability[[:space:]]of[[:space:]]Language[[:space:]]Models/b54716d2-96e1-420a-981e-86a27693e497_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Implicit[[:space:]]Memory[[:space:]]Transformer[[:space:]]for[[:space:]]Computationally[[:space:]]Efficient[[:space:]]Simultaneous[[:space:]]Speech[[:space:]]Translation/67204848-392e-4761-bcf4-3a168be6e98f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Importance[[:space:]]of[[:space:]]Synthesizing[[:space:]]High-quality[[:space:]]Data[[:space:]]for[[:space:]]Text-to-SQL[[:space:]]Parsing/cd97ebf7-b695-4707-b5bf-357c671f0c08_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improved[[:space:]]Logical[[:space:]]Reasoning[[:space:]]of[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Differentiable[[:space:]]Symbolic[[:space:]]Programming/a3b32712-0fc6-409b-b027-c0f54131f88f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improved[[:space:]]Visual[[:space:]]Story[[:space:]]Generation[[:space:]]with[[:space:]]Adaptive[[:space:]]Context[[:space:]]Modeling/6027a132-d9ee-4656-b4da-85df0a85a689_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Autoregressive[[:space:]]Grammatical[[:space:]]Error[[:space:]]Correction[[:space:]]with[[:space:]]Non-autoregressive[[:space:]]Models/3fd97eae-0b66-43ab-a8c2-6db714a9b22c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Classroom[[:space:]]Dialogue[[:space:]]Act[[:space:]]Recognition[[:space:]]from[[:space:]]Limited[[:space:]]Labeled[[:space:]]Data[[:space:]]with[[:space:]]Self-Supervised[[:space:]]Contrastive[[:space:]]Learning[[:space:]]Classifiers/b6413ac7-f15b-410f-9dfa-dd65e39b5f0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Contrastive[[:space:]]Learning[[:space:]]of[[:space:]]Sentence[[:space:]]Embeddings[[:space:]]from[[:space:]]AI[[:space:]]Feedback/8930019d-cd04-44ea-a08a-812fc0aef694_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Cross-task[[:space:]]Generalization[[:space:]]of[[:space:]]Unified[[:space:]]Table-to-text[[:space:]]Models[[:space:]]with[[:space:]]Compositional[[:space:]]Task[[:space:]]Configurations/c4fe3aff-0902-4948-b26d-fcae326e24e5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Diachronic[[:space:]]Word[[:space:]]Sense[[:space:]]Induction[[:space:]]with[[:space:]]a[[:space:]]Nonparametric[[:space:]]Bayesian[[:space:]]method/650dc53b-8e23-4a7c-9f4e-10f8ef2f245a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Embedding-based[[:space:]]Unsupervised[[:space:]]Keyphrase[[:space:]]Extraction[[:space:]]by[[:space:]]Incorporating[[:space:]]Structural[[:space:]]Information/a6ac9ba6-311e-4129-8f7e-5207177224bf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Empathetic[[:space:]]Dialogue[[:space:]]Generation[[:space:]]by[[:space:]]Dynamically[[:space:]]Infusing[[:space:]]Commonsense[[:space:]]Knowledge/2b8d89f7-6db0-432d-b63a-f299bf6fa55c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Grammatical[[:space:]]Error[[:space:]]Correction[[:space:]]with[[:space:]]Multimodal[[:space:]]Feature[[:space:]]Integration/2e774ee1-d0ce-47d5-9c20-f9aaf0c37e7f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Completion[[:space:]]with[[:space:]]Generative[[:space:]]Hard[[:space:]]Negative[[:space:]]Mining/f5b3a40c-d0b9-4194-abb7-efe173696136_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Language[[:space:]]Model[[:space:]]Integration[[:space:]]for[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/51d542e5-15e3-4012-8392-dc91118c7ba9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Long[[:space:]]Dialogue[[:space:]]Summarization[[:space:]]with[[:space:]]Semantic[[:space:]]Graph[[:space:]]Representation/d4936ed2-3014-49e9-bca7-7dcf4f82fb91_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition[[:space:]]via[[:space:]]Bridge-based[[:space:]]Domain[[:space:]]Adaptation/f0b40895-e4b4-4c92-9500-8775d746daf3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Radiology[[:space:]]Summarization[[:space:]]with[[:space:]]Radiograph[[:space:]]and[[:space:]]Anatomy[[:space:]]Prompts/df1b87a8-8bf7-40ea-9ef5-c0aaecfccc7c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Improving[[:space:]]Zero-shot[[:space:]]Multilingual[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation[[:space:]]by[[:space:]]Leveraging[[:space:]]Cross-lingual[[:space:]]Consistency[[:space:]]Regularization/f81f1029-8753-4ab8-b826-1b551f690a53_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/In-context[[:space:]]Examples[[:space:]]Selection[[:space:]]for[[:space:]]Machine[[:space:]]Translation/b26c7f48-156a-4cf2-88bd-81e965e7cc2e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Incomplete[[:space:]]Utterance[[:space:]]Rewriting[[:space:]]as[[:space:]]Sequential[[:space:]]Greedy[[:space:]]Tagging/e758f53b-0cf6-4813-bf70-b03c9632fb15_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Incomplete[[:space:]]Utterance[[:space:]]Rewriting[[:space:]]by[[:space:]]A[[:space:]]Two-Phase[[:space:]]Locate-and-Fill[[:space:]]Regime/a35c7d6d-31c3-44cb-8e96-331ff9970f92_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Incorporating[[:space:]]Factuality[[:space:]]Inference[[:space:]]to[[:space:]]Identify[[:space:]]Document-level[[:space:]]Event[[:space:]]Factuality/03b8706b-ab3b-4e19-81dd-78877764d28d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Incorporating[[:space:]]Graph[[:space:]]Information[[:space:]]in[[:space:]]Transformer-based[[:space:]]AMR[[:space:]]Parsing/46499372-cd0a-4eed-a52b-aa44835df7b0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Inducing[[:space:]]Character-level[[:space:]]Structure[[:space:]]in[[:space:]]Subword-based[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Type-level[[:space:]]Interchange[[:space:]]Intervention[[:space:]]Training/f08fb3ea-5207-414d-b547-727bf3231e9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/InfoSync_[[:space:]]Information[[:space:]]Synchronization[[:space:]]across[[:space:]]Multilingual[[:space:]]Semi-structured[[:space:]]Tables/ab64ee58-4ba1-4be8-9a95-3869d4d57a44_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Injecting[[:space:]]Comparison[[:space:]]Skills[[:space:]]in[[:space:]]Task-Oriented[[:space:]]Dialogue[[:space:]]Systems[[:space:]]for[[:space:]]Database[[:space:]]Search[[:space:]]Results[[:space:]]Disambiguation/4e52600d-d17b-48a2-a8f2-3552f0973f03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Intent[[:space:]]Discovery[[:space:]]with[[:space:]]Frame-guided[[:space:]]Semantic[[:space:]]Regularization[[:space:]]and[[:space:]]Augmentation/dd486390-6ab3-4639-ada5-3ce7488328ec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Interactive[[:space:]]Concept[[:space:]]Learning[[:space:]]for[[:space:]]Uncovering[[:space:]]Latent[[:space:]]Themes[[:space:]]in[[:space:]]Large[[:space:]]Text[[:space:]]Collections/dd91ac5a-8a4e-4837-82a6-e74822b9da65_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Interpretable[[:space:]]Automatic[[:space:]]Fine-grained[[:space:]]Inconsistency[[:space:]]Detection[[:space:]]in[[:space:]]Text[[:space:]]Summarization/a78ad68c-e9a1-4071-b6a1-07c75e883530_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Interpretable[[:space:]]Multimodal[[:space:]]Misinformation[[:space:]]Detection[[:space:]]with[[:space:]]Logic[[:space:]]Reasoning/652db0af-97b0-41b3-88ef-15183dd96e41_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Interpreting[[:space:]]Sentiment[[:space:]]Composition[[:space:]]with[[:space:]]Latent[[:space:]]Semantic[[:space:]]Tree/5fa65e7d-4d28-4550-bf66-fc31c0cbb5d6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Investigating[[:space:]]Glyph-Phonetic[[:space:]]Information[[:space:]]for[[:space:]]Chinese[[:space:]]Spell[[:space:]]Checking_[[:space:]]What[[:space:]]Works[[:space:]]and[[:space:]]What’s[[:space:]]Next_/8e9f3a1e-fe4d-47a1-a235-c2364876018f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Investigating[[:space:]]Transformer-Guided[[:space:]]Chaining[[:space:]]for[[:space:]]Interpretable[[:space:]]Natural[[:space:]]Logic[[:space:]]Reasoning/b24b755b-0a27-43c4-908d-85d37857f4c3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Investigating[[:space:]]the[[:space:]]Saliency[[:space:]]of[[:space:]]Sentiment[[:space:]]Expressions[[:space:]]in[[:space:]]Aspect-Based[[:space:]]Sentiment[[:space:]]Analysis/357cf293-0845-47f4-8e02-1aaa416e43d7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Is[[:space:]]Continuous[[:space:]]Prompt[[:space:]]a[[:space:]]Combination[[:space:]]of[[:space:]]Discrete[[:space:]]Prompts_[[:space:]]Towards[[:space:]]a[[:space:]]Novel[[:space:]]View[[:space:]]for[[:space:]]Interpreting[[:space:]]Continuous[[:space:]]Prompts/c195b192-7b29-4011-9b6f-d5396331cfbd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/It[[:space:]]Takes[[:space:]]Two[[:space:]]to[[:space:]]Tango_[[:space:]]Navigating[[:space:]]Conceptualizations[[:space:]]of[[:space:]]NLP[[:space:]]Tasks[[:space:]]and[[:space:]]Measurements[[:space:]]of[[:space:]]Performance/223eca96-b6fe-4d2f-8619-d27ca18666d9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/It[[:space:]]is[[:space:]]a[[:space:]]Bird[[:space:]]Therefore[[:space:]]it[[:space:]]is[[:space:]]a[[:space:]]Robin_[[:space:]]On[[:space:]]BERT’s[[:space:]]Internal[[:space:]]Consistency[[:space:]]Between[[:space:]]Hypernym[[:space:]]Knowledge[[:space:]]and[[:space:]]Logical[[:space:]]Words/b9317111-5df3-4211-8373-b338831cf1c1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Iterative[[:space:]]Nearest[[:space:]]Neighbour[[:space:]]Machine[[:space:]]Translation[[:space:]]for[[:space:]]Unsupervised[[:space:]]Domain[[:space:]]Adaptation/4617e2c4-a5af-48b9-a7b2-0e0093c450a1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/It’s[[:space:]]not[[:space:]]Sexually[[:space:]]Suggestive;[[:space:]]It’s[[:space:]]Educative[[:space:]]_[[:space:]]Separating[[:space:]]Sex[[:space:]]Education[[:space:]]from[[:space:]]Suggestive[[:space:]]Content[[:space:]]on[[:space:]]TikTok[[:space:]]videos/4cd7989f-f759-400e-8ebe-083d97616f9c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/JECC_[[:space:]]Commonsense[[:space:]]Reasoning[[:space:]]Tasks[[:space:]]Derived[[:space:]]from[[:space:]]Interactive[[:space:]]Fictions/2df83e9c-9242-472b-9b12-9e046958f4bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Joint[[:space:]]Generator-Ranker[[:space:]]Learning[[:space:]]for[[:space:]]Natural[[:space:]]Language[[:space:]]Generation/e39ca44c-9a7b-4ab8-b049-91a860ff12a3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Joint[[:space:]]Speech[[:space:]]Transcription[[:space:]]and[[:space:]]Translation_[[:space:]]Pseudo-Labeling[[:space:]]with[[:space:]]Out-of-Distribution[[:space:]]Data/693f491a-8803-4b99-a045-5ff0f19b99d2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Jointly[[:space:]]Reparametrized[[:space:]]Multi-Layer[[:space:]]Adaptation[[:space:]]for[[:space:]]Efficient[[:space:]]and[[:space:]]Private[[:space:]]Tuning/5a923175-8d2d-49b0-9a77-7a05deb114da_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/K-UniMorph_[[:space:]]Korean[[:space:]]Universal[[:space:]]Morphology[[:space:]]and[[:space:]]its[[:space:]]Feature[[:space:]]Schema/2b0be061-024d-4756-937c-16533c36d002_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/KNSE_[[:space:]]A[[:space:]]Knowledge-aware[[:space:]]Natural[[:space:]]Language[[:space:]]Inference[[:space:]]Framework[[:space:]]for[[:space:]]Dialogue[[:space:]]Symptom[[:space:]]Status[[:space:]]Recognition/5ddb0ce8-a93d-4483-a48a-646c0191cf8b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Kanbun-LM_[[:space:]]Reading[[:space:]]and[[:space:]]Translating[[:space:]]Classical[[:space:]]Chinese[[:space:]]in[[:space:]]Japanese[[:space:]]Methods[[:space:]]by[[:space:]]Language[[:space:]]Models/e8e59fee-6460-4690-8021-2736aa4f5014_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Know[[:space:]]What[[:space:]]I[[:space:]]don’t[[:space:]]Know_[[:space:]]Handling[[:space:]]Ambiguous[[:space:]]and[[:space:]]Unknown[[:space:]]Questions[[:space:]]for[[:space:]]Text-to-SQL/8f767dea-c3d9-40a4-a39b-fec6ccbf3e58_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Know[[:space:]]Where[[:space:]]You’re[[:space:]]Going_[[:space:]]Meta-Learning[[:space:]]for[[:space:]]Parameter-Efficient[[:space:]]Fine-Tuning/4d7b23fc-e05e-4a85-8a72-c24618f33d8d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Knowing-how[[:space:]]&[[:space:]]Knowing-that_[[:space:]]A[[:space:]]New[[:space:]]Task[[:space:]]for[[:space:]]Machine[[:space:]]Comprehension[[:space:]]of[[:space:]]User[[:space:]]Manuals/cb5ca73e-133f-473a-894c-4fc8d0c7abad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Knowledge[[:space:]]Graph[[:space:]]Embeddings[[:space:]]using[[:space:]]Neural[[:space:]]Ito[[:space:]]Process_[[:space:]]From[[:space:]]Multiple[[:space:]]Walks[[:space:]]to[[:space:]]Stochastic[[:space:]]Trajectories/08dfb287-8409-46bc-a5ed-61a2b871388f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/KoRC_[[:space:]]Knowledge[[:space:]]Oriented[[:space:]]Reading[[:space:]]Comprehension[[:space:]]Benchmark[[:space:]]for[[:space:]]Deep[[:space:]]Text[[:space:]]Understanding/7b1ca5ce-41e7-456f-a62c-913e9f77ca98_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LABO_[[:space:]]Towards[[:space:]]Learning[[:space:]]Optimal[[:space:]]Label[[:space:]]Regularization[[:space:]]via[[:space:]]Bi-level[[:space:]]Optimization/bfe8af4c-5b1e-4694-91ce-c2993d7da291_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LEDA_[[:space:]]a[[:space:]]Large-Organization[[:space:]]Email-Based[[:space:]]Decision-Dialogue-Act[[:space:]]Analysis[[:space:]]Dataset/0c970994-cbed-4037-a48c-b8d990d06573_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LET_[[:space:]]Leveraging[[:space:]]Error[[:space:]]Type[[:space:]]Information[[:space:]]for[[:space:]]Grammatical[[:space:]]Error[[:space:]]Correction/bcb7f9b0-0dee-40cd-bcdb-a16c906d8e92_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LMCap_[[:space:]]Few-shot[[:space:]]Multilingual[[:space:]]Image[[:space:]]Captioning[[:space:]]by[[:space:]]Retrieval[[:space:]]Augmented[[:space:]]Language[[:space:]]Model[[:space:]]Prompting/6e842d66-5be3-4c79-9097-9bab814073b5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LMentry_[[:space:]]A[[:space:]]Language[[:space:]]Model[[:space:]]Benchmark[[:space:]]of[[:space:]]Elementary[[:space:]]Language[[:space:]]Tasks/c9ad5eaf-cc50-4fac-8e34-cd4f17bfa752_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LMs[[:space:]]stand[[:space:]]their[[:space:]]Ground_[[:space:]]Investigating[[:space:]]the[[:space:]]Effect[[:space:]]of[[:space:]]Embodiment[[:space:]]in[[:space:]]Figurative[[:space:]]Language[[:space:]]Interpretation[[:space:]]by[[:space:]]Language[[:space:]]Models/75df301f-38b5-4494-8f9f-34b334a36b62_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LR-Sum_[[:space:]]Summarization[[:space:]]for[[:space:]]Less-Resourced[[:space:]]Languages/3dfa95f6-bfae-4363-aff6-7d297ea57ca1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LaSQuE_[[:space:]]Improved[[:space:]]Zero-Shot[[:space:]]Classification[[:space:]]from[[:space:]]Explanations[[:space:]]Through[[:space:]]Quantifier[[:space:]]Modeling[[:space:]]and[[:space:]]Curriculum[[:space:]]Learning/d5f98dc8-9816-4e0b-920d-986b9ca089d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Label[[:space:]]Agnostic[[:space:]]Pre-training[[:space:]]for[[:space:]]Zero-shot[[:space:]]Text[[:space:]]Classification/8d06ae10-f325-4c9d-b5e0-e209b563f051_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Language[[:space:]]Agnostic[[:space:]]Multilingual[[:space:]]Information[[:space:]]Retrieval[[:space:]]with[[:space:]]Contrastive[[:space:]]Learning/1700c1e6-3005-4899-b7c9-dc817003f695_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Language[[:space:]]Anisotropic[[:space:]]Cross-Lingual[[:space:]]Model[[:space:]]Editing/3298279b-c7b3-4699-9111-51fd47d06867_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Language[[:space:]]Model[[:space:]]Analysis[[:space:]]for[[:space:]]Ontology[[:space:]]Subsumption[[:space:]]Inference/6fa31862-b1ea-4de8-8da1-a4f2eba6ef5f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Language[[:space:]]Modeling[[:space:]]with[[:space:]]Latent[[:space:]]Situations/916ebf86-4821-46f4-a642-81f73035dbd7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Language[[:space:]]Models[[:space:]]for[[:space:]]German[[:space:]]Text[[:space:]]Simplification_[[:space:]]Overcoming[[:space:]]Parallel[[:space:]]Data[[:space:]]Scarcity[[:space:]]through[[:space:]]Style-specific[[:space:]]Pre-training/79849c60-d2cf-40cf-8f0a-9db2151769d3_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Importance of Synthesizing High-quality Data for Text-to-SQL Parsing/cd97ebf7-b695-4707-b5bf-357c671f0c08_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c5704fb1bfd2d020924648f8da672698b790ce18797ad9e5f29f6b30d7b265f
+size 590227

2023/Importance of Synthesizing High-quality Data for Text-to-SQL Parsing/full.md

@@ -0,0 +1,477 @@
+# Importance of Synthesizing High-quality Data for Text-to-SQL Parsing
+
+Yiqun Hu, Yiyun Zhao,\* Jiarong Jiang, Wuwei Lan, Henry Zhu, Anuj Chauhan, Alexander Li, Lin Pan, Jun Wang, Chung-Wei Hang, Sheng Zhang, Jiang Guo, Marvin Dong, Joe Lilien, Patrick Ng, Zhiguo Wang, Vittorio Castelli, Bing Xiang
+
+AWS AI Labs
+
+*yiyunzhao@arizona.edu
+
+{yiqunhu, jiarongj, lanwuwei, henghui, chaanj, hanboli, linpan, juwanga, cwhang, zshe, gujiang, mingwd, lilienj, patricng, zhiguow, vittorca, bxiang} @amazon.com
+
+# Abstract
+
+There has been increasing interest in synthesizing data to improve downstream text-to-SQL tasks. In this paper, we examined the existing synthesized datasets and discovered that state-of-the-art text-to-SQL algorithms did not further improve on popular benchmarks when trained with augmented synthetic data. We observed three shortcomings: illogical synthetic SQL queries from independent column sampling, arbitrary table joins, and language gaps between the synthesized SQL and natural language question (NLQ) pair. To address these issues, we propose a novel synthesis framework that imposes strong typing constraints, incorporates key relationships from schema, and conducts schema-distance-weighted column sampling. We also adopt an intermediate representation (IR) for the SQL-to-text task to further improve the quality of the generated NLQ. When existing powerful text-to-SQL parsers are pretrained on our high-quality synthesized data, these models have significant accuracy boosts and achieve new state-of-the-art performance on Spider. We also demonstrate the effectiveness of our techniques with ablation studies.
+
+# 1 Introduction
+
+Text-to-SQL parsing refers to the semantic parsing task that translates a natural language question (NLQ) to a corresponding SQL query. In recent decades, many industries have adopted high-level digitalization in their workflow and possessed large-scale datasets—many of which are stored as relational databases. Extracting insights from these relational databases to further drive business decisions is an important task. However, due to the complexity of these relational databases, query language experts are often needed to extract valuable insights. Thus a high-performing text-to-SQL system with a natural language interface would greatly
+
+lower the barrier for business users to query their databases.
+
+In order to obtain high-quality training data for the text-to-SQL parser, human annotators with SQL expertise are needed to construct NLQ-SQL parallel data, which are difficult and expensive to scale. Thus data scarcity is a well-known bottleneck in the text-to-SQL task (Yu et al., 2018b). To address the data scarcity issue, there is an increasing interest in leveraging synthetic data to improve the downstream performance. Yu et al. (2021) handcrafted high-quality rules to synthesize SQL and NLQ simultaneously, but these grammar rules need to be carefully designed through expensive manual work. To automate the synthesis process, recent attempts (Wang et al., 2021; Wu et al., 2021; Shi et al., 2021; Zhong et al., 2020) utilize a two-stage approach that synthesizes SQL first and then composes NLQ with a SQL-to-text generator. Alternatively, Yang et al. (2021) proposed a reversed pipeline that uses an entity-to-question model to generate natural language queries and then a text-to-SQL parser to generate SQL queries.
+
+In this paper, we delve into the two-stage synthesizing method that first synthesizes SQL queries and then generates NLQs. We first provide a comprehensive literature review and evaluate the contribution of existing synthetic datasets for the text-to-SQL task on popular benchmarks, such as Spider (Yu et al., 2018b). Results show that they only have negligible impact on boosting recent parsing models accuracy (e.g. T5 + PICARD proposed by Scholak et al. (2021)) when being augmented to the original training set. We analyze the quality of these synthetic datasets and find they suffer from three major issues, including illogical synthetic SQL queries from independent column sampling, arbitrary table joins, and language gaps between the synthesized SQL and NLQ pairs. We then propose a novel framework<sup>1</sup> aiming to reduce
+
+these problems present in existing methods. During the stage of SQL synthesis, we employ template synthesis with strong typing constraints, template key relationship preservation, and schema-distance-weighted column sampling. As for the SQL-to-text generation step, we adopt an intermediate representation (IR) to reduce the gap between SQL and target NLQ. We show that the top-performing text-to-SQL parsers can have significant accuracy improvements when being pretrained using our high-quality synthesized data and achieve new state-of-the-art performance on Spider.
+
+In summary, our main contributions are:
+
+- We systematically compare the existing text-to-SQL synthesis methods and identify three shortcomings that lead to the low quality;  
+- we propose three novel techniques for generating synthetic data and demonstrate its augmentation benefits for various text-to-SQL parers, underscoring the importance of high synthesis quality;  
+- we adopt an intermediate representation (IR) for the SQL-to-text task, which further improves the quality of the NLQ generation.
+
+# 2 Existing Synthesis Methods and Limitations
+
+We first conduct a detailed investigation towards the existing text-to-SQL synthesis frameworks to understand each of their advantages and shortcomings, the details of which can be found in Appendix A. In particular, Figure 5 summarizes and compares the key characteristics from different dimensions.
+
+In this section, we experiment with two recent synthetic datasets (Wang et al., 2021) and (Wu et al., 2021), and leverage the latest state-of-the-art text-to-SQL model T5 + PICARD (Scholak et al., 2021) to assess their effectiveness. We find that they only bring negligible impact on the performance when being used to train the parsing model. We then discuss three main shortcomings in these synthetic datasets based our manual inspection and analysis.
+
+# 2.1 Synthetic Data Effectiveness Assessment
+
+As a pilot study, we use T5-Large + PICARD as the text-to-SQL parser to examine the synthetic data quality. The baseline model is trained on Spider training set only. To add synthetic data during training, we setup a two-stage process. In Stage 1, the model is (pre-)trained using only the syn
+
+![](images/4ffbf79c20bc7b60edde1305fe27225284518bea6fa1975cf09046e34c56ff73.jpg)  
+Figure 1: Comparison of different synthetic data on Spider dev set performance. The baseline model is trained using Spider data only. Previous synthetic datasets (Wu et al., 2021; Wang et al., 2021) fail to bring additional performance boost to the model. Our synthetic data (SQL2NLQ and IR2NLQ) yield better results on both stages. (Stage 1 is training with synthetic data only and Stage 2 is finetuning with real Spider training set.)
+
+thetic data. In Stage 2, we finetune the model with the original Spider training set (also see Training Configurations in Section 4.1).
+
+As shown in Figure 1, the exact match (EM) accuracy of the models trained on both synthetic datasets (blue and light blue curves) are less than  $20\%$  during Stage 1, in contrast to  $60\%$  (green curve) when trained with real Spider data. This gap indicates the limited transferability from existing synthetic data to real data. Further finetuning on Spider training data in Stage 2 does not outperform the baseline model which is only trained using Spider data, indicating the ineffectiveness of these synthetic datasets. However, our synthetic data (red and gray curves for IR2NLQ and  $\mathrm{SQL2NLQ^2}$ ) can still yield better performance for both two stages. We will discuss in detail our methodologies in Section 3.
+
+# 2.2 Synthetic Data Quality Analysis
+
+We now provide an analysis towards the previous synthesis methodologies and identify three probable causes for obsolescence.
+
+# 2.2.1 Illogical Synthetic SQLs from Invalid Grammars or Templates
+
+Both Wang et al. (2021) and Wu et al. (2021) adopted context-free grammars (CFG) to generate SQL queries. The CFG designed by Wu et al. (2021) is constrained and they limited SQL generation to one table. While Wang et al. (2021) designed flexible grammars, they neglected
+
+the constraints between operators and column types. This negligence leads to mistakes such as SUM(student.name), where an aggregation operator is applied to a text column.
+
+Furthermore, CFG or probabilistic CFG (PCFG) generated SQL queries often fail to capture foreignkey and key relations between columns. This leads to invalid SQLs such as SELECT name, age FROM student INTERSECT SELECT address FROM teacher, where it intersects two sub-queries with different number of columns. In fact, designing a grammar to produce high coverage and logical SQLs is a difficult task due to the implicit dependencies of SQL elements.
+
+Alternatively, SQL templates extracted from training data better preserves column typing information (Zhong et al., 2020). This approach drastically reduces the invalid SQLs caused by a misalignment between operators and column types. However, existing work still misses the critical key relations in the templates.
+
+# 2.2.2 Over-Complex SQLs from Arbitrary Multi-table Joins
+
+When SQLs are materialized, the column/table selection from existing work is independent and results in SQL queries with unnecessary complexity. Those queries often have unclear intents and thus are difficult to be correctly translated to natural language questions. An example is presented in Table 2, where a simple template that requires only two columns can be turned into a complicated and nonsensical SQL query with three table joins.
+
+# 2.2.3 Language Gap between SQL and NLQ
+
+Recent work typically trains a sequence-to-sequence model to obtain corresponding NLQs from synthetic SQLs (Wang et al., 2021; Shi et al., 2021). The gap between SQL-NLQ pairs are well recognized in text-to-SQL task and intermediate representation (IR) is commonly used to reduce such mismatch (Gan et al., 2021b; Guo et al., 2019a; Yu et al., 2018a; Shi et al., 2021). However, the reverse of the source and target in SQL-to-text brings in its own challenge, such as incorrect references for SELECT *, missing conditions within long and complex SQL queries, and misinterpretation of ORDER phrases.
+
+# 3 Proposed Method
+
+Our synthesis framework builds on top of the template-based SQL synthesis approach similar to
+
+![](images/4a337a7a7b3ebedf1a8558210378c99cfae050161c00afdbc7134ff8e9cd0d4e.jpg)  
+Figure 2: Our NLQ-SQL synthesis framework. Novel components include strong-typing, key relations, schema-distance-weighted column sampler, and SQL  $\rightarrow$  IR converter.
+
+Zhong et al. (2020); Zhang et al. (2019) and generates SQL-NLQ pairs with a sequence-to-sequence model. Figure 2 provides a general overview of our pipeline. We develop the following techniques to address the issues in existing synthesis methods discussed in the previous sections:
+
+- for SQL synthesis:
+
+- introducing strong typing for columns w.r.t. semantic types and key properties,  
+- encoding the key relation with the extracted templates for more logical SQLs,  
+- proposing a schema-distance-weighted column sampling strategy to avoid overcomplex joins,
+
+- for NLQ synthesis:
+
+- designing an improved IR to bridge the gap between SQL and NLQ specifically for the SQL-to-text task.
+
+# 3.1 SQL Synthesis
+
+To create new SQLs on training data schemas, we utilize a template-based approach following Zhong et al. (2020): First, a pool of SQL templates are created by normalizing the schema-related mentions (column and value) and removing JOIN phrases. During SQL generation, a template is sampled based on the training distribution, and columns are sampled with constraints to fill in the normalized slots of the template. We highlight several improvements made to the existing approaches.
+
+# 3.1.1 Strong Typing
+
+When normalizing columns, we enforce strong typing of a template by enriching and preserving the data type (e.g., text, number, date, etc) as well as key identity (key or not) for each column. For example, in Table 1, we use textkey instead of key to normalize artist_name because operators such as MAX can be applied to number key but usually not to other text key.
+
+<table><tr><td rowspan="2">SQL</td><td>SELECT artist_name FROM song INTERSECT</td></tr><tr><td>SELECT artist_name FROM artist</td></tr><tr><td>Previous</td><td>SELECT col1_key INTERSECT col2_key</td></tr><tr><td>Ours</td><td>SELECT col1_textkey INTERSECT col2_textkey_fk1</td></tr></table>
+
+Table 1: Our modifications for template extraction: strong typing is highlighted in blue and key relation preservation is highlighted in pink.
+
+# 3.1.2 Template Key Relationship Preservation
+
+A foreign key is a column in a table referring to the primary key (unique identifier) of another table. In multiple table join scenarios, key and foreign key are the most common columns to be joined on. Restricting a column to be a foreign key to another key column is critical for a SQL to be valid especially in the following two cases: 1) queries including INTERSECT, EXCEPT, UNION, and 2) queries containing nested queries in WHERE conditions. For instance, the query in Table 1 implies the constraint that songartist_name should be a subset of artist.artist_name. In our template, FK1 captures the constraint of the key relationship between the two artist_name columns, which prevents the template from generating nonsensical queries such as SELECT gender FROM artist INTERSECT SELECT country FROM artist.
+
+# 3.1.3 Schema-distance-weighted Column Sampling
+
+To mitigate the issue of arbitrary multi-table joins, we implement a weighted sampling function biased toward columns that are close, in terms of table distance (defined below), to the columns already selected in a SQL template.
+
+For a given database  $d$ , we first establish an undirected graph for all the tables in  $d$ . Each table represents a node in the graph. The distance between any two tables,  $e(\cdot, \cdot)$ , is the least number of joins necessary to join the two tables (i.e. shortest path distance) under the restriction that table join can only take place with qualified primary key and foreign key pairs. See Appendix B for a detailed example demonstrating how table distances are computed.
+
+Define a template  $t$  as  $(q, \mathbf{c}, \mathbf{v})$ , where  $q$  is the flat template string,  $\mathbf{c} = [c_1, \ldots, c_m]$  is the set of column placeholders, and  $\mathbf{v} = [v_1, \ldots, v_n]$  is the set of value placeholders in  $q$ . Denote  $T_c$  to represent the table that contains column  $c$  and  $S_d(\tau)$  as the set of columns in  $d$  with the strong type  $\tau$ . Given a template  $t$  and a qualified database
+
+# Algorithm 1: Single SQL Synthesis with Schema-Weighted Column Sampling
+
+Input: template  $t = (q,\mathbf{c},\mathbf{v})$  database  $d$  decay rate  $\gamma$
+
+Output: SQL query  $y$
+
+1 Let  $y = q$  
+2 Random sample  $z_{1}$  from  $S_{d}(\tau_{c_{1}})$  and replace  $c_{1}$  with  $z_{1}$  in  $y$
+
+3 Compute sampling weights
+
+$$
+w (z) = \left\{ \begin{array}{l l} 1, & \text {i f} T _ {z} = T _ {c _ {1}} \\ \frac {1}{\gamma^ {\delta_ {c _ {1}}} (z)}, & \text {o . w .} \end{array} \right., \quad \forall z
+$$
+
+where  $\delta_c(z) = e(T_c,T_z)$
+
+4 for  $c\gets c_2:c_m$  do
+
+5 Compute sampling distribution
+
+$$
+p (z) = \left\{ \begin{array}{l l} \frac {w (z)}{\sum_ {z ^ {\prime} : \tau_ {z ^ {\prime}} = \tau_ {c}} w (z ^ {\prime})}, & \text {i f} \tau_ {z} = \tau_ {c} \\ 0, & \text {o . w .} \end{array} \right.
+$$
+
+6 Sample  $z$  from  $S_{d}(\tau_{c})$  with  $p$  
+7 Replace  $c$  with  $z$  in  $y$  
+8 Update sampling weights,  $\forall z$
+
+$$
+w (z) \leftarrow w (z) + \left\{ \begin{array}{l l} 1, & \text {i f} T _ {z} = T _ {c} \\ \frac {1}{\gamma^ {\delta_ {c} (z)}}, & \text {o . w .} \end{array} \right.,
+$$
+
+9 end
+
+10 for  $v\gets v_{1}:v_{n}$  do  
+11 Identify relevant columns w.r.t.  $v$  and retrieve a set of possible values for  $v$  from the  $d$  
+12 Random sample one value from the set and replace  $v$  with the value in  $y$
+
+13 end
+
+$d$ , the fundamental algorithm of SQL synthesis is described in Algorithm 1.
+
+The intuition behind the schema-weighted column sampling algorithm is as follows: after we select the first column for the given template (Line 2), we want to choose other columns in the database that are more relevant to the first column, so as to boost the chance of synthesizing more realistic SQL queries. We do so by sampling columns, for the remaining column placeholders in the template, according to a particular sampling probability (Line 5-6), which is a monotonically decreasing function of the column number  $n$ .
+
+<table><tr><td>Template</td><td>SELECT col1_numberkey WHERE col2_name = VALUE</td></tr><tr><td>Random</td><td>SELECT T1.CClub_ID FROM club AS T1 JOIN coach as T2
+ON T1.CClub_ID = T2.CClub_ID JOIN player_coach AS T3
+ON T2.Coach_ID = T3.Coach_ID JOIN player AS T4
+on T3 PLAYER_ID = T4 PLAYER_ID where T4.Rank = &quot;3rd&quot;</td></tr><tr><td>Ours</td><td>SELECT Club_ID FROM club WHERE Club_Name=&quot;AIK&quot;</td></tr></table>
+
+ing function of the edge value in the table graph for type-qualified column candidates (Line 3, 8), and 0 for non-qualified column candidate. Such implementation is motivated from the observation that over-lengthy SQLs resulted from multiple tables joins are rare in real world scenarios under the only-join-on-primary-key-foreign-key assumption. Table 2 shows an example of how adopting the schema-weighted sampling can help reduce the unrealistic SQLs in the random case.
+
+Value of  $\gamma$  in Algorithm 1.  $\gamma$  is a hyperparameter that controls the decay rate in the sampling probability for columns that are farther away from the columns that have already been selected. Under the restricted join condition, we look at the number of tables in a query as a proxy to the table distance. To determine the value of  $\gamma$ , we randomly sample 7000 synthetic SQL queries with replacement and calculate the average number of tables from the samples. We repeat this process for 1000 times and plot the distribution. Then we perform the same steps for the real Spider training data. We choose  $\gamma$  so that the distribution of the average number of tables in the synthetic data is close to the real data. This helps prevent generating over-simplified or over-complicated SQL queries.
+
+Based on this experiment, we chose  $\gamma$  to be 5 for the Spider benchmark. Figure 3 displays the distri
+
+![](images/3a9cede73d2c53474fded4b1c7df5438e16b59688fb4c67d551b91f98c0ad586.jpg)  
+Figure 3: Histogram of the average table count (i.e. number of joins) for three types of datasets with  $\gamma = 5$ . Our schema-distance-weighted column sampling reduces the table number of synthetic SQLs and better matches the training distribution.
+
+bution for three types of datasets: Spider training, synthetic dataset with schema-distance-weighted column sampling, and synthetic dataset with random column sampling. The figure demonstrates that the weighted sampling process, which provides an interface to tune the value of  $\gamma$ , can generate synthetic SQL queries that better match the real training data.
+
+# 3.2 NLQ Synthesis
+
+Intermediate representation (IR) has been employed to simplify the SQL query with minimum information loss (Gan et al., 2021a; Guo et al., 2019b; Gan et al., 2021b; Guo et al., 2019a; Yu et al., 2018a; Shi et al., 2021). Common operations include removing FROM/JOIN clauses and GROUP BY clauses, and merging WHERE clauses and HAVING clauses. Previous work found the use of IR often improves text-to-SQL performance.
+
+In this section, we explore whether the SQL-to-text generation could also benefit from an IR. According to a prior research by Wu et al. (2021), altering the query's linearization order could already affect the synthetic text quality. The objective of an IR here is to convert SQL to a representation that more closely resembles the NLQ. This conversion involves both simplifications (such as removal of redundant information) and specification (such as introducing information using heuristics).
+
+We outline the main new rules to transform SQLs into IRs and explain the rationale (examples in Table 3):
+
+Table 2: Random sampling vs our schema-distance-weighted column sampling for a given template. The former produced a query with three joins while ours have both columns from the same table.  
+
+<table><tr><td rowspan="3">EX1</td><td>SQL</td><td>SELECT T1.name FROM student AS T1 JOIN has.pet AS T2
+ON T1.student_id = T2.has.pet/student_id</td></tr><tr><td>IR</td><td>SELECT name of student FROM has.pet</td></tr><tr><td>NLQ</td><td>Find the name of students who have pets.</td></tr><tr><td rowspan="3">EX2</td><td>SQL</td><td>SELECT T2.name, count(*) FROM concert AS T1 JOIN stadium AS
+T2 ON T1.stadium_id = T2.stadium_id GROUP BY T1.stadium_id</td></tr><tr><td>IR</td><td>SELECT name of stadium, count (record of concert)
+GROUP BY (stadium_id of concert)</td></tr><tr><td>NLQ</td><td>Show the stadium name and the number of concerts in each stadium.</td></tr><tr><td rowspan="4">EX3</td><td>SQL</td><td>SELECT T1.neighbourhood_name neighbourhood AS T1 JOIN
+business AS T2 ON T1.business_id = T2.business_id</td></tr><tr><td>IR</td><td>WHERE T2.city = &quot;Madison&quot; GROUP BY T1.neighbourhood_name
+ORDER BY COUNT (DISTINCT T2.name) DESC LIMIT 1</td></tr><tr><td>IL</td><td>SELECT neighborhood_name of neighbourhood
+WITH most Count (DISTINCT name of business)
+WHERE city of business = &quot;Madison&quot;</td></tr><tr><td>NLQ</td><td>Which neighbourhood has the most number of businesses in Madison?</td></tr><tr><td rowspan="4">EX6</td><td>SQL</td><td>SELECT T2.name FROM USER AS T2 JOIN review AS T1
+ON T2.user_id = T1.user_id GROUP BY T2.name</td></tr><tr><td>IR</td><td>HAVING AVG (T1.rating) &lt; 3</td></tr><tr><td>IL</td><td>SELECT EACH (name of user) WITH Avg
+(rating of review) &lt; 3</td></tr><tr><td>NLQ</td><td>Find users whose average review rating is below 3.</td></tr></table>
+
+Table 3: IR examples that illustrate the examples of removing tables, enriching * columns, specifying most/least intent, removing redundant GROUP BY. Unwanted intents are in grey, redundant intents are in green. Texts related to IR operations are highlighted with yellow.
+
+- Only drop tables in the FROM/JOIN phrase if they appear in other SQL elements (EX2-EX4). Removal of tables can simplify queries but tables in JOIN can also behave as filters and need to be preserved to avoid information loss (EX1).  
+- Replace  $\star$  in count  $(\star)$  with the table whose columns in JOIN act as foreign key to provide explicit context for counting. This is because, in multi-table join queries, foreign key represents the many of the one-to-many relations and thus the rows from the table is more meaningful to be aggregated (see EX2 replaces  $\star$  with concert rather than stadium).  
+- When SQL contains ORDER BY COUNT (...), LIMIT ..., rewrite the query to explicitly express the most or least intent for better intent alignment (EX3).  
+- Drop GROUP BY phrase if the column grouped by appears in SELECT and attach EACH to the specific column if the query does not express the most/least intent (see GROUP dropped in EX3 - EX4 but not EX2). This aims to distinguish SQLs with GROUP BY and SELECT on the same column from those without SELECT.
+
+Similar to previous IR designs, we also removed repeated text in EXCEPT/INTERSECT/UNION queries and made lexical adjustments.
+
+# 4 Experiments
+
+We conduct experiments on the Spider benchmark and demonstrate the effectiveness of our synthesis framework for the text-to-SQL task.
+
+# 4.1 Experiment Setup
+
+Spider Benchmark Spider (Yu et al., 2018b) is a large-scale text-to-SQL dataset. It has 10,181 annotated questions, 5693 unique complex SQLs, and 200 databases with multiple tables. The train/train-others/dev/test sets contain 7000/1659/1034/2147 examples and 140/6/20/40 databases, respectively. Spider has a challenging and realistic evaluation setting, where SQL queries and databases do not appear across different splits, posing a generalization challenge for text-to-SQL semantic parser. Since Spider test set is not publicly available, we use dev set for evaluation and train-others for checkpoint selection.
+
+SQL-to-text Generator We finetune T5-large models on Spider training set for NLQ synthesis using both SQL and IR as input, named SQL2NLQ and IR2NLQ respectively. The best checkpoints are selected with the highest BLEU score on train-others.
+
+Text-to-SQL Parser We test our synthetic data with several top-performing text-to-SQL parsers, including T5-3B, RASAT, and T5-3B + PICARD. We use T5-3B (Raffel et al., 2020) as our base parser, since previous work (Shaw et al., 2021) has shown that T5-3B can achieve competitive performance for Text-to-SQL semantic parsing. RASAT leverages the advantage of T5 and integrates relational structures into the pretrained model. Recently, PICARD (Scholak et al., 2021) demonstrates that constraint decoding on top of T5-3B can produce state-of-the-art performance on Spider.
+
+Training Configurations We adopt a two-stage text-to-SQL training mechanism (Wang et al., 2021) in our experiment. In Stage 1, we use synthetic data only for model pre-finetuning. In Stage 2, we initialize the model weights with the first stage checkpoint, and then finetune it on the real data only. Both stages share the same hyperparameters. We train T5 with Adafactor and learning rate of 1e-4, and use gradient accumulation batch size 2050 for T5-3B model.
+
+# 4.2 Spider Results and Analysis
+
+The overall results<sup>3</sup> are shown in Table 4. We can see our synthetic data can further improve all three text-to-SQL parsing model, and achieve the best results with T5-3B + PICARD on Spider development set. Specifically, we have 4.4 points of EM score improvement on top of T5-3B model, while previous work (Wu et al., 2021; Wang et al., 2021) has marginal gain or even hurts the performance, demonstrating the effectiveness of our proposed method. More importantly, T5-3B has shown SOTA or near SOTA performance on 21 knowledge grounding tasks (Xie et al., 2022), our success of improving T5-3B with synthetic data for text-to-SQL can potentially generalize to other semantic parsing tasks with different logical forms. PICARD is an incremental parsing method for constraint decoding, which can reduce the syntax errors of language models for SQL generation. From Table 4,
+
+<table><tr><td>Model</td><td>EM</td><td>EX</td></tr><tr><td>DT-Fixup SQL-SP (Xu et al., 2021)</td><td>75.0</td><td>-</td></tr><tr><td>LGESQL + ELECTRA (Cao et al., 2021)</td><td>75.1</td><td>-</td></tr><tr><td>S2SQL + ELECTRA (Hui et al., 2022)</td><td>76.4</td><td>-</td></tr><tr><td>DT-Fixup + Syn (Yang et al., 2021)</td><td>76.4</td><td>-</td></tr><tr><td>T5-3B (Shaw et al., 2021)</td><td>70.0</td><td>-</td></tr><tr><td>T5-3B + Syn data (Wu et al., 2021)</td><td>69.1</td><td>-</td></tr><tr><td>T5-3B + Syn data (Wang et al., 2021)</td><td>70.3</td><td>-</td></tr><tr><td>T5-3B + Syn data (ours)</td><td>74.4</td><td>-</td></tr><tr><td>T5-3B + PICARD (Scholak et al., 2021)</td><td>74.1</td><td>-</td></tr><tr><td>T5-3B + PICARD + Syn data (ours)</td><td>76.9</td><td>-</td></tr><tr><td>SmBoP + GraPPa (Rubin and Berant, 2021)</td><td>69.5</td><td>71.1</td></tr><tr><td>GAP + NatSQL (Gan et al., 2021a)</td><td>73.7</td><td>75.0</td></tr><tr><td>RASAT (Qi et al., 2022)</td><td>72.6</td><td>76.6</td></tr><tr><td>RASAT + Syn data (ours)</td><td>74.4</td><td>78.8</td></tr><tr><td>T5-3B† (Scholak et al., 2021)</td><td>71.5</td><td>74.4</td></tr><tr><td>T5-3B† + Syn data (ours)</td><td>74.5</td><td>78.6</td></tr><tr><td>T5-3B† + PICARD (Scholak et al., 2021)</td><td>75.5</td><td>79.3</td></tr><tr><td>T5-3B† + PICARD + Syn data (ours)</td><td>76.1</td><td>81.4</td></tr></table>
+
+we see that T5-3B combined with PICARD and our synthetic data performs the best, implying the orthogonality of synthetic data augmentation and constraint decoding.
+
+We also submitted our model to the official Spider website for evaluation on the hidden test set and received 76.6 for EX and 73.1 for EM. Without developing new model architectures for the text-to-SQL task, we achieved best performance by only augmenting synthetic data on top of the training set, compared to all other non-anonymous evaluation submission. Since our approach is generic for the text-to-SQL task and model-independent, we can easily apply our framework to other model submissions for additional improvements.
+
+In Figure 1, we plot the training curves with different synthetic datasets. Compared with previous work (Wu et al., 2021; Wang et al., 2021), our synthetic data demonstrates significant improvement in Stage 1, with NLQ synthesized from both SQL2NLQ and IR2NLQ generator models, proving the high-quality of our synthesized SQLs. We also compare the generated NLQs with different automatic measurements in Table 5, where we can see IR benefits the NLQ generation process and produces the text closer to groundtruth NLQs.
+
+Table 4: Comparison of the top-performing text-to-SQL models in Spider leaderboard, as well as models trained with synthetic data (where synthetic are generated by training schema only). We report exact set match (EM) and execution accuracy (EX) for Spider dev set.  $\dagger$  means T5-3B is trained with database content. For each of the parsing model (RASAT, T5-3B, T5-3B + PICARD), we observe performance improvements when trained with our synthetic data.  
+
+<table><tr><td>Settings</td><td>BLEU</td><td>R-1</td><td>R-2</td><td>P-BERT</td><td>R-BERT</td></tr><tr><td>SQL→NLQ</td><td>27.7</td><td>59.6</td><td>35.3</td><td>93.6</td><td>93.2</td></tr><tr><td>SQL→IR→NLQ</td><td>29.3</td><td>60.5</td><td>36.8</td><td>93.9</td><td>93.3</td></tr></table>
+
+# 4.3 Ablation Study
+
+We conduct ablation study to understand the efficacy from each of the techniques we propose. In particular, we remove each technique and regenerate the synthetic data during Stage 1 training (using T5-3B) and evaluate on the Spider dev set. Results are summarized in Table 6.
+
+Table 5: Generated NLQ quality evaluations on the Spider dev set between SQL  $\rightarrow$  NLQ and SQL  $\rightarrow$  IR  $\rightarrow$  NLQ. The BLEU (Papineni et al., 2002), ROUGE (Lin, 2004), and P-BERT/R-BERT (Zhang et al., 2020) scores show that IR helps generate NLQs that are closer to the groundtruth.  
+
+<table><tr><td>Setting</td><td>EM</td><td>EX</td></tr><tr><td>All techniques included</td><td>74.5</td><td>78.6</td></tr><tr><td>w/o strong typing</td><td>72.3 (-2.2)</td><td>76.7 (-1.9)</td></tr><tr><td>w/o template key relation preserv</td><td>71.5 (-3.0)</td><td>77.2 (-1.4)</td></tr><tr><td>w/o schema-dist-weighted col sample</td><td>73.7 (-0.8)</td><td>76.6 (-2.0)</td></tr></table>
+
+Table 6: Ablation study on the three proposed synthesis techniques. We generate the same amount of synthetic data with each of the proposed technique removed and repeat the same training and evaluation process. Both EX and EM are worse on the dev set.
+
+We can see that removing any technique will result in a decrease in the final evaluation for the Spider dev set, indicating that each of the proposed three synthesis strategies contributes to a good quality synthetic data. In addition, Figure 1 shows that IR generated NLQ in general yields better results compared to SQL generated NLQ, which demonstrate the effectiveness of the proposed IR technique during NLQ synthesis.
+
+# 4.4 Discussions
+
+# Few-shot setting: How much real data do we need before achieving acceptable performance?
+
+Annotating text-to-SQL dataset takes extremely high human effort. In practice, it is hard to create a large-scale corpus with a limited annotation budget. Table 7 presents the text-to-SQL semantic parsing results with a limited number of training examples. We also generate our synthetic data on top of the corresponding subset. Interestingly, as training size decreases from 7K to 128, our synthetic data becomes more essential, and the performance gain increases from 4.4 points to 27.2 points. Even with only 512 training examples, our synthetic data can assist the T5-3B model to achieve  $\sim 60\%$  accuracy
+
+<table><tr><td>Model</td><td>f-shot: # tpl # syn</td><td>128 68 7839</td><td>256 116 10775</td><td>512 205 14457</td><td>1024 318 17002</td><td>full (7k) 746 21851</td></tr><tr><td rowspan="2">T5-3B</td><td>real only</td><td>19.1</td><td>32.3</td><td>43.6</td><td>53.2</td><td>70.0</td></tr><tr><td>real + syn</td><td>46.3</td><td>54.4</td><td>59.9</td><td>62.2</td><td>74.4</td></tr></table>
+
+Table 7: Text-to-SQL experiment with the few-shot setting, where we sampled a subset from the original Spider training set with size varying from 128 to 1024, then created synthetic data with templates only from the subset. # tpl and # syn represent the number of templates and synthesized NLQ-SQL pairs for the corresponding training subset. We report exact set match on the Spider dev set.
+
+level. These few-shot setting results are encouraging, as we can annotate a small-scale training set and still achieve acceptable performance with the help of synthetic data.
+
+Generator size: How big of the generator model do we need to produce high-quality NLQs? Since our proposed IR is to reduce the gap between NLQs and SQLs, we hypothesize that the NLQ generation process should have less reliance on model size. As shown in Figure 4, our synthetic data (with IR2NLQ) still presents comparable performance even with a small size T5-base generator, implying the effectiveness and robustness of our proposed IR. As comparison, SQL2NLQ has larger divergence between T5-Large and T5-Base, indicating some difficulty of translating SQL to NLQ.
+
+![](images/4b96a84337d6a13fc237670cd011fe3944df1fcd8f93c8e095018718ffceea73.jpg)  
+Figure 4: Comparison of different T5 model sizes for NLQ generations. On top of T5-Base (220M parameters) and T5-Large (770M parameters), we finetune generators for both SQL-to-text and IR-to-text, then evaluate the effectiveness with text-to-SQL semantic parsing in Spider.
+
+Seen schema: How good of the synthetic data if we consider a broader coverage of database schema? Since the cross-database evaluation setting presents generalization challenge for text-to-SQL parsers, our synthetic framework can potentially overcome this by utilizing more public database schemas, or even ones that can implicitly cover the evaluation set. In addition to using schema from training set, we can take advantage
+
+of more public schemas for data synthesis, for example, WikiTables (Bhagavatula et al., 2015), GitTables (Hulsebos et al., 2021), WikiSQL (Zhong et al., 2017) and SQL tutorial websites. Some of them are even schema source for Spider benchmark. We simply added 20 databases from dev set into our synthetic data generation, then trained text-to-SQL parser on top of T5-Large. With this setting, we observed  $\sim 2$  points of performance improvement compared to that with training schema only. This pilot study implies the potential helpfulness of synthesizing data with targeting database schemas to further improve the downstream performance. While this setting breaks the cross-schema setting, we believe it still has practical values.
+
+Single-table: How effective is our method on the single-table text-to-SQL parsing? Although our SQL synthesis is mainly designed for multtable operations, it should also be compatible with the single table, but with foreign key preservation ineffective. WikiSQL (Zhong et al., 2017) and SQUALL (Shi et al., 2020) are two popular datasets for single-table text-to-SQL parsing. Compared to multi-table case, the single-table is much easier, for example, most text-to-SQL parsers are above  $90\%$  accuracy level in WikiSQL<sup>4</sup>. We took a relatively challenging SQUALL dataset for experiment. From 9K training examples, we created 30K synthetic NLQ-SQL pairs. We observe a smaller performance gain and hypothesize several reasons: 1) foreign key relationship is not applicable in single table, but critical to our data synthesis framework (as shown in the ablation study); 2) 9k examples are sufficient for model training, especially for SQLs without JOIN clause. Therefore, the effect of synthetic data is further diluted.
+
+# 5 Conclusion
+
+In this work, we proposed a data synthesis framework for text-to-SQL semantic parsing. After incorporating key relationships from schema, imposing strong typing, conducting schemadistance-weighted column sampling, and bridging SQL  $\rightarrow$  NLQ generation with an intermediate representation, we synthesized high-quality dataset that can further improve the state-of-the-art parser on Spider benchmark. We also revealed the efficiency of the synthetic data and pointed out the potential usefulness of reducing human annotations for text-to-SQL parsing.
+
+# Limitations
+
+The proposed synthesis framework has been targeted at text-to-SQL task, which may not generalize to other tasks that require large amount of synthetic data without major modification. For instance, other popular tasks involving converting natural language questions to some sort of logic forms are in natural very similar to text-to-SQL, yet all techniques relying on the "key" property in the database might no longer be applicable. On the other hand, the template based synthesis method currently relies on templates extracted from the real data. By incorporating some carefully designed grammar (e.g. PCFG), we may be able to further enrich the template set.
+
+# Ethics Statement
+
+The training and evaluation of our experiments rely on many compute resources, which may not be environment-friendly. For example, each parsing model requires training using NVIDIA A100-SXM4-40GB GPUs for many hours, which can inevitably cause more CO2 emission.
+
+# References
+
+Chandra Bhagavatula, Thanapon Noraset, and Doug Downey. 2015. Tabel: Entity linking in web tables. In SEMWEB.  
+Ruisheng Cao, Lu Chen, Zhi Chen, Yanbin Zhao, Su Zhu, and Kai Yu. 2021. LGESQL: Line graph enhanced text-to-SQL model with mixed local and non-local relations. 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 2541-2555, Online. Association for Computational Linguistics.  
+Yujuan Gan, Xinyun Chen, Jinxia Xie, Matthew Purver, John R. Woodward, John Drake, and Qiaofu Zhang. 2021a. Natural SQL: Making SQL easier to infer from natural language specifications. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2030-2042, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Yujuan Gan, Xinyun Chen, Jinxia Xie, Matthew Purver, John R. Woodward, John H. Drake, and Qiaofu Zhang. 2021b. Natural SQL: making SQL easier to infer from natural language specifications. CoRR, abs/2109.05153.  
+Daya Guo, Yibo Sun, Duyu Tang, Nan Duan, Jian Yin, Hong Chi, James Cao, Peng Chen, and Ming Zhou.
+
+2018. Question generation from SQL queries improves neural semantic parsing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1597-1607, Brussels, Belgium. Association for Computational Linguistics.  
+Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, Jian-Guang Lou, Ting Liu, and Dongmei Zhang. 2019a. Towards complex text-to-sql in cross-domain database with intermediate representation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 4524-4535. Association for Computational Linguistics.  
+Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, Jian-Guang Lou, Ting Liu, and Dongmei Zhang. 2019b. Towards complex text-to-SQL in cross-domain database with intermediate representation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4524-4535, Florence, Italy. Association for Computational Linguistics.  
+Binyuan Hui, Ruiying Geng, Lihan Wang, Bowen Qin, Yanyang Li, Bowen Li, Jian Sun, and Yongbin Li. 2022. S²SQL: Injecting syntax to question-schema interaction graph encoder for text-to-SQL parsers. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1254–1262, Dublin, Ireland. Association for Computational Linguistics.  
+Madelon Hulsebos, Căgatay Demiralp, and Paul Groth. 2021. Gittables: A large-scale corpus of relational tables. arXiv preprint arXiv:2106.07258.  
+Robin Jia and Percy Liang. 2016. Data recombination for neural semantic parsing. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12-22, Berlin, Germany. Association for Computational Linguistics.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+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.  
+Jiexing Qi, Jingyao Tang, Ziwei He, Xiangpeng Wan, Chenghu Zhou, Xinbing Wang, Quanshi Zhang, and Zhouhan Lin. 2022. Rasat: Integrating relational structures into pretrained seq2seq model for text-to-sql.
+
+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.  
+Ohad Rubin and Jonathan Berant. 2021. SmBoP: Semi-autoregressive bottom-up semantic parsing. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 311-324, Online. Association for Computational Linguistics.  
+Torsten Scholak, Nathan Schucher, and Dzmitry Bahdanau. 2021. PICARD: Parsing incrementally for constrained auto-regressive decoding from language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.  
+Peter Shaw, Ming-Wei Chang, Panupong Pasupat, and Kristina Toutanova. 2021. Compositional generalization and natural language variation: Can a semantic parsing approach handle both? 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 922-938, Online. Association for Computational Linguistics.  
+Peng Shi, Patrick Ng, Zhiguo Wang, Henghui Zhu, Alexander Hanbo Li, Jun Wang, Cicero Nogueira dos Santos, and Bing Xiang. 2021. Learning contextual representations for semantic parsing with generation-augmented pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13806-13814.  
+Tianze Shi, Chen Zhao, Jordan Boyd-Graber, Hal Daumé III, and Lillian Lee. 2020. On the potential of lexico-logical alignments for semantic parsing to SQL queries. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1849–1864, Online. Association for Computational Linguistics.  
+Bailin Wang, Wenpeng Yin, Xi Victoria Lin, and Caiming Xiong. 2021. Learning to synthesize data for semantic parsing. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2760-2766, Online. Association for Computational Linguistics.  
+Kun Wu, Lijie Wang, Zhenghua Li, Ao Zhang, Xinyan Xiao, Hua Wu, Min Zhang, and Haifeng Wang. 2021. Data augmentation with hierarchical SQL-to-question generation for cross-domain text-to-SQL parsing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8974-8983, Online and Punta Cana, Dominican Republic. 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. Unifiedsgk: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. arXiv preprint arXiv:2201.05966.  
+Peng Xu, Dhruv Kumar, Wei Yang, Wenjie Zi, Keyi Tang, Chenyang Huang, Jackie Chi Kit Cheung, Simon J.D. Prince, and Yanshuai Cao. 2021. Optimizing deeper transformers on small datasets. 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 2089-2102, Online. Association for Computational Linguistics.  
+Wei Yang, Peng Xu, and Yanshuai Cao. 2021. Hierarchical neural data synthesis for semantic parsing. arXiv preprint arXiv:2112.02212.  
+Tao Yu, Chien-Sheng Wu, Xi Victoria Lin, Bailin Wang, Yi Chern Tan, Xinyi Yang, Dragomir Radev, Richard Socher, and Caiming Xiong. 2021. Grappa: Grammar-augmented pre-training for table semantic parsing. In International Conference on Learning Representations.  
+Tao Yu, Michihiro Yasunaga, Kai Yang, Rui Zhang, Dongxu Wang, Zifan Li, and Dragomir R. Radev. 2018a. Syntaxsqlnet: Syntax tree networks for complex and cross-domain text-to-sql task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 1653-1663. Association for Computational Linguistics.  
+Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018b. Spider: A largescale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3911-3921, Brussels, Belgium. Association for Computational Linguistics.  
+Rui Zhang, Tao Yu, Heyang Er, Sungrok Shim, Eric Xue, Xi Victoria Lin, Tianze Shi, Caiming Xiong, Richard Socher, and Dragomir Radev. 2019. Editing-based SQL query generation for cross-domain context-dependent questions. 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 5338-5349, Hong Kong, China. 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.  
+Victor Zhong, Mike Lewis, Sida I. Wang, and Luke Zettlemoyer. 2020. Grounded adaptation for zero-shot executable semantic parsing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6869-6882, Online. Association for Computational Linguistics.  
+Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2sql: Generating structured queries from natural language using reinforcement learning. CoRR, abs/1709.00103.
+
+# A Existing Synthesis Frameworks
+
+Inspired by prior work by Jia and Liang (2016) in semantic parsing, Yu et al. (2021) extended a synchronous context-free grammar (SCFG) approach to the text-to-SQL task where they manually crafted about 90 high-quality SQL-NLQ aligned patterns to generate new SQL-NLQ pairs. They found pretraining on the synthetic dataset leads to a significant improvement even tested with a very strong text-to-SQL parser RAT-SQL on the Spider benchmark.
+
+While SCFG usually creates high-quality data because patterns are carefully designed and aligned, the coverage of the patterns is limited, and expert knowledge is required to design such patterns. Thus, more efforts are devoted to automating the procedure. Guo et al. (2018) utilized a two-stage approach by first sampling SQL queries from a simple pattern and then generating questions using a copy-based RNN encoder-decoder structure find the synthetic data that can improve the existing state-of-the-art model on the WikiSQL benchmark. Zhong et al. (2020) followed the same two-stage approach but used templates extracted from training to generate SQL and augmented the NLQ generator with pretrained transformer BERT and iteratively updated the parser and generator. Only the synthetic dataset that was created using target schemas filtered with cycle consistency can facilitate the downstream performance.
+
+Along the same approach, Wang et al. (2021) identified problems with fixed SQL synthesis rules and employed a full-fledged probabilistic context-free grammar (PCFG) that enabled generating SQLs with varying structures. They synthesized natural language queries with a BART SQL-NLQ generator. Their synthesis method has been shown to boost the RAT-SQL parser performance on the Spider benchmark, though the improvement is not as significant as pretraining using SCFG generated synthetic data (Yu et al., 2021). The gap might be due to the quality of the synthetic dataset as the independent selection of generation step in PCFG introduces substantial noise such as illogical SQL queries.
+
+To improve the quality of synthetic data, Wu et al. (2021) introduced a clause-level synthesis framework: first decomposing a query into sub-clauses and translating sub-SQL clauses into sub-questions, and finally assembling sub-questions into a whole question. They found clause-based synthesis method is better than flat synthesis.
+
+Alternatively, Yang et al. (2021) proposed to improve the quality of synthetic data by incorporating domain information in question generation. Specifically, they learned an entity sampler and synthesized questions using an entity-to-question generator with entities sampled from the sampler, followed by generating pairing SQL queries through a baseline parser. For this approach, they also attractively updated the parser and generator, in a similar fashion as in Zhong et al. (2020). Their synthetic dataset can significantly improve a DT-Fixup parser on the Spider benchmark.
+
+This work seeks to investigate value of synthetic dataset with current state-of-the-art PICARD model and refine a synthetic method in an automate and non-iterative manner. Thus, we examine two synthetic datasets from recent work (Wang et al., 2021; Wu et al., 2021) that demonstrate improvement of downstream performance with previous state-of-the-art text-to-SQL parser (RAT-SQL) over Spider benchmark without iterative training.
+
+<table><tr><td rowspan="2">Paper</td><td rowspan="2">Method</td><td colspan="2">SQL Synthesis</td><td colspan="2">NLQ Synthesis</td><td rowspan="2">SQL-NLQ 
+Bridging</td><td rowspan="2">Manual 
+Effort</td></tr><tr><td>Abstraction</td><td>Limitation</td><td>Procedure</td><td>Generator</td></tr><tr><td>Guo et al 
+(2018)</td><td>Two stage</td><td>Template</td><td>Single template, no JOIN</td><td>SQL → NLQ</td><td>copy-based 
+RNN</td><td>-</td><td>minimal</td></tr><tr><td>GAZP 
+(Zhong et al 
+2020)</td><td>Iterative 
+two stage</td><td>Template</td><td>Violating foreign key 
+relations, limit to training 
+templates</td><td>SQL → NLQ</td><td>BERT + 
+point 
+decoder</td><td>-</td><td>minimal</td></tr><tr><td>Wang et al 
+(2021)</td><td>Two stage</td><td>PCFG</td><td>OP/COL incompatibility, 
+invalid SQL structure</td><td>SQL → NLQ</td><td>BART</td><td>-</td><td>minimal</td></tr><tr><td>Wu et al 
+(2021)</td><td>Two stage</td><td>CFG</td><td>No support for IEU, 
+no JOIN</td><td>SQL → Sub-SQL 
+→ NLQ fragment 
+→NLQ</td><td>copy-based 
+RNN</td><td>SQL clause 
+/ NLQ 
+fragment</td><td>combination 
+rules</td></tr><tr><td>Yang et al 
+(2021)</td><td>Iterative 
+reversed 
+two stage</td><td>-</td><td>Dependent on base parser</td><td>schema → entity 
+→ NLQ</td><td>T5</td><td>-</td><td>minimal</td></tr><tr><td>Grappa 
+(Yu et al 2021)</td><td>Synchronous</td><td>Template</td><td>Limit to training templates</td><td colspan="2">simultaneous instantiation of 
+SQL-NLQ template</td><td>Aligned 
+SQL-NLQ 
+template</td><td>alignment</td></tr><tr><td>Ours</td><td>Two stage</td><td>Template</td><td>Limit to training templates</td><td>SQL→ IR→ NLQ</td><td>T5</td><td>IR</td><td>minimal</td></tr></table>
+
+Figure 5: Comparison of different data synthesis methods for text-to-SQL task. Synchronous refers to generating SQL and NLQ together, Two-stage first synthesizes SQL then generates NLQ, reversed two-stage first generates NLQ then synthesizes SQL. SQL-NLQ Bridging refers to intermediate operations or representations for matching SQL and NLQ.
+
+# B Details on Schema-distance-weighted Column Sampling
+
+# B.1 Table Distance.
+
+For a given database  $d$ , we first establish an undirected graph for all the tables in  $d$ . We can then compute the distance between any two tables,  $e(\cdot, \cdot)$ , defined as the least number of joins necessary to join the two tables under the restriction that table join can only take place with qualified primary key and foreign key information. In other words, we disable arbitrary join of two tables if they lack key and foreign key relationship.
+
+We give some examples using one of the databases (id: college_1) in the Spider benchmark, as shown in Table 8.
+
+-  $e(\mathrm{T}1, \mathrm{T}2) = 1$  because the column class code in table class (T1) is a foreign key in table course (T2). We can also observe from the table graph in Figure 6: there is a direct path between table node class and table node course.  
+-  $e(\mathrm{T}2, \mathrm{T}7) = 2$  since we first need to join table course (T2) with table department (T3), followed by joining table department with table student (T7). Note that even though we can also join using the path  $T2 \to T1 \to T5 \to T7$ , this is not the least number of joins between the two tables.
+
+Table 8: Example database (id: college_1)  
+
+<table><tr><td rowspan="2">Alias</td><td rowspan="2">Table Name</td><td rowspan="2">Primary Key</td><td colspan="2">Foreign Key</td></tr><tr><td>Table</td><td>Column</td></tr><tr><td>T1</td><td>class</td><td>class code</td><td>enroll</td><td>class code</td></tr><tr><td>T2</td><td>course</td><td>course code</td><td>class</td><td>class code</td></tr><tr><td rowspan="3">T3</td><td rowspan="3">department</td><td rowspan="3">department code</td><td>course</td><td>department code</td></tr><tr><td>professor</td><td>department code</td></tr><tr><td>student</td><td>department code</td></tr><tr><td rowspan="3">T4</td><td rowspan="3">employee</td><td rowspan="3">employee number</td><td>class</td><td>professor employee number</td></tr><tr><td>department</td><td>employee number</td></tr><tr><td>professor</td><td>employee number</td></tr><tr><td>T5</td><td>enroll</td><td>-</td><td>-</td><td>-</td></tr><tr><td>T6</td><td>professor</td><td>-</td><td>-</td><td>-</td></tr><tr><td>T7</td><td>student</td><td>student num</td><td>enroll</td><td>student number</td></tr></table>
+
+![](images/87f6708872560bf1048512dae466de1736488c11cfd4e7452caebefcc7d3eeb2.jpg)  
+Figure 6: Example table graph (id: college_1)
+
+The reason we introduce the concept of table distance is that we want to leverage this value to promote table joins with appropriate relationships while discouraging illogical joins when two tables are irrelevant. During the process of column sampling, we will choose columns that have smaller table distance with the other columns that have already been selected with the objective to create more realistic synthetic SQL queries. In the example above, assume we have first sampled a column from the table student (T7). For the next column placeholder, we are more likely to sample a column from table enroll (T5) than table professor (T6) — it is more natural to ask questions like "how many students enrolled in class X" compared to asking "how many students enrolled in classes taught by professors who were employed before year YYYYY".
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+See the Limitations section.
+
+A2. Did you discuss any potential risks of your work?
+
+See the Ethics Statement section.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+See Abstract and Section 1 Introduction.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+We used the Spider data (Section 2 and Section 4)
+
+B1. Did you cite the creators of artifacts you used?
+
+Section 2, 4 and References.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Spider is one of the top used datasets for text-to-sql tasks and have been accepted by the community for scientific research.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Not applicable. Left blank.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Not applicable. Left blank.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Not applicable. Left blank.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+See Section 4.1.
+
+C Did you run computational experiments?
+
+Section 4.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Ethics Statement and Section 4.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Section 4.
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+Not applicable. Left blank.
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+Section 4.
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+Not applicable. Left blank.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+Not applicable. Left blank.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+Not applicable. Left blank.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+Not applicable. Left blank.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+Not applicable. Left blank.

2023/Importance of Synthesizing High-quality Data for Text-to-SQL Parsing/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6dad0a6a73224cb0c5b2905f854f02c235d8c981f1c4462b3415f8cae080a2c
+size 530988

2023/Improved Logical Reasoning of Language Models via Differentiable Symbolic Programming/a3b32712-0fc6-409b-b027-c0f54131f88f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ebeeff79471c1fc75f8546a7bd1c5dce7882de035b54f5e786a38d67a3f9a30
+size 671148

2023/Improved Logical Reasoning of Language Models via Differentiable Symbolic Programming/full.md

@@ -0,0 +1,421 @@
+# Improved Logical Reasoning of Language Models via Differentiable Symbolic Programming
+
+Hanlin Zhang $^{1,*}$  Jiani Huang $^{2,*}$  Ziyang Li $^{2}$  Mayur Naik $^{2}$  Eric Xing $^{1,3,4}$
+
+$^{1}$ Carnegie Mellon University,  $^{2}$ University of Pennsylvania,
+
+<sup>3</sup>Mohamed Bin Zayed University of Artificial Intelligence, <sup>4</sup>Petuum Inc.
+
+# Abstract
+
+Pre-trained large language models (LMs) struggle to perform logical reasoning reliably despite advances in scale and compositionality. In this work, we tackle this challenge through the lens of symbolic programming. We propose DSR-LM, a Differentiable Symbolic Reasoning framework where pre-trained LMs govern the perception of factual knowledge, and a symbolic module performs deductive reasoning. In contrast to works that rely on hand-crafted logic rules, our differentiable symbolic reasoning framework efficiently learns weighted rules and applies semantic loss to further improve LMs. DSR-LM is scalable, interpretable, and allows easy integration of prior knowledge, thereby supporting extensive symbolic programming to robustly derive a logical conclusion. The results of our experiments suggest that DSR-LM improves the logical reasoning abilities of pre-trained language models, resulting in a significant increase in accuracy of over  $20\%$  on deductive reasoning benchmarks. Furthermore, DSR-LM outperforms a variety of competitive baselines when faced with systematic changes in sequence length. $^{1}$
+
+# 1 Introduction
+
+Complex applications in natural language processing involve dealing with two separate challenges. On one hand, there is the richness, nuances, and extensive vocabulary of natural language. On the other hand, one needs logical connectives, long reasoning chains, and domain-specific knowledge to draw logical conclusions. The systems handling these two challenges are complementary to each other and are likened to psychologist Daniel Kahneman's human "system 1" and "system 2" (Kahneman, 2011): while the former makes fast and intuitive decisions, akin to neural networks, the latter
+
+<table><tr><td>Language Model</td><td>Symbolic Reasoner</td></tr><tr><td>·Rapid reasoning
+·Sub-symbolic knowledge
+·Handling noise, ambiguities, and naturalness
+·Process open domain text
+·Can learn in-context</td><td>·Multi-hop reasoning
+·Compositionality
+·Interpretability
+·Data efficiency
+·Can incorporate domain-specific knowledge</td></tr></table>
+
+Table 1: Respective advantages of language models and symbolic reasoners.
+
+thinks more rigorously and methodically. Considering LMs as "system 1" and symbolic reasoners as "system 2", we summarize their respective advantages in Table 1.
+
+Although pre-trained LMs have demonstrated remarkable predictive performance, making them an effective "system 1", they fall short when asked to perform consistent logical reasoning (Kassner et al., 2020; Helwe et al., 2021; Creswell et al., 2022), which usually requires "system 2". In part, this is because LMs largely lack capabilities of systematic generalization (Elazar et al., 2021; Hase et al., 2021; Valmeekam et al., 2022).
+
+In this work, we seek to incorporate deductive logical reasoning with LMs. Our approach has the same key objectives as neuro-symbolic programming (Chaudhuri et al., 2021): compositionality, consistency, interpretability, and easy integration of prior knowledge. We present DSR-LM, which tightly integrates a differentiable symbolic reasoning module with pre-trained LMs in an end-to-end fashion. With DSR-LM, the underlying LMs govern the perception of natural language and are finetuned to extract relational triplets with only weak supervision. To overcome a common limitation of symbolic reasoning systems, the reliance on human-crafted logic rules (Huang et al., 2021; Nye et al., 2021), we adapt DSR-LM to induce and finetune rules automatically. Further, DSR-LM allows incorporation of semantic loss obtained by logical integrity constraints given as prior knowledge,
+
+which substantially helps the robustness.
+
+We conduct extensive experiments showing that DSR-LM can consistently improve the logical reasoning capability upon pre-trained LMs. Even if DSR-LM uses a RoBERTa backbone with much less parameters and does not explicitly take triplets as supervision, it can still outperform various baselines by large margins. Moreover, we show that DSR-LM can induce logic rules that are amenable to human understanding to explain decisions given only higher-order predicates. As generalization over long-range dependencies is a significant weakness of transformer-based language models (Lake and Baroni, 2018; Tay et al., 2020), we highlight that in systematic, long-context scenarios, where most pre-trained or neural approaches fail to generalize compositionally, DSR-LM can still achieve considerable performance gains.
+
+# 2 Related Work
+
+Logical reasoning with LMs. Pre-trained LMs have been shown to struggle with logical reasoning over factual knowledge (Kassner et al., 2020; Helwe et al., 2021; Talmor et al., 2020a). There is encouraging recent progress in using transformers for reasoning tasks (Zhou et al., 2020; Clark et al., 2021; Wei et al., 2022; Chowdhery et al., 2022; Zelikman et al., 2022) but these approaches usually require a significant amount of computation for re-training or human annotations on reasoning provenance (Camburu et al., 2018; Zhou et al., 2020; Nye et al., 2021; Wei et al., 2022). Moreover, their entangled nature with natural language makes it fundamentally hard to achieve robust inference over factual knowledge (Greff et al., 2020; Saparov and He, 2022; Zhang et al., 2022).
+
+There are other obvious remedies for LMs' poor reasoning capability. Ensuring that the training corpus contains a sufficient amount of exemplary episodes of sound reasoning reduces the dependency on normative biases and annotation artifacts (Talmor et al., 2020b; Betz et al., 2020; Hase et al., 2021). Heuristics like data augmentation are also shown to be effective (Talmor et al., 2020b). But the above works require significant efforts for crowdsourcing and auditing training data. Our method handily encodes a few prototypes/templates of logic rules and is thus more efficient in terms of human effort. Moreover, our goal is fundamentally different from theirs in investigating the tight integration of neural and symbolic models in an end-to-end manner.
+
+Neuro-symbolic reasoning. Neuro-symbolic approaches are proposed to integrate the perception of deep neural components and the reasoning of symbolic components. Representative works can be briefly categorized into regularization (Xu et al., 2018), program synthesis (Mao et al., 2018), and proof-guided probabilistic programming (Evans and Grefenstette, 2018; Rocktäschel and Riedel, 2017; Manhaeve et al., 2018; Zhang et al., 2019; Huang et al., 2021). To improve compositionality of LMs, previous works propose to parameterize grammatical rules (Kim, 2021; Shaw et al., 2021) but show that those hybrid models are inefficient and usually underperform neural counterparts. In contrast to the above works, DSR-LM focuses on improving LMs' reasoning over logical propositions with tight integration of their pre-trained knowledge in a scalable and automated way.
+
+# 3 Methodology
+
+# 3.1 Problem Formulation
+
+Each question answering (QA) example in the dataset is a triplet containing input text  $x$ , query  $q$ , and the answer  $y$ . Figure 1 shows an instance that we will use as our running example. The input text  $x$  is a natural language passage within which there will be a set of entities, possibly referenced by 3rd person pronouns. The sentences hint at the relationships between entities. For example, "Dorothy went to her brother Rich's birthday party" implies that Rich is Dorothy's brother and Dorothy is Rich's sister. The query  $q$  is a tuple of two entities, representing the people with whom we want to infer the relation. The expected relation is stored in the answer  $y$ , which will be one of a confined set of possible relations  $\mathcal{R}$ , allowing us to treat the whole problem as an  $|\mathcal{R}|$ -way classification problem. We focus only on the problems where the desired relation is not explicitly stated in the context but need to be deduced through a sequence of reasoning.
+
+# 3.2 Methodology Overview
+
+The design of DSR-LM concerns tightly integrating a perceptive model for relation extraction with a symbolic engine for logical reasoning. While we apply LMs for low-level perception and relation extraction, we employ a symbolic reasoning module to consistently and logically reason about the extracted relations. With a recent surge in neuro-symbolic methods, reasoning engines are made differentiable, allowing us to differentiate through
+
+![](images/379cd4cbe60883df6dd5990227187abd721009a8eecfa3338679ec44f52fbef7.jpg)  
+Figure 1: Overview of DSR-LM with a motivating example where "Anne is the niece of Dorothy" should be logically inferred from the context. We abbreviate the names with their first initials in the relational symbols.
+
+the logical reasoning process. In particular, we employ Scallop (Huang et al., 2021) as our reasoning engine. We propose two add-ons to the existing neuro-symbolic methodology. First, some rules used for logical deduction are initialized using language models and further tuned by our end-to-end pipeline, alleviating human efforts. Secondly, we employ integrity constraints on the extracted relation graphs and the logical rules, to improve the logical consistency of LMs and the learned rules.
+
+Based on this design, we formalize our method as follows. We adopt pretrained LMs to build relation extractors, denoted  $\mathcal{M}_{\theta}$ , which take in the natural language input  $x$  and return a set of probabilistic relational symbols  $\mathbf{r}$ . Next, we employ a differentiable deductive reasoning program,  $\mathcal{P}_{\phi}$  where  $\phi$  represents the weights of the learned logic rules. It takes as input the probabilistic relational symbols and the query  $q$  and returns a distribution over  $\mathcal{R}$  as the output  $\hat{y}$ . Overall, the deductive model is written as
+
+$$
+\hat {y} = \mathcal {P} _ {\phi} (\mathcal {M} _ {\theta} (x), q). \tag {1}
+$$
+
+Additionally, we have the semantic loss (s1) derived by another symbolic program  $\mathcal{P}_{\mathrm{s1}}$  computing the probability of violating the integrity constraints:
+
+$$
+l _ {\mathrm {s l}} = \mathcal {P} _ {\mathrm {s l}} \left(\mathcal {M} _ {\theta} (x), \phi\right) \tag {2}
+$$
+
+Combined, we aim to minimize the objective  $J$  over training set  $\mathcal{D}$  with loss function  $\mathcal{L}$ :
+
+$$
+\begin{array}{l} J (\theta , \phi) = \frac {1}{| \mathcal {D} |} \sum_ {(x, q, y) \in \mathcal {D}} w _ {1} \mathcal {L} (\mathcal {P} _ {\phi} (\mathcal {M} _ {\theta} (x), q), y) \\ + w _ {2} \mathcal {P} _ {\mathrm {s l}} \left(\mathcal {M} _ {\theta} (x), \phi\right), \tag {3} \\ \end{array}
+$$
+
+where  $w_{1}$  and  $w_{2}$  are tunable hyper-parameters to balance the deduction loss and semantic loss.
+
+# 3.3 Relation Extraction
+
+Since pre-trained LMs have strong pattern recognition capabilities for tasks like Named-EntityRecognition (NER) and Relation Extraction (RE) (Tenney et al., 2019; Soares et al., 2019), we adopt them as our neural components in DSR-LM. To ensure that LMs take in strings of similar length, we divide the whole context into multiple windows. The goal is to extract the relations between every pair of entities in each windowed context. Concretely, our relation extractor  $\mathcal{M}_{\theta}$  comprises three components: 1) a Named-Entity Recognizer (NER) to obtain the entities in the input text, 2) a pretrained language model, to be fine-tuned, that converts windowed text into embeddings, and 3) a classifier that takes in the embedding of entities and predicts the relationship between them. The set of parameters  $\theta$  contains the parameters of both the LM and the classifier.
+
+We assume the relations to be classified come from a finite set of relations  $\mathcal{R}$ . For example in CLUTRR (Sinha et al., 2019), we have 20 kinship relations including mother, son, uncle, father-in-law, etc. In practice, we perform  $(|\mathcal{R}| + 1)$ -way classification over each pair of entities, where the extra class stands for "n/a". The windowed contexts are split based on simple heuristics of "contiguous one to three sentences that contain at least two entities", to account for coreference resolution. The windowed contexts can be overlapping and we allow the reasoning module to deal with noisy and redundant data. Overall, assuming that there are  $m$  windows in the context  $x$ , we extract  $mn(n - 1)(|\mathcal{R}| + 1)$  probabilistic relational symbols. Each symbol is denoted as an atom of the form  $p(s,o)$ , where  $p \in \mathcal{R} \cup \{\mathrm{n/a}\}$  is the relational predicate, and  $s, o$  are the two entities connected by the predicate. We denote
+
+the probability of such symbol extracted by the LM and relational classifier as  $\operatorname{Pr}(p(s,o) \mid \theta)$ . All these probabilities combined form the output vector  $\mathbf{r} = \mathcal{M}_{\theta}(x) \in \mathbb{R}^{mn(n-1)(|\mathcal{R}|+1)}$ .
+
+# 3.4 Differentiable Symbolic Inference
+
+The symbolic inference modules  $\mathcal{P}_{\phi}$  and  $\mathcal{P}_{\mathrm{s1}}$  are responsible for processing the extracted relations to deduce 1) an expected output relation in  $\mathcal{R}$ , and 2) a semantic loss encoding the probability of constraint violation. There are two main objectives for these modules. First, they need to logically reason about the output relation and the semantic loss based on the extracted relational symbols  $\mathbf{r}$ , the query  $q$ , and the rule weights  $\phi$ . Second, they need to compute the gradients of  $\hat{y}$  and  $l_{\mathrm{s1}}$  with respect to  $\theta$  and  $\phi$ , namely  $\frac{\partial \hat{y}}{\partial \theta}, \frac{\partial \hat{y}}{\partial \phi}, \frac{\partial l_{\mathrm{s1}}}{\partial \phi}$ , and  $\frac{\partial l_{\mathrm{s1}}}{\partial \theta}$ , in order for the fine-tuning and rule learning to happen.
+
+Logical deduction. Logic rules can be applied to known facts to deduce new ones. For example, below is a horn clause, which reads "if  $b$  is  $a$ 's brother and  $c$  is  $b$ 's daughter, then  $c$  is  $a$ 's niece":
+
+$$
+\operatorname {n i e c e} (a, c) \leftarrow \operatorname {b r o t h e r} (a, b) \wedge \operatorname {d a g u t h e r} (b, c).
+$$
+
+Note that the structure of the above rule can be captured by a higher-order logical predicate called "composite" (abbreviated as comp). This allows us to express many other similarly structured rules with ease. For instance, we can have comp(brother, daughter, niece) and comp(father, mother, grandmother). With this set of rules, we may derive more facts based on known kinship relations. In fact, composition is the only kind of rule we need for kinship reasoning. In general, there are many other useful higher-order predicates to reason over knowledge bases, which we list out in Table 2.
+
+<table><tr><td>Predicate</td><td>Example</td></tr><tr><td>transitive</td><td>transitive(relative)</td></tr><tr><td>symmetric</td><td>symmetric(spouse)</td></tr><tr><td>inverse</td><td>inverse(husband, wife)</td></tr><tr><td>implies</td><td>implies(mother, parent)</td></tr></table>
+
+Table 2: Higher-order predicate examples.
+
+Probability propagation. We seek to have the deduced facts to also be associated with probabilities computed using probabilities predicted by the underlying relation extractor  $\mathcal{M}_{\theta}$ . This is achieved by allowing the propagation of probabilities. For
+
+example, we have the proof tree with probabilities:
+
+$$
+\frac {0 . 9 : \operatorname {b r o t h e r} (D , R) \quad 0 . 8 : \operatorname {d a g h t e r} (R , K)}{0 . 7 2 : \operatorname {n i c e} (D , K)}
+$$
+
+In practice, there could be multiple steps in the proof tree (multi-hop) and one fact can be derived by multiple proof trees. We employ the inference algorithms based on approximated weighted model counting (WMC) presented in (Manhaeve et al., 2018) to account for probabilistic inference under complex scenarios. Since the WMC procedure is augmented for differentiation, we can obtain the gradient  $\frac{\partial\hat{y}}{\partial\mathbf{r}}$ . From here, we can obtain  $\frac{\partial\hat{y}}{\partial\theta} = \frac{\partial\hat{y}}{\partial\mathbf{r}}\frac{\partial\mathbf{r}}{\partial\theta}$ , where the second part can be automatically derived from differentiating  $\mathcal{M}_{\theta}$ .
+
+Rule learning. Hand-crafted rules could be expensive or even impossible to obtain. To alleviate this issue, DSR-LM applies LMs to help automatically extract rules, and further utilizes the differentiable pipeline to fine-tune the rules. Each rule such as comp(brother, daughter, niece) is attached a weight, initialized by prompting an underlying LM. For example, the prompt we use for extracting comp(r,p,q) is "one's \( r \)'s \( p \) is their \( <q:mask>"). Given that the relations \( r, p, q \in \mathcal{R} \), DSR-LM automatically enumerates \( r \) and \( p \) from \( \mathcal{R} \) while querying for LM to unmask the value of \( q \). LM then returns a distribution of words, which we take an intersection with \( \mathcal{R} \). The probabilities combined form the initial rule weights \( \phi \). This type of rule extraction strategy is different from existing approaches in inductive logic programming since we are exploiting LMs for existing knowledge about relationships.
+
+Note that LMs often make simple mistakes answering such prompt. In fact, with the above prompt, even GPT-3 can only produce  $62\%$  of composition rules correctly. While we can edit prompt to include few-shot examples, in this work we consider fine-tuning such rule weights  $\phi$  within our differentiable reasoning pipeline. The gradient with respect to  $\phi$  is also derived with the WMC procedure, giving us  $\frac{\partial \hat{y}}{\partial \phi}$ . In practice, we use two optimizers with different hyper-parameters to update the rule weights  $\phi$  and the underlying model parameter  $\theta$ , in order to account for optimizing different types of weights.
+
+Semantic loss and integrity constraints. In general, learning with weak supervision label is hard, not to mention that the deductive rules are learnt as well. We thereby introduce an additional semantic
+
+loss during training. Here, semantic loss is derived by a set of integrity constraints used to regularize the predicted entity-relation graph as well as the learnt logic rules. In particular, we consider rules that detect violations of integrity constraints. For example, "if A is B's father, then B should be A's son or daughter" is an integrity constraint for relation extractor—if the model predicts a father relationship between A and B, then it should also predict a son or daughter relationship between B and A. Encoded in first order logic, it is
+
+$$
+\forall a, b, \text {f a t h e r} (a, b) \Rightarrow (\text {s o n} (b, a) \vee \text {d a g u t h e r} (b, a)).
+$$
+
+Through differentiable reasoning, we evaluate the probability of such constraint being violated, yielding our expected semantic loss. In practice, arbitrary number of constraints can be included, though too many interleaving ones could hinder learning.
+
+# 4 Experiments
+
+We evaluate DSR-LM on both CLUTRR and DBpedia-INF. We show that DSR-LM has accurate and generalizable long-range reasoning capability.
+
+# 4.1 Datasets
+
+CLUTRR (Sinha et al., 2019) consists of kinship reasoning questions. Given a context that describes a family's routine activity, the goal is to deduce the relationship between two family members that is not explicitly mentioned in the story. Although the dataset is synthetic, the sentences are crowdsourced and hence there is a considerable amount of naturalness inside the dataset. The family kinship graph is synthetic and the names of the family members are randomized. For ablation study, we manually crafted 92 kinship composition rules as an external symbolic knowledge base. This yields the following symbolic information for each datapoint: 1) the full kinship graph corresponding to the story, 2) the symbolic knowledge base (KB), and 3) a query representing the question. The CLUTRR dataset is divided into different difficulties measured by  $k$ , the number of facts used in the reasoning chain. For training, we only have 10K data points with  $5\mathrm{K}k = 2$  and another 5K  $k = 3$ , meaning that we can only receive supervision on data with short reasoning chains. The test set, on the other hand, contains 1.1K examples with  $k \in \{2,\dots ,10\}$ .
+
+DBpedia-INF is a curated subset of the evaluation dataset used in RuleBert (Saeed et al., 2021). Similar to CLUTRR, it is generated synthetically
+
+to test the reasoning capability of LMs. Given a synthetic passage describing the relation between entities, and soft deductive logic rules, we aim to deduce the relationship between any two entities. The symbolic program of DBpedia-INF consists of 26 predicates, 161 soft rules mined from DBpedia, and 16 rules defining the negation and symmetry between the predicates. The difficulty of the questions is represented in terms of reasoning length from  $k \in \{0, \dots, 5\}$ . Larger  $k$  implies harder question. Compared to the exact dataset used in Rulebert, we clean it in order to ensure the question-answer pairs are logically consistent and probabilistically correct.
+
+# 4.2 Experimental Setup
+
+Implementation. We employ Scallop (Huang et al., 2021) as the differentiable symbolic inference module. We show the program used for CLUTRR reasoning task in Figure 2. It comprises relation type declarations, deductive rules for kinship reasoning, and integrity constraints for computing semantic loss (attached in the Appendix). The program used for DBpedia-INF is written in a similar manner with additional high-order predicates listed in Table 2.
+
+Pre-trained LMs for fine-tuning. We used the HuggingFace (Wolf et al., 2019) pre-trained  $w2v$ -google-news-300, RoBERTa-base, and DeBERTa-base as the pretrained language models. We fine-tune RoBERTa-base and DeBERTa-base during training with binary cross entropy loss. Our relation extraction module is implemented by adding an MLP classifier after the LM, accepting a concatenation of the embedding of the two entities and the embedding of the whole windowed context.
+
+Our model. Our main model, DSR-LM, uses RoBERTa as the underlying LM. The relation classifier is a 2-layer fully connected MLP. For training, we initialize  $\phi$  by prompting the LM. To accelerate the learning process, we use multinomial sampling to retrieve 150 rules for symbolic reasoning. During testing, we will instead pick the top 150 rules. We use two Adam optimizer to update  $\theta$  and  $\phi$ , with learning rate  $10^{-5}$  and  $10^{-2}$  respectively.
+
+For ablation studies, we present a few other models. First, we ablate on back-bone LMs. Specifically, we have DSR-LM-DeBERTa which uses De
+
+```txt
+// Relation declaration
+type kinship(rela: String, subject: String, object: String)
+type query(subject: String, object: String)
+type composite(r1: String, r2: String, r3: String)
+// Rules to derive the final answer
+rel kinship(r3,a,c) = kinship(r1,a,b), kinship(r2,b,c), composite(r1,r2,r3), a != c
+rel answer(r) = query(s,o), derive(r,s,o)
+// Integrity constraints (6 for kinship graph and 2 for rule learning)
+rel violation(!r) = r :=forall(a,b: kinship(FATHER,a,b) => kinship(SON,b,a) or kinship(DAUGHTER,b,a)) // Other constraints are omitted...
+```
+
+BERTa as the back-bone LM. DSR-w2v-BiLSTM, on the other hand, uses as back-bone the word2vec (Mikolov et al., 2013) model for word embedding and BiLSTM (Huang et al., 2015) for sequential encoding. For DSR-LM-with-Manual-Rule we treat the logic rules as given, meaning that we provide 92 composition rules for CLUTRR and around 180 rules for DBpedia-INF. In this case, we set ground truth rules to have 1.0 weight and therefore  $\phi$  is not learnt. Then, we have DSR-LM-without-IC which does not have integrity constraints and semantic loss. Lastly, we have DSR-without-LM that takes ground truth structured entity relation graph as input. This way, we do not need the underlying relation extractor and only  $\phi$  needs to be learned.
+
+Baselines. We compare DSR-LM with a spectrum of baselines from purely neural to logically structured. The baselines include pretrained large language models (BERT (Kenton and Toutanova, 2019) and RoBERTa (Liu et al., 2019)), non-LM counterparts (BiLSTM (Hochreiter and Schmidhuber, 1997; Cho et al., 2014) and BERT-LSTM), structured models (GAT (Veličković et al., 2018), RN (Santoro et al., 2017), and MAC (Hudson and Manning, 2018)), and other neuro-symbolic models (CTP (Minervini et al., 2020), RuleBert (Saeed et al., 2021)). The structured models include those models with relational inductive biases, while the neuro-symbolic model uses logic constraints.
+
+Baseline setup. We highlight a few baselines we include for completeness but are treated as unfair comparison to us: GAT, CTP, and GPT-3 variants. All baselines other than GAT and CTP take as input natural language stories and the question to produce the corresponding answer. GAT and CTP, on the contrary, takes entity relation graph rather than natural language during training and testing.
+
+The model sizes are different across baselines as well. Model size generally depends on two parts, the backbone pre-trained LM, and the classifica
+
+tion network built upon the LM. GPT-3 contains 175B parameters, and RoBERTa uses 123M parameters. The classification model of our method has 2.97M parameters (assuming using embeddings from RoBERTa). With extra 10K parameters for rule weights, our DSR-LM framework has around 127M parameters.
+
+For GPT-3 variants, we conduct experiments on CLUTRR with GPT-3 under the Zero-Shot (GPT-3 ZS), GPT-3 Fine-Tuned (GPT-3 FT), and Few(5)-Shot (GPT-3 5S) (Brown et al., 2020), as well as Zero-Shot-CoT (GPT-3 ZS-CoT) (Kojima et al., 2022a) settings. For fair comparison, we also include the ground truth kinship composition knowledge in GPT-3 zero shot (GPT-3 ZS w/ Rule), and 5 shot (GPT-3 5S w/ Rule). We include the prompts we used and additional details in Appendix A.
+
+![](images/320f49040380ff6274845ada3842cad27dfafc7873a5504fb49fc0a0f547e1f3.jpg)  
+Figure 2: The Scallop program used in the CLUTRR reasoning task.  
+Figure 3: DSR-LM's performance on CLUTRR compared with various baselines
+
+# 4.3 Experimental Results
+
+DSR-LM systematically outperforms a wide range of baselines by a large margin. We evaluate DSR-LM and baselines on both CLUTRR and DBpedia-INF, as reported in Figure 3 and Table 3.
+
+In the CLUTRR experiment, DSR-LM achieves the best performance among all the models (Figure 3). Next, we examine how models trained on stories generated from clauses of length  $k \leq 3$  and
+
+<table><tr><td>Test Length</td><td>DSR-LM</td><td>RuleBert</td></tr><tr><td>Overall</td><td>95.87</td><td>72.59</td></tr><tr><td>0</td><td>100.0</td><td>98.40</td></tr><tr><td>1</td><td>100.0</td><td>54.80</td></tr><tr><td>2</td><td>98.4</td><td>75.20</td></tr><tr><td>3</td><td>89.2</td><td>64.00</td></tr><tr><td>4</td><td>88.1</td><td>69.89</td></tr><tr><td>5</td><td>100.0</td><td>72.29</td></tr></table>
+
+Table 3: DBpedia-INF generalization evaluation under different test reasoning length. Models are trained on 10K reasoning length  $k = 0$  sequences, and tested on sequences of reasoning length  $k = [0,5]$ .  
+
+<table><tr><td>Confidence</td><td>Learnt Rules</td></tr><tr><td>1.154</td><td>mother(a,c) ← sister(a,b) ∧ mother(b,c)</td></tr><tr><td>1.152</td><td>daughter(a,c) ← daughter(a,b) ∧ sister(b,c)</td></tr><tr><td>1.125</td><td>sister(a,c) ← daughter(a,b) ∧ aunt(b,c)</td></tr><tr><td>1.125</td><td>father(a,c) ← brother(a,b) ∧ father(b,c)</td></tr><tr><td>1.123</td><td>granddaughter(a,c) ← grandson(a,b) ∧ sister(b,c)</td></tr><tr><td>1.120</td><td>brother(a,c) ← sister(a,b) ∧ brother(b,c)</td></tr><tr><td>1.117</td><td>brother(a,c) ← son(a,b) ∧ uncle(b,c)</td></tr><tr><td>1.105</td><td>brother(a,c) ← daughter(a,b) ∧ uncle(b,c)</td></tr><tr><td>1.104</td><td>daughter(a,c) ← wife(a,b) ∧ daughter(b,c)</td></tr><tr><td>1.102</td><td>mother(a,c) ← brother(a,b) ∧ mother(b,c)</td></tr><tr><td>...</td><td>...</td></tr></table>
+
+Table 4: The learnt top-10 confident logic rules over CLUTRR.
+
+![](images/c81fbb683627de2a449f600e96b4f555e3ccabb942a46db9df46d008b2249cde.jpg)  
+(a) Comparison to common baselines
+
+![](images/db05178039dc9182baf0307c28e993e31013f0d89b5731ca45313bcc5e141c3b.jpg)  
+(b) Comparison to GPT-3 related baselines  
+Figure 4: Systematic generalization performance comparison on CLUTRR dataset. Models except GPT-3-ZS*, GPT-3-FS are trained (or fine-tuned) on  $k \in \{2, 3\}$ . All models are tested on  $k \in \{2, \dots, 10\}$ .
+
+evaluated on stories generated from larger clauses of length  $k \geq 4$ . A fine-grained generalizability study reveals that although all models' performances decline as the reasoning length of the test sequence increases, pure neural-based models decrease the fastest (Figure 4a and 4b). It manifests the systematic issue that language models alone are still not robust for length generalization (Lake and Baroni, 2018). On the other hand, the performance of DSR-LM decreases much slower as test reasoning length increases and outperforms all the baselines when  $k \geq 4$ .
+
+In the DBpedia-INF experiment, DSR-LM outperforms RuleBert by  $37\%$  in terms of overall performance (Table 3), showing that DSR-LM has much more robust generalization. Recall that RuleBert aims to improve the logical reasoning of LMs by straightforward fine-tuning with soft rules and facts. Our results show that augmenting data alone for fine-tuning do not effectively improve systematicity. Meanwhile, DSR-LM imbues reasoning inductive biases throughout training and learns useful rules to generalize to longer reasoning lengths.
+
+Learning interpretable logic rules. DSR-LM is capable of producing explicit logic rules as part of
+
+the learning process. For presentation, we show the top-10 rules learnt from DSR-LM model in Table 4. We compare the top-92 most likely prompted and fine-tuned rules against the 92 hand-crafted rules, and 70 of them match. Additionally, we find that our rule weight fine-tuning helps correct 11 of the incorrect rules produced by LM. Through this qualitative analysis, it is clear that DSR-LM provides an interface to probe and interpret the intermediate steps, enhancing the interpretability.
+
+GPT-3 variants are inferior in long-range reasoning. Interestingly, ZS scores  $28.6\%$  accuracy on CLUTRR while ZS-CoT scores  $25.6\%$ , suggesting that the chain-of-thought prompting might not work in long-range reasoning (Figure 3). In fact, there are many cases where GPT-3 favors complication over simplicity: GPT-3 frequently answers "stepdaughter", "stepmother", and "adopted son", while the real answers are simply "daughter", "mother", and "son". Additionally, GPT-3 could derive the correct result for the wrong reason, e.g. "Jeffrey is Gabrielle's son, which would make William her grandson, and Jeffrey's brother." While we count the final answer to be correct (William is Jeffrey's brother), there is a clear in
+
+consistency in the reasoning chain: William cannot be Gabrielle's grandson and Jeffrey's brother simultaneously, given that Jeffrey is Gabrielle's son. Lastly, we observe that, both GPT-3 FT and many other methods have an accuracy drop as  $k$  becomes larger (Figure 4b), ZS and ZS-CoT stay relatively consistent, suggesting that the size of context and the reasoning chain may have a low impact on GPT-3's performance.
+
+# 4.4 Analyses and Ablation Studies
+
+Symbolic reasoner consistently improves LMs and word embeddings. Since DSR-LM has a model agnostic architecture, we study how the choice of different LMs impacts the reasoning performance. As shown in Table 5, the two transformer-based models have on-par performance and outperform the word2vec one. However, note that the word2vec-based model still has better performance than all other baselines. Besides higher final accuracy, the pre-trained transformer-based language model also accelerates the training process. Both DSR-LM-RoBERTa and DSR-LM-DeBERTa reach their best performance within 20 epochs, while it takes DSR-w2v-BiLSTM 40 epochs to peak.
+
+<table><tr><td>Model</td><td>Accuracy (%)</td></tr><tr><td>DSR-LM (RoBERTa)</td><td>60.98 ± 2.64</td></tr><tr><td>DSR-LM-DeBERTa</td><td>60.92 ± 2.72</td></tr><tr><td>DSR-w2v-BiLSTM</td><td>40.39 ± 0.06</td></tr></table>
+
+Incorporate domain knowledge. DSR-LM allows injecting domain specific knowledge. In DSR-LM-with-Rule, we manually crafted 92 rules for kinship reasoning to replace the learnt rules. As shown in Table 6, it obtained a  $0.36\%$  performance gain over DSR-LM. The fact that the improvement is marginal implies our method extracts useful rules to obtain on-par performance with manually crafted ones. DSR-LM-without-IC, our model without integrity constraints specified on predicted relations and rules, performs worse than DSR-LM, suggesting that logical integrity constraints are essential component for improving the model robustness.
+
+The impact of the relation extractor. To understand what causes the failure case of DSR-LM, we study the performance of our relation classification model separately. We isolate the trained relation
+
+Table 5: Ablation study about neural backbones of DSR-LM. We compare the CLUTRR performance of DSR-LM using different LMs.  
+
+<table><tr><td>Model</td><td>Accuracy (%)</td></tr><tr><td>DSR-LM</td><td>60.98 ± 2.64</td></tr><tr><td>DSR-LM-without-IC</td><td>51.48 ± 0.57</td></tr><tr><td>DSR-LM-with-Manual-Rule</td><td>61.34 ± 1.56</td></tr></table>
+
+extractor and found that it reaches  $84.69\%$  accuracy on the single relation classification task. For comparison, we train a relation extractor using all the intermediate labels in the training dataset, and it reaches  $85.32\%$  accuracy. It shows that even using only weak supervision (i.e., the final answers to multi-hop questions), our approach can reach on-par performance as supervised relation extraction.
+
+Reasoning over structured KBs. To understand the rule learning capability of our approach, we design our ablation model DSR-without-LM to take as input ground-truth KBs instead of natural language. In this case, rule weights are not initialized by LM but randomized. As shown in Table 7, our model outperforms GAT and CTP which also operates on structured KBs. It demonstrates that our differentiable rule learning paradigm learns rules to reason about KBs consistently.
+
+Table 6: Ablation study. We compare our model's performance on CLUTRR with different setups.  
+
+<table><tr><td>Model</td><td>Accuracy (%)</td></tr><tr><td>GAT</td><td>39.05</td></tr><tr><td>CTP</td><td>95.57</td></tr><tr><td>DSR-without-LM</td><td>98.81</td></tr></table>
+
+Table 7: DSR-without-LM compared against GAT and CTP on reasoning with ground truth KBs. For this comparison we train on  $k \in [2,3]$  and test on  $k \in [4,10]$ .
+
+Failure cases of DSR-LM. We showcase in Appendix Table 8 that even state-of-the-art large LMs are prone to logical fallacies. On the other hand, the failure case of our method usually occurs in the stage of relation extraction. For example, for the following sentence "Christopher and Guillermina are having a father-daughter dance", our RoBERTa based relation extractor fails to recognize the father-daughter relationship but rather thinks C and G have a husband-wife relationship. We require most of the relation extraction to be correct in order to avoid cascading error. As the error rate on individual relation extraction accumulates, it leads to the observed drop in accuracy as  $k$  becomes larger.
+
+# 5 Concluding Remarks
+
+We investigate how to improve LMs' logical reasoning capability using differentiable symbolic rea
+
+soning. Through extensive experiments, we demonstrate the effectiveness of DSR-LM over challenging scenarios where widely deployed large LMs fail to reason reliably. We hope our work can lay the groundwork for exploring neuro-symbolic programming techniques to improve the robustness of LMs on reasoning problems.
+
+# Limitations
+
+The primary limitation of DSR-LM is the need for a confined problem space. It requires a well-defined relational schema to perform logical reasoning, and thus will not be suited for an open-ended problem setup. Nevertheless, DSR-LM is suitable for many domain specific problems within Natural Language Understanding and Reasoning, allowing domain experts to freely inject domain-specific knowledge in a structured and logical manner.
+
+# References
+
+Gregor Betz, Christian Voigt, and Kyle Richardson. 2020. Critical thinking for language models. arXiv preprint arXiv:2009.07185.  
+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. NeurIPS.  
+Oana-Maria Camburu, Tim Roktaschel, Thomas Lukasiewicz, and Phil Blunsom. 2018. e-snli: Natural language inference with natural language explanations. Advances in Neural Information Processing Systems, 31.  
+Swarat Chaudhuri, Kevin Ellis, Oleksandr Polozov, Rishabh Singh, Armando Solar-Lezama, Yisong Yue, et al. 2021. Neurosymbolic Programming.  
+Kyunghyun Cho, Bart van Merrienboer, Caglar Güliquehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rn encoder-decoder for statistical machine translation. In EMNLP.  
+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.  
+Peter Clark, Oyvind Tafjord, and Kyle Richardson. 2021. Transformers as soft reasoners over language. In IJCAI.  
+Antonia Creswell, Murray Shanahan, and Irina Higgins. 2022. Selection-inference: Exploiting large language
+
+models for interpretable logical reasoning. arXiv preprint arXiv:2205.09712.  
+Yanai Elazar, Nora Kassner, Shauli Ravfogel, Abhilasha Ravichander, Eduard Hovy, Hinrich Schütze, and Yoav Goldberg. 2021. Measuring and improving consistency in pretrained language models. TACL.  
+Richard Evans and Edward Grefenstette. 2018. Learning explanatory rules from noisy data. Journal of Artificial Intelligence Research, 61:1-64.  
+Klaus Greff, Sjoerd Van Steenkiste, and Jurgen Schmidhuber. 2020. On the binding problem in artificial neural networks. arXiv preprint arXiv:2012.05208.  
+Peter Hase, Mona Diab, Asli Celikyilmaz, Xian Li, Zornitsa Kozareva, Veselin Stoyanov, Mohit Bansal, and Srinivasan Iyer. 2021. Do language models have beliefs? methods for detecting, updating, and visualizing model beliefs. arXiv preprint arXiv:2111.13654.  
+Chadi Helwe, Chloe Clavel, and Fabian M. Suchanek. 2021. Reasoning with transformer-based models: Deep learning, but shallow reasoning. In AKBC.  
+Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural computation*.  
+Jiani Huang, Ziyang Li, Binghong Chen, Karan Samel, Mayur Naik, Le Song, and Xujie Si. 2021. Scallop: From probabilistic deductive databases to scalable differentiable reasoning. NeurIPS.  
+Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirectional lstm-crf models for sequence tagging. arXiv preprint arXiv:1508.01991.  
+Drew A Hudson and Christopher D Manning. 2018. Compositional attention networks for machine reasoning. In *ICLR*.  
+Daniel Kahneman. 2011. Thinking, fast and slow. Macmillan.  
+Nora Kassner, Benno Krojer, and Hinrich Schütze. 2020. Are pretrained language models symbolic reasoners over knowledge? CoNLL.  
+Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL.  
+Yoon Kim. 2021. Sequence-to-sequence learning with latent neural grammars. NeurIPS.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022a. Large language models are zero-shot reasoners. NeurIPS.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022b. Large language models are zero-shot reasoners. NeurIPS.  
+Brenden Lake and Marco Baroni. 2018. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In ICML.
+
+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.  
+Robin Manhaeve, Sebastijan Dumancic, Angelika Kim-mig, Thomas Demeester, and Luc De Raedt. 2018. Deepproblog: Neural probabilistic logic program-ming. NeurIPS.  
+Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. 2018. The neurosymbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. In ICLR.  
+Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. NeurIPS.  
+Pasquale Minervini, Sebastian Riedel, Pontus Stenetorp, Edward Grefenstette, and Tim Rocktäschel. 2020. Learning reasoning strategies in end-to-end differentiable proving. In ICML.  
+Maxwell Nye, Michael Tessler, Josh Tenenbaum, and Brenden M Lake. 2021. Improving coherence and consistency in neural sequence models with dual-system, neuro-symbolic reasoning. NeurIPS.  
+Tim Rocktäschel and Sebastian Riedel. 2017. End-to-end differentiable proving. NeurIPS.  
+Mohammed Saeed, Naser Ahmadi, Preslav Nakov, and Paolo Papotti. 2021. Rulebert: Teaching soft rules to pre-trained language models. In *EMNLP*.  
+Adam Santoro, David Raposo, David G Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia, and Timothy Lillicrap. 2017. A simple neural network module for relational reasoning. NeurIPS.  
+Abulhair Saparov and He He. 2022. Language models are greedy reasoners: A systematic formal analysis of chain-of-thought. arXiv preprint arXiv:2210.01240.  
+Peter Shaw, Ming-Wei Chang, Panupong Pasupat, and Kristina Toutanova. 2021. Compositional generalization and natural language variation: Can a semantic parsing approach handle both? In ACL.  
+Koustuv Sinha, Shagun Sodhani, Jin Dong, Joelle Pineau, and William L. Hamilton. 2019. Clutrr: A diagnostic benchmark for inductive reasoning from text. EMNLP.  
+Livio Baldini Soares, Nicholas Fitzgerald, Jeffrey Ling, and Tom Kwiatkowski. 2019. Matching the blanks: Distributional similarity for relation learning. In ACL.  
+Alon Talmor, Yanai Elazar, Yoav Goldberg, and Jonathan Berant. 2020a. olympics-on what language model pre-training captures. TACL.
+
+Alon Talmor, Oyvind Tafjord, Peter Clark, Yoav Goldberg, and Jonathan Berant. 2020b. Leap-of-thought: Teaching pre-trained models to systematically reason over implicit knowledge. *NeurIPS*.  
+Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. 2020. Long range arena: A benchmark for efficient transformers. In ICLR.  
+Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. Bertrediscoverstheclassicalnlppipeline.InACL.  
+Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. 2022. Large language models still can't plan (a benchmark for llms on planning and reasoning about change). arXiv preprint arXiv:2206.10498.  
+Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. In ICLR.  
+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. NeurIPS.  
+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. EMNLP.  
+Jingyi Xu, Zilu Zhang, Tal Friedman, Yitao Liang, and Guy Broeck. 2018. A semantic loss function for deep learning with symbolic knowledge. In ICML.  
+Eric Zelikman, Yuhuai Wu, and Noah D Goodman. 2022. Star: Bootstrapping reasoning with reasoning. NeurIPS.  
+Hanlin Zhang, Yi-Fan Zhang, Li Erran Li, and Eric Xing. 2022. The impact of symbolic representations on in-context learning for few-shot reasoning. arXiv preprint.  
+Yuyu Zhang, Xinshi Chen, Yuan Yang, Arun Ramamurthy, Bo Li, Yuan Qi, and Le Song. 2019. Efficient probabilistic logic reasoning with graph neural networks. In ICLR.  
+Wangchunshu Zhou, Jinyi Hu, Hanlin Zhang, Xiaodan Liang, Maosong Sun, Chenyan Xiong, and Jian Tang. 2020. Towards interpretable natural language understanding with explanations as latent variables. NeurIPS.
+
+# A Implementation Details
+
+Reasoner details. The learning of rules and the fine-tuning of the underlying LM should happen separately with different learning rates – fine-tuning LM is an intricate process that requires a very small learning rate, while rules should be learned with larger learning rates since gradients are directly back-propagated onto the weights. This can be realized by employing two separate optimizers, one for fine-tuning and the other for rule learning. During training time, we rotate training the two parts by toggling one and the other optimizer for every 10 batches of data points.
+
+Rule learning training setup. For rule learning, we can initialize the transitivity tensor using the language model provided composite rules. Since the CLUTRR dataset consists of 20 different relations and a transitivity relationship is defined over 3 relations, there are 8K possible transitivity facts over these relations. Specifically, we give every predicted composite rule by the GPT with a 0.5 weight, while initializing the other rules with a range such as [0, 0.1], since otherwise, an insensible transitive fact may be getting a random high weight while it effectively does nothing for reasoning. The learning process encourages the rules that yield the correct query result and suppresses the rules that lead to wrong answers. To avoid the exponential blow-up caused by injecting all the 8K rules in the reasoning engine, we sample 200 rules according to their weights during the training time and deterministically use the top 200 learned rules during the test time. For the QA-No-Rule setup, the confidence score of rules, the MLP classifier for relation extraction, and the underlying LM are learned and updated simultaneously during training. To account for their difference, we employ two Adam optimizers  $A_{\mathrm{RL}}$  and  $A_{\mathrm{RE}}$ .  $A_{\mathrm{RE}}$  is used for optimizing models for relation extraction, and thus will take as parameters the MLP classifier and the underlying LM. It has a low learning rate 0.00001 since it needs to fine-tune LMs.  $A_{\mathrm{RL}}$ , on the other hand, will take as a parameter the confidence score tensor for the transitive rules, and is set to have a higher learning rate of 0.001. For the integrity constraints, we set the result integrity violation loss with the weight 0.1, and set the rule integrity constraint violation loss with the weight 0.01. We set the batch size to 16 and train for 20 epochs.
+
+To obtain the initial rule weights for the composition rule in our CLUTRR experiment, the prompt
+
+we use is "Mary's P's Q is her  $<\text{mask}>$ ." where P and Q are enumerations of all possible relationships, and the unmasked value is treated as the answer R, producing composite(P, Q, R). For the other rule templates we used, the prompts are
+
+1. transitive: "is R's R one's R? <mask>"; the probability of the unmasked word being "yes" is treated the rule weight for transitive(R).  
+2. symmetric: "does A is R of B means B is R of A? <mask>"; the probability of the un-masked word being "yes" is treated the rule weight for symmetric(R).  
+3. inverse: "A is R of B means B is <mask> of A"; the unmasked value is treated as the answer P, producing inverse (R, P).  
+4. implies: "does R imply P? <mask>"; the probability of unmasked value being "yes" is treated as the rule weight for implies (R, P).
+
+GPT-3 Prompt Setups. For Zero-Shot, we use the prompt "So  $B$  is  $A$ 's:" for the query pair  $(A, B)$  to ask GPT-3 to complete the relationship between  $A$  and  $B$ . We pick the phrase in the first line or before the first period from the completed text and compare it directly with the ground truth relation. For the Few(5)-Shot setting, we randomly select 5 examples from the training dataset used for other models ( $k \in [2, 3]$ ) to serve as examples. We use the same prompt for Few-Shot and Fine-Tuned as the Zero-Shot and the automated GPT-3 fine-tuning setup for our training dataset, trained for 4 epochs. To add in the transitive KB, we simply include 92 hand-crafted rules in natural language as a part of the prompt, and we performed Zero-shot with KB, and Few(5)-shot with KB experiments. For the Zero-Shot-CoT setting, we use the prompt "Who is  $B$  to  $A$ ? Let's think step by step" to suggest GPT-3 to auto-complete while working out a reasoning chain. Under this setup, it is impossible to compare the answer to the ground truth automatically. Therefore, we manually check through the whole test dataset of CLUTRR.
+
+# B Additional Experimental Results
+
+In Table 8, we showcase the failure cases of large LMs for logical inference, where Zero-shot-CoT denotes zero-shot chain-of-thoughts (Kojima et al., 2022b).
+
+<table><tr><td>k</td><td>Prompt</td><td>GPT-3 CoT Output</td><td>Ground Truth</td></tr><tr><td>2</td><td>Jeffrey asked his brother William if he would come help him fix his car next weekend. Gabrielle was delighted to see her grandson, William, at her birthday party. Who is Jeffrey to Gabrielle? Let&#x27;s do it step by step:</td><td>Jeffrey is Gabrielle&#x27;s grandson. William is Jeffrey&#x27;s brother.</td><td>grandson</td></tr><tr><td>3</td><td>Elliott felt lonely when his wife Marian was gone to see her mother Molly. Elliott and his daughter Karen took a day off school to go to the zoo. Who is Karen to Molly? Let&#x27;s do it step by step:</td><td>Karen is Elliott&#x27;s daughter. Molly is Elliott&#x27;s wife&#x27;s mother. There-fore, Karen is Molly&#x27;s grand-daughter.</td><td>granddaughter</td></tr><tr><td>4</td><td>Frank took his daughter Julia to the park. Beverly, Emilie&#x27;s sister went with them. Kimberley was excited because today she was going to the zoo with her uncle Frank. Jenny spent a great day shopping with her daughter, Kimberley. Who is Beverly to Jenny? Let&#x27;s do it step by step:</td><td>Beverly is Jenny&#x27;s sister-in-law.</td><td>niece</td></tr><tr><td>5</td><td>Melissa and her father, Stephen, went to the marina. Stephen&#x27;s daughter, Margaret, had purchased a boat, and they were eager to see it. Annie and her aunt, Evelyn, went to the deli. They got a half a pound of corned beef, and two pounds of salami. Melissa and her sister Annie were arguing over a toy. While they were fighting, Melissa&#x27;s sister Margaret grabbed the toy, and ran off to play with it by herself. Who is Evelyn to Melissa? Let&#x27;s do it step by step:</td><td>Melissa&#x27;s father is Stephen. Stephen has a daughter, Margaret. Margaret is Annie&#x27;s aunt. There-fore, Evelyn is Annie&#x27;s aunt, and Melissa&#x27;s great-aunt.</td><td>aunt</td></tr><tr><td>6</td><td>James took his grandson Davis to the baseball game. Sharon and her son Brian went out to lunch together yesterday. Brian and his brother Oliver went to the park to play basketball. Davis played scrabble with his mother Sharon and his brother Oliver. Davis asked his brother Oliver if he would come help him fix his car next weekend. Who is James to Oliver? Let&#x27;s do it step by step:</td><td>James is Davis&#x27;s grandfather, so he is Oliver&#x27;s great-grandfather.</td><td>grandfather</td></tr><tr><td>7</td><td>George rushed to the hospital to find out that his wife and already given birth to a boy and had named him David. Larry was playing checkers with his brother Craig. Larry&#x27;s sister Jean was waiting to play. Judy made mixed drinks for her son Craig. Judy&#x27;s daughter Jean was jealous because she was n&#x27;t offered one. Jean planned a trip to the zoo for her brother, Craig. They had a great time. George was having trouble with his divorce because he only got to see his daughter Jean a few times a week. Who is David to Craig? Let&#x27;s do it step by step:</td><td>George&#x27;s wife gave birth to a boy and named him David. David is Craig&#x27;s nephew.</td><td>brother</td></tr><tr><td>8</td><td>Anthony went to the bakery with his uncle Roger to pick up some bread for lunch. Anthony loved going to the store with his mom Mary. She always bought him snacks Charles would n&#x27;t let his son Floyd go to the park by himself. Floyd&#x27;s brother Anthony offered to go with him. Faye was hoping to travel out of state to visit her son Cecil. Ellie&#x27;s brother, Charles, was doing his best to find the best flight. Mary and her son Anthony went out to lunch together yesterday. Cecil took her sister, Christine, out to dinner for her birthday. Who is Roger to Christine? Let&#x27;s do it step by step:</td><td>Cecil is Faye&#x27;s son, so Faye is Christine&#x27;s mother-in-law. Christine is married to Faye&#x27;s son, so Christine is Faye&#x27;s daughter-in-law. Charles is Ellie&#x27;s brother, so Charles is Christine&#x27;s brother-in-law. Anthony is Charles&#x27;s son, so Anthony is Christine&#x27;s nephew. Roger is Anthony&#x27;s uncle, so Roger is Christine&#x27;s brother-in-law.</td><td>uncle</td></tr><tr><td>9</td><td>Gino bought gifts for his lovely sister, Jami. Michele and William left that evening to go bowling. Ethel and her son, Gordon, went to the park, and had a wonderful time. Manuel&#x27;s father, Kenneth, thought his son was a monster. Meanwhile Michele thought her son Manuel was a little angel. Gino, Kenneth&#x27;s other son, agreed with his dad. William asked his brother Larry if he would come help him fix his car next weekend. Joseph was so happy for his brother, Larry, getting a promotion. Gordon and his uncle Joseph went to the movies Sunday after church and got popcorn and candy while they were there. Who is Jami to Ethel? Let&#x27;s do it step by step:</td><td>1. Jami is Gino&#x27;s sister. 2. Ethel is Gino&#x27;s mother. 3. Therefore, Jami is Ethel&#x27;s daughter-in-law.</td><td>niece</td></tr></table>
+
+Table 8: Qualitative analysis of GPT-3 Zero-Shot-CoT on the CLUTRR dataset. The novelty comes from the sentence marked in orange. Queries that are of interest are marked in blue. Correct answer in the output is marked green and incorrect ones are marked red.
+
+<table><tr><td>Confidence</td><td>Rule</td></tr><tr><td>1.154</td><td>mother(A,B) ← sister(A,C) ∧ mother(C,B)</td></tr><tr><td>1.152</td><td>daughter(A,B) ← daughter(A,C) ∧ sister(C,B)</td></tr><tr><td>1.125</td><td>sister(A,B) ← daughter(A,C) ∧ aunt(C,B)</td></tr><tr><td>1.125</td><td>father(A,B) ← brother(A,C) ∧ father(C,B)</td></tr><tr><td>1.123</td><td>granddaughter(A,B) ← grandson(A,C) ∧ sister(C,B)</td></tr><tr><td>1.120</td><td>brother(A,B) ← sister(A,C) ∧ brother(C,B)</td></tr><tr><td>1.117</td><td>brother(A,B) ← son(A,C) ∧ uncle(C,B)</td></tr><tr><td>1.105</td><td>brother(A,B) ← daughter(A,C) ∧ uncle(C,B)</td></tr><tr><td>1.104</td><td>daughter(A,B) ← wife(A,C) ∧ daughter(C,B)</td></tr><tr><td>1.102</td><td>mother(A,B) ← brother(A,C) ∧ mother(C,B)</td></tr><tr><td>1.102</td><td>brother(A,B) ← father(A,C) ∧ son(C,B)</td></tr><tr><td>1.096</td><td>sister(A,B) ← mother(A,C) ∧ daughter(C,B)</td></tr><tr><td>1.071</td><td>sister(A,B) ← father(A,C) ∧ daughter(C,B)</td></tr><tr><td>1.071</td><td>son(A,B) ← son(A,C) ∧ brother(C,B)</td></tr><tr><td>1.070</td><td>uncle(A,B) ← father(A,C) ∧ brother(C,B)</td></tr><tr><td>1.066</td><td>daughter(A,B) ← son(A,C) ∧ sister(C,B)</td></tr><tr><td>1.061</td><td>brother(A,B) ← brother(A,C) ∧ brother(C,B)</td></tr><tr><td>1.056</td><td>grandson(A,B) ← husband(A,C) ∧ grandson(C,B)</td></tr><tr><td>1.055</td><td>sister(A,B) ← son(A,C) ∧ aunt(C,B)</td></tr><tr><td>1.053</td><td>grandmother(A,B) ← sister(A,C) ∧ grandmother(C,B)</td></tr><tr><td>1.050</td><td>granddaughter(A,B) ← granddaughter(A,C) ∧ sister(C,B)</td></tr><tr><td>1.050</td><td>grandmother(A,B) ← brother(A,C) ∧ grandmother(C,B)</td></tr><tr><td>1.047</td><td>grandson(A,B) ← granddaughter(A,C) ∧ brother(C,B)</td></tr><tr><td>1.046</td><td>grandfather(A,B) ← mother(A,C) ∧ father(C,B)</td></tr><tr><td>1.036</td><td>son(A,B) ← daughter(A,C) ∧ brother(C,B)</td></tr><tr><td>1.035</td><td>sister(A,B) ← brother(A,C) ∧ sister(C,B)</td></tr><tr><td>1.029</td><td>grandmother(A,B) ← mother(A,C) ∧ mother(C,B)</td></tr><tr><td>1.027</td><td>grandfather(A,B) ← sister(A,C) ∧ grandfather(C,B)</td></tr><tr><td>1.019</td><td>brother(A,B) ← mother(A,C) ∧ son(C,B)</td></tr><tr><td>1.017</td><td>granddaughter(A,B) ← wife(A,C) ∧ granddaughter(C,B)</td></tr></table>
+
+Table 9: Showcase of the learnt logic rules with top@30 confidence of DSR-LM rule learning.
+
+```txt
+// question :: (sub, obj) represents a question asking about relation
+// between `sub` and `obj`
+type question(sub: String, obj: String)
+// context :: (rela, sub, obj) represents there is a `rela`
+// between `sub` and `obj`
+type kinship(rela: usize, sub: String, obj: String)
+// Composition rule :: (r1, r2, r3) represents compositing r1 and r2 yields r3
+type composite(r1: usize, r2: usize, r3: usize)
+// Constants used for defining relation properties
+const DAUGHTER = 0, SISTER = 1, ..., MOTHER_IN_LAW = 19
+const MALE = 0, FEMALE = 1
+type gender(r: usize, gender_id: i32)
+rel gender = { (DAUGHTER, FEMALE), (SISTER, FEMALE), ..., (MOTHER_IN_LAW, FEMALE)}
+```
+
+Figure 5: Full Scallop program including deductive rules and integrity constraints
+
+# A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Left blank.
+
+A2. Did you discuss any potential risks of your work?
+
+Left blank.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Left blank.
+
+□ A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+# B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+Left blank.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Left blank.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Left blank.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Left blank.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Left blank.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+Left blank.
+
+# C Did you run computational experiments?
+
+Left blank.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Left blank.
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank.
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank.

2023/Improved Logical Reasoning of Language Models via Differentiable Symbolic Programming/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3cd6946d5dfd79d838779dde4bc4f098a4d275b2a3685330207d3d195e475ca
+size 1001197

2023/Improved Visual Story Generation with Adaptive Context Modeling/6027a132-d9ee-4656-b4da-85df0a85a689_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fdc57888aee81a1c606f3553b97f2ebce59b81311810117daca2ec8285b937b
+size 3341825

2023/Improved Visual Story Generation with Adaptive Context Modeling/full.md

@@ -0,0 +1,557 @@
+# Improved Visual Story Generation with Adaptive Context Modeling
+
+Zhangyin Feng<sup>1</sup>, Yuchen Ren<sup>2</sup>, Xinmiao Yu<sup>1</sup>, Xiaocheng Feng<sup>1,3</sup>, Duyu Tang, Shuming Shi, Bing Qin<sup>1,3</sup>
+
+<sup>1</sup> Harbin Institute of Technology,
+
+2 Renmin University of China,
+
+3 Peng Cheng Laboratory
+
+{zyfeng, xmyu, xcfeng, qinb} @ir.hit.edu.cn
+
+siriusren@ruc.edu.cn
+
+# Abstract
+
+Diffusion models developed on top of powerful text-to-image generation models like Stable Diffusion achieve remarkable success in visual story generation. However, the best-performing approach considers historically generated results as flattened memory cells, ignoring the fact that not all preceding images contribute equally to the generation of the characters and scenes at the current stage. To address this, we present a simple method that improves the leading system with adaptive context modeling, which is not only incorporated in the encoder but also adopted as additional guidance in the sampling stage to boost the global consistency of the generated story. We evaluate our model on PororoSV and FlintstonesSV datasets and show that our approach achieves state-of-the-art FID scores on both story visualization and continuation scenarios. We conduct detailed model analysis and show that our model excels at generating semantically consistent images for stories.
+
+# 1 Introduction
+
+Diffusion models trained on broad text-image data (Rombach et al., 2022; Bao et al., 2022; Feng et al., 2022; Ramesh et al., 2022; Sahara et al., 2022; Balaji et al., 2022; Nichol et al., 2021) achieved remarkable success in text-to-image generation and showed strong abilities to synthesize photorealistic images of high resolution and great semantic consistency to text prompts. Such a huge success drives the extension of modern diffusion text-to-image models into more scenarios like visual story generation, which is to generate a series of images for a story of multiple sentences.
+
+A recent work, AR-LDM (Pan et al., 2022), which is built upon open-sourced Stable Diffusion, achieves the state-of-the-art FID on the benchmark datasets for visual story generation. AR-LDM encodes previous text-image context as a sequence of additional conditions, which is then attended by
+
+![](images/6740c8ae4606cadeb805e01ab28b87fb1697c1b43471309869c0a0cc212a9477.jpg)  
+Figure 1: A motivating example of a story with five sentences. Blue and purple lines indicate the dependencies between images.
+
+the UNet decoder for image generation. Despite its remarkable success, one limitation is that previous text-image pairs of the same story are all flattened as conditioning memories. This is different from the fact that not all the scenes/characters of sentences in the same story are closely related. Take Figure 1 as an example. The scene of the fourth sentence is not related to either the second or the third sentence. On the contrary, the generation of the fifth image should depend more on the second and third images than others. From this example, we can see that the dependency between images could be largely measured by the semantic relations between sentences.
+
+In this work, we present a simple approach that selectively adopts historical text-image data from the same story in the generation of an image. Specifically, we freeze the text and image representations produced by off-the-shell encoders, and adaptively compute conditioning vectors of context
+
+by considering the semantic relation between the current sentence and all the history. Such resulting conditioning vectors will be queried by UNet in a traditional way. Furthermore, based on the consideration that images should have similar scenes and characters if their corresponding sentences are similar, we further add context-aware guidance like the use of classifier guidance or CLIP guidance (Nichol et al., 2021) in standard text-to-image generation.
+
+To validate the effectiveness of our approach, we evaluate our model on story visualization and continuation tasks. Experimental results on PororoSV and FlintstonesSV datasets show that both adaptive encoder and guidance improve the quality of the generated images as well as the global consistency of the visual story. The contributions of this work are as follows:
+
+- We present a diffusion model that adaptively uses context information in the encoder and sampling guidance for visual story generation.  
+- Our approach achieves state-of-the-art results on benchmark datasets for both story visualization and continuation tasks.  
+- We show that our model excels at generating semantically consistent images for stories.
+
+# 2 Related work
+
+# 2.1 Text-to-Image Generation
+
+We group modern text-to-image generation approaches into three categories. The first category is generative adversarial network (Goodfellow et al., 2014; Reed et al., 2016; Zhang et al., 2017). They jointly learn a generator and a discriminator, where the generator is trained to generate images to fool the discriminator and the discriminator is trained to distinguish between real and (generated) fake images. The second category is encoder-decoder plus discrete variational autoencoder (dVAE). Methods are developed based on a well-trained discrete variational autoencoder (Van Den Oord et al., 2017), which is capable of mapping an image to discrete tokens and reconstructing an image from discrete tokens. Thus, the task of text-to-image generation could be viewed as a special translation task that converts natural language tokens to image tokens. Autoregressive models (Ramesh et al., 2021a; Ding et al., 2021; Gafni et al., 2022; Yu et al., 2022) typically use Transformer (Vaswani et al., 2017) to
+
+generate a visual token conditioned on the previously generated tokens, resulting in high latency in the inference stage. Muse (Chang et al., 2023) is a non-autoregressive model that tremendously speeds up the inference stage by generating image tokens in parallel. The third category is diffusion models — image generation is considered as an iterative refinement process, where two ends of the spectrum are the Gaussian noise and the real image, respectively. Some studies adopt a variational autoencoder to compress an image to the latent space and learn the diffusion process in the latent space of images (Rombach et al., 2022; Bao et al., 2022; Feng et al., 2022). Some works (Ramesh et al., 2022; Sahara et al., 2022; Balaji et al., 2022) directly learn the diffusion model over pixels and typically include cascaded up-sampling models (e.g., from  $64 \times 64$  to  $256 \times 256$  and from  $256 \times 256$  to  $1024 \times 1024$ ) to produce high-resolution images.
+
+# 2.2 Visual Story Generation
+
+Visual story generation includes two settings: story visualization and story continuation. Story visualization was firstly introduced by Li et al. (2019), who proposes the StoryGAN model for sequential text-to-image generation. Based on the GAN network, they proposed to combine image and story discriminators for adversarial learning. To improve the global consistency across dynamic scenes and characters in the story, Zeng et al. (2019) jointly considers story-to-image-sequence, sentence-to-image, and word-to-image-patch alignment by proposing an aligned sentence encoder and attentional word encoder. Li et al. (2020) includes dilated convolution in the discriminators to expand the receptive field of the convolution kernel in the feature maps and weighted activation degree to provide a robust evaluation between images and stories. To improve the visual quality, coherence and relevance of generated images, Maharana et al. (2021a) extends the GAN structure by including a dual learning framework that utilizes video captioning to reinforce the semantic alignment between the story and generated images, and a copy-transform mechanism for sequentially consistent story visualization. Maharana and Bansal (2021a) improves the generation quality by incorporating constituency parse trees, commonsense knowledge, and visual structure via bounding boxes and dense captioning. Unlike the story visualization task, whose input only contains the text story, the story continuation
+
+![](images/c7bf92c824373fae2c2a8218d997695e88f1519568251ded9c7d67410548ff3e.jpg)  
+Figure 2: An overview of our model architecture. Based on the latent diffusion model (Rombach et al., 2022), we propose an adaptive encoder and an adaptive guidance. Adaptive encoder is used to get the adaptive context vectors. Conditional diffusion module transforms context vectors to image. Adaptive guidance aims to guide diffusion sampling process with adaptive context information.
+
+task also includes the first image as input. Maharana et al. (2022) introduces story continuation and modifies the pre-trained text-to-image model DALL-E (Ramesh et al., 2021b) by adding a cross attention module for story continuation. Pan et al. (2022) employs a history-aware encoding to incorporate previously generated text-image history to diffusion model for visual story generation.
+
+# 3 Model
+
+We introduce our approach in this section. We first present the model architecture of our approach (§3.1), and then describe three important components: adaptive encoder (§3.2), conditional diffusion model (§3.3) and adaptive guidance (§3.4).
+
+# 3.1 Model Architecture
+
+An overview of the approach is depicted in Figure 2. It includes an adaptive encoder, a conditional diffusion model, and an adaptive guidance. Based on current text prompt and historical text-image context, the adaptive encoder represents them as conditional vectors. Then the conditional diffusion model transforms these vectors into the correspond
+
+ing image. During the diffusion sampling process, the adaptive guidance component further guides each diffusion step by comparing it to similar preceding images in the current story to enhance the global consistency of the generated images.
+
+# 3.2 Adaptive Encoder
+
+Given a story  $\mathbf{S}$  which consists of a sequence of text prompts:  $\mathbf{S} = \{s_1, s_2, \dots, s_L\}$ . Story visualization aims to generate a sequence of images  $\mathbf{X} = \{x_1, x_2, \dots, x_L\}$ . Each image corresponds to a text prompt. Different from text-to-image generation, which only generates one isolated image for the text prompt, story visualization requires global consistency between the generated images. A natural idea is to combine historical text-image context when generating the current image.
+
+$$
+\begin{array}{l} P (\mathbf {X} | \mathbf {S}) = \prod_ {i = 1} ^ {L} P (x _ {i} | \hat {x} _ {<   i}, \mathbf {S}) \\ = \prod_ {i = 1} ^ {L} P (x _ {i} | \tau_ {\theta} (\hat {x} _ {<   i}, s _ {\leq i})) \\ \end{array}
+$$
+
+![](images/cfbd7c405fb2458cda24fb4a655a8c00f3b342014e28786c0f5f671d983728f1.jpg)  
+Figure 3: Adaptive encoder consists of three modules: (1) CLIP text encoder is used to encode text prompt. (2) BLIP encoder is used to encode historical text-image pair. (3) Cross attention module is used to filter useful information adaptively.
+
+where  $\tau_{\theta}$  denotes the history-aware conditioning encoder.
+
+As shown in Figure 1, we find that some images in the history of the same story are similar to the current image, and some images are even completely irrelevant. The purpose of the adaptive encoder is to automatically find the relevant historical text-image pairs, and then encode them into the condition vectors. As shown in Figure 3, adaptive encoder consists of a CLIP text encoder, a BLIP text-image encoder and a cross attention module. Both CLIP (Radford et al., 2021) and BLIP (Li et al., 2022a) are multimodal pre-trained models. The difference is that CLIP encodes text and image respectively, and BLIP can jointly represent text-image pair. We use CLIP to get the current text prompt vector  $v_{i}$ , and BLIP to get the historical vectors  $\{h_0,\dots,h_{i - 1}\}$ . Then a cross attention is equipped to filter history information and we can obtain the updated vectors  $\{\hat{h}_0,\dots,h_{i - 1}\}$ . In the cross attention module, the text vector  $v_{i}$  is the query, and each historical vector  $h_{< i}$  is the key and value. Finally, we concatenate the current text vector and history vectors to get the final condition vector  $c = [v_i;\hat{h}_0;\dots;\hat{h}_{i - 1}]$ .
+
+# 3.3 Conditional Diffusion Model
+
+Denoising diffusion probabilistic models are a class of score-based generative models, which have recently gained traction in the field of text-to-image generation (Ho et al., 2020). A diffusion model typically contains forward and reverse processes. Given a data  $x_0$  sampled from a real-world data distribution  $q(x)$ , the forward process is imple
+
+mented as a predefined Markov chain that gradually corrupts  $x_0$  into an isotropic Gaussian distribution  $x_{T} \sim \mathcal{N}(0, I)$  in  $T$  steps:
+
+$$
+x _ {t} = \sqrt {\alpha_ {t}} x _ {t - 1} + \sqrt {1 - \alpha_ {t}} \epsilon_ {t}, t \in \{1, \ldots , T \}
+$$
+
+where  $\epsilon_t \sim \mathcal{N}(0, I)$ , and  $\{\alpha_t \in (0, 1)\}_{t=1}^T$  is a predefined noise variance schedule. The reverse process aims to learn a denoising network  $\epsilon_\theta(\cdot)$  to reconstruct the data distribution  $x_0$  from the Gaussian noise  $x_T \sim \mathcal{N}(0, I)$ . We can express an arbitrary sample  $x_t$  from the initial data  $x_0$ :
+
+$$
+x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon ,
+$$
+
+where  $\bar{\alpha}_{t} = \prod_{i=1}^{t} \alpha_{i}$  and  $\epsilon \sim \mathcal{N}(0, I)$ . The denoising network  $\epsilon_{\theta}(\cdot)$  is trained to recover  $x_{0}$  by predicting the noise  $\epsilon_{\theta}(x_{t}, t)$ . The corresponding learning objective can be formalized as a simple mean-squared error loss between the true noise and the predicted noise:
+
+$$
+\mathcal {L} = \mathbb {E} _ {x _ {0}, \epsilon , t, c} \left[ | | \epsilon - \epsilon_ {\theta} (x _ {t}, t, c) | | _ {2} ^ {2} \right],
+$$
+
+where  $t$  is uniformly sampled from  $\{1,\dots,T\}$ ,  $c$  is condition and  $\epsilon \sim \mathcal{N}(0,I)$ .
+
+The denoising network  $\epsilon_{\theta}(\cdot)$  is typically implemented by U-Net (Ho et al., 2020). To make the diffusion process conditional on the input, condition  $c$  is fed into  $\epsilon_{\theta}(\cdot)$  via a cross-attention layer implementing  $Attention(Q,K,V) = softmax(\frac{QKT}{\sqrt{d}})\cdot V$ , where the intermediate representations of the U-Net acting as the query  $Q$ , and the condition embeddings  $c$  acting as the key  $K$  and value  $V$ .
+
+Classifier-free guidance (Ho and Salimans, 2022) is a widely used technique to improve sample quality while reducing diversity in conditional diffusion models, which jointly trains a single diffusion model on conditional and unconditional objectives via randomly dropping  $c$  during training (e.g. with  $10\%$  probability). During sampling, the output of the model is extrapolated further in the direction of  $\epsilon_{\theta}(x_t|c)$  and away from  $\epsilon_{\theta}(x_t|\emptyset)$  as follows:
+
+$$
+\hat {\epsilon} _ {\theta} (x _ {t} | c) = \epsilon_ {\theta} (x _ {t} | \emptyset) + \gamma \cdot (\epsilon_ {\theta} (x _ {t} | c) - \epsilon_ {\theta} (x _ {t} | \emptyset))
+$$
+
+where  $\gamma \geq 1$  is the guidance scale.
+
+# 3.4 Adaptive Guidance
+
+Previous work (Dhariwal and Nichol, 2021; Nichol et al., 2021; Li et al., 2022b) in text-to-image generation have explored to utilize a classifier or a CLIP (Radford et al., 2021) model to improve a diffusion generator. A CLIP model consists of two separate pieces: an image encoder and a caption encoder. The model optimizes a contrastive cross-entropy loss that encourages a high dot-product if the image is paired with the given caption, or a low dot-product if the image and caption correspond to different parts in the training data. The denoising diffusion process can be perturbed by the gradient of the dot product of the image and caption.
+
+One of the primary challenges of visual story generation is to maintain consistent background and character appearances throughout the story. In order to achieve this goal, during the diffusion sample stage, we propose an adaptive guidance, which explicitly requires that the image generated later should be consistent with the preceding generated images. Considering that images whose corresponding sentences are similar should have similar scenes/characters. When generating the image  $x_{i}$  in the story, we first use clip text encoder to calculate the similarity score for each historical text in  $\{s_1,\dots,s_{i - 1}\}$  with the current text  $s_i$ . After that, we select the text-image pair with the highest similarity score. When the similarity score exceeds the threshold, we believe that the selected image and the image to be generated currently have high similarity, and we can use this image to guide the sampling process of the diffusion model. When the similarity score is lower than the threshold, we think that images in history are not similar to the image to be generated at present, and do not add sampling guidance.
+
+The previous CLIP guided model (Nichol et al., 2021) needs to train an additional noisy CLIP model. It's time and computation costly, and difficult to classify noisied image. Following UPainting (Li et al., 2022b), we use normal CLIP for guidance, and modify the CLIP inputs as follows:
+
+$$
+\hat {x} _ {0} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} (x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t})
+$$
+
+$$
+x _ {i n} = \sqrt {1 - \bar {\alpha} _ {t}} \hat {x} _ {0} + (1 - \sqrt {1 - \bar {\alpha} _ {t}}) x _ {t}
+$$
+
+The denoising diffusion process can be formulated as follows:
+
+$$
+\begin{array}{l} \hat {\epsilon} _ {\theta} ^ {\prime} \left(x _ {t} | c\right) = \epsilon_ {\theta} \left(x _ {t} | \emptyset\right) + \gamma \cdot \left(\epsilon_ {\theta} \left(x _ {t} | c\right) - \epsilon_ {\theta} \left(x _ {t} | \emptyset\right)\right) \\ - g \sqrt {1 - \bar {\alpha} _ {t}} \nabla_ {x _ {t}} (f (x _ {i n}) \cdot f (x _ {h})) \\ \end{array}
+$$
+
+where  $\gamma \geq 1$  is the classifier weight,  $g \geq 0$  is adaptive guidance weight,  $f(.)$  is the CLIP image encoder and  $x_{h}$  is the most similar image to the current image  $x_{t}$ .
+
+# 4 Experiments
+
+# 4.1 Datasets and Metrics
+
+We carried out experiments on both story visualization and story continuation tasks. Given a sequence of sentences forming a narrative, story visualization is the task of generating a corresponding sequence of images. Story continuation, including an initial ground truth image as input, is a variant of story visualization. We use two popular datasets, PororoSV (Li et al., 2019) and FlintstonesSV (Gupta et al., 2018), to evaluate our model. We give the statistics of two datasets in Table 1 and show main characters in Figure 4 and Figure 5 to help the reader understand our examples.
+
+<table><tr><td></td><td>Train</td><td>Valid</td><td>Test</td></tr><tr><td>PororoSV</td><td>10,191</td><td>2,334</td><td>2,208</td></tr><tr><td>FlintstonesSV</td><td>20,132</td><td>2,071</td><td>2,309</td></tr></table>
+
+Table 1: Statistics for PororoSV and FlintstonesSV datasets.
+
+We adopt the automatic evaluation metrics following existing works and report results using the evaluation script provided in prior work<sup>2</sup>.
+
+Frechet Inception Distance (FID) captures the level of similarity between two groups based on statistical analysis of visual features in their respective
+
+![](images/c5e584e079b192f27845aa6fec06774c277eb1e57ab359f5fa1186f45e868066.jpg)  
+Figure 4: Main characters in PororoSV dataset.
+
+![](images/8968d8bb9c38d96d095630d8758434f4e9b72a9e5a1525127158b39849dfebd5.jpg)  
+Figure 5: Main characters in FlintstonesSV dataset.
+
+raw images, using the inception v3 model. Lower FID scores indicate higher resemblance between the predicted images and the ground-truth images.
+
+Character F1 score calculates the proportion of characters present in the generated images that exactly match the characters in the story inputs. To achieve this, a pretrained inception v3 model (Szegedy et al., 2015) is fine-tuned on each dataset using a multi-label classification loss, enabling it to make predictions of characters in test images.
+
+Frame Accuracy evaluates whether all characters from a story are correctly represented in the corresponding images, utilizing the same model employed in the Character F1 score. While the Character F1 score measures the proportion of characters captured in a story, Frame Accuracy quantifies the percentage of samples where all characters are appropriately included.
+
+# 4.2 Implementation Details
+
+Our model is fine-tuned from the pre-traiend Stable Diffusion text-to-image generation model. We use CLIP base model and BLIP base model. We only train the parameters of diffusion model and cross attention module, and freeze the parameters of variational auto-encoder, CLIP and BLIP, which could speed up training and save GPU memory. Follow previous work, we train our model for 50 epochs. We use Adam optimizer and set learning rate to 1e-4. For  $\gamma$ , we used the default value in stable diffusion without adjustment. For the threshold of similarity score, we randomly sampled 50 stories
+
+to calculate the similarity score, then manually observed the relationship between the similarity score and the image similarity, and finally set the threshold to 0.65. For the  $g$ , we chose the best value 0.15 from 0.1, 0.15, 0.2, 0.5.
+
+# 4.3 Baselines
+
+StoryGAN (Li et al., 2019) uses the standard GAN technique, which includes a recurrent text encoder, an image generation module, and two discriminators - image and story discriminator.
+
+StoryGANc (Maharana et al., 2022) follows the general framework of the StoryGAN model and adds the source image as input for the story continuation task.
+
+CP-CSV (Song et al., 2020) tries to better preserve character information with three modules: story and context encoder, figure-ground segmentation, and figure-ground aware generation.
+
+DUCO-StoryGAN (Maharana et al., 2021b) utilizes a video captioning model to generate an additional learning signal forcing the alignment of image and text, and a memory-augmented transformer to model complex interactions between frames.
+
+VLC-StoryGAN (Maharana and Bansal, 2021b) incorporates constituency parse trees, commonsense information and visual information, including bounding boxes and dense captioning, to enhance the visual quality and image consistency.
+
+Word-Level (Li and Lukasiewicz, 2022) incorporates word information and extends word-level spatial attention to focus on all words and visual spatial locations in the entire story.
+
+StoryDALL-E (Maharana et al., 2022) modifies the pre-trained text-to-image model DALL-E by adding a cross attention module for story continuation.
+
+AR-LDM (Pan et al., 2022) employs a history-aware encoding module to incorporate the current text prompt and previously generated text-image history to diffusion model for visual story generation.
+
+# 5 Results
+
+# 5.1 Story Visualization
+
+We evaluate our model on PororoSV dataset for story visualization task. Results are shown in Table 3. We can observe that diffusion-based model
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">PororoSV</td><td colspan="3">FlintstonesSV</td></tr><tr><td>FID ↓</td><td>Char-F1↑</td><td>F-Acc↑</td><td>FID ↓</td><td>Char-F1↑</td><td>F-Acc↑</td></tr><tr><td>StoryGANc(BERT)</td><td>72.98</td><td>43.22</td><td>17.09</td><td>91.37</td><td>70.45</td><td>55.78</td></tr><tr><td>StoryGANc (CLIP)</td><td>74.63</td><td>39.68</td><td>16.57</td><td>90.29</td><td>72.80</td><td>58.39</td></tr><tr><td>StoryDALL-E(prompt)</td><td>61.23</td><td>29.68</td><td>11.65</td><td>53.71</td><td>42.48</td><td>32.54</td></tr><tr><td>StoryDALL-E (finetuning)</td><td>25.90</td><td>36.97</td><td>17.26</td><td>26.49</td><td>73.43</td><td>55.19</td></tr><tr><td>MEGA-StoryDALL-E</td><td>23.48</td><td>39.91</td><td>18.01</td><td>23.58</td><td>74.26</td><td>54.68</td></tr><tr><td>AR-LDM</td><td>17.40</td><td>-</td><td>-</td><td>19.28</td><td>-</td><td>-</td></tr><tr><td>ACM-VSG (Ours)</td><td>15.36</td><td>45.71</td><td>22.62</td><td>18.41</td><td>94.95</td><td>88.89</td></tr></table>
+
+Table 2: Results on the test sets of PororoSV and FlintstonesSV datasets from various models. Scores are based on FID, character classification F1, and frame accuracy evaluations.  
+
+<table><tr><td>Id</td><td>Text</td><td>Gold</td><td>Ours</td><td>AR-LDM</td></tr><tr><td>1</td><td>Barney is standing in a room. He speaks and looks tired.</td><td></td><td></td><td></td></tr><tr><td>2</td><td>Fred and Barney are in jail. Fred is explaining something to Barney while the two of them are standing in a cell behind bars.</td><td></td><td></td><td></td></tr><tr><td>3</td><td>Fred and Barney are in jail. Fred opens his arms and speaks. Then Barney responds.</td><td></td><td></td><td></td></tr><tr><td>4</td><td>Wilma and Betty are walking through a yard together.</td><td></td><td></td><td></td></tr><tr><td>5</td><td>Wilma and Betty are happily walking next to each other outside. Betty is talking while Wilma is listening.</td><td></td><td></td><td></td></tr></table>
+
+Figure 6: Example of generated images from previous model AR-LDM and our proposed model.  
+
+<table><tr><td>Model</td><td>FID ↓</td></tr><tr><td>StoryGAN</td><td>158.06</td></tr><tr><td>CP-CSV</td><td>149.29</td></tr><tr><td>DUCO-StoryGAN</td><td>96.51</td></tr><tr><td>VLC-StoryGAN</td><td>84.96</td></tr><tr><td>VP-CSV</td><td>65.51</td></tr><tr><td>Word-Level SV</td><td>56.08</td></tr><tr><td>AR-LDM</td><td>16.59</td></tr><tr><td>ACM-VSG (Ours)</td><td>15.48</td></tr></table>
+
+Table 3: Story visualization FID score results on PororoSV dataset.
+
+outperforms the prior methods by a large margin, and our proposed ACM-VSG achieves the best FID score 15.48, indicating our model is able to generate high-quality images.
+
+# 5.2 Story Continuation
+
+Table 2 shows the results for story continuation task. As we can see, our model can achieve the best results on both datasets, 15.36 and 18.41 FID for PororoSV and FlintstonesSV, respectively. And our model can greatly preserve characters to improve the consistency of the story. In addition, we show an example on FlintstonesSV and pororoSV dataset in Figure 6 and Figure 7. We can observe that our model is able to maintain the text-image alignment and consistency across images.
+
+# 5.3 Ablation Study
+
+Table 4 shows ablation studies to ensure that each component in the our proposed method benefits visual story generation. -Guidance means removing the adaptive guidance. -Attention means removing the cross attention module in the adaptive encoder.
+
+<table><tr><td>Id</td><td>Text</td><td>Gold</td><td>Ours</td><td>AR-LDM</td></tr><tr><td>1</td><td>Eddy keep holding his picture and explains his idea to Poby. Because of the snowy weather we have no choice but to rescue Pororo and Crong by airship. Therefore Eddy resorts to Poby that they need Poby&#x27;s help.</td><td></td><td></td><td></td></tr><tr><td>2</td><td>Poby seems surprised because Poby doesn&#x27;t expect that Poby will be needed in this situation. After hearing from Eddy Poby turns his head to the left side.</td><td></td><td></td><td></td></tr><tr><td>3</td><td>Pororo and Crong are in the middle of the mountain. They seem tired and exhausted. Pororo close his eyes with long hard thinking.</td><td></td><td></td><td></td></tr><tr><td>4</td><td>Pororo closes his eyes with long hard thinking. The weather is snowy and it becomes worse. Pororo can&#x27;t find any other solutions except rope to get out of this mountain. Pororo and Crong are stuck in this mountain so Pororo tries to use rope.</td><td></td><td></td><td></td></tr><tr><td>5</td><td>Pororo and Crong try to pull the rope to overcome this situation. However it is hard to fully apply their force.</td><td></td><td></td><td></td></tr></table>
+
+Figure 7: Example of generated images from previous model AR-LDM and our proposed model.  
+
+<table><tr><td>Model</td><td>FID ↓</td><td>Char-F1↑</td><td>F-ACC↑</td></tr><tr><td>ACM-VSG</td><td>15.36</td><td>45.71</td><td>22.62</td></tr><tr><td>- Guidance</td><td>15.96</td><td>44.56</td><td>22.13</td></tr><tr><td>- Attention</td><td>16.88</td><td>44.27</td><td>20.25</td></tr></table>
+
+Table 4: Ablation study results for story continuation task on PororoSV.  
+
+<table><tr><td>Id</td><td>Text</td><td>Gold</td><td>Ours</td></tr><tr><td>1</td><td>The green lizard is in the kitchen. He is being held by someone using him as a knife. He speaks to the camera.</td><td></td><td></td></tr><tr><td>2</td><td>Betty is talking on the phone in the bedroom, lying in bed.</td><td></td><td></td></tr><tr><td>3</td><td>Betty is laying in bed in the bedroom.</td><td></td><td></td></tr><tr><td>4</td><td>Wilma is in the room speaking on the phone.</td><td></td><td></td></tr><tr><td>5</td><td>Wilma is in the living room. She is standing by the phone while speaking into the phone receiver.</td><td></td><td></td></tr></table>
+
+# 5.4 Error Analysis
+
+Our model significantly improves the performance of visual story generation, but there are still some limits. In order to solve these limitations in fu
+
+Figure 8: The generated image story has global inconsistency.  
+
+<table><tr><td></td><td>Text</td><td>Gold</td><td>Ours</td></tr><tr><td>1</td><td>Loopy pushes her drink. loopy says Loopy won&#x27;t drink juice.</td><td></td><td></td></tr><tr><td>2</td><td>Fred and Wilma are standing in a room. Fred speaks while Wilma holds onto his shoulder.</td><td></td><td></td></tr></table>
+
+Figure 9: The bad cases for repetitive character and character action error.
+
+ture work, we analyze the generated images. We randomly sample hundreds of stories from PororoSV and FlintstonesSV datasets and summarize the errors.
+
+Story Inconsistency. As shown in Figure 8, there are inconsistencies across generated images. The style of the bed and the color of the quilt between the second and third generated images are inconsistent. The environment around Wilma is inconsistent between the fourth and fifth images.
+
+Repetitive Character. As shown in Figure 9 case 1, the model may generate the same character repeatedly if it appears multiple times in the text.
+
+Character Action Error. As shown in Figure 9 case 2, the subtle actions of characters in the image are misaligned with the text. In case 2, Wilma holds
+
+onto Fred's shoulder in text, while Fred holds onto Wilma's shoulder in the generated image.
+
+# 6 Conclusion
+
+In this paper, we explore an effective adaptive context modeling method to improve visual story generation. First, we use an adaptive encoder to select the context closely related to the current image from the historical text-image pairs using the current text. Then we fed the context vectors and the text vector to the diffusion model, and use an adaptive guidance to guide the generation of the current image. Experimental results verify that adaptive context modeling could help generate higher quality images and more consistent stories. In addition, we analyze the generated images and find potential research directions in the future: (1) focus on the global consistency of the story, (2) pay attention to the action and expression of the character, (3) obtain the exact semantics of long stories.
+
+# Limitations
+
+A limitation of this work is that it is only evaluated on synthesized datasets of cartoons with limited characters and scenes. In the real world application, there might be many different scenes/characters, posing new challenges to the proposed approach. Another limitation is the requirement of supervised training data and resources. Despite the number of trainable parameters of our approach (850M) is less than AR-LDM ( $\sim 1.5\mathrm{B}$ ), the model still needs many story-level training data and computing resources.
+
+# Acknowledgements
+
+Xiaocheng Feng is the corresponding author of this work. We thank the anonymous reviewers for their insightful comments. Zhangyin Feng, Xinmiao Yu, Xiaocheng Feng and Bing Qin are supported by the National Key R&D Program of China via grant 2020AAA0106502, National Natural Science Foundation of China (NSFC) via grant 62276078, the Key R&D Program of Heilongjiang via grant 2022ZX01A32 and the International Cooperation Project of PCL, PCL2022D01.
+
+# References
+
+Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. 2022. edifi: Text-to-image diffusion models with
+
+an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324.  
+Fan Bao, Chongxuan Li, Yue Cao, and Jun Zhu. 2022. All are worth words: a vit backbone for score-based diffusion models. arXiv preprint arXiv:2209.12152.  
+Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. 2023. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704.  
+Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794.  
+Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. 2021. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835.  
+Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. 2022. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257.  
+Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. 2022. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131.  
+Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. Advances in neural information processing systems, 27.  
+Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. 2018. Imagine this! scripts to compositions to videos. arXiv: Computer Vision and Pattern Recognition.  
+Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851.  
+Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598.  
+Bowen Li and Thomas Lukasiewicz. 2022. Word-level fine-grained story visualization.  
+Chunye Li, Liya Kong, and Zhiping Zhou. 2020. Improved-storygan for sequential images visualization. Journal of Visual Communication and Image Representation, 73:102956.
+
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022a. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086.  
+Wei Li, Xue Xu, Xinyan Xiao, Jiachen Liu, Hu Yang, Guohao Li, Zhanpeng Wang, Zhifan Feng, Qiaoqiao She, Yajuan Lyu, et al. 2022b. Upainting: Unified text-to-image diffusion generation with cross-modal guidance. arXiv preprint arXiv:2210.16031.  
+Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. 2019. Storygan: A sequential conditional gan for story visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338.  
+Adyasha Maharana and Mohit Bansal. 2021a. Integrating visuospatial, linguistic, and commonsense structure into story visualization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6772-6786, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Adyasha Maharana and Mohit Bansal. 2021b. Integrating visuospatial, linguistic and commonsense structure into story visualization. arXiv: Computation and Language.  
+Adyasha Maharana, Darryl Hannan, and Mohit Bansal. 2021a. Improving generation and evaluation of visual stories via semantic consistency. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2427-2442, Online. Association for Computational Linguistics.  
+Adyasha Maharana, Darryl Hannan, and Mohit Bansal. 2021b. Improving generation and evaluation of visual stories via semantic consistency. arXiv: Computation and Language.  
+Adyasha Maharana, Darryl Hannan, and Mohit Bansal. 2022. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. In European Conference on Computer Vision, pages 70-87. Springer.  
+Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. 2021. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741.  
+Xichen Pan, Pengda Qin, Yuhong Li, Hui Xue, and Wenhu Chen. 2022. Synthesizing coherent story with auto-regressive latent diffusion models. arXiv preprint arXiv:2211.10950.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models
+
+from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR.  
+Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. 2022. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125.  
+Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021a. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR.  
+Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021b. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR.  
+Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. 2016. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR.  
+Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695.  
+Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487.  
+Yun Zhu Song, Zhi Rui Tam, Hung Jen Chen, Huiao Han Lu, and Hong-Han Shuai. 2020. Character-preserving coherent story visualization. Springer International Publishing eBooks.  
+Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. 2015. Rethinking the inception architecture for computer vision. arXiv: Computer Vision and Pattern Recognition.  
+Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning. Advances in neural information processing systems, 30.  
+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.  
+Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. 2022. Scaling autoregressive models for
+
+content-rich text-to-image generation. arXiv preprint arXiv:2206.10789.
+
+Gangyan Zeng, Zhaohui Li, and Yuan Zhang. 2019. Pororogan: an improved story visualization model on pororo-sv dataset. In Proceedings of the 2019 3rd International Conference on Computer Science and Artificial Intelligence, pages 155-159.
+
+Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. 2017. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915.
+
+# A PororoSV Cases
+
+# Case 1:
+
+1. Tongtong opens the door. Crong is now on the Pororo's car. They are entering the Tongtong's house. Tongtong tries to find where the magic wand is.  
+2. The pink magic wand is located behind the chair. Then the magic wand becomes to come out.  
+3. Tongtong finally finds out the magic wands.  
+4. As Tongtong finds out the magic wand Tongtong is confident to make Pororo normal. Tongtong says that Pororo will turn back to normal.  
+5. Pororo jumps high after Tongtong's promise. Tongtong asks Pororo not to mess up.
+
+# Case 2:
+
+1. Pororo and Crong is in Pororo's house. They are standing next to the bed. Pororo is pointing Crong. Crong looks sad.  
+2. Pororo and Crong is in Pororo's house standing next to the bed. Pororo is pointing drawer and Crong looks at it with a sad face.  
+3. Poby is in Pororo's house. Poby is approaching a drawer. Above the drawer there is a book which is slightly open.  
+4. Poby is in Pororo's house searching for something. Poby leans his head down to take a look.  
+5. Poby is in Pororo's house. Poby is thinking something standing still putting his right hand on his jaw.
+
+# Case 3:
+
+1. Pororo dust the snow off from Poby.  
+2. Petty is holding the green block.  
+3. Harry explains situation. Pororo walks toward chair.  
+4. Pororo sits and joins play.  
+5. Eddy and Crong are yelling.
+
+# Case 4:
+
+1. Poby looks at Harry and Harry is talking to Poby while sitting on Poby's shoulder.  
+2. Petty talks and smiles and mops the floor.  
+3. Petty is smiling and mobbing the floor.  
+4. Petty smiles and puts the stuff on the table.  
+5. Poby talks and opens Poby mouth.
+
+# Case 5:
+
+1. Pororo says it was so stinky.  
+2. feeling embarrassed Poby waved Poby hands.  
+3. Pororo thinks who farted just before.  
+4. everyone saw Crong pinching everyone noses.  
+5. Crong is sitting on the toilet.
+
+# Case 6:
+
+1. Poby is leaving Loopy house.  
+2. Poby says bye to Loopy.  
+3. Poby thinks Pororo might have fixed broken chair.  
+4. Poby smiles. Loopy walks toward the chair.  
+5. Loopy is satisfied with fixed chair.
+
+# Case 7:
+
+1. Poby is tired so Poby says to Harry that Poby wants to go to bed with sleepy eyes.  
+2. Harry is surprised. Harry looks at the window  
+3. there are two cactus on the shelf. outside the window is dark already.  
+4. light is turned off and Poby and Harry finish ready to sleep.  
+Harry say good night to Poby.  
+5. light is turned off and Poby and Harry finish ready to sleep. Poby lays down on the bed.
+
+# Case 8:
+
+1. Poby notices that someone is skiing down.  
+2. Pororo is skiing away. Loopy is chasing.  
+3. Pororo notices that Poby is waiting for Pororo.  
+4. Loopy and Poby are lying down. Eddy approaches.  
+5. Poby and Loopy stands up.
+
+# Case 9:
+
+1. Poby comes out of the house and find out his friends.  
+2. Poby explains to the friends that Poby must have fallen asleep inside.  
+3. Eddy is happy to see that Poby is also already at Eddy's house along with other friends.  
+4. Eddy is inviting his friends to come into his house. everyone follows him.  
+5. Eddy is leading his friends into his house. everyone is getting inside.
+
+# Case 10:
+
+1. Pororo and friends came running toward Poby. Poby is watching them.  
+2. Pororo and friends are talking to Poby. Poby has no idea why his friends are acting like this.  
+3. Harry sat on Poby's head. Harry is saying sorry to Poby. Poby looks surprised. Harry looks sad.  
+4. Harry is feeling guilty. meanwhile Poby has no idea what Harry is talking about. Harry is sitting on Poby's head.  
+5. Harry is feeling very sorry to Poby. Harry is talking to Poby on Poby's head.
+
+# Case 11:
+
+1. Poby is brushing pole to Poby's noses for making Poby
+
+itchy. Pororo tries sneezing Poby to take out of the air.
+
+2. Pororo shows that small helicopter is still working properly.  
+3. Pororo continually suggests Poby trying to sneeze again.  
+4. Poby tries to sneeze to get out of air from his body. However sneezing with Poby's free will is really difficult.  
+5. Poby tries to sneeze to get out of air from his body. However sneezing with Poby's free will is really difficult. Poby gives up sneezing and says Poby can't sneeze anymore.
+
+# Case 12:
+
+1. Poby was keep singing. Suddenly Poby falls down.  
+2. Poby feels ashamed and wants that nobody saw him falling down.  
+3. Seeing Poby through the telescope Eddy secretly smiles and talks to himself that Eddy saw Poby falling down.  
+4. Eddy is interested in seeing things and friends through telescope. Eddy brings telescope and goes to the mountain to observe his friends more.  
+5. Up on the mountain Eddy chooses a target. It is Pororo. Eddy looks through the telescope.
+
+# B FlintstonesSV Cases
+
+# Case 1:
+
+1.fred and barney stand outside holding blue lunch boxes.fred talks to barney  
+2.Fred stands in the kitchen having a friendly conversation with someone.  
+3.Fred is standing in a room. He is speaking while looking over his shoulder and smiling.  
+4.Fred is trying to kiss Wilma in a room. Wilma is holding a type of plant in her hand.  
+5. Wilma is in a living room adjusting the leafs on a house plant that is sitting on a table while talking then she stands up straight and turns her head.
+
+# Case 2:
+
+1.A Lounging creature is lounging around the room and talking.  
+2. Wilma is ironing a shirt in the laundry room.  
+3. Wilma is in the living room. Wilma is ironing. Wilma is bobbing her head.  
+4. Wilma is in a room. She talks to someone.  
+5. Wilma is in a room. She is talking.
+
+# Case 3:
+
+1.Fred sits in the living room and speaks to Barney, who waits to respond.  
+2.Barney stands and talks with someone in the living room.  
+3.Fred and Barney are in the quarry. Fred is speaking to Barney.  
+4. Fred and Barney look worried. Fred and Barney are behind the wall in the yard. Barney and Fred are talking to each other.
+
+5.Barney is talking to Fred outside behind a stone fence. Fred begins to slump down and look sad.
+
+# Case 4:
+
+1.Fred is riding in the car thinking and talking to himself.  
+2.Fred touches his chin then crosses his arm while outside.  
+3.Fred is walking outside while speaking out loud.  
+4.Fred and Barney are riding in the car with golf clubs strapped to the bumper.  
+5.Fred is standing in the room, talking to someone off screen left.
+
+# Case 5:
+
+1. Fred and Barney are standing outside next to the stone wall. Barney is wearing an outfit that makes him look like a boy scout. Fred says something to Barney and then points at him.  
+2.Fred and Barney are standing on a sidewalk. Barney is speaking to Fred, while Fred listens silently with his hands on his hips.  
+3.Barney is outside talking.  
+4.Fred and Barney stand in the yard. They speak to each other.  
+5. Fred and Barney are standing outside talking. They are in front of the wall and Barney has a hate on.
+
+# Case 6:
+
+1.Fred sits in the living room and speaks to Barney, who waits to respond.  
+2.Barney stands and talks with someone in the living room.  
+3.Fred and Barney are in the quarry. Fred is speaking to Barney.  
+4. Fred and Barney look worried. Fred and Barney are behind the wall in the yard. Barney and Fred are talking to each other.  
+5.Barney is talking to Fred outside behind a stone fence. Fred begins to slump down and look sad.
+
+# Case 7:
+
+1. Wilma is in the room, she is talking.  
+2. Wilma is in the dining room talking to someone then she starts to laugh.  
+3.Barney slides towards doorway, and opens door.  
+4. Wilma is sitting in the dining room at the table while talking on the phone.  
+5. Wilma is in the dining room. She sits at the table on the phone. Fred is wheeled into the room laying in a bed. As Fred enters, Wilma lowers the phone and looks at him with concern and surprise.
+
+# Case 8:
+
+1.Fred is in a living room kneeling next to a blue chair.  
+2.Pebbles and Fred are standing in a room talking. Pebbles turns her head and Fred shrugs his shoulders.  
+3.Fred and wilma talk in the bedroom, fred laughs in response to what Wilma says.  
+4.Fred makes an angry comment while sitting in the room.
+
+5.Fred stands in the kitchen with an ice block on his head. Then someone reaches up and pats the ice. Fred makes a face and the ice starts melting.
+
+# Case 9:
+
+1.Barney is talking in the living room.  
+2.Betty is standing in a room, hanging up balloons. She is talking to someone, as she hangs up a green balloon.  
+3.Betty is in a room. She stands on a stool and holds a balloon in one hand while talking to someone on the ground. The room is decorated for a party.  
+4. Wilma is in the living room. Wilma is talking.  
+5.There is a bird that has its head lowered in the room.
+
+# Case 10:
+
+1. Wilma and Betty are in the room. They are talking to one another while standing.  
+2.Betty and Wilma are standing in a room. Wilma has bones in her hair. Betty is talking to Wilma.  
+3. Wilma is wearing a bone curler in her hair while Betty talks to her in a room.  
+4.Betty and wilma are talking in a room.  
+5. Wilma and Betty are standing in a room by the window. They keep looking out the window while Wilma holds the curtain.
+
+# Case 11:
+
+1.Mr slate is driving his car and laughing.  
+2.A police officer in a police station sits at a desk and talks into a speaker while looking at a stack of papers. He turns the speaker away from his mouth.  
+3.A Small Policeman behind wheel and a Policeman with Brown Mustache sit in their police car blinking.  
+4. The officer that is driving the car is speaking to the officer with mustache.  
+5.Fred and Barney talk as they sit in the car.
+
+# Case 12:
+
+1.Barney is sitting outside in a chair reading out loud.  
+2.Barney is reading the news papers in the backyard.  
+3. The scene begins with no one in the picture. Barney emerges from hiding behind a stone wall that is in front of him. He is standing outside in the yard. The house is to his right. He says something and then points to himself with his thumb.  
+4.Barney is outside. He is sitting on a stone wall and is talking.  
+5.Fred and Barney are in the yard. Fred is yelling at Barney. Fred is holding Barney with a fist raised.
+
+![](images/8cfb626c4b6d21774281873bab3ed81b9cd29d5556b755bc6cc0a19cdaa74eef.jpg)  
+Figure 10: Example of ground truths (left 5 frames) and corresponding generated visual stories (right 5 frames) on PororoSV. These cases are under story continuation setting.
+
+![](images/8d505469491cbc79c07a6052dc03f55f444b5368b3fea5c8f841a7259b57a857.jpg)
+
+![](images/e1d66097c21eea5a6da4692e86a9eabb65b101f7c05b59ac2fb45378371aebe3.jpg)  
+Figure 11: Example of ground truths (left 5 frames) and corresponding generated visual stories (right 5 frames) on FlintstonesSV. These cases are under story continuation setting.
+
+![](images/2f04d93e8d057a280c5c7107b6e10753e3488b02e7688c972118c18dfb6e9978.jpg)
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? limitation  
+A2. Did you discuss any potential risks of your work? 4  
+A3. Do the abstract and introduction summarize the paper's main claims?  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used? No response.  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response.
+
+C Did you run computational experiments?
+
+4
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 4
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 4  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response.

2023/Improved Visual Story Generation with Adaptive Context Modeling/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ceedab79f4a3a5940165187122a56494396aa0de660033a5b5f8b286b5ea814
+size 1447808

2023/Improving Autoregressive Grammatical Error Correction with Non-autoregressive Models/3fd97eae-0b66-43ab-a8c2-6db714a9b22c_content_list.json

@@ -0,0 +1,1945 @@
+[
+    {
+        "type": "text",
+        "text": "Improving Autoregressive Grammatical Error Correction with Non-autoregressive Models",
+        "text_level": 1,
+        "bbox": [
+            174,
+            79,
+            823,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hang Cao<sup>1</sup>, Zhiquan Cao<sup>1</sup>, Chi Hu<sup>1</sup>, Baoyu Hou<sup>1</sup>, Tong Xiao<sup>1,2*</sup>, Jingbo Zhu<sup>1,2</sup>",
+        "bbox": [
+            287,
+            124,
+            719,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ NLP Lab, School of Computer Science and Engineering",
+        "bbox": [
+            268,
+            159,
+            732,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Northeastern University, Shenyang, China",
+        "bbox": [
+            329,
+            177,
+            670,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ NiuTrans Research, Shenyang, China",
+        "bbox": [
+            344,
+            193,
+            657,
+            209
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "caohang_neu@outlook.com",
+        "bbox": [
+            384,
+            210,
+            618,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{xiaotong,zhujingbo}@mail.neu.edu.cn",
+        "bbox": [
+            314,
+            225,
+            677,
+            242
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Grammatical Error Correction (GEC) aims to correct grammatical errors in sentences. We find that autoregressive models tend to assign low probabilities to tokens that need corrections. Here we introduce additional signals to the training of GEC models so that these systems can learn to better predict at ambiguous positions. To do this, we use a non-autoregressive model as an auxiliary model, and develop a new regularization term of training by considering the difference in predictions between the autoregressive and non-autoregressive models. We experiment with this method on both English and Chinese GEC tasks. Experimental results show that our GEC system outperforms the baselines on all the data sets significantly.",
+        "bbox": [
+            141,
+            279,
+            460,
+            521
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            533,
+            258,
+            549
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Grammatical Error Correction (GEC) has attracted much attention in recent years, which aims to correct grammatical errors in a given text automatically. It is widely applied to natural language processing scenarios such as Automatic Speech Recognition (ASR) (Kubis et al., 2020; Wang et al., 2020), writing assistant and language learning platforms, etc. The GEC task is characterized by a significant overlap between input and output sentences with only a few errors requiring modification.",
+        "bbox": [
+            112,
+            558,
+            489,
+            719
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Since the transformer-based autoregressive (Vaswani et al., 2017) (AR) model with sequence-to-sequence (seq2seq) architecture has been successful in many generation tasks, a few works (Chollampatt and Ng, 2018) have applied it to the GEC task by taking the incorrect text as the source language and the text without errors as the target language, which has become a mainstream paradigm. However, in the GEC task, the overlap of source and target sentences makes the AR model simply copy most of the tokens",
+        "bbox": [
+            112,
+            721,
+            489,
+            897
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/5d2520ac3d05f0056bd1f7dcad40a451fba27b0c9e65f1bc0c89a357658f82a7.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Incorrect</td><td colspan=\"2\">Correct</td></tr><tr><td rowspan=\"3\">Sub</td><td rowspan=\"3\">X: I have a apple.</td><td colspan=\"2\">Y: I have an apple.</td></tr><tr><td>0.92</td><td>0.85</td></tr><tr><td></td><td>0.42</td></tr><tr><td rowspan=\"3\">Del</td><td rowspan=\"3\">X: I have the an apple.</td><td colspan=\"2\">Y: I have an apple.</td></tr><tr><td>0.90</td><td>0.89</td></tr><tr><td></td><td>0.35</td></tr><tr><td rowspan=\"3\">Ins</td><td rowspan=\"3\">X: I have apple.</td><td colspan=\"2\">Y: I have an apple.</td></tr><tr><td>0.86</td><td>0.91</td></tr><tr><td></td><td>0.25</td></tr></table>",
+        "bbox": [
+            515,
+            250,
+            877,
+            350
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: Illustration for the confidence in different types of errors, where Sub denotes substitution, Del means deletion and Ins is insertion.",
+        "bbox": [
+            507,
+            360,
+            882,
+            403
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "over from the input to the output. We further find that the AR has high confidence for the tokens that are unchanged between the source and target sentence, while it usually has low confidence for correcting operations such as insertion, deletion, and substitution. Figure 1 is an example to illustrate this phenomenon. Intuitively, we believe that the reasonable cause of this phenomenon is the class imbalance issue (Li and Shi, 2021). With the influence of this problem, the AR model cannot confidently predict these incorrect tokens according to only the local context. Therefore, a natural idea is to improve the model performance by exploiting the global information, which can be captured by the non-autoregressive (NAR) (Gu et al., 2018; Lee et al., 2018) model. Although prior works have explored combining the two approaches through joint training, a combination for the GEC task is still missing. Besides, due to the inconsistency between AR and NAR output, a simple combination of them will lead to poor performance.",
+        "bbox": [
+            507,
+            434,
+            884,
+            785
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a simple yet novel approach to focus on incorrect tokens and integrate global information with the non-autoregressive model. Specifically, by masking the tokens in the golden target sentence corresponding to the low confidence positions of the AR output, we construct the input for NAR method. We combine the AR and NAR generation mechanisms to effectively",
+        "bbox": [
+            507,
+            790,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author.",
+        "bbox": [
+            141,
+            904,
+            289,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12014",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 12014-12027 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            220,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "utilize global information by constraining the consistency of their output distribution.",
+        "bbox": [
+            112,
+            84,
+            490,
+            116
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments on standard English GEC datasets and evaluate the system against strong baselines. Experimental results show that our approach can consistently achieve better results without relying on any resources other than the training data. Furthermore, we compare with a combination method of AR and NAR to verify whether the proposed model is more favorable for the GEC task. Here we use the Chinese GEC dataset as a benchmark to validate the generalization ability of the model. Meanwhile, we also conduct comparative ablation studies to illustrate the effectiveness of our proposed method.",
+        "bbox": [
+            112,
+            118,
+            489,
+            326
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            344,
+            270,
+            360
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Seq2seq for GEC In recent years, a number of Transformer-based AR methods have been developed for GEC tasks. Junczys-Dowmunt et al. (2018) adapt several methods from low-resource machine translation to GEC by regarding GEC as low-resource machine translation. Zhao et al. (2019) aim to copy the words that overlap between the source and target sentence. They propose a copy-augmented architecture for GEC task which is pre-trained with unlabeled data. A series of work focus on data augmentation (Grundkiewicz et al., 2019; Ge et al., 2018; Lichtarge et al., 2019), Xie et al. (2018) propose to synthesize \"realistic\" parallel corpus with grammatical errors by back-translation. Zhao and Wang (2020) add a dynamic masking method to the original source sentence during training, which enhances the model performance without requiring additional data. With the help of large pre-trained language models (Kaneko et al., 2020), the performance of Transformer based AR models can be improved effectively. Meanwhile, the NAR approach emerges as a competitive alternative, which can correct the errors by modeling the whole sentence information. Li and Shi (2021) apply a Conditional Random Fields (CRF) layer to conduct non-autoregressive sequence prediction by modeling the dependencies among neighbor tokens.",
+        "bbox": [
+            112,
+            374,
+            490,
+            826
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Combination of AR and NAR The combination of AR and NAR modeling mechanisms has been discussed in other tasks. Wei et al. (2019) use a pre-trained AR model to supervise the decoding state of NAR, which can alleviate the problem of",
+        "bbox": [
+            112,
+            838,
+            489,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "large search space. Li et al. (2019) propose that learning the hidden representation and attention distribution of AR by hints from the hidden representation can effectively improve the performance of NAR. Several approaches (Guo et al., 2020; Liu et al., 2020) are proposed to gradually guide the model transition from AR to NAR by designing the decoder input and semi-autoregressive tasks as courses. Some other works (Sun and Yang, 2020; Hao et al., 2021; Wang et al., 2022) attempt to utilize a unified framework to train AR and NAR jointly so that the NAR can be enhanced. Besides, Zhou et al. (2020) have also explored using the output of NAR to improve the AR performance. Unlike them, we focus on the GEC task and introduce the NAR model to utilize the global information to help the model understand the context around incorrect tokens.",
+        "bbox": [
+            505,
+            84,
+            885,
+            374
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            507,
+            388,
+            658,
+            406
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we elaborate on our proposed framework for GEC. As shown in figure 2, we introduce the CMLM-based NAR model to integrate more context information into our single model.",
+        "bbox": [
+            507,
+            414,
+            885,
+            480
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 Overview",
+        "text_level": 1,
+        "bbox": [
+            507,
+            492,
+            633,
+            507
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given the training dataset  $(X,Y)$ , the definition of the GEC task is to correct the original erroneous source  $X$  and generate a sentence  $Y$  without grammatical error, where  $X = (x_{1},x_{2},\\dots,x_{K})$  and  $Y = (y_{1},y_{2},\\dots,y_{N})$ . Specifically, the transformer encoder takes the source sentence  $X$  as input. Different from previous seq2seq works, our decoder consists of two components: AR decoder and NAR decoder. We keep the AR decoder as a traditional seq2seq decoder without any change. For the NAR decoder, we mask the tokens in the input that corresponds to the positions of the low confidence tokens from the output distribution of the AR decoder. Then, we regenerate tokens in masked positions more accurately by bidirectional semantic modeling of the NAR decoder. Finally, we try to decrease the output distribution distance which is in masked positions between two manners to further improve the model performance during the training stage.",
+        "bbox": [
+            505,
+            514,
+            885,
+            837
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.2 Mask low Confidence",
+        "text_level": 1,
+        "bbox": [
+            507,
+            848,
+            726,
+            863
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Here, it is important that the output probability represents whether the model is confident in the prediction. As for the GEC task, there are only a",
+        "bbox": [
+            507,
+            870,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12015",
+        "bbox": [
+            477,
+            927,
+            524,
+            941
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/d7262101471a593e2cf569a989db337c4d35cb0fbfe5d4269bfa883c713187e5.jpg",
+        "image_caption": [
+            "Figure 2: An illustration of our proposed model. The results  $\\{y_1', y_2', y_3', y_4', y_5', y_6', y_7'\\}$  are predicted by AR Decoder with the golden target. The mask set  $\\{m_3, m_5\\}$  indicates the position in the golden target that needs to be masked with a special token  $<\\text{mask}>$ . After the mask operation, the input of the NAR decoder is  $\\{y_1, y_2, m_3, y_4, m_5, y_6, y_7\\}$ , and the output  $\\{y_3'', y_5''\\}$  are re-predicted by CMLM based NAR decoder, and then the output are used to calibrate the corresponding AR decoder output."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            139,
+            80,
+            860,
+            294
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1 Mask strategy",
+        "text_level": 1,
+        "bbox": [
+            114,
+            403,
+            322,
+            419
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Input: The AR decoder output  $y_{AR}$ , the golden target  $y_{tgt}$ , the mask ratio  $\\delta$",
+        "bbox": [
+            114,
+            422,
+            485,
+            454
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Output: The NAR input  $x_{NAR}$",
+        "bbox": [
+            114,
+            455,
+            356,
+            470
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1: while not converged do",
+            "2: Select the max probability for each token;",
+            "3: max_logit  $\\leftarrow$  get_max_logit(yAR);",
+            "4:  $L_{t}\\gets \\mathrm{get\\_target\\_length}(y_{tgt})$",
+            "5:  $L_{m} = L_{t}\\times \\delta$",
+            "6: Get the  $L_{m}$  index of the low confidence positions in maxim logit;",
+            "7: Index  $\\leftarrow$  select_index(max_logit,  $L_{m}$ );",
+            "8: Replace with  $<\\mathrm{mask}>$  in the  $y_{tgt}$  corresponding position;",
+            "9:  $x_{NAR} \\gets \\text{mask\\_target}(Index, y_{tgt})$ ;"
+        ],
+        "bbox": [
+            126,
+            470,
+            487,
+            648
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "10: end while",
+        "text_level": 1,
+        "bbox": [
+            119,
+            649,
+            226,
+            661
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "few tokens that need to be modified (About  $10\\%$ ). Therefore, the model tends to focus on high confidence correct tokens that need to be kept, but not so much on low confidence tokens that need to be modified. As mentioned above, we choose the low confidence positions in the AR output distribution and substitute them with special symbols  $\\langle \\text{mask} \\rangle$  as the input of the NAR decoder. In this way, the NAR decoder is forced to learn the knowledge of low confidence tokens from the bidirectional context in hidden layers, which helps to boost the performance.",
+        "bbox": [
+            112,
+            692,
+            489,
+            883
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To construct the input effectively, we design a special mask strategy. Details are described in Al-",
+        "bbox": [
+            112,
+            887,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "**Algorithm 1.** Specifically, we select the maximum probability of each token from the AR decoder output distribution. Then, we reorder each token in the output sentence from low confidence to high confidence to get a specific number of positions for the low confidence tokens. We introduce a special token  $<\\mathrm{mask}>$  to mask the token at the corresponding position in the golden target, which serves as a placeholder to represent the position where the target token needs to be regenerated. The golden target after the masking operation is used as input to the NAR decoder to introduce bidirectional contextual information.",
+        "bbox": [
+            507,
+            403,
+            884,
+            611
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3 Restrict Output Consistency",
+        "text_level": 1,
+        "bbox": [
+            507,
+            623,
+            778,
+            639
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The objective of our model is to overcome the limitations of AR models by introducing the NAR generation mechanism, and then correct sentences with grammatical errors. A common way is to implicitly pass the information learned by the NAR branch to the AR branch using the parameter-sharing method. Specifically, we share the parameters of the Transformer layer in both manners. However, there is a huge difference between the AR manner and the NAR manner in the training process, as shown in Equation 1 and Equation 2, where the AR generation process is more concerned with local dependencies, while the NAR generation process is more concerned with global dependencies.",
+        "bbox": [
+            507,
+            643,
+            884,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP (Y | X) = \\prod_ {i = 1} ^ {N} P _ {A R} \\left(y _ {i} | X, Y _ {<   i}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            571,
+            879,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12016",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP (Y | X) = \\prod_ {i = 1} ^ {N} P _ {N A R} \\left(y _ {i} | X\\right). \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            188,
+            82,
+            487,
+            124
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The inconsistency between the two generation methods can lead to direct parameter sharing between the two branches without enabling the AR manner to obtain the exact information provided by the NAR manner. This sharing method only implicitly considers the correlation of model parameters and ignores the inconsistency between the two generation methods, which seriously hinders performance.",
+        "bbox": [
+            112,
+            137,
+            487,
+            281
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In contrast, in our work, to make the AR manner learn the information from NAR in a way that is more adapted to AR generation, we take an explicit approach to constrain the two manners. This approach can avoid the inconsistency caused by the different ways of AR and NAR generation and break the performance bottleneck. In practice, we accomplish explicit information modeling by using bidirectional Kullback-Leibler (KL) divergence to force the AR and NAR output distributions at the mask positions to be consistent with each other. Fortunately, Liang et al. (2022) also use KL divergence to combine the advantages of AR and NAR, which gives us much inspiration.",
+        "bbox": [
+            112,
+            284,
+            489,
+            508
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4 Training and Inference",
+        "text_level": 1,
+        "bbox": [
+            112,
+            526,
+            344,
+            542
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Multi-Task Framework We learn the GEC model under the multi-task learning framework, including an AR primary task and a NAR auxiliary task. It should be noted that AR and NAR manners are regarded as two different tasks. For the AR task, we employ the negative log-likelihood (NLL) as the loss function which is akin to the traditional seq2seq. Therefore, the optimization objective is:",
+        "bbox": [
+            112,
+            550,
+            487,
+            680
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {A R} = - \\sum_ {i = 1} ^ {N} \\log P _ {A R} \\left(y _ {i} \\mid X, Y _ {<   i}\\right), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            696,
+            487,
+            740
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $N$  is the target length, and  $Y_{< i}$  represents the tokens before the  $i$ -th time step.  $P_{AR}(y_i|X,Y_{< i})$  represents the output probability of the AR decoder, which will be used in the later process.",
+        "bbox": [
+            112,
+            756,
+            487,
+            820
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For the NAR task, we obtain the positions of the specified number of low confidence tokens based on the mask ratio  $\\delta$ , and replace the tokens with the special symbols  $<\\mathrm{mask}>$  at the corresponding positions of the golden target. The loss function  $\\mathcal{L}_{NAR}$  for NAR task is to minimize the sum of",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "negative log-likelihood in masked positions:",
+        "bbox": [
+            507,
+            84,
+            840,
+            101
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {N A R} = - \\sum_ {i = 1} ^ {M} \\log P _ {N A R} \\left(y _ {i} \\mid X, Y _ {\\text {m a s k}}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            532,
+            108,
+            882,
+            149
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $M$  is the number of the masked tokens, and  $Y_{mask}$  is the set of the tokens in masked. In this way, the NAR decoder regenerates the masked tokens with more context information to help the AR task. Then the loss function of the multi-task framework is:",
+        "bbox": [
+            507,
+            158,
+            884,
+            253
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {m} = \\lambda_ {t} \\mathcal {L} _ {N A R} + (1 - \\lambda_ {t}) \\mathcal {L} _ {A R}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            571,
+            263,
+            882,
+            280
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda_{t}$  is the important factor to balance the weight of AR and NAR tasks during training. We will present the design details in the following paragraphs.",
+        "bbox": [
+            507,
+            288,
+            882,
+            354
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Curriculum Learning Compared with the AR task, the NAR task is more complex, and unreasonable weight setting will make training difficult. For example, the excessive weight of the NAR task will disturb the parameter learning of the AR primary task at the beginning. Inspired by curriculum learning (Bengio et al., 2009), which is to imitate the human learning process, we propose the dynamic weight strategy. More concretely, we start with  $\\lambda_{t} = 0$  and gradually increase the NAR task weight  $\\lambda_{t}$  to introduce learning signals. The dynamic weight scheme is:",
+        "bbox": [
+            507,
+            365,
+            884,
+            558
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\lambda_ {t} = \\frac {t}{T}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            660,
+            565,
+            882,
+            596
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $t$  and  $T$  are the current and total steps of training. We increase the weight linearly in all the experiments.",
+        "bbox": [
+            507,
+            601,
+            882,
+            648
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It is not enough to use only the hard parameter sharing method mentioned above, we regularize the two output distributions  $P_{AR}$  and  $P_{NAR}$  for unconfident words with the token-level bidirectional Kullback-Leibler divergence to further transfer the knowledge of NAR:",
+        "bbox": [
+            507,
+            649,
+            882,
+            746
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {K L} = \\sum_ {Y _ {\\text {m a s k}}} K L \\left(P _ {A R} \\| P _ {N A R}\\right) + \\\\ \\sum_ {Y _ {\\text {m a s k}}} K L \\left(P _ {N A R} \\| P _ {A R}\\right). \\end{array} \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            766,
+            884,
+            839
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The final training objective for our GEC model is a combination of the three terms reviewed above as:",
+        "bbox": [
+            507,
+            845,
+            882,
+            892
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\lambda_ {t} \\mathcal {L} _ {N A R} + (1 - \\lambda_ {t}) \\mathcal {L} _ {A R} + \\alpha \\mathcal {L} _ {K L}. \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            534,
+            903,
+            882,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12017",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/e09ba183edf186948070f1cfffc4559b6785402a3889f157c23e0012421bf831.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Architecture</td><td>Precision</td><td>Recall</td><td>F0.5</td></tr><tr><td>Transformer Big†</td><td>1024-1024-4096</td><td>65.26</td><td>27.19</td><td>50.98</td></tr><tr><td>LaserTagger* (Malmi et al., 2019)</td><td>-</td><td>50.9</td><td>26.9</td><td>43.2</td></tr><tr><td>Adversarial-GE (Raheja and Alikaniotis, 2020)</td><td>512-512-2048</td><td>64.68</td><td>22.57</td><td>47.10</td></tr><tr><td>ESD+ESC* (Chen et al., 2020)</td><td>1024-1024-4096</td><td>66.0</td><td>24.7</td><td>49.5</td></tr><tr><td>SAD(9+3) (Sun et al., 2021)</td><td>1024-1024-4096</td><td>58.8</td><td>33.1</td><td>50.9</td></tr><tr><td>S2A (Li et al., 2022)</td><td>1024-1024-4096</td><td>65.9</td><td>28.9</td><td>52.5</td></tr><tr><td>CMLM† (Ghazvininejad et al., 2019)</td><td>1024-1024-4096</td><td>46.3</td><td>27.17</td><td>40.59</td></tr><tr><td>Levenshtein Transformer* (Gu et al., 2019)</td><td>1024-1024-4096</td><td>39.9</td><td>24.4</td><td>35.4</td></tr><tr><td>JANUS† (Liang et al., 2022)</td><td>1024-1024-4096</td><td>66.22</td><td>27.76</td><td>51.85</td></tr><tr><td>Ours-base</td><td>512-512-2048</td><td>66.63</td><td>28.70</td><td>52.70</td></tr><tr><td>Ours</td><td>1024-1024-4096</td><td>65.10</td><td>32.29</td><td>54.11</td></tr></table>",
+        "bbox": [
+            127,
+            80,
+            870,
+            298
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: The results of systems on the CoNLL-2014 English GEC task. For the models with  $\\star$ , their performance is from (Chen et al., 2020).  $\\dagger$  indicates the models are implemented by us with the released codes of the original papers. The Architecture column represents the embedding, hidden, and FFN size of the model. Here we bold the best results of the models.",
+        "bbox": [
+            112,
+            307,
+            882,
+            365
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inference During the inference stage, we use the AR decoder to generate the correct sentences, and the inference efficiency is the same as the traditional seq2seq model since the NAR decoder is only used in training.",
+        "bbox": [
+            112,
+            391,
+            487,
+            470
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            486,
+            321,
+            502
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            514,
+            228,
+            527
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To validate the effectiveness of our proposed GEC model, we conduct a set of experiments on both the restricted track of the BEA-2019 GEC shared task (Bryant et al., 2019) and NLPCC 2018 Task 2 (Zhao et al., 2018).",
+        "bbox": [
+            112,
+            536,
+            487,
+            617
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "BEA-2019 GEC shared task This is a public dataset for the English GEC task, we follow the setting of (Chollampatt and Ng, 2018) and take the FCE training set (Yannakoudakis et al., 2011), Lang-8 Corpus of Learner English (Mizumoto et al., 2011), NUCLE (Dahlmeier et al., 2013) and W&I+LOCNESS (Granger, 2014; Bryant et al., 2019) as the training set. The development set is a subset of NUCLE, and our model is evaluated on the CoNLL-2014 (Ng et al., 2014), which is a well-known English GEC benchmark test set. Specifically, we use pre-processed script in (Chollampatt and Ng, 2018) to obtain the parallel corpus.",
+        "bbox": [
+            112,
+            629,
+            489,
+            839
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "NLPCC 2018 Task 2 It is the first and latest benchmark dataset for Chinese GEC. We combine",
+        "bbox": [
+            112,
+            851,
+            487,
+            882
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "the incorrect sentence with each corrected sentence to build the parallel sentence pairs as described in (Zhao and Wang, 2020) and get 1.2 million sentence pairs in all. Next, we randomly sample 5,000 training instances as the development set. The official test set extracted from PKU Chinese Learner Corpus contains 2,000 samples. We use the combination of two group annotations that mark the golden edits of grammatical errors in these sentences to evaluate our model. Following the setting of NLPCC 2018 Task (Zhao et al., 2018), the tokenization of training data is implemented with the PKUNLP tool<sup>2</sup>.",
+        "bbox": [
+            507,
+            391,
+            884,
+            599
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            613,
+            618,
+            629
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "While in the training process, we use the base model configuration of the Transformer for the Chinese GEC task, with 6 layers, the number of self-attention heads is set to 8, the embedding dimension is 512 and the size of FFN layer is 2048, the dropout and weight decay is 0.3 and 0.01 respectively. In the English GEC task, we use the big Transformer setting, which contains 6 layers and 16 self-attention heads, the size of word vectors on the source side and the target side are 1024, the FFN layer size is 4096, the dropout is applied with a probability of 0.1 and the weight decay value is set to be 0.0001. We adopt Adam (Kingma and Ba, 2015) optimizer with initial learning rate 0.0005 and 0.0007 for Chinese and English GEC tasks respectively, and a beta value of (0.9, 0.98). We use",
+        "bbox": [
+            507,
+            636,
+            882,
+            892
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<https://github.com/nusnlp/mlconvgec2018/tree/master/data>",
+        "bbox": [
+            112,
+            891,
+            485,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://github.com/zhaoyyoo/NLPCC2018_GEC",
+        "bbox": [
+            529,
+            903,
+            850,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12018",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/b11d041b41a7e89f182ecc9336203eef82f81411053fed93af56204d7f57953d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Model type</td><td>Precision</td><td>Recall</td><td>F0.5</td></tr><tr><td>Transformer</td><td>Single</td><td>36.91</td><td>15.57</td><td>28.97</td></tr><tr><td>YouDao (Fu et al., 2018)</td><td>Ensemble</td><td>35.24</td><td>18.64</td><td>29.91</td></tr><tr><td>AliGM (Zhou et al., 2018)</td><td>Ensemble</td><td>41.00</td><td>13.75</td><td>29.36</td></tr><tr><td>BLCU (Ren et al., 2018)</td><td>Ensemble</td><td>47.63</td><td>12.56</td><td>30.57</td></tr><tr><td>ESD+ESC (Chen et al., 2020)</td><td>Single</td><td>37.3</td><td>14.5</td><td>28.4</td></tr><tr><td>SAD(9+3) (Sun et al., 2021)</td><td>Single</td><td>33.0</td><td>20.5</td><td>29.4</td></tr><tr><td>S2A (Li et al., 2022)</td><td>Single</td><td>36.57</td><td>18.25</td><td>30.46</td></tr><tr><td>Ours</td><td>Single</td><td>41.90</td><td>15.24</td><td>31.04</td></tr></table>",
+        "bbox": [
+            216,
+            80,
+            783,
+            250
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: The results of systems on the NLPCC-2018 Chinese GEC task. For a fair comparison, all the results are produced by training on the original NLPCC-2018 training data. We **bold** the best results.",
+        "bbox": [
+            112,
+            259,
+            882,
+            291
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "learning rate schedule as in (Vaswani et al., 2017), 10,000 warmup steps for the Chinese GEC task and 4,000 for the English GEC task. Lable smoothing is added with an epsilon value of 0.1. We use 32K Byte Pair Encoding (BPE) (Sennrich et al., 2016) for tokenization on Chinese and English GEC tasks. We save the checkpoint for each epoch and select the best checkpoint based on the loss on the development set. The beam size is 5 during the inference stage. All experiments are based on fairseq (Ott et al., 2019).",
+        "bbox": [
+            112,
+            313,
+            489,
+            491
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            507,
+            235,
+            521
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare the performance of the proposed model with several representative baseline methods on both English and Chinese GEC tasks. Specifically, for the English GEC task, Transformer Big is the typical AR model. LaserTagger proposes to predict tags with a smaller vocabulary (Malmi et al., 2019). Adversarial-GEc presents an adversarial learning approach to generate realistic texts in a generator-discriminator framework (Raheja and Alikaniotis, 2020). ESD+ESC is a pipeline model (Chen et al., 2020). SAD employs a new decoding method with a shallow decoder to conduct the prediction (Sun et al., 2021). S2A proposes to integrate action probabilities into token prediction probabilities to obtain the final results (Li et al., 2022). Levenshtein Transformer (Gu et al., 2019) and CMLM (Ghazvininejad et al., 2019) are NAR models, which achieve excellent performance with an iterative generation paradigm. In addition, we also compare with JANUS (Liang et al., 2022), which joints AR and NAR training for sequence generation.",
+        "bbox": [
+            112,
+            531,
+            489,
+            885
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For the Chinese GEC task, we compare our model to all previous systems in the NLPCC 2018",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "dataset. YouDao corrects the sentences independently by utilizing five different mixture models (Fu et al., 2018). AliGM combines three approaches, including NMT-based, SMT-based, and rule-based together (Zhou et al., 2018). BLCU is based on a multi-layer convolutional seq2seq model (Ren et al., 2018).",
+        "bbox": [
+            507,
+            313,
+            884,
+            426
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            439,
+            705,
+            454
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Following the typical previous works (Chen et al., 2020; Li et al., 2022), we use the official MaxMatch  $(M^2)$  (Dahlmeier and Ng, 2012) scorer for evaluation of our grammatical error correction system.  $M^2$  scorer computes the sequence of phrase-level edits between a source sentence and a system hypothesis that achieves the maximal overlap with the gold standard annotation. Given the set of system edits and the set of gold edits for all sentences, the value of precision, recall, and  $F_{0.5}$  are computed by m2scorer<sup>3</sup>.",
+        "bbox": [
+            507,
+            461,
+            884,
+            638
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            651,
+            608,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Main Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            678,
+            660,
+            693
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The results of our proposed approach and recent models on English GEC task are shown in Table 1. We can see that our approach significantly outperforms the baselines mentioned above. Our model achieves an improvement above Transformer Big by nearly 3.1 in  $F_{0.5}$  score, and performs better than the strong baseline S2A, by a large margin of 1.6  $F_{0.5}$ . Moreover, the proposed model surpasses the recent JANUS model by  $F_{0.5}$  score of 2.3, which shows excellent performance on multiple tasks by combining AR and NAR. This result implies that our designed joint training method is more suit-",
+        "bbox": [
+            507,
+            701,
+            884,
+            894
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://github.com/nusnlp/m2scorer",
+        "bbox": [
+            529,
+            903,
+            798,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12019",
+        "bbox": [
+            477,
+            927,
+            524,
+            941
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/8f3ea12a776c21dea54164bbdb1d80236690191c2609ee6ac6ab97c39412cff4.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Mask Ratio</td><td colspan=\"4\">BEA-2019</td><td colspan=\"4\">NLPCC-2018</td></tr><tr><td>Precision</td><td>Recall</td><td>F0.5</td><td>F1</td><td>Precision</td><td>Recall</td><td>F0.5</td><td>F1</td></tr><tr><td>10%</td><td>61.70</td><td>31.31</td><td>51.67</td><td>41.61</td><td>38.99</td><td>14.57</td><td>29.20</td><td>21.22</td></tr><tr><td>15%</td><td>62.32</td><td>31.82</td><td>52.30</td><td>40.65</td><td>41.90</td><td>15.24</td><td>31.04</td><td>22.35</td></tr><tr><td>20%</td><td>65.10</td><td>32.29</td><td>54.11</td><td>43.21</td><td>42.24</td><td>14.95</td><td>30.94</td><td>22.09</td></tr><tr><td>25%</td><td>64.87</td><td>30.61</td><td>53.01</td><td>41.68</td><td>40.01</td><td>14.48</td><td>29.58</td><td>21.26</td></tr><tr><td>30%</td><td>63.35</td><td>29.68</td><td>51.63</td><td>40.56</td><td>41.18</td><td>13.68</td><td>29.36</td><td>20.54</td></tr><tr><td>35%</td><td>64.51</td><td>30.43</td><td>52.71</td><td>41.48</td><td>41.73</td><td>13.70</td><td>29.61</td><td>20.63</td></tr></table>",
+        "bbox": [
+            164,
+            83,
+            831,
+            218
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/053fc08311b39d285d47dca0440fe472e1272e9458c16169cf19180e7e3ff207.jpg",
+        "table_caption": [
+            "Table 3: Effect of Mask Ratio. The mask ratio represents the percentage of low-confidence tokens in a sentence that are masked. Best results of the Chinese GEC task and the English GEC task are bold separately."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Sub</td><td>Del</td><td>Ins</td></tr><tr><td>AR</td><td>58.12%</td><td>77.50%</td><td>73.82%</td></tr><tr><td>Ours</td><td>52.30%</td><td>52.84%</td><td>61.70%</td></tr></table>",
+        "bbox": [
+            149,
+            284,
+            453,
+            338
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: The Correction Coincidence Rate of the AR model and our method on the English CoNLL-2014 test set.",
+        "bbox": [
+            112,
+            351,
+            487,
+            394
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "able for the GEC task. It is noteworthy that our model with Transformer base settings still consistently exceeds the baselines with Transformer big settings. These results all support that our proposed approach can effectively improve the AR GEC by using a NAR model.",
+        "bbox": [
+            112,
+            420,
+            487,
+            516
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To validate the effectiveness of our approach, we conduct experiments on the Chinese GEC task and present the results in Table 2. These results demonstrate that the Chinese GEC task is more challenging than the English GEC. Despite this, the proposed model yields a higher  $F_{0.5}$  than the listed methods. Moreover, we can observe that all the top three models are ensemble models, including YouDao, AliGM, and BLCU, but our single model still surpasses them. This result means that our model is generalizable.",
+        "bbox": [
+            112,
+            517,
+            487,
+            694
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Fix more Grammar Errors",
+        "text_level": 1,
+        "bbox": [
+            112,
+            705,
+            371,
+            720
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We carefully investigate the number of different types of errors corrected in the two datasets, and find that most of the corrected grammar errors are the same between the proposed method and the AR model. To show the advantages of our model intuitively, we propose a Correction Coincidence Rate, which is the number of overlaps of correction errors to the total number of respective correction errors. The results are summarized in Table 4. For computational convenience, the errors are broadly categorized into insertion, deletion, and substitution. From Table 4, the overlap rate of our proposed",
+        "bbox": [
+            112,
+            726,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "method on all types of error modifications is much lower. For instance, the percentage of deletion decreases by  $25\\%$ . This indicates that our model is able to correct more grammar errors while maintaining the ability of the AR model.",
+        "bbox": [
+            507,
+            285,
+            884,
+            365
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Ablation Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            376,
+            695,
+            391
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effect of Mask Ratio In this section, we present exhaustive investigations on the impact of mask ratio. Here we vary the mask ratios in  $\\{0.1, 0.15, 0.2, 0.3, 0.35\\}$  and conduct experiments in BEA-2019 and NLPCC 2018. The corresponding results are provided in Table 3. It can be observed that all mask ratios outperform the Transformer baseline. A reasonable reason is that the masking operation makes the model focus more on incorrect tokens, and the model is forced to capture more context information, which facilitates error correction. On the other hand, a small mask ratio (e.g., 0.1) cannot perform as well as a large one (e.g., 0.15), which means that there is a fraction of incorrect tokens that are not focused on. However, too much masking ratio is also not good. It will result in many correct words being masked, which may prevent the correction of incorrect tokens. Note that the choice of mask ratio is distinct for different datasets, and the best balance choices for BEA-2019 and NLPCC-2018 are 0.2 and 0.15 respectively.",
+        "bbox": [
+            507,
+            409,
+            884,
+            762
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effect of KL Loss Weight  $\\alpha$  We explore the effect of KL-divergence loss weight  $\\alpha$  in Equation 8. The result is illustrated in Table 6. By comparing the performance with KL loss and without KL loss, we can see that the performance of the former is consistently better than the performance of the latter, which suggests that KL loss can be further combined with information from AR and NAR to correct errors. In addition, the performance is lower",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12020",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/3109970c58ef07bd54defa0924b85bb8ffce8eaa7ef6ad8cdaca1f5ea0919f81.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>Samples</td></tr><tr><td>SRC</td><td>I think the family will stay mentally healthy as it is, without having emotional stress.</td></tr><tr><td>TGT</td><td>I think the family will stay mentally healthy as it is, without having emotional stress.</td></tr><tr><td>Transformer</td><td>I think the family will stay mentally healthy as it is, without having emotional stress.</td></tr><tr><td>Ours</td><td>I think the family will stay mentally healthy as it is, without having emotional stress.</td></tr><tr><td>SRC</td><td>While we do know that we should not discriminate them based on their limitations...</td></tr><tr><td>TGT</td><td>While we do know that we should not discriminate against them based on their limitations...</td></tr><tr><td>Transformer</td><td>While we do know that we should not discriminate them based on their limitations...</td></tr><tr><td>Ours</td><td>While we do know that we should not discriminate against them based on their limitations...</td></tr><tr><td>SRC</td><td>First and foremost, I would like to share on the advantages of using such social media...</td></tr><tr><td>TGT</td><td>First and foremost, I would like to share the advantages of using such social media...</td></tr><tr><td>Transformer</td><td>First and foremost, I would like to share on the advantages of using such social media...</td></tr><tr><td>Ours</td><td>First and foremost, I would like to share the advantages of using such social media...</td></tr></table>",
+        "bbox": [
+            121,
+            80,
+            877,
+            307
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/2ed90113ef834f7dc1e89ddd1d8ec9fe400bfa76ab5cfecb9153ec9d0edba3ba.jpg",
+        "table_caption": [
+            "Table 5: Case studies of the original Transformer model and our proposed model on the English CoNLL-2014 test set. The tokens in red and wave line are errors, while tokens with underline and in green are the corrections made by the gold target or our model."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>TrainSet</td><td>α</td><td>Precision</td><td>Recall</td><td>F0.5</td></tr><tr><td rowspan=\"6\">BEA-2019</td><td>0</td><td>59.47</td><td>31.82</td><td>50.66</td></tr><tr><td>0.3</td><td>61.72</td><td>32.21</td><td>52.16</td></tr><tr><td>0.4</td><td>60.68</td><td>31.27</td><td>51.07</td></tr><tr><td>0.5</td><td>65.10</td><td>32.29</td><td>54.11</td></tr><tr><td>0.6</td><td>60.51</td><td>31.88</td><td>51.30</td></tr><tr><td>0.7</td><td>65.03</td><td>31.29</td><td>53.50</td></tr><tr><td rowspan=\"6\">NLPCC-2018</td><td>0</td><td>37.58</td><td>14.34</td><td>28.38</td></tr><tr><td>0.8</td><td>40.53</td><td>15.54</td><td>30.66</td></tr><tr><td>0.9</td><td>39.46</td><td>14.24</td><td>29.14</td></tr><tr><td>1.0</td><td>41.90</td><td>15.24</td><td>31.04</td></tr><tr><td>1.1</td><td>39.34</td><td>14.23</td><td>29.07</td></tr><tr><td>1.2</td><td>40.32</td><td>13.88</td><td>29.20</td></tr></table>",
+        "bbox": [
+            136,
+            382,
+            467,
+            569
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6: The results with different weight.  $\\alpha$  equal 0 represents the model without KL loss. Weights are adjusted according to different datasets. Best results are bold.",
+        "bbox": [
+            112,
+            582,
+            487,
+            639
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "than the baseline when the value of  $\\alpha$  is 0, i.e., the model is fused using only the simple method of parameter sharing. It indicates that simple fusion will lead to poor performance. We also find that the performance is not optimal when  $\\alpha$  is set too small or too large. We believe that the model does not learn enough information when  $\\alpha$  is set too small, while setting it too large leads to the introduction of too much noise.",
+        "bbox": [
+            112,
+            668,
+            487,
+            810
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4 Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            829,
+            248,
+            846
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In order to qualitatively show the effectiveness of global context information, we conduct case studies with Transformer and our proposed model. We pick the cases from the CoNLL-2014 English GEC",
+        "bbox": [
+            112,
+            854,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "test set. The results are listed in Table 5. Generally, it is easy to see that both approaches can copy most of the correct tokens from the source to the target. Nevertheless, when correcting grammatical errors, our approach can predict more accurately by considering more context information. For example, as shown in the third sample in Table 5, the AR model generates the phase \"share on\" which tends to be consistent with the source language, while our model can delete the token \"on\" by utilizing more context information. This again confirms that our method can make use of the global information to correct errors.",
+        "bbox": [
+            507,
+            384,
+            884,
+            592
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            627,
+            640,
+            643
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose a joint AR and NAR learning objective for the GEC, using a multi-task learning framework. To better predict tokens at low-confidence positions, we introduce additional signals to the training of GEC models by using the NAR model as an auxiliary model. Meanwhile, we develop a new regularization term of training to constrain the inconsistency between the two manners. Through our experiments in the English and Chinese GEC task, the proposed approach can significantly improve the GEC model performance without additional inference costs.",
+        "bbox": [
+            507,
+            670,
+            884,
+            864
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In the future, we are also interested in introducing syntax and lexical knowledge to focus on incorrect tokens to further improve performance.",
+        "bbox": [
+            507,
+            870,
+            884,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12021",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            250,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we achieve a noticeable improvement in the GEC task by introducing additional context information with a NAR model. However, in order to focus on incorrect tokens, the input of the NAR is required to be constructed based on the AR output distribution. In this way, the AR and NAR model perform sequentially, which leads to much time consumption in the training stage. In the future, we will apply a layer dropout strategy to speed up model training. On the other hand, due to the limitation of computation resources, all experiments are conducted on two Nvidia Titan V GPUs with 12GB VRAM. Therefore, we could not compare with the state-of-the-art models which are pre-trained with 100M synthetic parallel examples (Li et al., 2022). We left it as our future work.",
+        "bbox": [
+            112,
+            109,
+            492,
+            382
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            394,
+            307,
+            410
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported in part by the National Science Foundation of China (No.62276056), the National Key R&D Program of China, the China HTRD Center Project (No.2020AAA0107904), the Natural Science Foundation of Liaoning Province of China (2022-KF-16-01), the Yunnan Provincial Major Science and Technology Special Plan Projects (No.202103AA080015), the Fundamental Research Funds for the Central Universities (Nos.N2216016, N2216001, and N2216002), and the Program of Introducing Talents of Discipline to Universities, Plan 111 (No.B16009).",
+        "bbox": [
+            112,
+            419,
+            489,
+            612
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            639,
+            213,
+            653
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning, ICML 2009, Montreal, Quebec, Canada, June 14-18, 2009, volume 382 of ACM International Conference Proceeding Series, pages 41-48. ACM.",
+            "Christopher Bryant, Mariano Felice, Øistein E. Andersen, and Ted Briscoe. 2019. The BEA-2019 shared task on grammatical error correction. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, BEA@ACL 2019, Florence, Italy, August 2, 2019, pages 52-75. Association for Computational Linguistics.",
+            "Mengyun Chen, Tao Ge, Xingxing Zhang, Furu Wei, and Ming Zhou. 2020. Improving the efficiency of grammatical error correction with erroneous span detection and correction. In Proceedings of the 2020"
+        ],
+        "bbox": [
+            115,
+            661,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 7162-7169. Association for Computational Linguistics.",
+            "Shamil Chollampatt and Hwee Tou Ng. 2018. A multilayer convolutional encoder-decoder neural network for grammatical error correction. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 5755-5762. AAAI Press.",
+            "Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In Human Language Technologies: Conference of the North American Chapter of the Association of Computational Linguistics, Proceedings, June 3-8, 2012, Montreal, Canada, pages 568-572. The Association for Computational Linguistics.",
+            "Daniel Dahlmeier, Hwee Tou Ng, and Siew Mei Wu. 2013. Building a large annotated corpus of learner english: The NUS corpus of learner english. In Proceedings of the Eighth Workshop on Innovative Use of NLP for Building Educational Applications, BEA@NAACL-HLT 2013, June 13, 2013, Atlanta, Georgia, USA, pages 22-31. The Association for Computer Linguistics.",
+            "Kai Fu, Jin Huang, and Yitao Duan. 2018. Youdao's winning solution to the NLPCC-2018 task 2 challenge: A neural machine translation approach to chinese grammatical error correction. In Natural Language Processing and Chinese Computing - 7th CCF International Conference, NLPCC 2018, Ho-hhot, China, August 26-30, 2018, Proceedings, Part I, volume 11108 of Lecture Notes in Computer Science, pages 341-350. Springer.",
+            "Tao Ge, Furu Wei, and Ming Zhou. 2018. Fluency boost learning and inference for neural grammatical error correction. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 1055-1065. Association for Computational Linguistics.",
+            "Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask-predict: Parallel decoding of conditional masked language models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 6111-6120. Association for Computational Linguistics.",
+            "Sylviane Granger. 2014. The computer learner corpus: a versatile new source of data for sola research. In Lerner English on computer, pages 3-18. Routledge."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12022",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Roman Grundkiewicz, Marcin Junczys-Dowmunt, and Kenneth Heafield. 2019. Neural grammatical error correction systems with unsupervised pre-training on synthetic data. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, BEA@ACL 2019, Florence, Italy, August 2, 2019, pages 252-263. Association for Computational Linguistics.",
+            "Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. 2018. Non-autoregressive neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.",
+            "Jiatao Gu, Changhan Wang, and Junbo Zhao. 2019. Levenshtein transformer. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 11179-11189.",
+            "Junliang Guo, Xu Tan, Linli Xu, Tao Qin, Enhong Chen, and Tie-Yan Liu. 2020. Fine-tuning by curriculum learning for non-autoregressive neural machine translation. 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 7839-7846. AAAI Press.",
+            "Yongchang Hao, Shilin He, Wenxiang Jiao, Zhaopeng Tu, Michael R. Lyu, and Xing Wang. 2021. Multi-task learning with shared encoder for non-autoregressive machine translation. 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 3989-3996. Association for Computational Linguistics.",
+            "Marcin Junczys-Dowmunt, Roman Grundkiewicz, Shubha Guha, and Kenneth Heafield. 2018. Approaching neural grammatical error correction as a low-resource machine translation task. In Proceedings of the 2018 Conference of the North American Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 595-606. Association for Computational Linguistics.",
+            "Masahiro Kaneko, Masato Mita, Shun Kiyono, Jun Suzuki, and Kentaro Inui. 2020. Encoder-decoder models can benefit from pre-trained masked language models in grammatical error correction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 4248-4254. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.",
+            "Marek Kubis, Zygmunt Vetulani, Mikolaj Wypych, and Tomasz Zietkiewicz. 2020. Open challenge for correcting errors of speech recognition systems. CoRR, abs/2001.03041.",
+            "Jason Lee, Elman Mansimov, and Kyunghyun Cho. 2018. Deterministic non-autoregressive neural sequence modeling by iterative refinement. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 1173-1182. Association for Computational Linguistics.",
+            "Jiquan Li, Junliang Guo, Yongxin Zhu, Xin Sheng, Deqiang Jiang, Bo Ren, and Linli Xu. 2022. Sequence-to-action: Grammatical error correction with action guided sequence generation. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 10974-10982. AAAI Press.",
+            "Piji Li and Shuming Shi. 2021. Tail-to-tail non-autoregressive sequence prediction for chinese grammatical error correction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 4973-4984. Association for Computational Linguistics.",
+            "Zhuohan Li, Zi Lin, Di He, Fei Tian, Tao Qin, Liwei Wang, and Tie-Yan Liu. 2019. Hint-based training for non-autoregressive machine translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 5707-5712. Association for Computational Linguistics.",
+            "Xiaobo Liang, Lijun Wu, Juntao Li, and Min Zhang. 2022. Janus: Joint autoregressive and nonautoregressive training with auxiliary loss for sequence generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, pages 8050-8060. The Association for Computer Linguistics.",
+            "Jared Lichtarge, Chris Alberti, Shankar Kumar, Noam Shazeer, Niki Parmar, and Simon Tong. 2019. Corpora generation for grammatical error correction. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7,"
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12023",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "2019, Volume 1 (Long and Short Papers), pages 3291-3301. Association for Computational Linguistics.",
+            "Jinglin Liu, Yi Ren, Xu Tan, Chen Zhang, Tao Qin, Zhou Zhao, and Tie-Yan Liu. 2020. Task-level curriculum learning for non-autoregressive neural machine translation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 3861-3867. ijcai.org.",
+            "Eric Malmi, Sebastian Krause, Sascha Rothe, Daniil Mirylenka, and Aliaksei Severyn. 2019. Encode, tag, realize: High-precision text editing. 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 2019, Hong Kong, China, November 3-7, 2019, pages 5053-5064. Association for Computational Linguistics.",
+            "Tomoya Mizumoto, Mamoru Komachi, Masaaki Nagata, and Yuji Matsumoto. 2011. Mining revision log of language learning SNS for automated japanese error correction of second language learners. In Fifth International Joint Conference on Natural Language Processing, IJCNLP 2011, Chiang Mai, Thailand, November 8-13, 2011, pages 147-155. The Association for Computer Linguistics.",
+            "Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. 2014. The conll-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, CoNLL 2014, Baltimore, Maryland, USA, June 26-27, 2014, pages 1-14. ACL.",
+            "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. 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, Demonstrations, pages 48-53. Association for Computational Linguistics.",
+            "Vipul Raheja and Dimitris Alikaniotis. 2020. Adversarial grammatical error correction. In Findings of the Association for Computational Linguistics: EMNLP 2020, Online Event, 16-20 November 2020, volume EMNLP 2020 of Findings of ACL, pages 3075-3087. Association for Computational Linguistics.",
+            "Hongkai Ren, Liner Yang, and Endong Xun. 2018. A sequence to sequence learning for chinese grammatical error correction. In Natural Language Processing and Chinese Computing - 7th CCF International Conference, NLPCC 2018, Hohhot, China, August 26-30, 2018, Proceedings, Part II, volume 11109 of Lecture Notes in Computer Science, pages 401-410. Springer.",
+            "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics.",
+            "Xin Sun, Tao Ge, Furu Wei, and Houfeng Wang. 2021. Instantaneous grammatical error correction with shallow aggressive decoding. 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 5937-5947. Association for Computational Linguistics.",
+            "Zhiqing Sun and Yiming Yang. 2020. An EM approach to non-autoregressive conditional sequence generation. 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 9249-9258. PMLR.",
+            "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 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.",
+            "Haoyu Wang, Shuyan Dong, Yue Liu, James Logan, Ashish Kumar Agrawal, and Yang Liu. 2020. ASR error correction with augmented transformer for entity retrieval. In *Interspeech* 2020, 21st Annual Conference of the International Speech Communication Association, Virtual Event, Shanghai, China, 25-29 October 2020, pages 1550-1554. ISCA.",
+            "Xinyou Wang, Zaixiang Zheng, and Shujian Huang. 2022. Helping the weak makes you strong: Simple multi-task learning improves non-autoregressive translators. CoRR, abs/2211.06075.",
+            "Bingzhen Wei, Mingxuan Wang, Hao Zhou, Junyang Lin, and Xu Sun. 2019. Imitation learning for non-autoregressive neural machine translation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 1304-1312. Association for Computational Linguistics.",
+            "Ziang Xie, Guillaume Genthial, Stanley Xie, Andrew Y. Ng, and Dan Jurafsky. 2018. Noising and denoising natural language: Diverse backtranslation for grammar correction. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 619-628. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "12024",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Helen Yannakoudakis, Ted Briscoe, and Ben Medlock. 2011. A new dataset and method for automatically grading ESOL texts. In The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, Proceedings of the Conference, 19-24 June, 2011, Portland, Oregon, USA, pages 180-189. The Association for Computer Linguistics.",
+            "Wei Zhao, Liang Wang, Kewei Shen, Ruoyu Jia, and Jingming Liu. 2019. Improving grammatical error correction via pre-training a copy-augmented architecture with unlabeled data. 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 156-165. Association for Computational Linguistics.",
+            "Yuanyuan Zhao, Nan Jiang, Weiwei Sun, and Xiaojun Wan. 2018. Overview of the NLPCC 2018 shared task: Grammatical error correction. In *Natural Language Processing and Chinese Computing - 7th CCF International Conference*, NLPCC 2018, Hohhot, China, August 26-30, 2018, Proceedings, Part II, volume 11109 of Lecture Notes in Computer Science, pages 439-445. Springer.",
+            "Zewei Zhao and Houfeng Wang. 2020. Maskgec: Improving neural grammatical error correction via dynamic masking. 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 1226-1233. AAAI Press.",
+            "Junpei Zhou, Chen Li, Hengyou Liu, Zuyi Bao, Guangwei Xu, and Linlin Li. 2018. Chinese grammatical error correction using statistical and neural models. In Natural Language Processing and Chinese Computing - 7th CCF International Conference, NLPCC 2018, Hohhot, China, August 26-30, 2018, Proceedings, Part II, volume 11109 of Lecture Notes in Computer Science, pages 117-128. Springer.",
+            "Long Zhou, Jiajun Zhang, and Chengqing Zong. 2020. Improving autoregressive NMT with non-autoregressive model. In Proceedings of the First Workshop on Automatic Simultaneous Translation, pages 24-29, Seattle, Washington. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            778
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "12025",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? 7",
+            "A2. Did you discuss any potential risks of your work? 7",
+            "A3. Do the abstract and introduction summarize the paper's main claims?",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            695,
+            288
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            300,
+            489,
+            316
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            215,
+            336
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? No response.",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response."
+        ],
+        "bbox": [
+            127,
+            347,
+            880,
+            753
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "5",
+        "bbox": [
+            134,
+            788,
+            146,
+            800
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 7",
+        "bbox": [
+            129,
+            812,
+            880,
+            858
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            866,
+            877,
+            889
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "12026",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?",
+        "bbox": [
+            129,
+            83,
+            878,
+            115
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "5.3",
+        "bbox": [
+            151,
+            117,
+            178,
+            130
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+        "bbox": [
+            129,
+            142,
+            882,
+            190
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "4",
+        "bbox": [
+            151,
+            192,
+            166,
+            204
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?",
+        "bbox": [
+            129,
+            217,
+            882,
+            265
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "4.4",
+        "bbox": [
+            151,
+            267,
+            179,
+            280
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "bbox": [
+            112,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?",
+        "bbox": [
+            127,
+            340,
+            882,
+            372
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            373,
+            248,
+            388
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?",
+        "bbox": [
+            127,
+            399,
+            882,
+            447
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            449,
+            248,
+            464
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?",
+        "bbox": [
+            127,
+            475,
+            882,
+            521
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            524,
+            248,
+            539
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?",
+        "bbox": [
+            127,
+            549,
+            873,
+            565
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            565,
+            248,
+            582
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?",
+        "bbox": [
+            127,
+            592,
+            880,
+            623
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "No response.",
+        "bbox": [
+            149,
+            626,
+            248,
+            640
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "12027",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/Improving Autoregressive Grammatical Error Correction with Non-autoregressive Models/3fd97eae-0b66-43ab-a8c2-6db714a9b22c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af2ffefc92cee1e2935e28173ab99230edd0fe656f8ac18bf7efd06f2a5a419d
+size 478174

2023/Improving Autoregressive Grammatical Error Correction with Non-autoregressive Models/full.md

@@ -0,0 +1,364 @@
+# Improving Autoregressive Grammatical Error Correction with Non-autoregressive Models
+
+Hang Cao<sup>1</sup>, Zhiquan Cao<sup>1</sup>, Chi Hu<sup>1</sup>, Baoyu Hou<sup>1</sup>, Tong Xiao<sup>1,2*</sup>, Jingbo Zhu<sup>1,2</sup>
+
+$^{1}$ NLP Lab, School of Computer Science and Engineering
+
+Northeastern University, Shenyang, China
+
+$^{2}$ NiuTrans Research, Shenyang, China
+
+caohang_neu@outlook.com
+
+{xiaotong,zhujingbo}@mail.neu.edu.cn
+
+# Abstract
+
+Grammatical Error Correction (GEC) aims to correct grammatical errors in sentences. We find that autoregressive models tend to assign low probabilities to tokens that need corrections. Here we introduce additional signals to the training of GEC models so that these systems can learn to better predict at ambiguous positions. To do this, we use a non-autoregressive model as an auxiliary model, and develop a new regularization term of training by considering the difference in predictions between the autoregressive and non-autoregressive models. We experiment with this method on both English and Chinese GEC tasks. Experimental results show that our GEC system outperforms the baselines on all the data sets significantly.
+
+# 1 Introduction
+
+Grammatical Error Correction (GEC) has attracted much attention in recent years, which aims to correct grammatical errors in a given text automatically. It is widely applied to natural language processing scenarios such as Automatic Speech Recognition (ASR) (Kubis et al., 2020; Wang et al., 2020), writing assistant and language learning platforms, etc. The GEC task is characterized by a significant overlap between input and output sentences with only a few errors requiring modification.
+
+Since the transformer-based autoregressive (Vaswani et al., 2017) (AR) model with sequence-to-sequence (seq2seq) architecture has been successful in many generation tasks, a few works (Chollampatt and Ng, 2018) have applied it to the GEC task by taking the incorrect text as the source language and the text without errors as the target language, which has become a mainstream paradigm. However, in the GEC task, the overlap of source and target sentences makes the AR model simply copy most of the tokens
+
+<table><tr><td colspan="2">Incorrect</td><td colspan="2">Correct</td></tr><tr><td rowspan="3">Sub</td><td rowspan="3">X: I have a apple.</td><td colspan="2">Y: I have an apple.</td></tr><tr><td>0.92</td><td>0.85</td></tr><tr><td></td><td>0.42</td></tr><tr><td rowspan="3">Del</td><td rowspan="3">X: I have the an apple.</td><td colspan="2">Y: I have an apple.</td></tr><tr><td>0.90</td><td>0.89</td></tr><tr><td></td><td>0.35</td></tr><tr><td rowspan="3">Ins</td><td rowspan="3">X: I have apple.</td><td colspan="2">Y: I have an apple.</td></tr><tr><td>0.86</td><td>0.91</td></tr><tr><td></td><td>0.25</td></tr></table>
+
+Figure 1: Illustration for the confidence in different types of errors, where Sub denotes substitution, Del means deletion and Ins is insertion.
+
+over from the input to the output. We further find that the AR has high confidence for the tokens that are unchanged between the source and target sentence, while it usually has low confidence for correcting operations such as insertion, deletion, and substitution. Figure 1 is an example to illustrate this phenomenon. Intuitively, we believe that the reasonable cause of this phenomenon is the class imbalance issue (Li and Shi, 2021). With the influence of this problem, the AR model cannot confidently predict these incorrect tokens according to only the local context. Therefore, a natural idea is to improve the model performance by exploiting the global information, which can be captured by the non-autoregressive (NAR) (Gu et al., 2018; Lee et al., 2018) model. Although prior works have explored combining the two approaches through joint training, a combination for the GEC task is still missing. Besides, due to the inconsistency between AR and NAR output, a simple combination of them will lead to poor performance.
+
+In this paper, we propose a simple yet novel approach to focus on incorrect tokens and integrate global information with the non-autoregressive model. Specifically, by masking the tokens in the golden target sentence corresponding to the low confidence positions of the AR output, we construct the input for NAR method. We combine the AR and NAR generation mechanisms to effectively
+
+utilize global information by constraining the consistency of their output distribution.
+
+We conduct experiments on standard English GEC datasets and evaluate the system against strong baselines. Experimental results show that our approach can consistently achieve better results without relying on any resources other than the training data. Furthermore, we compare with a combination method of AR and NAR to verify whether the proposed model is more favorable for the GEC task. Here we use the Chinese GEC dataset as a benchmark to validate the generalization ability of the model. Meanwhile, we also conduct comparative ablation studies to illustrate the effectiveness of our proposed method.
+
+# 2 Related Work
+
+Seq2seq for GEC In recent years, a number of Transformer-based AR methods have been developed for GEC tasks. Junczys-Dowmunt et al. (2018) adapt several methods from low-resource machine translation to GEC by regarding GEC as low-resource machine translation. Zhao et al. (2019) aim to copy the words that overlap between the source and target sentence. They propose a copy-augmented architecture for GEC task which is pre-trained with unlabeled data. A series of work focus on data augmentation (Grundkiewicz et al., 2019; Ge et al., 2018; Lichtarge et al., 2019), Xie et al. (2018) propose to synthesize "realistic" parallel corpus with grammatical errors by back-translation. Zhao and Wang (2020) add a dynamic masking method to the original source sentence during training, which enhances the model performance without requiring additional data. With the help of large pre-trained language models (Kaneko et al., 2020), the performance of Transformer based AR models can be improved effectively. Meanwhile, the NAR approach emerges as a competitive alternative, which can correct the errors by modeling the whole sentence information. Li and Shi (2021) apply a Conditional Random Fields (CRF) layer to conduct non-autoregressive sequence prediction by modeling the dependencies among neighbor tokens.
+
+Combination of AR and NAR The combination of AR and NAR modeling mechanisms has been discussed in other tasks. Wei et al. (2019) use a pre-trained AR model to supervise the decoding state of NAR, which can alleviate the problem of
+
+large search space. Li et al. (2019) propose that learning the hidden representation and attention distribution of AR by hints from the hidden representation can effectively improve the performance of NAR. Several approaches (Guo et al., 2020; Liu et al., 2020) are proposed to gradually guide the model transition from AR to NAR by designing the decoder input and semi-autoregressive tasks as courses. Some other works (Sun and Yang, 2020; Hao et al., 2021; Wang et al., 2022) attempt to utilize a unified framework to train AR and NAR jointly so that the NAR can be enhanced. Besides, Zhou et al. (2020) have also explored using the output of NAR to improve the AR performance. Unlike them, we focus on the GEC task and introduce the NAR model to utilize the global information to help the model understand the context around incorrect tokens.
+
+# 3 Methodology
+
+In this section, we elaborate on our proposed framework for GEC. As shown in figure 2, we introduce the CMLM-based NAR model to integrate more context information into our single model.
+
+# 3.1 Overview
+
+Given the training dataset  $(X,Y)$ , the definition of the GEC task is to correct the original erroneous source  $X$  and generate a sentence  $Y$  without grammatical error, where  $X = (x_{1},x_{2},\dots,x_{K})$  and  $Y = (y_{1},y_{2},\dots,y_{N})$ . Specifically, the transformer encoder takes the source sentence  $X$  as input. Different from previous seq2seq works, our decoder consists of two components: AR decoder and NAR decoder. We keep the AR decoder as a traditional seq2seq decoder without any change. For the NAR decoder, we mask the tokens in the input that corresponds to the positions of the low confidence tokens from the output distribution of the AR decoder. Then, we regenerate tokens in masked positions more accurately by bidirectional semantic modeling of the NAR decoder. Finally, we try to decrease the output distribution distance which is in masked positions between two manners to further improve the model performance during the training stage.
+
+# 3.2 Mask low Confidence
+
+Here, it is important that the output probability represents whether the model is confident in the prediction. As for the GEC task, there are only a
+
+![](images/d7262101471a593e2cf569a989db337c4d35cb0fbfe5d4269bfa883c713187e5.jpg)  
+Figure 2: An illustration of our proposed model. The results  $\{y_1', y_2', y_3', y_4', y_5', y_6', y_7'\}$  are predicted by AR Decoder with the golden target. The mask set  $\{m_3, m_5\}$  indicates the position in the golden target that needs to be masked with a special token  $<\text{mask}>$ . After the mask operation, the input of the NAR decoder is  $\{y_1, y_2, m_3, y_4, m_5, y_6, y_7\}$ , and the output  $\{y_3'', y_5''\}$  are re-predicted by CMLM based NAR decoder, and then the output are used to calibrate the corresponding AR decoder output.
+
+# Algorithm 1 Mask strategy
+
+Input: The AR decoder output  $y_{AR}$ , the golden target  $y_{tgt}$ , the mask ratio  $\delta$
+
+Output: The NAR input  $x_{NAR}$
+
+1: while not converged do  
+2: Select the max probability for each token;  
+3: max_logit  $\leftarrow$  get_max_logit(yAR);  
+4:  $L_{t}\gets \mathrm{get\_target\_length}(y_{tgt})$  
+5:  $L_{m} = L_{t}\times \delta$  
+6: Get the  $L_{m}$  index of the low confidence positions in maxim logit;  
+7: Index  $\leftarrow$  select_index(max_logit,  $L_{m}$ );  
+8: Replace with  $<\mathrm{mask}>$  in the  $y_{tgt}$  corresponding position;  
+9:  $x_{NAR} \gets \text{mask\_target}(Index, y_{tgt})$ ;
+
+# 10: end while
+
+few tokens that need to be modified (About  $10\%$ ). Therefore, the model tends to focus on high confidence correct tokens that need to be kept, but not so much on low confidence tokens that need to be modified. As mentioned above, we choose the low confidence positions in the AR output distribution and substitute them with special symbols  $\langle \text{mask} \rangle$  as the input of the NAR decoder. In this way, the NAR decoder is forced to learn the knowledge of low confidence tokens from the bidirectional context in hidden layers, which helps to boost the performance.
+
+To construct the input effectively, we design a special mask strategy. Details are described in Al-
+
+**Algorithm 1.** Specifically, we select the maximum probability of each token from the AR decoder output distribution. Then, we reorder each token in the output sentence from low confidence to high confidence to get a specific number of positions for the low confidence tokens. We introduce a special token  $<\mathrm{mask}>$  to mask the token at the corresponding position in the golden target, which serves as a placeholder to represent the position where the target token needs to be regenerated. The golden target after the masking operation is used as input to the NAR decoder to introduce bidirectional contextual information.
+
+# 3.3 Restrict Output Consistency
+
+The objective of our model is to overcome the limitations of AR models by introducing the NAR generation mechanism, and then correct sentences with grammatical errors. A common way is to implicitly pass the information learned by the NAR branch to the AR branch using the parameter-sharing method. Specifically, we share the parameters of the Transformer layer in both manners. However, there is a huge difference between the AR manner and the NAR manner in the training process, as shown in Equation 1 and Equation 2, where the AR generation process is more concerned with local dependencies, while the NAR generation process is more concerned with global dependencies.
+
+$$
+P (Y | X) = \prod_ {i = 1} ^ {N} P _ {A R} \left(y _ {i} | X, Y _ {<   i}\right), \tag {1}
+$$
+
+$$
+P (Y | X) = \prod_ {i = 1} ^ {N} P _ {N A R} \left(y _ {i} | X\right). \tag {2}
+$$
+
+The inconsistency between the two generation methods can lead to direct parameter sharing between the two branches without enabling the AR manner to obtain the exact information provided by the NAR manner. This sharing method only implicitly considers the correlation of model parameters and ignores the inconsistency between the two generation methods, which seriously hinders performance.
+
+In contrast, in our work, to make the AR manner learn the information from NAR in a way that is more adapted to AR generation, we take an explicit approach to constrain the two manners. This approach can avoid the inconsistency caused by the different ways of AR and NAR generation and break the performance bottleneck. In practice, we accomplish explicit information modeling by using bidirectional Kullback-Leibler (KL) divergence to force the AR and NAR output distributions at the mask positions to be consistent with each other. Fortunately, Liang et al. (2022) also use KL divergence to combine the advantages of AR and NAR, which gives us much inspiration.
+
+# 3.4 Training and Inference
+
+Multi-Task Framework We learn the GEC model under the multi-task learning framework, including an AR primary task and a NAR auxiliary task. It should be noted that AR and NAR manners are regarded as two different tasks. For the AR task, we employ the negative log-likelihood (NLL) as the loss function which is akin to the traditional seq2seq. Therefore, the optimization objective is:
+
+$$
+\mathcal {L} _ {A R} = - \sum_ {i = 1} ^ {N} \log P _ {A R} \left(y _ {i} \mid X, Y _ {<   i}\right), \tag {3}
+$$
+
+where  $N$  is the target length, and  $Y_{< i}$  represents the tokens before the  $i$ -th time step.  $P_{AR}(y_i|X,Y_{< i})$  represents the output probability of the AR decoder, which will be used in the later process.
+
+For the NAR task, we obtain the positions of the specified number of low confidence tokens based on the mask ratio  $\delta$ , and replace the tokens with the special symbols  $<\mathrm{mask}>$  at the corresponding positions of the golden target. The loss function  $\mathcal{L}_{NAR}$  for NAR task is to minimize the sum of
+
+negative log-likelihood in masked positions:
+
+$$
+\mathcal {L} _ {N A R} = - \sum_ {i = 1} ^ {M} \log P _ {N A R} \left(y _ {i} \mid X, Y _ {\text {m a s k}}\right), \tag {4}
+$$
+
+where  $M$  is the number of the masked tokens, and  $Y_{mask}$  is the set of the tokens in masked. In this way, the NAR decoder regenerates the masked tokens with more context information to help the AR task. Then the loss function of the multi-task framework is:
+
+$$
+\mathcal {L} _ {m} = \lambda_ {t} \mathcal {L} _ {N A R} + (1 - \lambda_ {t}) \mathcal {L} _ {A R}, \tag {5}
+$$
+
+where  $\lambda_{t}$  is the important factor to balance the weight of AR and NAR tasks during training. We will present the design details in the following paragraphs.
+
+Curriculum Learning Compared with the AR task, the NAR task is more complex, and unreasonable weight setting will make training difficult. For example, the excessive weight of the NAR task will disturb the parameter learning of the AR primary task at the beginning. Inspired by curriculum learning (Bengio et al., 2009), which is to imitate the human learning process, we propose the dynamic weight strategy. More concretely, we start with  $\lambda_{t} = 0$  and gradually increase the NAR task weight  $\lambda_{t}$  to introduce learning signals. The dynamic weight scheme is:
+
+$$
+\lambda_ {t} = \frac {t}{T}, \tag {6}
+$$
+
+where  $t$  and  $T$  are the current and total steps of training. We increase the weight linearly in all the experiments.
+
+It is not enough to use only the hard parameter sharing method mentioned above, we regularize the two output distributions  $P_{AR}$  and  $P_{NAR}$  for unconfident words with the token-level bidirectional Kullback-Leibler divergence to further transfer the knowledge of NAR:
+
+$$
+\begin{array}{l} \mathcal {L} _ {K L} = \sum_ {Y _ {\text {m a s k}}} K L \left(P _ {A R} \| P _ {N A R}\right) + \\ \sum_ {Y _ {\text {m a s k}}} K L \left(P _ {N A R} \| P _ {A R}\right). \end{array} \tag {7}
+$$
+
+The final training objective for our GEC model is a combination of the three terms reviewed above as:
+
+$$
+\mathcal {L} = \lambda_ {t} \mathcal {L} _ {N A R} + (1 - \lambda_ {t}) \mathcal {L} _ {A R} + \alpha \mathcal {L} _ {K L}. \tag {8}
+$$
+
+<table><tr><td>Model</td><td>Architecture</td><td>Precision</td><td>Recall</td><td>F0.5</td></tr><tr><td>Transformer Big†</td><td>1024-1024-4096</td><td>65.26</td><td>27.19</td><td>50.98</td></tr><tr><td>LaserTagger* (Malmi et al., 2019)</td><td>-</td><td>50.9</td><td>26.9</td><td>43.2</td></tr><tr><td>Adversarial-GE (Raheja and Alikaniotis, 2020)</td><td>512-512-2048</td><td>64.68</td><td>22.57</td><td>47.10</td></tr><tr><td>ESD+ESC* (Chen et al., 2020)</td><td>1024-1024-4096</td><td>66.0</td><td>24.7</td><td>49.5</td></tr><tr><td>SAD(9+3) (Sun et al., 2021)</td><td>1024-1024-4096</td><td>58.8</td><td>33.1</td><td>50.9</td></tr><tr><td>S2A (Li et al., 2022)</td><td>1024-1024-4096</td><td>65.9</td><td>28.9</td><td>52.5</td></tr><tr><td>CMLM† (Ghazvininejad et al., 2019)</td><td>1024-1024-4096</td><td>46.3</td><td>27.17</td><td>40.59</td></tr><tr><td>Levenshtein Transformer* (Gu et al., 2019)</td><td>1024-1024-4096</td><td>39.9</td><td>24.4</td><td>35.4</td></tr><tr><td>JANUS† (Liang et al., 2022)</td><td>1024-1024-4096</td><td>66.22</td><td>27.76</td><td>51.85</td></tr><tr><td>Ours-base</td><td>512-512-2048</td><td>66.63</td><td>28.70</td><td>52.70</td></tr><tr><td>Ours</td><td>1024-1024-4096</td><td>65.10</td><td>32.29</td><td>54.11</td></tr></table>
+
+Table 1: The results of systems on the CoNLL-2014 English GEC task. For the models with  $\star$ , their performance is from (Chen et al., 2020).  $\dagger$  indicates the models are implemented by us with the released codes of the original papers. The Architecture column represents the embedding, hidden, and FFN size of the model. Here we bold the best results of the models.
+
+Inference During the inference stage, we use the AR decoder to generate the correct sentences, and the inference efficiency is the same as the traditional seq2seq model since the NAR decoder is only used in training.
+
+# 4 Experimental Setup
+
+# 4.1 Datasets
+
+To validate the effectiveness of our proposed GEC model, we conduct a set of experiments on both the restricted track of the BEA-2019 GEC shared task (Bryant et al., 2019) and NLPCC 2018 Task 2 (Zhao et al., 2018).
+
+BEA-2019 GEC shared task This is a public dataset for the English GEC task, we follow the setting of (Chollampatt and Ng, 2018) and take the FCE training set (Yannakoudakis et al., 2011), Lang-8 Corpus of Learner English (Mizumoto et al., 2011), NUCLE (Dahlmeier et al., 2013) and W&I+LOCNESS (Granger, 2014; Bryant et al., 2019) as the training set. The development set is a subset of NUCLE, and our model is evaluated on the CoNLL-2014 (Ng et al., 2014), which is a well-known English GEC benchmark test set. Specifically, we use pre-processed script in (Chollampatt and Ng, 2018) to obtain the parallel corpus.
+
+NLPCC 2018 Task 2 It is the first and latest benchmark dataset for Chinese GEC. We combine
+
+the incorrect sentence with each corrected sentence to build the parallel sentence pairs as described in (Zhao and Wang, 2020) and get 1.2 million sentence pairs in all. Next, we randomly sample 5,000 training instances as the development set. The official test set extracted from PKU Chinese Learner Corpus contains 2,000 samples. We use the combination of two group annotations that mark the golden edits of grammatical errors in these sentences to evaluate our model. Following the setting of NLPCC 2018 Task (Zhao et al., 2018), the tokenization of training data is implemented with the PKUNLP tool<sup>2</sup>.
+
+# 4.2 Settings
+
+While in the training process, we use the base model configuration of the Transformer for the Chinese GEC task, with 6 layers, the number of self-attention heads is set to 8, the embedding dimension is 512 and the size of FFN layer is 2048, the dropout and weight decay is 0.3 and 0.01 respectively. In the English GEC task, we use the big Transformer setting, which contains 6 layers and 16 self-attention heads, the size of word vectors on the source side and the target side are 1024, the FFN layer size is 4096, the dropout is applied with a probability of 0.1 and the weight decay value is set to be 0.0001. We adopt Adam (Kingma and Ba, 2015) optimizer with initial learning rate 0.0005 and 0.0007 for Chinese and English GEC tasks respectively, and a beta value of (0.9, 0.98). We use
+
+<table><tr><td>Model</td><td>Model type</td><td>Precision</td><td>Recall</td><td>F0.5</td></tr><tr><td>Transformer</td><td>Single</td><td>36.91</td><td>15.57</td><td>28.97</td></tr><tr><td>YouDao (Fu et al., 2018)</td><td>Ensemble</td><td>35.24</td><td>18.64</td><td>29.91</td></tr><tr><td>AliGM (Zhou et al., 2018)</td><td>Ensemble</td><td>41.00</td><td>13.75</td><td>29.36</td></tr><tr><td>BLCU (Ren et al., 2018)</td><td>Ensemble</td><td>47.63</td><td>12.56</td><td>30.57</td></tr><tr><td>ESD+ESC (Chen et al., 2020)</td><td>Single</td><td>37.3</td><td>14.5</td><td>28.4</td></tr><tr><td>SAD(9+3) (Sun et al., 2021)</td><td>Single</td><td>33.0</td><td>20.5</td><td>29.4</td></tr><tr><td>S2A (Li et al., 2022)</td><td>Single</td><td>36.57</td><td>18.25</td><td>30.46</td></tr><tr><td>Ours</td><td>Single</td><td>41.90</td><td>15.24</td><td>31.04</td></tr></table>
+
+Table 2: The results of systems on the NLPCC-2018 Chinese GEC task. For a fair comparison, all the results are produced by training on the original NLPCC-2018 training data. We **bold** the best results.
+
+learning rate schedule as in (Vaswani et al., 2017), 10,000 warmup steps for the Chinese GEC task and 4,000 for the English GEC task. Lable smoothing is added with an epsilon value of 0.1. We use 32K Byte Pair Encoding (BPE) (Sennrich et al., 2016) for tokenization on Chinese and English GEC tasks. We save the checkpoint for each epoch and select the best checkpoint based on the loss on the development set. The beam size is 5 during the inference stage. All experiments are based on fairseq (Ott et al., 2019).
+
+# 4.3 Baselines
+
+We compare the performance of the proposed model with several representative baseline methods on both English and Chinese GEC tasks. Specifically, for the English GEC task, Transformer Big is the typical AR model. LaserTagger proposes to predict tags with a smaller vocabulary (Malmi et al., 2019). Adversarial-GEc presents an adversarial learning approach to generate realistic texts in a generator-discriminator framework (Raheja and Alikaniotis, 2020). ESD+ESC is a pipeline model (Chen et al., 2020). SAD employs a new decoding method with a shallow decoder to conduct the prediction (Sun et al., 2021). S2A proposes to integrate action probabilities into token prediction probabilities to obtain the final results (Li et al., 2022). Levenshtein Transformer (Gu et al., 2019) and CMLM (Ghazvininejad et al., 2019) are NAR models, which achieve excellent performance with an iterative generation paradigm. In addition, we also compare with JANUS (Liang et al., 2022), which joints AR and NAR training for sequence generation.
+
+For the Chinese GEC task, we compare our model to all previous systems in the NLPCC 2018
+
+dataset. YouDao corrects the sentences independently by utilizing five different mixture models (Fu et al., 2018). AliGM combines three approaches, including NMT-based, SMT-based, and rule-based together (Zhou et al., 2018). BLCU is based on a multi-layer convolutional seq2seq model (Ren et al., 2018).
+
+# 4.4 Evaluation Metrics
+
+Following the typical previous works (Chen et al., 2020; Li et al., 2022), we use the official MaxMatch  $(M^2)$  (Dahlmeier and Ng, 2012) scorer for evaluation of our grammatical error correction system.  $M^2$  scorer computes the sequence of phrase-level edits between a source sentence and a system hypothesis that achieves the maximal overlap with the gold standard annotation. Given the set of system edits and the set of gold edits for all sentences, the value of precision, recall, and  $F_{0.5}$  are computed by m2scorer<sup>3</sup>.
+
+# 5 Results
+
+# 5.1 Main Results
+
+The results of our proposed approach and recent models on English GEC task are shown in Table 1. We can see that our approach significantly outperforms the baselines mentioned above. Our model achieves an improvement above Transformer Big by nearly 3.1 in  $F_{0.5}$  score, and performs better than the strong baseline S2A, by a large margin of 1.6  $F_{0.5}$ . Moreover, the proposed model surpasses the recent JANUS model by  $F_{0.5}$  score of 2.3, which shows excellent performance on multiple tasks by combining AR and NAR. This result implies that our designed joint training method is more suit-
+
+<table><tr><td rowspan="2">Mask Ratio</td><td colspan="4">BEA-2019</td><td colspan="4">NLPCC-2018</td></tr><tr><td>Precision</td><td>Recall</td><td>F0.5</td><td>F1</td><td>Precision</td><td>Recall</td><td>F0.5</td><td>F1</td></tr><tr><td>10%</td><td>61.70</td><td>31.31</td><td>51.67</td><td>41.61</td><td>38.99</td><td>14.57</td><td>29.20</td><td>21.22</td></tr><tr><td>15%</td><td>62.32</td><td>31.82</td><td>52.30</td><td>40.65</td><td>41.90</td><td>15.24</td><td>31.04</td><td>22.35</td></tr><tr><td>20%</td><td>65.10</td><td>32.29</td><td>54.11</td><td>43.21</td><td>42.24</td><td>14.95</td><td>30.94</td><td>22.09</td></tr><tr><td>25%</td><td>64.87</td><td>30.61</td><td>53.01</td><td>41.68</td><td>40.01</td><td>14.48</td><td>29.58</td><td>21.26</td></tr><tr><td>30%</td><td>63.35</td><td>29.68</td><td>51.63</td><td>40.56</td><td>41.18</td><td>13.68</td><td>29.36</td><td>20.54</td></tr><tr><td>35%</td><td>64.51</td><td>30.43</td><td>52.71</td><td>41.48</td><td>41.73</td><td>13.70</td><td>29.61</td><td>20.63</td></tr></table>
+
+Table 3: Effect of Mask Ratio. The mask ratio represents the percentage of low-confidence tokens in a sentence that are masked. Best results of the Chinese GEC task and the English GEC task are bold separately.  
+
+<table><tr><td>Model</td><td>Sub</td><td>Del</td><td>Ins</td></tr><tr><td>AR</td><td>58.12%</td><td>77.50%</td><td>73.82%</td></tr><tr><td>Ours</td><td>52.30%</td><td>52.84%</td><td>61.70%</td></tr></table>
+
+Table 4: The Correction Coincidence Rate of the AR model and our method on the English CoNLL-2014 test set.
+
+able for the GEC task. It is noteworthy that our model with Transformer base settings still consistently exceeds the baselines with Transformer big settings. These results all support that our proposed approach can effectively improve the AR GEC by using a NAR model.
+
+To validate the effectiveness of our approach, we conduct experiments on the Chinese GEC task and present the results in Table 2. These results demonstrate that the Chinese GEC task is more challenging than the English GEC. Despite this, the proposed model yields a higher  $F_{0.5}$  than the listed methods. Moreover, we can observe that all the top three models are ensemble models, including YouDao, AliGM, and BLCU, but our single model still surpasses them. This result means that our model is generalizable.
+
+# 5.2 Fix more Grammar Errors
+
+We carefully investigate the number of different types of errors corrected in the two datasets, and find that most of the corrected grammar errors are the same between the proposed method and the AR model. To show the advantages of our model intuitively, we propose a Correction Coincidence Rate, which is the number of overlaps of correction errors to the total number of respective correction errors. The results are summarized in Table 4. For computational convenience, the errors are broadly categorized into insertion, deletion, and substitution. From Table 4, the overlap rate of our proposed
+
+method on all types of error modifications is much lower. For instance, the percentage of deletion decreases by  $25\%$ . This indicates that our model is able to correct more grammar errors while maintaining the ability of the AR model.
+
+# 5.3 Ablation Analysis
+
+Effect of Mask Ratio In this section, we present exhaustive investigations on the impact of mask ratio. Here we vary the mask ratios in  $\{0.1, 0.15, 0.2, 0.3, 0.35\}$  and conduct experiments in BEA-2019 and NLPCC 2018. The corresponding results are provided in Table 3. It can be observed that all mask ratios outperform the Transformer baseline. A reasonable reason is that the masking operation makes the model focus more on incorrect tokens, and the model is forced to capture more context information, which facilitates error correction. On the other hand, a small mask ratio (e.g., 0.1) cannot perform as well as a large one (e.g., 0.15), which means that there is a fraction of incorrect tokens that are not focused on. However, too much masking ratio is also not good. It will result in many correct words being masked, which may prevent the correction of incorrect tokens. Note that the choice of mask ratio is distinct for different datasets, and the best balance choices for BEA-2019 and NLPCC-2018 are 0.2 and 0.15 respectively.
+
+Effect of KL Loss Weight  $\alpha$  We explore the effect of KL-divergence loss weight  $\alpha$  in Equation 8. The result is illustrated in Table 6. By comparing the performance with KL loss and without KL loss, we can see that the performance of the former is consistently better than the performance of the latter, which suggests that KL loss can be further combined with information from AR and NAR to correct errors. In addition, the performance is lower
+
+<table><tr><td>Type</td><td>Samples</td></tr><tr><td>SRC</td><td>I think the family will stay mentally healthy as it is, without having emotional stress.</td></tr><tr><td>TGT</td><td>I think the family will stay mentally healthy as it is, without having emotional stress.</td></tr><tr><td>Transformer</td><td>I think the family will stay mentally healthy as it is, without having emotional stress.</td></tr><tr><td>Ours</td><td>I think the family will stay mentally healthy as it is, without having emotional stress.</td></tr><tr><td>SRC</td><td>While we do know that we should not discriminate them based on their limitations...</td></tr><tr><td>TGT</td><td>While we do know that we should not discriminate against them based on their limitations...</td></tr><tr><td>Transformer</td><td>While we do know that we should not discriminate them based on their limitations...</td></tr><tr><td>Ours</td><td>While we do know that we should not discriminate against them based on their limitations...</td></tr><tr><td>SRC</td><td>First and foremost, I would like to share on the advantages of using such social media...</td></tr><tr><td>TGT</td><td>First and foremost, I would like to share the advantages of using such social media...</td></tr><tr><td>Transformer</td><td>First and foremost, I would like to share on the advantages of using such social media...</td></tr><tr><td>Ours</td><td>First and foremost, I would like to share the advantages of using such social media...</td></tr></table>
+
+Table 5: Case studies of the original Transformer model and our proposed model on the English CoNLL-2014 test set. The tokens in red and wave line are errors, while tokens with underline and in green are the corrections made by the gold target or our model.  
+
+<table><tr><td>TrainSet</td><td>α</td><td>Precision</td><td>Recall</td><td>F0.5</td></tr><tr><td rowspan="6">BEA-2019</td><td>0</td><td>59.47</td><td>31.82</td><td>50.66</td></tr><tr><td>0.3</td><td>61.72</td><td>32.21</td><td>52.16</td></tr><tr><td>0.4</td><td>60.68</td><td>31.27</td><td>51.07</td></tr><tr><td>0.5</td><td>65.10</td><td>32.29</td><td>54.11</td></tr><tr><td>0.6</td><td>60.51</td><td>31.88</td><td>51.30</td></tr><tr><td>0.7</td><td>65.03</td><td>31.29</td><td>53.50</td></tr><tr><td rowspan="6">NLPCC-2018</td><td>0</td><td>37.58</td><td>14.34</td><td>28.38</td></tr><tr><td>0.8</td><td>40.53</td><td>15.54</td><td>30.66</td></tr><tr><td>0.9</td><td>39.46</td><td>14.24</td><td>29.14</td></tr><tr><td>1.0</td><td>41.90</td><td>15.24</td><td>31.04</td></tr><tr><td>1.1</td><td>39.34</td><td>14.23</td><td>29.07</td></tr><tr><td>1.2</td><td>40.32</td><td>13.88</td><td>29.20</td></tr></table>
+
+Table 6: The results with different weight.  $\alpha$  equal 0 represents the model without KL loss. Weights are adjusted according to different datasets. Best results are bold.
+
+than the baseline when the value of  $\alpha$  is 0, i.e., the model is fused using only the simple method of parameter sharing. It indicates that simple fusion will lead to poor performance. We also find that the performance is not optimal when  $\alpha$  is set too small or too large. We believe that the model does not learn enough information when  $\alpha$  is set too small, while setting it too large leads to the introduction of too much noise.
+
+# 5.4 Case Study
+
+In order to qualitatively show the effectiveness of global context information, we conduct case studies with Transformer and our proposed model. We pick the cases from the CoNLL-2014 English GEC
+
+test set. The results are listed in Table 5. Generally, it is easy to see that both approaches can copy most of the correct tokens from the source to the target. Nevertheless, when correcting grammatical errors, our approach can predict more accurately by considering more context information. For example, as shown in the third sample in Table 5, the AR model generates the phase "share on" which tends to be consistent with the source language, while our model can delete the token "on" by utilizing more context information. This again confirms that our method can make use of the global information to correct errors.
+
+# 6 Conclusion
+
+In this work, we propose a joint AR and NAR learning objective for the GEC, using a multi-task learning framework. To better predict tokens at low-confidence positions, we introduce additional signals to the training of GEC models by using the NAR model as an auxiliary model. Meanwhile, we develop a new regularization term of training to constrain the inconsistency between the two manners. Through our experiments in the English and Chinese GEC task, the proposed approach can significantly improve the GEC model performance without additional inference costs.
+
+In the future, we are also interested in introducing syntax and lexical knowledge to focus on incorrect tokens to further improve performance.
+
+# 7 Limitations
+
+In this work, we achieve a noticeable improvement in the GEC task by introducing additional context information with a NAR model. However, in order to focus on incorrect tokens, the input of the NAR is required to be constructed based on the AR output distribution. In this way, the AR and NAR model perform sequentially, which leads to much time consumption in the training stage. In the future, we will apply a layer dropout strategy to speed up model training. On the other hand, due to the limitation of computation resources, all experiments are conducted on two Nvidia Titan V GPUs with 12GB VRAM. Therefore, we could not compare with the state-of-the-art models which are pre-trained with 100M synthetic parallel examples (Li et al., 2022). We left it as our future work.
+
+# 8 Acknowledgments
+
+This work was supported in part by the National Science Foundation of China (No.62276056), the National Key R&D Program of China, the China HTRD Center Project (No.2020AAA0107904), the Natural Science Foundation of Liaoning Province of China (2022-KF-16-01), the Yunnan Provincial Major Science and Technology Special Plan Projects (No.202103AA080015), the Fundamental Research Funds for the Central Universities (Nos.N2216016, N2216001, and N2216002), and the Program of Introducing Talents of Discipline to Universities, Plan 111 (No.B16009).
+
+# References
+
+Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning, ICML 2009, Montreal, Quebec, Canada, June 14-18, 2009, volume 382 of ACM International Conference Proceeding Series, pages 41-48. ACM.  
+Christopher Bryant, Mariano Felice, Øistein E. Andersen, and Ted Briscoe. 2019. The BEA-2019 shared task on grammatical error correction. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, BEA@ACL 2019, Florence, Italy, August 2, 2019, pages 52-75. Association for Computational Linguistics.  
+Mengyun Chen, Tao Ge, Xingxing Zhang, Furu Wei, and Ming Zhou. 2020. Improving the efficiency of grammatical error correction with erroneous span detection and correction. In Proceedings of the 2020
+
+Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 7162-7169. Association for Computational Linguistics.  
+Shamil Chollampatt and Hwee Tou Ng. 2018. A multilayer convolutional encoder-decoder neural network for grammatical error correction. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 5755-5762. AAAI Press.  
+Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In Human Language Technologies: Conference of the North American Chapter of the Association of Computational Linguistics, Proceedings, June 3-8, 2012, Montreal, Canada, pages 568-572. The Association for Computational Linguistics.  
+Daniel Dahlmeier, Hwee Tou Ng, and Siew Mei Wu. 2013. Building a large annotated corpus of learner english: The NUS corpus of learner english. In Proceedings of the Eighth Workshop on Innovative Use of NLP for Building Educational Applications, BEA@NAACL-HLT 2013, June 13, 2013, Atlanta, Georgia, USA, pages 22-31. The Association for Computer Linguistics.  
+Kai Fu, Jin Huang, and Yitao Duan. 2018. Youdao's winning solution to the NLPCC-2018 task 2 challenge: A neural machine translation approach to chinese grammatical error correction. In Natural Language Processing and Chinese Computing - 7th CCF International Conference, NLPCC 2018, Ho-hhot, China, August 26-30, 2018, Proceedings, Part I, volume 11108 of Lecture Notes in Computer Science, pages 341-350. Springer.  
+Tao Ge, Furu Wei, and Ming Zhou. 2018. Fluency boost learning and inference for neural grammatical error correction. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 1055-1065. Association for Computational Linguistics.  
+Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask-predict: Parallel decoding of conditional masked language models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 6111-6120. Association for Computational Linguistics.  
+Sylviane Granger. 2014. The computer learner corpus: a versatile new source of data for sola research. In Lerner English on computer, pages 3-18. Routledge.
+
+Roman Grundkiewicz, Marcin Junczys-Dowmunt, and Kenneth Heafield. 2019. Neural grammatical error correction systems with unsupervised pre-training on synthetic data. In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, BEA@ACL 2019, Florence, Italy, August 2, 2019, pages 252-263. Association for Computational Linguistics.  
+Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. 2018. Non-autoregressive neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.  
+Jiatao Gu, Changhan Wang, and Junbo Zhao. 2019. Levenshtein transformer. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 11179-11189.  
+Junliang Guo, Xu Tan, Linli Xu, Tao Qin, Enhong Chen, and Tie-Yan Liu. 2020. Fine-tuning by curriculum learning for non-autoregressive neural machine translation. 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 7839-7846. AAAI Press.  
+Yongchang Hao, Shilin He, Wenxiang Jiao, Zhaopeng Tu, Michael R. Lyu, and Xing Wang. 2021. Multi-task learning with shared encoder for non-autoregressive machine translation. 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 3989-3996. Association for Computational Linguistics.  
+Marcin Junczys-Dowmunt, Roman Grundkiewicz, Shubha Guha, and Kenneth Heafield. 2018. Approaching neural grammatical error correction as a low-resource machine translation task. In Proceedings of the 2018 Conference of the North American Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 595-606. Association for Computational Linguistics.  
+Masahiro Kaneko, Masato Mita, Shun Kiyono, Jun Suzuki, and Kentaro Inui. 2020. Encoder-decoder models can benefit from pre-trained masked language models in grammatical error correction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 4248-4254. Association for Computational Linguistics.
+
+Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Marek Kubis, Zygmunt Vetulani, Mikolaj Wypych, and Tomasz Zietkiewicz. 2020. Open challenge for correcting errors of speech recognition systems. CoRR, abs/2001.03041.  
+Jason Lee, Elman Mansimov, and Kyunghyun Cho. 2018. Deterministic non-autoregressive neural sequence modeling by iterative refinement. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 1173-1182. Association for Computational Linguistics.  
+Jiquan Li, Junliang Guo, Yongxin Zhu, Xin Sheng, Deqiang Jiang, Bo Ren, and Linli Xu. 2022. Sequence-to-action: Grammatical error correction with action guided sequence generation. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 10974-10982. AAAI Press.  
+Piji Li and Shuming Shi. 2021. Tail-to-tail non-autoregressive sequence prediction for chinese grammatical error correction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 4973-4984. Association for Computational Linguistics.  
+Zhuohan Li, Zi Lin, Di He, Fei Tian, Tao Qin, Liwei Wang, and Tie-Yan Liu. 2019. Hint-based training for non-autoregressive machine translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 5707-5712. Association for Computational Linguistics.  
+Xiaobo Liang, Lijun Wu, Juntao Li, and Min Zhang. 2022. Janus: Joint autoregressive and nonautoregressive training with auxiliary loss for sequence generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, pages 8050-8060. The Association for Computer Linguistics.  
+Jared Lichtarge, Chris Alberti, Shankar Kumar, Noam Shazeer, Niki Parmar, and Simon Tong. 2019. Corpora generation for grammatical error correction. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7,
+
+2019, Volume 1 (Long and Short Papers), pages 3291-3301. Association for Computational Linguistics.  
+Jinglin Liu, Yi Ren, Xu Tan, Chen Zhang, Tao Qin, Zhou Zhao, and Tie-Yan Liu. 2020. Task-level curriculum learning for non-autoregressive neural machine translation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 3861-3867. ijcai.org.  
+Eric Malmi, Sebastian Krause, Sascha Rothe, Daniil Mirylenka, and Aliaksei Severyn. 2019. Encode, tag, realize: High-precision text editing. 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 2019, Hong Kong, China, November 3-7, 2019, pages 5053-5064. Association for Computational Linguistics.  
+Tomoya Mizumoto, Mamoru Komachi, Masaaki Nagata, and Yuji Matsumoto. 2011. Mining revision log of language learning SNS for automated japanese error correction of second language learners. In Fifth International Joint Conference on Natural Language Processing, IJCNLP 2011, Chiang Mai, Thailand, November 8-13, 2011, pages 147-155. The Association for Computer Linguistics.  
+Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. 2014. The conll-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, CoNLL 2014, Baltimore, Maryland, USA, June 26-27, 2014, pages 1-14. ACL.  
+Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. 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, Demonstrations, pages 48-53. Association for Computational Linguistics.  
+Vipul Raheja and Dimitris Alikaniotis. 2020. Adversarial grammatical error correction. In Findings of the Association for Computational Linguistics: EMNLP 2020, Online Event, 16-20 November 2020, volume EMNLP 2020 of Findings of ACL, pages 3075-3087. Association for Computational Linguistics.  
+Hongkai Ren, Liner Yang, and Endong Xun. 2018. A sequence to sequence learning for chinese grammatical error correction. In Natural Language Processing and Chinese Computing - 7th CCF International Conference, NLPCC 2018, Hohhot, China, August 26-30, 2018, Proceedings, Part II, volume 11109 of Lecture Notes in Computer Science, pages 401-410. Springer.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with
+
+subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics.  
+Xin Sun, Tao Ge, Furu Wei, and Houfeng Wang. 2021. Instantaneous grammatical error correction with shallow aggressive decoding. 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 5937-5947. Association for Computational Linguistics.  
+Zhiqing Sun and Yiming Yang. 2020. An EM approach to non-autoregressive conditional sequence generation. 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 9249-9258. PMLR.  
+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 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.  
+Haoyu Wang, Shuyan Dong, Yue Liu, James Logan, Ashish Kumar Agrawal, and Yang Liu. 2020. ASR error correction with augmented transformer for entity retrieval. In *Interspeech* 2020, 21st Annual Conference of the International Speech Communication Association, Virtual Event, Shanghai, China, 25-29 October 2020, pages 1550-1554. ISCA.  
+Xinyou Wang, Zaixiang Zheng, and Shujian Huang. 2022. Helping the weak makes you strong: Simple multi-task learning improves non-autoregressive translators. CoRR, abs/2211.06075.  
+Bingzhen Wei, Mingxuan Wang, Hao Zhou, Junyang Lin, and Xu Sun. 2019. Imitation learning for non-autoregressive neural machine translation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 1304-1312. Association for Computational Linguistics.  
+Ziang Xie, Guillaume Genthial, Stanley Xie, Andrew Y. Ng, and Dan Jurafsky. 2018. Noising and denoising natural language: Diverse backtranslation for grammar correction. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 619-628. Association for Computational Linguistics.
+
+Helen Yannakoudakis, Ted Briscoe, and Ben Medlock. 2011. A new dataset and method for automatically grading ESOL texts. In The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, Proceedings of the Conference, 19-24 June, 2011, Portland, Oregon, USA, pages 180-189. The Association for Computer Linguistics.  
+Wei Zhao, Liang Wang, Kewei Shen, Ruoyu Jia, and Jingming Liu. 2019. Improving grammatical error correction via pre-training a copy-augmented architecture with unlabeled data. 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 156-165. Association for Computational Linguistics.  
+Yuanyuan Zhao, Nan Jiang, Weiwei Sun, and Xiaojun Wan. 2018. Overview of the NLPCC 2018 shared task: Grammatical error correction. In *Natural Language Processing and Chinese Computing - 7th CCF International Conference*, NLPCC 2018, Hohhot, China, August 26-30, 2018, Proceedings, Part II, volume 11109 of Lecture Notes in Computer Science, pages 439-445. Springer.  
+Zewei Zhao and Houfeng Wang. 2020. Maskgec: Improving neural grammatical error correction via dynamic masking. 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 1226-1233. AAAI Press.  
+Junpei Zhou, Chen Li, Hengyou Liu, Zuyi Bao, Guangwei Xu, and Linlin Li. 2018. Chinese grammatical error correction using statistical and neural models. In Natural Language Processing and Chinese Computing - 7th CCF International Conference, NLPCC 2018, Hohhot, China, August 26-30, 2018, Proceedings, Part II, volume 11109 of Lecture Notes in Computer Science, pages 117-128. Springer.  
+Long Zhou, Jiajun Zhang, and Chengqing Zong. 2020. Improving autoregressive NMT with non-autoregressive model. In Proceedings of the First Workshop on Automatic Simultaneous Translation, pages 24-29, Seattle, Washington. Association for Computational Linguistics.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? 7  
+A2. Did you discuss any potential risks of your work? 7  
+A3. Do the abstract and introduction summarize the paper's main claims?  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used? No response.  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response.
+
+C Did you run computational experiments?
+
+5
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 7
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+5.3
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+4
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+4.4
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+No response.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+No response.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+No response.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+No response.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+No response.

2023/Improving Autoregressive Grammatical Error Correction with Non-autoregressive Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60fe8e5c5c75d4f652934022de9fd958757888716aee5c290aae177ea7af07de
+size 519943

2023/Improving Classroom Dialogue Act Recognition from Limited Labeled Data with Self-Supervised Contrastive Learning Classifiers/b6413ac7-f15b-410f-9dfa-dd65e39b5f0b_content_list.json

@@ -0,0 +1,1907 @@
+[
+    {
+        "type": "text",
+        "text": "Improving Classroom Dialogue Act Recognition from Limited Labeled Data with Self-Supervised Contrastive Learning Classifiers",
+        "text_level": 1,
+        "bbox": [
+            134,
+            78,
+            867,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Vikram Kumaran Jonathan Rowe Bradford Mott",
+        "bbox": [
+            260,
+            124,
+            742,
+            142
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "North Carolina State University",
+        "bbox": [
+            369,
+            143,
+            630,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{vkumara, jprowe, bwmott}@ncsu.edu",
+        "bbox": [
+            329,
+            160,
+            673,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Snigdha Chaturvedi",
+        "bbox": [
+            240,
+            187,
+            421,
+            203
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "UNC Chapel Hill",
+        "bbox": [
+            260,
+            204,
+            403,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "snigdha@cs.unc.edu",
+        "bbox": [
+            240,
+            221,
+            423,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "James Lester",
+        "bbox": [
+            606,
+            187,
+            724,
+            202
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "North Carolina State University",
+        "bbox": [
+            536,
+            204,
+            794,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "lester@ncsu.edu",
+        "bbox": [
+            589,
+            222,
+            742,
+            235
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            342,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recognizing classroom dialogue acts has significant promise for yielding insight into teaching, student learning, and classroom dynamics. However, obtaining K-12 classroom dialogue data with labels is a significant challenge, and therefore, developing data-efficient methods for classroom dialogue act recognition is essential. This work addresses the challenge of classroom dialogue act recognition from limited labeled data using a contrastive learning-based self-supervised approach (SSCon). SSCon uses two independent models that iteratively improve each other's performance by increasing the accuracy of dialogue act recognition and minimizing the embedding distance between the same dialogue acts. We evaluate the approach on three complementary dialogue act recognition datasets: the TalkMoves dataset (annotated K-12 mathematics lesson transcripts), the DailyDialog dataset (multi-turn daily conversation dialogues), and the Dialogue State Tracking Challenge 2 (DSTC2) dataset (restaurant reservation dialogues). Results indicate that our self-supervised contrastive learning-based model outperforms competitive baseline models when trained with limited examples per dialogue act. Furthermore, SSCon outperforms other few-shot models that require considerably more labeled data<sup>1</sup>.",
+        "bbox": [
+            141,
+            282,
+            460,
+            694
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            709,
+            260,
+            725
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dialogue analysis offers significant potential for improving our understanding of classroom learning and teaching by modeling dialogue between students and teachers. Studies of classroom dialogue can provide deep insight into how students learn most effectively and engage with each other and with teachers (Mercer et al., 2019; Mercer, 2010; Resnick et al., 2010; Hmelo-Silver, 2004). A long-standing goal in analyzing classroom dialogue",
+        "bbox": [
+            112,
+            736,
+            490,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "is to understand how student-student and student-teacher dialogues lead to better student learning outcomes (Wendel and Konert, 2016). This work addresses the problem of dialogue act recognition in K-12 classroom dialogues.",
+        "bbox": [
+            507,
+            253,
+            884,
+            332
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dialogue act recognition has garnered considerable attention and is useful for many tasks such as dialogue generation and understanding (Chen et al., 2022; Lin et al., 2021; Goo and Chen, 2018). Recent efforts in dialogue act recognition are built on large-scale pre-trained language models (Qin et al., 2021, 2020; Wang et al., 2020; Raheja and Tetreault, 2019; Chen et al., 2018). These models demonstrate high performance on standard datasets but require substantial labeled training data and, in some cases, combine other corroborative labels such as sentiment. Finding labeled public datasets of K-12 classroom dialogues is challenging for several reasons. First, there are concerns about participants' privacy and security. Second, researchers often develop individualized coding schemes specific to their design framework, and research context (Mercer, 2010; Song et al., 2019; Hao et al., 2020; Song et al., 2021). Therefore, even when labeled datasets are available, the assortment of coding schemes makes it challenging to cross-train across datasets. Third, classroom dialogue utterances are usually specific to a given subject matter, making generic labeled dialogue datasets less useful as auxiliary data. It is, therefore, essential to develop the capability to build dialogue act recognition models from limited labeled training data.",
+        "bbox": [
+            507,
+            336,
+            884,
+            770
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Our research addresses the lack of large-scale labeled classroom dialogue datasets by using a self-supervised contrastive learning-based model (SSCon) trained using limited labeled data. SSCon uses contrastive learning to transform the dialogue utterance representation into a new embedding space where identical dialogue acts cluster together, and distinct dialogue acts are pushed further apart. The system iteratively improves per",
+        "bbox": [
+            507,
+            774,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ Our code is available at https://gitlab.com/vkumara/SSCon",
+        "bbox": [
+            112,
+            891,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10978",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 10978-10992",
+        "bbox": [
+            220,
+            945,
+            778,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            959,
+            700,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "formance, as the contrastive learning step benefits self-supervision, even when presented with limited labeled data. Experiments show that SSCon outperforms competitive baselines with just tens of labeled examples per dialog act in both a K-12 mathematics classroom dataset and an everyday conversation dataset, DailyDialog. Our key contributions are the following:",
+        "bbox": [
+            112,
+            84,
+            489,
+            212
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a novel self-supervised contrastive learning dialogue act recognizer.",
+            "- We test our model on multiple datasets in distinctly different domains under label-scarce settings. Our experiments show that our model outperforms strong baselines.",
+            "- We illustrate with an ablation study why our model outperforms the baseline."
+        ],
+        "bbox": [
+            134,
+            230,
+            489,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Contrastive Learning Model",
+        "text_level": 1,
+        "bbox": [
+            112,
+            406,
+            394,
+            422
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Problem Definition",
+        "text_level": 1,
+        "bbox": [
+            112,
+            434,
+            310,
+            448
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "A dialogue  $D = (u_{1}, u_{2}, \\dots, u_{N})$  in a dialogue dataset consists of a sequence of  $N$  utterances, and a set of dialogue acts  $A$ . Dialogue act recognition (DAR) is defined as a classification problem, that involves recognizing the dialogue act  $da_{i} = DAR(u_{i}|u_{i-1}, \\dots, u_{i-m})$  for utterance  $u_{i}$ , given its context, a set of previous  $m$  utterances, where  $da_{i} \\subseteq A$ . The task is formalized as a multiclass classification problem and sometimes a multi-label classification problem, depending on the coding scheme used.",
+        "bbox": [
+            112,
+            457,
+            489,
+            633
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Approach and Assumptions",
+        "text_level": 1,
+        "bbox": [
+            112,
+            649,
+            378,
+            665
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "While our approach is primarily evaluated on the multi-class classification problem, we also test our model on the multi-label dataset DSTC2 (Henderson et al., 2014), with promising results.",
+        "bbox": [
+            112,
+            671,
+            487,
+            734
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Since we test a few-shot learning scenario, we start with a small set of labeled utterance examples I and the remaining training set  $U$  of unlabeled utterances. We also have a set of labeled utterances  $T$  set aside as a validation set.",
+        "bbox": [
+            112,
+            737,
+            487,
+            816
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 SSCon Overview",
+        "text_level": 1,
+        "bbox": [
+            112,
+            831,
+            294,
+            845
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we describe the architecture of our model, the self-supervised contrastive learning (SS-Con) based multi-class classifier, shown in Figure 1. Our model operates in multiple stages. We",
+        "bbox": [
+            112,
+            854,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "begin by finetuning a large pre-trained transformer-based language model (PLM) trained on large publicly available dialogue datasets using our domain-specific dialogue dataset. We use the finetuned PLM to generate dialogue embeddings for our model. In Stage 1, we use the finetuned PLM to encode the utterance and its dialogue history. We also use sentence-BERT (Reimers and Gurevych, 2019) to create a latent representation of each utterance. A classifier built on the latent representation makes an initial soft prediction of the dialogue act. We distill the initial high-confidence predictions as soft labels for Stage 2. In Stage 2, the soft labels from Stage 1 train an encoder using contrastive learning. It translates the latent representations from Stage 1 into a vector in a different encoding space where identical dialogue acts cluster together while distinct dialogue acts are separated. In Stage 3, we pass the utterance representations through the encoder trained in Stage 2 to get the embeddings, which will be the input to classify the dialogue acts. The high probability soft labels from the Stage 3 classifier are sent back to Stage 1 as input to self-supervise the model in the next iteration.",
+        "bbox": [
+            505,
+            84,
+            885,
+            470
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.4 Pretraining",
+        "text_level": 1,
+        "bbox": [
+            507,
+            481,
+            648,
+            495
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For our pretraining stage, we finetune DialogGPT (Zhang et al., 2020), a dialogue PLM built on GPT2 (Radford et al.) and trained on 147M Reddit and other online conversations, licensed under MIT License. Given an utterance  $u_{i}$  and its context of a set of previous m utterances, the dialogue PLM is finetuned to maximize the conditional probability for the subsequent utterance,  $P(u_{i + 1}|u_i,u_{i - 1},\\dots ,u_{i - m})$ . For pretraining, we create a dataset consisting of target utterances and its context consisting of preceding  $m$  utterances. SSCon uses the finetuned PLM's hidden state as the representation of the utterance  $u_{i}$  given its dialogue context  $(u_{i - 1}\\dots u_{i - m})$ .",
+        "bbox": [
+            507,
+            502,
+            885,
+            728
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.5 Stage 1: Context Classifier",
+        "text_level": 1,
+        "bbox": [
+            507,
+            737,
+            764,
+            753
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The Stage 1 classifier concatenates the pretrained dialogue PLM's last layer embedding  $H_{i}\\in R^{768}$  with the sentence-BERT embedding  $S_{i}\\in R^{384}$  of the utterance  $u_{i}$  and predicts the dialog act,  $y_{i}^{stage1}$  for the utterance,",
+        "bbox": [
+            507,
+            758,
+            882,
+            837
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ny _ {i} ^ {\\text {s t a g e 1}} = C L _ {\\text {s t a g e 1}} \\left(X _ {i}, \\Theta\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            848,
+            882,
+            869
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nX _ {i} = \\left[ H _ {i}; S _ {i} \\right] \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            642,
+            878,
+            882,
+            896
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH _ {i} = \\text {D i a l o G P T} \\left(u _ {i}, u _ {i - 1}, \\dots , u _ {i - m}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            549,
+            903,
+            882,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10979",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/2063650e1fe7cc04479a1be6f1c79c2dc554171972558e39a0079d8d3ba1ab7b.jpg",
+        "image_caption": [
+            "Figure 1: Self-supervised contrastive learning (SSCon) multiclass classifier"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            84,
+            823,
+            280
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS _ {i} = \\text {S e n t e n c e B E R T} (u _ {i}) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            329,
+            487,
+            346
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Where  $y_{i}^{stage1}$  is the predicted dialogue act class for the last utterance  $u_{i}$  and  $CL_{stage1}$  is the Stage 1 multi-class classifier with trainable weights  $\\Theta$ . The concatenated vector  $X_{i}$  represents the utterance  $u_{i}$  as an independent sentence and in the context of its dialogue  $(u_{i-1}, \\dots, u_{i-m})$ . The output of Stage 1 is a dialogue act label (and the corresponding prediction probability) for each utterance in the dataset. For the choice of classifier, we explored both MLP classifiers (Haykin, 1994) and XGBoost (Chen and Guestrin, 2016); however, any reasonable classifier can be utilized. We chose the XGBoost classifier as it was faster to train and run without any impact on the performance of our model. However, for one of the datasets (DSTC2), where an utterance could have multiple labels, we used an MLP-based multi-label classifier.",
+        "bbox": [
+            112,
+            353,
+            487,
+            626
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the first iteration of the self-supervision process, the training data used for the classifier is limited to the number of available labeled samples, usually between ten to a hundred examples per class. In subsequent iterations, the number of samples increases based on the soft labels from earlier iterations.",
+        "bbox": [
+            112,
+            629,
+            489,
+            740
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.6 Stage 2:Contrastive Encoder",
+        "text_level": 1,
+        "bbox": [
+            112,
+            753,
+            393,
+            768
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The prediction from the Stage 1 classifier,  $CL_{stage1}$ , is used as (soft) ground truth for Stage 2. Specifically, we use the dialogue act labels with the highest confidence in terms of prediction probability as new soft labels along with the initial labeled examples used in Stage 1. Including high confidence, soft-labeled samples increases the effective size of the training set with each iteration. Using this data, we adopt a contrastive training approach.",
+        "bbox": [
+            112,
+            774,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This training process creates a network that can encode the finetuned PLM latent representation into a space where utterances with the identical dialogue act class are close together while utterances with distinct dialogue act labels are farther apart. The labeled samples are paired to generate positive and negative triplets  $P = (X_{i}, X_{j}, P_{ij})$  where  $X_{i}$  (Equation 2) is the concatenated encoding defined in the previous section for a given utterance  $u_{i}$ , and  $P_{ij}$  is 1 if both the utterances map to the same dialogue act or -1 if they map to different dialogue acts. In the case of multi-label utterances,  $P_{ij}$  is the cosine distance between the one-hot encoding of the dialogue act vectors of the two utterances. The encoder is a five-layer MLP that transforms the  $X_{i} \\in R^{1152}$  concatenated latent representation into an  $E_{i} \\in R^{384}$  encoding. We use a twin encoder network to train on the positive/negative triplets  $\\mathbf{P}$  using cosine similarity between output embeddings as the similarity score. Given a triplet  $(X_{i}, X_{j}, P_{ij})$  the network trains,",
+        "bbox": [
+            507,
+            329,
+            885,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {i} = B _ {\\text {e n c o d e r}} \\left(X _ {i}\\right) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            621,
+            678,
+            882,
+            696
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {j} = B _ {\\text {e n c o d e r}} \\left(X _ {j}\\right) \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            620,
+            706,
+            884,
+            724
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nl _ {i j} = \\operatorname {c o s i n e \\_ s i m i l a r i t y} \\left(E _ {i}, E _ {j}\\right) \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            728,
+            882,
+            746
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nl o s s _ {i j} = M S E \\left(l _ {i j}, P _ {i j}\\right) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            601,
+            752,
+            882,
+            770
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The trained encoder transforms each utterance, into a new encoding,  $E_{i} = B_{encoder}(X_{i})$ , where  $E_{i} \\in R^{384}$  is the generated embedding of the MLP encoder ( $B_{encoder}$ ).",
+        "bbox": [
+            507,
+            775,
+            882,
+            840
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.7 Stage 3: Embedding Classifier",
+        "text_level": 1,
+        "bbox": [
+            507,
+            850,
+            794,
+            866
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this final stage of our model, we use the encoder from the Contrastive Encoder network that was trained in the previous stage to convert each  $X_{i}$",
+        "bbox": [
+            507,
+            871,
+            882,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10980",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "into the embedding  $E_{i}$  used as input to the Stage 3 classifier.",
+        "bbox": [
+            112,
+            84,
+            487,
+            115
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ny _ {i} ^ {\\text {s t a g e 3}} = C L _ {\\text {s t a g e 3}} \\left(E _ {i}, \\Phi\\right) \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            198,
+            128,
+            487,
+            149
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $y_{i}^{state3}$  is the dialogue act labels for the utterance  $u_{i}$  and  $CL_{state3}$  is a multi-class classifier with trainable weights  $\\Phi$ . Like Stage 1, we use an XGBoost classifier. We use the dialogue act labels with the highest confidence in terms of prediction probability from Stage 1 as the soft label for training the classifier.",
+        "bbox": [
+            112,
+            161,
+            487,
+            275
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The training starts with a small set of labeled utterance examples  $L$ , and the remaining training set  $U$  of unlabeled utterances. During every iteration of the self-supervision process, the model labels the unlabeled  $U$  utterances in our training set. We filter out the utterances labeled with low confidence (low prediction probability of  $CL_{\\text{stage3}}$ ). The distilled silver-labeled instances are moved from  $U$  to  $L$ , along with the initial labeled examples. The updated  $L$  is the training set for Stage 1, starting the next iteration of the self-supervision process.",
+        "bbox": [
+            112,
+            275,
+            487,
+            454
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            467,
+            321,
+            483
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            494,
+            344,
+            508
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The pretraining stage involves fine-tuning the DialogPT model with utterance data from a given dialogue dataset. The training set uses nine previous utterances as context. We finetune a HuggingFace pre-trained \"dialogPT-small\" base model on a single GPU for four epochs. The version of DialogPT we used is a 12-layer transformer. We use the hidden state vector  $H_{i}$  of the end-of-sentence tag in the  $12^{\\mathrm{th}}$  layer of the transformer as the embedding representing the last utterance in the sequence.",
+        "bbox": [
+            112,
+            514,
+            487,
+            677
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Stage 1 of the model uses a dialog act classifier. We implemented an MLP-based classifier and an XGBoost classifier. There was no difference in performance between the two classifiers, so we picked XGBoost as it is a standard baseline. We use the hidden state,  $H_{i}$ , of the last layer as the embedding to represent the dialogue context of the utterance. For the sentence-BERT embedding  $S_{i}$  of the utterance we used a pre-trained network (\"all-MiniLM-L6-v2\") that is trained on roughly 1B sentence pairs (Reimers and Gurevych, 2019). The input to the model is the concatenated vector of  $H_{i}$  and  $S_{i}$ . After running the model for various distillation thresholds, a threshold of 0.85 was used for high-probability soft-label distillation based on",
+        "bbox": [
+            112,
+            678,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d2704382bee809b0eff47e5086ed9b80fe3c02d2407170d56677d670cf4369b9.jpg",
+        "image_caption": [
+            "Figure 2: Improvement of performance over a baseline XGBoost classifier by our SSCon classifier."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            534,
+            82,
+            850,
+            284
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "a simple grid search on the threshold parameter as discussed in the appendix.",
+        "bbox": [
+            507,
+            350,
+            880,
+            381
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Stage 2 of the model is a contrastive encoder network. Each encoder network in the twin network is a 5-layer MLP with a  $20\\%$  dropout between the layers and an output embedding vector of dimension 384. About 1.2 million similarity pairs are used to train the twin network for 4-6 epochs on a single GPU. Stage 3 is an XGBoost classifier head on top of the encoder trained in Stage 2. The iterative self-supervision process continues until we show no improvement in the validation data. We run 5-10 iterations. In our experiments, we use a validation set to determine when to stop the iterations. We report the results on the test set.",
+        "bbox": [
+            507,
+            382,
+            882,
+            592
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Datasets",
+        "text_level": 1,
+        "bbox": [
+            509,
+            607,
+            621,
+            620
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We use three datasets in our experiments: the TalkMoves, the DailyDialog, and the Dialogue State Tracking Challenge 2 datasets. The TalkMoves dataset (Suresh et al., 2022a) consists of 567 human-annotated class video transcripts of K-12 mathematics lessons between teachers and students. A human-transcribed dataset consists of 174,186 teacher utterances and 59,874 student utterances. The dialogue act labels in the dataset have an inter-rater agreement score above  $90\\%$  for all labels. The dialogue acts for student utterances include 'relating to another student', 'asking for more information', 'making a claim', and 'providing evidence'. The dialogue acts for teacher utterances include: 'keeping everyone together', 'getting students to relate', 'restating', 'revoicing', 'pressing for accuracy', and 'pressing for a reason'. For the TalkMoves dataset, we train our model on student",
+        "bbox": [
+            507,
+            629,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10981",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "utterances.",
+        "bbox": [
+            114,
+            86,
+            196,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "DailyDialog is a multi-turn dialogue dataset (Li et al., 2017) consisting of everyday conversations. The dataset consists of 13,118 conversations between multiple people. Four categories of dialogue acts are coded in the dataset. They are 'inform'  $(45\\%)$ , 'questions'  $(29\\%)$ , 'directives'  $(17\\%)$ , and 'commissive'  $(9\\%)$ .",
+        "bbox": [
+            112,
+            102,
+            487,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To evaluate our approach on multi-label data set, we train on the DSTC2 dataset (Henderson et al., 2014). This dataset contains dialogues between crowdsourced workers and automated dialogue systems in the restaurant reservations domain, with 1000 train and test dialogues and about 21 dialogue acts.",
+        "bbox": [
+            112,
+            217,
+            489,
+            328
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            347,
+            233,
+            361
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We run our experiments on different datasets described above. The TalkMoves dataset does not yet have many published baselines for dialogue act recognition, so we compare our approach against multiple baseline models. One is a baseline XGBoost classifier using the same utterance representation input as our model. This classifier is the same as our Stage 1 classifier. The second baseline is a self-supervised XGBoost classifier similar to SSCon without the contrastive learning in Stage 2. For the third baseline, we use an embedding prototype distance-based classifier. The average sentence embedding vector for labeled examples from each class represents a prototype for each label. We then measure the cosine distance of every utterance's sentence embedding in our test set against each class-prototype embedding. An utterance is assigned the class label of the closest prototype in the embedding space. To validate SS-Con against current state-of-the-art dialogue act recognition models, we compare our results on the DailyDialog dataset against a few-shot learning model and a model that uses all available labeled training data. The state-of-the-art Co-GAT model (Qin et al., 2021) uses all training data available, including related sentiment labels on the utterance. Trained on very limited data, we do not expect our model to beat Co-GAT but use its performance as an upper bound. ProtoSeq (Guibon et al., 2021) is a sequential prototypical network trained to work in a few-shot fashion. However, they do use all training data to train their network. We use this model for comparison purposes as our approach is a type of few-shot learner using a small number of samples.",
+        "bbox": [
+            112,
+            370,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/93b74bb9d96f7dd951916098b2bc3241f28203c0b49c1d087915696d3752d5af.jpg",
+        "image_caption": [
+            "Figure 3: Heatmaps show the median cosine distance between embedding examples (TalkMoves dataset) for each dialog act type (8-relating to another student; 9-Asking for more info; 10-making a claim; 11-providing evidence/reasoning). The left heatmap corresponds to the input embeddings. The right heatmap corresponds to the Contrastive Encoder embeddings"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            80,
+            680,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/0a22d0461e1f1917e5dc45e866e5a90cc9ec0030c408ddf5aac49238875b468c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            682,
+            80,
+            865,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We also compare SSCon to standard XGBoost as we do for the TalkMoves dataset. We also compare SSCon dialogue act recognition results on DSTC2, a multi-label problem, against the self-supervised student-teacher approach by (Mi et al., 2021a).",
+        "bbox": [
+            507,
+            350,
+            882,
+            431
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            442,
+            705,
+            456
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The various baselines are compared with SSCon using four metrics. The Matthews Correlation Coefficient (MCC) for multi-class classification has a range of -1 to 1 and handles imbalanced datasets well. In some cases, the baselines only report the F1 (macro) score, the arithmetic mean of individual class F1 scores giving equal weight to all classes. We also report macro-averaged precision and recall.",
+        "bbox": [
+            507,
+            463,
+            882,
+            605
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            619,
+            606,
+            634
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The first dataset we consider is the TalkMoves dataset. To evaluate the models in a few-shot learning scenario, we experiment using 10-70 labeled instances per dialogue act type (less than  $1\\%$  of the overall data). Table 1 shows our results. SSCon shows an improvement of about  $7 - 10\\%$  over the Utterance Embedding XGBoost baseline and the Prototype Distance Classifier when trained on 70 labeled examples. When SSCon is compared with the self-supervised XGBoost classifier without contrastive learning, it shows an improvement of  $1\\%$  for the 70 labeled examples case but an improvement of more than  $60\\%$  for the 10 labeled example case. This difference in performance suggests that contrastive learning is more beneficial when working with fewer labeled examples. Besides the three baselines, we also report the performance of a pub",
+        "bbox": [
+            507,
+            645,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10982",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/3dee35fefcca03771335e804691f5e2efcdcc520b95e211d91ac329eeeaa2ad5.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Model Type</td><td>Labeled Examples per Class</td><td>MCC</td><td>F1 (macro)</td><td>Precision (macro)</td><td>Recall (macro)</td></tr><tr><td rowspan=\"9\">TalkMoves Student</td><td rowspan=\"3\">SSCon Classifier</td><td>10</td><td>0.295±0.042</td><td>0.412±0.044</td><td>0.454±0.034</td><td>0.512±0.016</td></tr><tr><td>30</td><td>0.383±0.035</td><td>0.495±0.033</td><td>0.492±0.032</td><td>0.588±0.012</td></tr><tr><td>70</td><td>0.457±0.015</td><td>0.566±0.015</td><td>0.534±0.014</td><td>0.652±0.011</td></tr><tr><td>Utterance Embedding XGBoost</td><td>70</td><td>0.421±0.025</td><td>0.507±0.023</td><td>0.502±0.019</td><td>0.615±0.015</td></tr><tr><td rowspan=\"2\">Self Supervised XGBoost with- out Contrastive Learning</td><td>10</td><td>0.183±0.047</td><td>0.327±0.037</td><td>0.361±0.024</td><td>0.397±0.031</td></tr><tr><td>70</td><td>0.451±0.021</td><td>0.514±0.021</td><td>0.535±0.015</td><td>0.609±0.016</td></tr><tr><td>Prototype Distance Classifier</td><td>70</td><td>0.369±0.015</td><td>0.463±0.022</td><td>0.457±0.009</td><td>0.558±0.011</td></tr><tr><td rowspan=\"2\">Transformer for Seq Classifier (Suresh et al., 2022b)</td><td>All training</td><td>0.6716</td><td>0.7312</td><td>-</td><td>-</td></tr><tr><td>70% training</td><td>0.588±0.005</td><td>0.653±0.005</td><td>-</td><td>-</td></tr><tr><td rowspan=\"6\">Daily Dialog</td><td rowspan=\"3\">SSCon Classifier</td><td>10</td><td>0.409±0.065</td><td>0.471±0.037</td><td>0.550±0.064</td><td>0.510±0.044</td></tr><tr><td>30</td><td>0.488±0.046</td><td>0.564±0.043</td><td>0.608±0.013</td><td>0.619±0.039</td></tr><tr><td>70</td><td>0.601±0.022</td><td>0.660±0.018</td><td>0.663±0.009</td><td>0.687±0.015</td></tr><tr><td>Utterance Embedding XGBoost</td><td>70</td><td>0.567±0.013</td><td>0.638±0.008</td><td>0.639±0.008</td><td>0.671±0.009</td></tr><tr><td>ProtoSeq (Guibon et al., 2021)</td><td>Few-shot</td><td>0.392±0.023</td><td>0.352±0.030</td><td>-</td><td>-</td></tr><tr><td>Co-GAT (Qin et al., 2021)</td><td>All training</td><td>-</td><td>0.794</td><td>0.81</td><td>0.781</td></tr><tr><td rowspan=\"4\">DSTC2</td><td rowspan=\"2\">SSCon Classifier</td><td>1%</td><td></td><td>0.293±0.025</td><td>0.322±0.041</td><td>0.293±0.007</td></tr><tr><td>10%</td><td></td><td>0.447±0.008</td><td>0.456±0.024</td><td>0.449±0.011</td></tr><tr><td rowspan=\"2\">ToD-BERT-ST (Mi et al., 2021b)</td><td>1%</td><td></td><td>0.285±0.040</td><td>-</td><td>-</td></tr><tr><td>10%</td><td></td><td>0.405±0.090</td><td>-</td><td>-</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            878,
+            419
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Comparison of results for the Daily Dialog, TalkMoves and DSTC2 dataset against baselines.",
+        "bbox": [
+            151,
+            429,
+            842,
+            445
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "lished model that uses  $100\\%$  (21K labeled samples) and  $70\\%$  of the labeled training data. SSCon works in a label-scarce scenario, so our results are lower than the model using the complete labeled data set. Figure 2 shows the impact of labeled example counts on the overall performance of the SSCon classifier and an XGBoost classifier. While our iterative approach increases the performance on average by almost  $50\\%$  over the baseline for a small training size (10 labels per class), it is about  $9\\%$  for a larger labeled set size (70 labels per class). We notice that the overall performance improvement using SSCon is more with small labeled sets, and the improvement tapers as the labeled examples count increases.",
+        "bbox": [
+            112,
+            468,
+            489,
+            709
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The second dataset we consider is the DailyDialog dataset with four possible dialogue acts. Like the TalkMoves dataset, we consider 10 to 70 labeled examples per class. Table 1 shows that our performance, with just 70 instances per class (0.6% of 21521 labeled examples), is within 16% of the state-of-the-art dialogue act recognition model (Qin et al., 2021) that uses 100% of all the training data labels plus auxiliary sentiment labels. We also show that we outperform the other few-shot learning model, ProtoSeq (Guibon et al., 2021), by a significant margin even though they train using the",
+        "bbox": [
+            112,
+            726,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "entire labeled dataset. We investigate the change in performance of SSCon with increasing size of the labeled training set (Figure 2). As before, the performance improvement of SSCon over a standard XGBoost classifier is more significant for smaller label sets and less so as the number of samples increases. With ten samples per class, SSCon performance improvement over the baseline classifier is almost  $41\\%$ . With seventy labels per class, the improvement is only about  $9\\%$ .",
+        "bbox": [
+            507,
+            468,
+            884,
+            630
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We also compare our results on the DSTC2 dataset on restaurant inquiries. This dataset differs from the other two as each utterance can have multiple dialogue act labels. SSCon performs better on the DSTC2 dataset than the baseline self-supervising student-teacher model ToD-BERT-ST (Mi et al., 2021a) by 3 to  $10\\%$ , depending on the size of the labeled training dataset. The ToD-BERT-ST model uses a data augmentation approach, while we use a contrastive learning approach. Both these approaches are complementary, and future work should explore using them together.",
+        "bbox": [
+            507,
+            709,
+            884,
+            903
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10983",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d1e52837370d8d2f4f79f55e2f8a8e7ca8042131a27300a135265a81419b4343.jpg",
+        "image_caption": [
+            "Figure 4: Confusion matrix with counts normalized by actual label counts (TalkMoves Dataset) for the SSCon classifier (70-samples per class training set). Dialog acts: 8-relating to another student; 9-Asking for more info; 10-making a claim; 11-providing evidence/reasoning."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            149,
+            84,
+            457,
+            265
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 SSCon Component Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            372,
+            391,
+            388
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Clustering in Embedding Space",
+        "text_level": 1,
+        "bbox": [
+            112,
+            398,
+            410,
+            416
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The fundamental intuition behind our approach is that contrastive learning brings similar utterances close to each other and dissimilar utterances further apart in the embedding space. Figure 3 shows the spatial clustering of labels in the embedding space of the trained Contrastive Encoder for one iteration with the TalkMoves dataset. Each square in the heatmap corresponds to the median cosine similarity distance between individual dialogue act class examples. To compute this distance, we take utterances from each dialogue act class and calculate the median cosine distance between their embeddings. Each cell in the heatmap shows a scaled distance, with the blue color corresponding to closer embeddings and red corresponding to the embeddings being farther away. The heatmap on the left corresponds to the distance matrix between embeddings provided as input to Stage 2. The heatmap to the right corresponds to the distance matrix for embeddings that the Contrastive Encoder has transformed. Please note that the figure uses scaled distance values, and we show four student dialog act classes.",
+        "bbox": [
+            112,
+            419,
+            489,
+            788
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "When looking at the heatmap on the left for the input feature space, instances belonging to the dialogue act \"relating to another student\" (first row, label 8) are closer to the instances belonging to the dialogue act, \"making a claim\" (third row, label 10) than among themselves. After being transformed by the Contrastive Encoder (heatmap on the right), the smallest distances for each dialogue act fall on",
+        "bbox": [
+            112,
+            790,
+            489,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2861d58d961ed45d795f497f4fc45ec482f4b1e2c114515c35fa8616e919837d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Classifier</td><td>LastUtterance(u_i)</td></tr><tr><td>&quot;relating to another&quot; predicted as &quot;making a claim&quot;</td><td>&quot;Reciprocal&quot; \n&quot;I don&#x27;t agree, I measured them&#x27; \n&quot;by five&quot;</td></tr><tr><td>&quot;relating to another&quot; predicted as &quot;relating to another&quot;</td><td>&quot;He didn&#x27;t show his work&quot; \n&quot;You don&#x27;t know that&#x27; \n&quot;The first one&quot;</td></tr><tr><td>&quot;making a claim&quot; predicted as &quot;making a claim&quot;</td><td>&quot;Two fifths&quot; \n&quot;Y intercept&#x27; \n&quot;power of three&quot;</td></tr></table>",
+        "bbox": [
+            514,
+            80,
+            880,
+            229
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Examples of true positive and false negative for a couple of the TalkMoves dialog acts",
+        "bbox": [
+            507,
+            237,
+            882,
+            267
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the diagonal (blue), as one expects with clustered labels. The diagonal corresponds to utterances with the same label. The non-diagonal terms, which are utterances with different labels, are pulled farther apart (red).",
+        "bbox": [
+            507,
+            293,
+            882,
+            374
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Classifier Performance by Dialogue Act",
+        "text_level": 1,
+        "bbox": [
+            507,
+            384,
+            865,
+            399
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 shows a confusion matrix with counts normalized by the ground truth label counts. In the TalkMoves dataset, the classifier struggles the most with the dialogue act corresponding to \"relating to another student\" (label 8). It is not able to clearly distinguish between \"relating to another student\" and \"making a claim\" (label 10). Label-scarce training suggests we only cover a limited variety of examples for each dialogue act class. In Table 2, we show some examples of correct and incorrect predictions. The first row shows utterances of the type \"relating to another\" mislabeled as \"making a claim\". The other two rows are true positive examples. We can see that the example utterances are hard to distinguish between \"making a claim\" or \"relating to another\" even for a human. The difference is in the context of the dialogue, and with limited training data, such distinctions are hard to make and might lead to overfitting.",
+        "bbox": [
+            507,
+            404,
+            884,
+            709
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Labeled Sample Selection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            721,
+            756,
+            736
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The initial set of labeled samples, in essence, drives the final performance of SSCon. Figure 5 shows the performance distribution of the baseline, XGBoost, for multiple runs for different sample sizes. Two trends are evident as the number of samples increases for every dataset. The first trend is that the performance improves as more samples are available for training. The improvement in performance with sample size is an expected trend as more examples indicate more information for the classifier to model. The second trend is that the variance",
+        "bbox": [
+            507,
+            741,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10984",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7763f39eb297626566414606b3937a837df4ce361efbdf81b77836fb3bd1b68f.jpg",
+        "image_caption": [
+            "Figure 5: Drift in performance as we increase the number of labeled samples (TalkMoves dataset)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            84,
+            453,
+            258
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "in performance decreases as the number of samples increases. One can assume a certain level of variability in the labeled training sets data's pattern of utterances for each dialogue act class. The dataset can only capture some of the variances in the data when the number of labeled examples is small. Hence, the classifier's performance can vary considerably depending on included examples for smaller training sets.",
+        "bbox": [
+            112,
+            326,
+            489,
+            470
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            488,
+            268,
+            502
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Kalchbrenner and Blunsom (2013) proposed a hierarchical network architecture that combines CNN and RNN models to capture discourse structure. Subsequent work on dialogue act recognition (DAR) followed a similar approach using hierarchical architectures combining CNN and LSTM (Lee and Dernoncourt, 2016; Ji et al., 2016; Liu et al., 2017). More recently, researchers enhanced the model architecture by adding a CRF layer for classification (Kumar et al., 2018; Chen et al., 2018). Raheja and Tetreault (2019) build on earlier work that solves DAR as a sequential labeling problem using deep hierarchical networks by adding context-aware self-attention with promising results on standard benchmark datasets. More recent work has built off publicly available large pre-trained language models, significantly reducing the required training data. Qin et al. (2020) showed that combining associated labels such as sentiment with DAR can improve performance. However, all the above architectures require significant training data to achieve peak performance. Our model (SS-Con) works with limited training data as it builds on dialogue context captured by a finetuned pretrained language model. We use an iterative self",
+        "bbox": [
+            112,
+            517,
+            490,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "supervised training approach combined with a contrastive learning step to accommodate the lack of large labeled training datasets.",
+        "bbox": [
+            507,
+            84,
+            884,
+            131
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The idea of using pre-trained models to learn a few examples has been shown to be successful for natural language processing (Miller et al., 2000; Fei-Fei et al., 2006; Brown et al., 2020), and specifically, task-oriented dialogue systems (Liu et al., 2021b; Wu et al., 2020). Using few-shot learning, Mi et al. (2022) show they can improve dialogue state tracking, intent recognition, and natural language generation with limited labeled data using talk-specific instructions through prompts. Guibon et al. (2021) propose a prototypical network for sequence labeling on conversational data. While their network is trained to support few-shot DAR, they still require significant data for episodic training. The approaches mentioned above are not suitable for label-scarce situations. We show that SS-Con outperforms their model's performance on the benchmark dataset with limited labeled data.",
+        "bbox": [
+            507,
+            137,
+            885,
+            426
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Pretrained language models have been used for classifying texts in the K-12 math education context. Shen et al. (2021) applied a BERT-based model to classify knowledge components in descriptive math texts. Loginova and Benoit (2022) employed an LSTM model to predict math problem difficulty trained on a question-answer dataset. These models work on descriptive text samples, a distinct use case from classroom dialogues. Okur et al. (2022) developed a speech dialogue system using MathBERT for natural language understanding, trained on significant pre-labeled data, differing from our limited labeled data scenario.",
+        "bbox": [
+            507,
+            432,
+            885,
+            640
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Self-supervision is a viable approach for DAR with limited training data. Mi et al. (2021a) use a teacher-student model iterative approach to improve performance. They use a novel text augmentation technique that adds to each iteration's training data. We show that SSCon improves upon their results on standard datasets. Our model uses contrastive learning (Hadsell et al., 2006) to cluster the different classes. A contemporary work by Tunstall et al. (2022) uses a similar contrastive learning-based approach for few-shot text classification. Liu et al.'s (2021a) trans-encoder combines two learning paradigms, cross, and bi-encoder, in a joined iterative framework to build state-of-the-art sentence similarity models. Our self-supervised contrastive learning-based (SSCon) based multiclass classifier also combines two learning paradigms",
+        "bbox": [
+            507,
+            645,
+            885,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10985",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "like the trans-encoder by Liu et al. (2021a). An embedding-based classifier, and a twin network, trained to leverage different training goals, help each other improve. One model uses the whole dialogue context captured in a hidden layer embedding to train a classifier. At the same time, the other is a twin neural network taking the contrastive representation learning approach, clustering same dialogue acts and separating distinct dialogue acts within the embedding space.",
+        "bbox": [
+            112,
+            84,
+            492,
+            247
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            256,
+            247,
+            272
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Classroom dialogue analysis can yield significant insight into student learning. However, collecting and coding classroom dialogue datasets is very labor-intensive. To address this problem, we introduce a novel self-supervised contrastive learning approach that can automate a portion of this process and make it more efficient even when limited labeled is available. We show that our approach improves on other methods that work on limited labeled datasets. The results also show that our approach can match and exceed the performance of some models trained on the fully labeled dataset.",
+        "bbox": [
+            112,
+            282,
+            489,
+            475
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            487,
+            220,
+            502
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Selecting a representative set of examples to label becomes essential when working with limited labeled data. In this work, we use uniform sampling for our results, which might not be the best approach. We discuss these limitations in more detail in the appendix.",
+        "bbox": [
+            112,
+            512,
+            489,
+            608
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While we evaluate our model on a multi-label dataset (DSTC2) and show improvement over standard baselines, the effectiveness of our approach on such problems needs more investigation.",
+        "bbox": [
+            112,
+            609,
+            489,
+            674
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethical Considerations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            684,
+            315,
+            699
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While our algorithm is primarily a tool for improving classifier performance in label-scarce settings and uses publicly available, anonymized datasets, we acknowledge the potential ethical implications it may carry. Despite the neutral nature of our tool, it could unintentionally propagate or amplify biases favoring certain styles of communication. If used in real-time settings or without proper checks in place, it could inadvertently alter the natural dynamics of the classroom as teachers or students might modify their behavior based on how they believe the classifier categorizes their utterances. Like any other, we recognize that our tool could make",
+        "bbox": [
+            112,
+            709,
+            492,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "mistakes or be misused by over-relying on quantitative aspects over qualitative aspects of instruction. Therefore, real-world application requires continuous vigilance and open dialogue with practitioners and stakeholders to ensure its use benefits teaching and learning.",
+        "bbox": [
+            507,
+            84,
+            885,
+            181
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            193,
+            680,
+            209
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by the National Science Foundation under award DRL-2112635. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.",
+        "bbox": [
+            507,
+            219,
+            885,
+            315
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            343,
+            608,
+            357
+        ],
+        "page_idx": 8
+    },
+    {
+        "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, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.",
+            "Maximillian Chen, Alexandros Papangelis, Chenyang Tao, Andy Rosenbaum, Seokhwan Kim, Yang Liu, Zhou Yu, and Dilek Hakkani-Tur. 2022. Weakly supervised data augmentation through prompting for dialogue understanding. arXiv preprint arXiv:2210.14169.",
+            "Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, pages 785-794, New York, NY, USA. ACM.",
+            "Zheqian Chen, Rongqin Yang, Zhou Zhao, Deng Cai, and Xiaofei He. 2018. Dialogue act recognition via crf-attentive structured network. In The 41st international acm SIGIR conference on research & development in information retrieval, pages 225-234.",
+            "Li Fei-Fei, Robert Fergus, and Pietro Perona. 2006. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 28(4):594-611.",
+            "Chih-Wen Goo and Yun-Nung Chen. 2018. Abstractive dialogue summarization with sentence-gated modeling optimized by dialogue acts. In 2018 IEEE Spoken Language Technology Workshop (SLT), pages 735-742. IEEE.",
+            "Gael Guibon, Matthieu Labeau, Hélène Flamein, Luce Lefeuvre, and Chloe Clavel. 2021. Few-shot emotion recognition in conversation with sequential prototypical networks. In *The 2021 Conference on Empirical Methods in Natural Language Processing* (EMNLP 2021)."
+        ],
+        "bbox": [
+            509,
+            366,
+            885,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10986",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE.",
+            "Tianyong Hao, Xieling Chen, and Yu Song. 2020. A topic-based bibliometric analysis of two decades of research on the application of technology in classroom dialogue. Journal of educational computing research, 58(7):1311-1341.",
+            "Simon Haykin. 1994. Neural networks: a comprehensive foundation. Prentice Hall PTR.",
+            "Matthew Henderson, Blaise Thomson, and Jason D Williams. 2014. The second dialog state tracking challenge. In Proceedings of the 15th annual meeting of the special interest group on discourse and dialogue (SIGDIAL), pages 263-272.",
+            "Cindy E Hmelo-Silver. 2004. Problem-based learning: What and how do students learn? Educational psychology review, 16(3):235-266.",
+            "Yangfeng Ji, Gholamreza Haffari, and Jacob Eisenstein. 2016. A latent variable recurrent neural network for discourse relation language models. In Proceedings of NAACL-HLT, pages 332-342.",
+            "Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent convolutional neural networks for discourse compositionality. In Proceedings of the 2013 Workshop on Continuous Vector Space Models and their Compositionality.",
+            "Harshit Kumar, Arvind Agarwal, Riddhiman Dasgupta, and Sachindra Joshi. 2018. Dialogue act sequence labeling using hierarchical encoder with crf. In Proceedings of the aai conference on artificial intelligence, volume 32.",
+            "Ji Young Lee and Franck Dernoncourt. 2016. Sequential short-text classification with recurrent and convolutional neural networks. In Proceedings of NAACL-HLT, pages 515-520.",
+            "Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986-995.",
+            "Zhaojiang Lin, Bing Liu, Seungwhan Moon, Paul Crook, Zhenpeng Zhou, Zhiguang Wang, Zhou Yu, Andrea Madotto, Eunjoon Cho, and Rajen Subba. 2021. 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, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Fangyu Liu, Yunlong Jiao, Jordan Massiah, Emine Yilmaz, and Serhii Havrylov. 2021a. Trans-encoder: Unsupervised sentence-pair modelling through self- and mutual-distillations. In International Conference on Learning Representations.",
+            "Jiexi Liu, Ryuichi Takanobu, Jiaxin Wen, Dazhen Wan, Hongguang Li, Weiran Nie, Cheng Li, Wei Peng, and Minlie Huang. 2021b. Robustness testing of language understanding in task-oriented dialog. 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 2467-2480.",
+            "Yang Liu, Kun Han, Zhao Tan, and Yun Lei. 2017. Using context information for dialog act classification in dnn framework. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2170-2178.",
+            "Ekaterina Loginova and Dries Benoit. 2022. Structural information in mathematical formulas for exercise difficulty prediction: a comparison of NLP representations. In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022), pages 101-106, Seattle, Washington. Association for Computational Linguistics.",
+            "Neil Mercer. 2010. The analysis of classroom talk: Methods and methodologies. British journal of educational psychology, 80(1):1-14.",
+            "Neil Mercer, Sara Hennessy, and Paul Warwick. 2019. Dialogue, thinking together and digital technology in the classroom: Some educational implications of a continuing line of inquiry. International Journal of Educational Research, 97:187-199.",
+            "Fei Mi, Yasheng Wang, and Yitong Li. 2022. Cins: Comprehensive instruction for few-shot learning in task-oriented dialog systems. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11076-11084.",
+            "Fei Mi, Wanhao Zhou, Lingjing Kong, Fengyu Cai, Minlie Huang, and Boi Faltings. 2021a. Self-training improves pre-training for few-shot learning in task-oriented dialog systems. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1887-1898.",
+            "Fei Mi, Wanhao Zhou, Lingjing Kong, Fengyu Cai, Minlie Huang, and Boi Faltings. 2021b. Self-training improves pre-training for few-shot learning in task-oriented dialog systems. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1887-1898, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Erik G Miller, Nicholas E Matsakis, and Paul A Viola. 2000. Learning from one example through shared"
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10987",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "densities on transforms. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), volume 1, pages 464-471. IEEE.",
+            "Eda Okur, Saurav Sahay, Roddy Fuentes Alba, and Lama Nachman. 2022. End-to-end evaluation of a spoken dialogue system for learning basic mathematics. In Proceedings of the 1st Workshop on Mathematical Natural Language Processing (MathNLP), pages 51-64, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.",
+            "Libo Qin, Wanxiang Che, Yangming Li, Mingheng Ni, and Ting Liu. 2020. Dcr-net: A deep co-interactive relation network for joint dialog act recognition and sentiment classification. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8665–8672.",
+            "Libo Qin, Zhouyang Li, Wanxiang Che, Minheng Ni, and Ting Liu. 2021. Co-gat: A co-interactive graph attention network for joint dialog act recognition and sentiment classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13709-13717.",
+            "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners.",
+            "Vipul Raheja and Joel Tetreault. 2019. Dialogue act classification with context-aware self-attention. 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 3727-3733.",
+            "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. 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 3982-3992.",
+            "Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. 2021. A survey of deep active learning. ACM computing surveys (CSUR), 54(9):1-40.",
+            "Lauren B Resnick, Sarah Michaels, and Catherine O'Connor. 2010. How (well structured) talk builds the mind. Innovations in educational psychology: Perspectives on learning, teaching and human development, pages 163-194.",
+            "Jia Tracy Shen, Michiharu Yamashita, Ethan Prihar, Neil Heffernan, Xintao Wu, Sean McGrew, and Dongwon Lee. 2021. Classifying math knowledge components via task-adaptive pre-trained bert. In Artificial Intelligence in Education: 22nd International Conference, AIED 2021, Utrecht, The Netherlands, June 14-18, 2021, Proceedings, Part I 22, pages 408-419. Springer."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yu Song, Tianyong Hao, Zhinan Liu, and Zixin Lan. 2019. A systematic review of frameworks for coding towards classroom dialogue. In International Symposium on Emerging Technologies for Education, pages 226-236. Springer.",
+            "Yu Song, Shunwei Lei, Tianyong Hao, Zixin Lan, and Ying Ding. 2021. Automatic classification of semantic content of classroom dialogue. Journal of Educational Computing Research, 59(3):496-521.",
+            "Abhijit Suresh, Jennifer Jacobs, Charis Harty, Margaret Perkoff, James H. Martin, and Tamara Sumner. 2022a. The TalkMoves dataset: K-12 mathematics lesson transcripts annotated for teacher and student discursive moves. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4654-4662, Marseille, France. European Language Resources Association.",
+            "Abhijit Suresh, Jennifer Jacobs, Margaret Perkoff, James H. Martin, and Tamara Sumner. 2022b. Finetuning transformers with additional context to classify discursive moves in mathematics classrooms. In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022), pages 71-81, Seattle, Washington. Association for Computational Linguistics.",
+            "Lewis Tunstall, Nils Reimers, Unso Eun Seo Jo, Luke Bates, Daniel Korat, Moshe Wasserblat, and Oren Pereg. 2022. Efficient few-shot learning without prompts. arXiv preprint arXiv:2209.11055.",
+            "Dong Wang, Ziran Li, Haitao Zheng, and Ying Shen. 2020. Integrating user history into heterogeneous graph for dialogue act recognition. In Proceedings of the 28th International Conference on Computational Linguistics, pages 4211-4221.",
+            "Viktor Wendel and Johannes Konert. 2016. Multiplayer serious games. In Serious Games, pages 211-241. Springer.",
+            "Chien-Sheng Wu, Steven CH 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.",
+            "Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and William B Dolan. 2020. Dialogpt: Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270-278."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            810
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            512,
+            831,
+            749,
+            846
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In the paper, we describe the iterative approach of our model, where we alternate between a contrastive representation learning model and a standard classification model to improve performance.",
+        "bbox": [
+            510,
+            854,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "10988",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/991cab92591ba3b956b3cbaa88c6eefe49f284c2b61cf17391336ed08edbd29d.jpg",
+        "image_caption": [
+            "Figure 6: The model's performance changes with each iteration for multiple runs. Each run starts with different labeled data sets of 10 examples per class (TalkMoves dataset)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            85,
+            472,
+            244
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Two aspects determine the decision to stop the iteration. Figure 6 shows the model's performance in subsequent iterations for the TalkMoves dataset, starting from a label set of size ten examples per dialog act. A held-out dataset was used to calibrate the model's performance. One can see that the performance improvement is non-monotonic. In most cases, stopping after just one or two iterations is necessary to get maximum performance. The system is trying to improve based on minimal information in the labeled set, so as the size of the label set is small, there is a potential for over-fitting to the information contained in the label set. Another implementation detail is the selection of threshold confidence levels to distill soft labels as we expand the training set of the next model at the end of an iteration. Figure 7 shows the performance of our model for various prediction probability distillation thresholds. As mentioned earlier in the main paper and observed in Figure 7,  $85\\%$  is the distillation threshold with the best performance distribution.",
+        "bbox": [
+            115,
+            335,
+            489,
+            674
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Sample Selection by Active Learning",
+        "text_level": 1,
+        "bbox": [
+            114,
+            684,
+            470,
+            702
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We experimented with using active learning techniques to pick labeled examples (Ren et al., 2021). We first trained a model with just ten initial labeled examples per class. We used the model to find a new set of ten samples for each class with the least confidence and added them to the labeled set using human labels. Figure 8 compares our active learning with uniform random selection approaches as a baseline. We can see that the random approach has much more variance for every label set size. However, we also note that the best-performing model was a random set for every label size. This difference in performance might be because our ap",
+        "bbox": [
+            112,
+            709,
+            490,
+            919
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/86966fc5971adbb39b44d3867f8765f100f2a30f13fdef1a662094191d2a4ca8.jpg",
+        "image_caption": [
+            "Figure 7: The figure shows our model's performance distribution for different soft label distillation thresholds. The runs were for 70 labeled samples per class as the initial training set from the TalkMoves dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            177,
+            867,
+            338
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/8b40287e4cbfc4d9f28b4391c8db8bdd186a4275c933040c162253d6019cc17b.jpg",
+        "image_caption": [
+            "Figure 8: Comparison between active learning selection vs uniform random selection of labeled samples."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            542,
+            602,
+            845,
+            782
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "10989",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "proach for active sample selection was naive. When we add a new sample based on low prediction confidence, there is no guarantee that we will improve the model. The new sample might be predicted with low confidence either because our model has limited information to decide or because the sample itself is an outlier. If it is the former adding the sample might help, but if it is the latter, adding the sample increases the number of outliers in our training set. As our training data is limited, outliers in the input data disrupt overall performance. We plan to experiment with more sophisticated active learning techniques for sample selection in the future.",
+        "bbox": [
+            115,
+            84,
+            489,
+            307
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "10990",
+        "bbox": [
+            478,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? It is in a separate section after section 7 (Conclusion)",
+            "A2. Did you discuss any potential risks of your work? It is in the Ethical Considerations section",
+            "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            695,
+            288
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            299,
+            489,
+            316
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "2, 3.2",
+        "bbox": [
+            132,
+            321,
+            179,
+            335
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? 2,3.2",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? 2.4, 3.2",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? 3.2",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3.2",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3.2"
+        ],
+        "bbox": [
+            127,
+            346,
+            880,
+            752
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "3",
+        "bbox": [
+            132,
+            787,
+            146,
+            799
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3",
+        "bbox": [
+            129,
+            810,
+            880,
+            859
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_footnote",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            866,
+            877,
+            889
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "10991",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 3",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?"
+        ],
+        "bbox": [
+            129,
+            83,
+            880,
+            279
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank."
+        ],
+        "bbox": [
+            127,
+            340,
+            880,
+            640
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "10992",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2023/Improving Classroom Dialogue Act Recognition from Limited Labeled Data with Self-Supervised Contrastive Learning Classifiers/b6413ac7-f15b-410f-9dfa-dd65e39b5f0b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5e3a2710274ce703736ddbc142ec9808f41647f56bb116815e588db4047cdc8
+size 1197278

2023/Improving Classroom Dialogue Act Recognition from Limited Labeled Data with Self-Supervised Contrastive Learning Classifiers/full.md

@@ -0,0 +1,365 @@
+# Improving Classroom Dialogue Act Recognition from Limited Labeled Data with Self-Supervised Contrastive Learning Classifiers
+
+Vikram Kumaran Jonathan Rowe Bradford Mott
+
+North Carolina State University
+
+{vkumara, jprowe, bwmott}@ncsu.edu
+
+Snigdha Chaturvedi
+
+UNC Chapel Hill
+
+snigdha@cs.unc.edu
+
+James Lester
+
+North Carolina State University
+
+lester@ncsu.edu
+
+# Abstract
+
+Recognizing classroom dialogue acts has significant promise for yielding insight into teaching, student learning, and classroom dynamics. However, obtaining K-12 classroom dialogue data with labels is a significant challenge, and therefore, developing data-efficient methods for classroom dialogue act recognition is essential. This work addresses the challenge of classroom dialogue act recognition from limited labeled data using a contrastive learning-based self-supervised approach (SSCon). SSCon uses two independent models that iteratively improve each other's performance by increasing the accuracy of dialogue act recognition and minimizing the embedding distance between the same dialogue acts. We evaluate the approach on three complementary dialogue act recognition datasets: the TalkMoves dataset (annotated K-12 mathematics lesson transcripts), the DailyDialog dataset (multi-turn daily conversation dialogues), and the Dialogue State Tracking Challenge 2 (DSTC2) dataset (restaurant reservation dialogues). Results indicate that our self-supervised contrastive learning-based model outperforms competitive baseline models when trained with limited examples per dialogue act. Furthermore, SSCon outperforms other few-shot models that require considerably more labeled data<sup>1</sup>.
+
+# 1 Introduction
+
+Dialogue analysis offers significant potential for improving our understanding of classroom learning and teaching by modeling dialogue between students and teachers. Studies of classroom dialogue can provide deep insight into how students learn most effectively and engage with each other and with teachers (Mercer et al., 2019; Mercer, 2010; Resnick et al., 2010; Hmelo-Silver, 2004). A long-standing goal in analyzing classroom dialogue
+
+is to understand how student-student and student-teacher dialogues lead to better student learning outcomes (Wendel and Konert, 2016). This work addresses the problem of dialogue act recognition in K-12 classroom dialogues.
+
+Dialogue act recognition has garnered considerable attention and is useful for many tasks such as dialogue generation and understanding (Chen et al., 2022; Lin et al., 2021; Goo and Chen, 2018). Recent efforts in dialogue act recognition are built on large-scale pre-trained language models (Qin et al., 2021, 2020; Wang et al., 2020; Raheja and Tetreault, 2019; Chen et al., 2018). These models demonstrate high performance on standard datasets but require substantial labeled training data and, in some cases, combine other corroborative labels such as sentiment. Finding labeled public datasets of K-12 classroom dialogues is challenging for several reasons. First, there are concerns about participants' privacy and security. Second, researchers often develop individualized coding schemes specific to their design framework, and research context (Mercer, 2010; Song et al., 2019; Hao et al., 2020; Song et al., 2021). Therefore, even when labeled datasets are available, the assortment of coding schemes makes it challenging to cross-train across datasets. Third, classroom dialogue utterances are usually specific to a given subject matter, making generic labeled dialogue datasets less useful as auxiliary data. It is, therefore, essential to develop the capability to build dialogue act recognition models from limited labeled training data.
+
+Our research addresses the lack of large-scale labeled classroom dialogue datasets by using a self-supervised contrastive learning-based model (SSCon) trained using limited labeled data. SSCon uses contrastive learning to transform the dialogue utterance representation into a new embedding space where identical dialogue acts cluster together, and distinct dialogue acts are pushed further apart. The system iteratively improves per
+
+formance, as the contrastive learning step benefits self-supervision, even when presented with limited labeled data. Experiments show that SSCon outperforms competitive baselines with just tens of labeled examples per dialog act in both a K-12 mathematics classroom dataset and an everyday conversation dataset, DailyDialog. Our key contributions are the following:
+
+- We propose a novel self-supervised contrastive learning dialogue act recognizer.  
+- We test our model on multiple datasets in distinctly different domains under label-scarce settings. Our experiments show that our model outperforms strong baselines.  
+- We illustrate with an ablation study why our model outperforms the baseline.
+
+# 2 Contrastive Learning Model
+
+# 2.1 Problem Definition
+
+A dialogue  $D = (u_{1}, u_{2}, \dots, u_{N})$  in a dialogue dataset consists of a sequence of  $N$  utterances, and a set of dialogue acts  $A$ . Dialogue act recognition (DAR) is defined as a classification problem, that involves recognizing the dialogue act  $da_{i} = DAR(u_{i}|u_{i-1}, \dots, u_{i-m})$  for utterance  $u_{i}$ , given its context, a set of previous  $m$  utterances, where  $da_{i} \subseteq A$ . The task is formalized as a multiclass classification problem and sometimes a multi-label classification problem, depending on the coding scheme used.
+
+# 2.2 Approach and Assumptions
+
+While our approach is primarily evaluated on the multi-class classification problem, we also test our model on the multi-label dataset DSTC2 (Henderson et al., 2014), with promising results.
+
+Since we test a few-shot learning scenario, we start with a small set of labeled utterance examples I and the remaining training set  $U$  of unlabeled utterances. We also have a set of labeled utterances  $T$  set aside as a validation set.
+
+# 2.3 SSCon Overview
+
+In this section, we describe the architecture of our model, the self-supervised contrastive learning (SS-Con) based multi-class classifier, shown in Figure 1. Our model operates in multiple stages. We
+
+begin by finetuning a large pre-trained transformer-based language model (PLM) trained on large publicly available dialogue datasets using our domain-specific dialogue dataset. We use the finetuned PLM to generate dialogue embeddings for our model. In Stage 1, we use the finetuned PLM to encode the utterance and its dialogue history. We also use sentence-BERT (Reimers and Gurevych, 2019) to create a latent representation of each utterance. A classifier built on the latent representation makes an initial soft prediction of the dialogue act. We distill the initial high-confidence predictions as soft labels for Stage 2. In Stage 2, the soft labels from Stage 1 train an encoder using contrastive learning. It translates the latent representations from Stage 1 into a vector in a different encoding space where identical dialogue acts cluster together while distinct dialogue acts are separated. In Stage 3, we pass the utterance representations through the encoder trained in Stage 2 to get the embeddings, which will be the input to classify the dialogue acts. The high probability soft labels from the Stage 3 classifier are sent back to Stage 1 as input to self-supervise the model in the next iteration.
+
+# 2.4 Pretraining
+
+For our pretraining stage, we finetune DialogGPT (Zhang et al., 2020), a dialogue PLM built on GPT2 (Radford et al.) and trained on 147M Reddit and other online conversations, licensed under MIT License. Given an utterance  $u_{i}$  and its context of a set of previous m utterances, the dialogue PLM is finetuned to maximize the conditional probability for the subsequent utterance,  $P(u_{i + 1}|u_i,u_{i - 1},\dots ,u_{i - m})$ . For pretraining, we create a dataset consisting of target utterances and its context consisting of preceding  $m$  utterances. SSCon uses the finetuned PLM's hidden state as the representation of the utterance  $u_{i}$  given its dialogue context  $(u_{i - 1}\dots u_{i - m})$ .
+
+# 2.5 Stage 1: Context Classifier
+
+The Stage 1 classifier concatenates the pretrained dialogue PLM's last layer embedding  $H_{i}\in R^{768}$  with the sentence-BERT embedding  $S_{i}\in R^{384}$  of the utterance  $u_{i}$  and predicts the dialog act,  $y_{i}^{stage1}$  for the utterance,
+
+$$
+y _ {i} ^ {\text {s t a g e 1}} = C L _ {\text {s t a g e 1}} \left(X _ {i}, \Theta\right) \tag {1}
+$$
+
+$$
+X _ {i} = \left[ H _ {i}; S _ {i} \right] \tag {2}
+$$
+
+$$
+H _ {i} = \text {D i a l o G P T} \left(u _ {i}, u _ {i - 1}, \dots , u _ {i - m}\right) \tag {3}
+$$
+
+![](images/2063650e1fe7cc04479a1be6f1c79c2dc554171972558e39a0079d8d3ba1ab7b.jpg)  
+Figure 1: Self-supervised contrastive learning (SSCon) multiclass classifier
+
+$$
+S _ {i} = \text {S e n t e n c e B E R T} (u _ {i}) \tag {4}
+$$
+
+Where  $y_{i}^{stage1}$  is the predicted dialogue act class for the last utterance  $u_{i}$  and  $CL_{stage1}$  is the Stage 1 multi-class classifier with trainable weights  $\Theta$ . The concatenated vector  $X_{i}$  represents the utterance  $u_{i}$  as an independent sentence and in the context of its dialogue  $(u_{i-1}, \dots, u_{i-m})$ . The output of Stage 1 is a dialogue act label (and the corresponding prediction probability) for each utterance in the dataset. For the choice of classifier, we explored both MLP classifiers (Haykin, 1994) and XGBoost (Chen and Guestrin, 2016); however, any reasonable classifier can be utilized. We chose the XGBoost classifier as it was faster to train and run without any impact on the performance of our model. However, for one of the datasets (DSTC2), where an utterance could have multiple labels, we used an MLP-based multi-label classifier.
+
+In the first iteration of the self-supervision process, the training data used for the classifier is limited to the number of available labeled samples, usually between ten to a hundred examples per class. In subsequent iterations, the number of samples increases based on the soft labels from earlier iterations.
+
+# 2.6 Stage 2:Contrastive Encoder
+
+The prediction from the Stage 1 classifier,  $CL_{stage1}$ , is used as (soft) ground truth for Stage 2. Specifically, we use the dialogue act labels with the highest confidence in terms of prediction probability as new soft labels along with the initial labeled examples used in Stage 1. Including high confidence, soft-labeled samples increases the effective size of the training set with each iteration. Using this data, we adopt a contrastive training approach.
+
+This training process creates a network that can encode the finetuned PLM latent representation into a space where utterances with the identical dialogue act class are close together while utterances with distinct dialogue act labels are farther apart. The labeled samples are paired to generate positive and negative triplets  $P = (X_{i}, X_{j}, P_{ij})$  where  $X_{i}$  (Equation 2) is the concatenated encoding defined in the previous section for a given utterance  $u_{i}$ , and  $P_{ij}$  is 1 if both the utterances map to the same dialogue act or -1 if they map to different dialogue acts. In the case of multi-label utterances,  $P_{ij}$  is the cosine distance between the one-hot encoding of the dialogue act vectors of the two utterances. The encoder is a five-layer MLP that transforms the  $X_{i} \in R^{1152}$  concatenated latent representation into an  $E_{i} \in R^{384}$  encoding. We use a twin encoder network to train on the positive/negative triplets  $\mathbf{P}$  using cosine similarity between output embeddings as the similarity score. Given a triplet  $(X_{i}, X_{j}, P_{ij})$  the network trains,
+
+$$
+E _ {i} = B _ {\text {e n c o d e r}} \left(X _ {i}\right) \tag {5}
+$$
+
+$$
+E _ {j} = B _ {\text {e n c o d e r}} \left(X _ {j}\right) \tag {6}
+$$
+
+$$
+l _ {i j} = \operatorname {c o s i n e \_ s i m i l a r i t y} \left(E _ {i}, E _ {j}\right) \tag {7}
+$$
+
+$$
+l o s s _ {i j} = M S E \left(l _ {i j}, P _ {i j}\right) \tag {8}
+$$
+
+The trained encoder transforms each utterance, into a new encoding,  $E_{i} = B_{encoder}(X_{i})$ , where  $E_{i} \in R^{384}$  is the generated embedding of the MLP encoder ( $B_{encoder}$ ).
+
+# 2.7 Stage 3: Embedding Classifier
+
+In this final stage of our model, we use the encoder from the Contrastive Encoder network that was trained in the previous stage to convert each  $X_{i}$
+
+into the embedding  $E_{i}$  used as input to the Stage 3 classifier.
+
+$$
+y _ {i} ^ {\text {s t a g e 3}} = C L _ {\text {s t a g e 3}} \left(E _ {i}, \Phi\right) \tag {9}
+$$
+
+where  $y_{i}^{state3}$  is the dialogue act labels for the utterance  $u_{i}$  and  $CL_{state3}$  is a multi-class classifier with trainable weights  $\Phi$ . Like Stage 1, we use an XGBoost classifier. We use the dialogue act labels with the highest confidence in terms of prediction probability from Stage 1 as the soft label for training the classifier.
+
+The training starts with a small set of labeled utterance examples  $L$ , and the remaining training set  $U$  of unlabeled utterances. During every iteration of the self-supervision process, the model labels the unlabeled  $U$  utterances in our training set. We filter out the utterances labeled with low confidence (low prediction probability of  $CL_{\text{stage3}}$ ). The distilled silver-labeled instances are moved from  $U$  to  $L$ , along with the initial labeled examples. The updated  $L$  is the training set for Stage 1, starting the next iteration of the self-supervision process.
+
+# 3 Experimental Setup
+
+# 3.1 Implementation Details
+
+The pretraining stage involves fine-tuning the DialogPT model with utterance data from a given dialogue dataset. The training set uses nine previous utterances as context. We finetune a HuggingFace pre-trained "dialogPT-small" base model on a single GPU for four epochs. The version of DialogPT we used is a 12-layer transformer. We use the hidden state vector  $H_{i}$  of the end-of-sentence tag in the  $12^{\mathrm{th}}$  layer of the transformer as the embedding representing the last utterance in the sequence.
+
+Stage 1 of the model uses a dialog act classifier. We implemented an MLP-based classifier and an XGBoost classifier. There was no difference in performance between the two classifiers, so we picked XGBoost as it is a standard baseline. We use the hidden state,  $H_{i}$ , of the last layer as the embedding to represent the dialogue context of the utterance. For the sentence-BERT embedding  $S_{i}$  of the utterance we used a pre-trained network ("all-MiniLM-L6-v2") that is trained on roughly 1B sentence pairs (Reimers and Gurevych, 2019). The input to the model is the concatenated vector of  $H_{i}$  and  $S_{i}$ . After running the model for various distillation thresholds, a threshold of 0.85 was used for high-probability soft-label distillation based on
+
+![](images/d2704382bee809b0eff47e5086ed9b80fe3c02d2407170d56677d670cf4369b9.jpg)  
+Figure 2: Improvement of performance over a baseline XGBoost classifier by our SSCon classifier.
+
+a simple grid search on the threshold parameter as discussed in the appendix.
+
+Stage 2 of the model is a contrastive encoder network. Each encoder network in the twin network is a 5-layer MLP with a  $20\%$  dropout between the layers and an output embedding vector of dimension 384. About 1.2 million similarity pairs are used to train the twin network for 4-6 epochs on a single GPU. Stage 3 is an XGBoost classifier head on top of the encoder trained in Stage 2. The iterative self-supervision process continues until we show no improvement in the validation data. We run 5-10 iterations. In our experiments, we use a validation set to determine when to stop the iterations. We report the results on the test set.
+
+# 3.2 Datasets
+
+We use three datasets in our experiments: the TalkMoves, the DailyDialog, and the Dialogue State Tracking Challenge 2 datasets. The TalkMoves dataset (Suresh et al., 2022a) consists of 567 human-annotated class video transcripts of K-12 mathematics lessons between teachers and students. A human-transcribed dataset consists of 174,186 teacher utterances and 59,874 student utterances. The dialogue act labels in the dataset have an inter-rater agreement score above  $90\%$  for all labels. The dialogue acts for student utterances include 'relating to another student', 'asking for more information', 'making a claim', and 'providing evidence'. The dialogue acts for teacher utterances include: 'keeping everyone together', 'getting students to relate', 'restating', 'revoicing', 'pressing for accuracy', and 'pressing for a reason'. For the TalkMoves dataset, we train our model on student
+
+utterances.
+
+DailyDialog is a multi-turn dialogue dataset (Li et al., 2017) consisting of everyday conversations. The dataset consists of 13,118 conversations between multiple people. Four categories of dialogue acts are coded in the dataset. They are 'inform'  $(45\%)$ , 'questions'  $(29\%)$ , 'directives'  $(17\%)$ , and 'commissive'  $(9\%)$ .
+
+To evaluate our approach on multi-label data set, we train on the DSTC2 dataset (Henderson et al., 2014). This dataset contains dialogues between crowdsourced workers and automated dialogue systems in the restaurant reservations domain, with 1000 train and test dialogues and about 21 dialogue acts.
+
+# 3.3 Baselines
+
+We run our experiments on different datasets described above. The TalkMoves dataset does not yet have many published baselines for dialogue act recognition, so we compare our approach against multiple baseline models. One is a baseline XGBoost classifier using the same utterance representation input as our model. This classifier is the same as our Stage 1 classifier. The second baseline is a self-supervised XGBoost classifier similar to SSCon without the contrastive learning in Stage 2. For the third baseline, we use an embedding prototype distance-based classifier. The average sentence embedding vector for labeled examples from each class represents a prototype for each label. We then measure the cosine distance of every utterance's sentence embedding in our test set against each class-prototype embedding. An utterance is assigned the class label of the closest prototype in the embedding space. To validate SS-Con against current state-of-the-art dialogue act recognition models, we compare our results on the DailyDialog dataset against a few-shot learning model and a model that uses all available labeled training data. The state-of-the-art Co-GAT model (Qin et al., 2021) uses all training data available, including related sentiment labels on the utterance. Trained on very limited data, we do not expect our model to beat Co-GAT but use its performance as an upper bound. ProtoSeq (Guibon et al., 2021) is a sequential prototypical network trained to work in a few-shot fashion. However, they do use all training data to train their network. We use this model for comparison purposes as our approach is a type of few-shot learner using a small number of samples.
+
+![](images/93b74bb9d96f7dd951916098b2bc3241f28203c0b49c1d087915696d3752d5af.jpg)  
+Figure 3: Heatmaps show the median cosine distance between embedding examples (TalkMoves dataset) for each dialog act type (8-relating to another student; 9-Asking for more info; 10-making a claim; 11-providing evidence/reasoning). The left heatmap corresponds to the input embeddings. The right heatmap corresponds to the Contrastive Encoder embeddings
+
+![](images/0a22d0461e1f1917e5dc45e866e5a90cc9ec0030c408ddf5aac49238875b468c.jpg)
+
+We also compare SSCon to standard XGBoost as we do for the TalkMoves dataset. We also compare SSCon dialogue act recognition results on DSTC2, a multi-label problem, against the self-supervised student-teacher approach by (Mi et al., 2021a).
+
+# 3.4 Evaluation Metrics
+
+The various baselines are compared with SSCon using four metrics. The Matthews Correlation Coefficient (MCC) for multi-class classification has a range of -1 to 1 and handles imbalanced datasets well. In some cases, the baselines only report the F1 (macro) score, the arithmetic mean of individual class F1 scores giving equal weight to all classes. We also report macro-averaged precision and recall.
+
+# 4 Results
+
+The first dataset we consider is the TalkMoves dataset. To evaluate the models in a few-shot learning scenario, we experiment using 10-70 labeled instances per dialogue act type (less than  $1\%$  of the overall data). Table 1 shows our results. SSCon shows an improvement of about  $7 - 10\%$  over the Utterance Embedding XGBoost baseline and the Prototype Distance Classifier when trained on 70 labeled examples. When SSCon is compared with the self-supervised XGBoost classifier without contrastive learning, it shows an improvement of  $1\%$  for the 70 labeled examples case but an improvement of more than  $60\%$  for the 10 labeled example case. This difference in performance suggests that contrastive learning is more beneficial when working with fewer labeled examples. Besides the three baselines, we also report the performance of a pub
+
+<table><tr><td>Dataset</td><td>Model Type</td><td>Labeled Examples per Class</td><td>MCC</td><td>F1 (macro)</td><td>Precision (macro)</td><td>Recall (macro)</td></tr><tr><td rowspan="9">TalkMoves Student</td><td rowspan="3">SSCon Classifier</td><td>10</td><td>0.295±0.042</td><td>0.412±0.044</td><td>0.454±0.034</td><td>0.512±0.016</td></tr><tr><td>30</td><td>0.383±0.035</td><td>0.495±0.033</td><td>0.492±0.032</td><td>0.588±0.012</td></tr><tr><td>70</td><td>0.457±0.015</td><td>0.566±0.015</td><td>0.534±0.014</td><td>0.652±0.011</td></tr><tr><td>Utterance Embedding XGBoost</td><td>70</td><td>0.421±0.025</td><td>0.507±0.023</td><td>0.502±0.019</td><td>0.615±0.015</td></tr><tr><td rowspan="2">Self Supervised XGBoost with- out Contrastive Learning</td><td>10</td><td>0.183±0.047</td><td>0.327±0.037</td><td>0.361±0.024</td><td>0.397±0.031</td></tr><tr><td>70</td><td>0.451±0.021</td><td>0.514±0.021</td><td>0.535±0.015</td><td>0.609±0.016</td></tr><tr><td>Prototype Distance Classifier</td><td>70</td><td>0.369±0.015</td><td>0.463±0.022</td><td>0.457±0.009</td><td>0.558±0.011</td></tr><tr><td rowspan="2">Transformer for Seq Classifier (Suresh et al., 2022b)</td><td>All training</td><td>0.6716</td><td>0.7312</td><td>-</td><td>-</td></tr><tr><td>70% training</td><td>0.588±0.005</td><td>0.653±0.005</td><td>-</td><td>-</td></tr><tr><td rowspan="6">Daily Dialog</td><td rowspan="3">SSCon Classifier</td><td>10</td><td>0.409±0.065</td><td>0.471±0.037</td><td>0.550±0.064</td><td>0.510±0.044</td></tr><tr><td>30</td><td>0.488±0.046</td><td>0.564±0.043</td><td>0.608±0.013</td><td>0.619±0.039</td></tr><tr><td>70</td><td>0.601±0.022</td><td>0.660±0.018</td><td>0.663±0.009</td><td>0.687±0.015</td></tr><tr><td>Utterance Embedding XGBoost</td><td>70</td><td>0.567±0.013</td><td>0.638±0.008</td><td>0.639±0.008</td><td>0.671±0.009</td></tr><tr><td>ProtoSeq (Guibon et al., 2021)</td><td>Few-shot</td><td>0.392±0.023</td><td>0.352±0.030</td><td>-</td><td>-</td></tr><tr><td>Co-GAT (Qin et al., 2021)</td><td>All training</td><td>-</td><td>0.794</td><td>0.81</td><td>0.781</td></tr><tr><td rowspan="4">DSTC2</td><td rowspan="2">SSCon Classifier</td><td>1%</td><td></td><td>0.293±0.025</td><td>0.322±0.041</td><td>0.293±0.007</td></tr><tr><td>10%</td><td></td><td>0.447±0.008</td><td>0.456±0.024</td><td>0.449±0.011</td></tr><tr><td rowspan="2">ToD-BERT-ST (Mi et al., 2021b)</td><td>1%</td><td></td><td>0.285±0.040</td><td>-</td><td>-</td></tr><tr><td>10%</td><td></td><td>0.405±0.090</td><td>-</td><td>-</td></tr></table>
+
+Table 1: Comparison of results for the Daily Dialog, TalkMoves and DSTC2 dataset against baselines.
+
+lished model that uses  $100\%$  (21K labeled samples) and  $70\%$  of the labeled training data. SSCon works in a label-scarce scenario, so our results are lower than the model using the complete labeled data set. Figure 2 shows the impact of labeled example counts on the overall performance of the SSCon classifier and an XGBoost classifier. While our iterative approach increases the performance on average by almost  $50\%$  over the baseline for a small training size (10 labels per class), it is about  $9\%$  for a larger labeled set size (70 labels per class). We notice that the overall performance improvement using SSCon is more with small labeled sets, and the improvement tapers as the labeled examples count increases.
+
+The second dataset we consider is the DailyDialog dataset with four possible dialogue acts. Like the TalkMoves dataset, we consider 10 to 70 labeled examples per class. Table 1 shows that our performance, with just 70 instances per class (0.6% of 21521 labeled examples), is within 16% of the state-of-the-art dialogue act recognition model (Qin et al., 2021) that uses 100% of all the training data labels plus auxiliary sentiment labels. We also show that we outperform the other few-shot learning model, ProtoSeq (Guibon et al., 2021), by a significant margin even though they train using the
+
+entire labeled dataset. We investigate the change in performance of SSCon with increasing size of the labeled training set (Figure 2). As before, the performance improvement of SSCon over a standard XGBoost classifier is more significant for smaller label sets and less so as the number of samples increases. With ten samples per class, SSCon performance improvement over the baseline classifier is almost  $41\%$ . With seventy labels per class, the improvement is only about  $9\%$ .
+
+We also compare our results on the DSTC2 dataset on restaurant inquiries. This dataset differs from the other two as each utterance can have multiple dialogue act labels. SSCon performs better on the DSTC2 dataset than the baseline self-supervising student-teacher model ToD-BERT-ST (Mi et al., 2021a) by 3 to  $10\%$ , depending on the size of the labeled training dataset. The ToD-BERT-ST model uses a data augmentation approach, while we use a contrastive learning approach. Both these approaches are complementary, and future work should explore using them together.
+
+![](images/d1e52837370d8d2f4f79f55e2f8a8e7ca8042131a27300a135265a81419b4343.jpg)  
+Figure 4: Confusion matrix with counts normalized by actual label counts (TalkMoves Dataset) for the SSCon classifier (70-samples per class training set). Dialog acts: 8-relating to another student; 9-Asking for more info; 10-making a claim; 11-providing evidence/reasoning.
+
+# 5 SSCon Component Analysis
+
+# 5.1 Clustering in Embedding Space
+
+The fundamental intuition behind our approach is that contrastive learning brings similar utterances close to each other and dissimilar utterances further apart in the embedding space. Figure 3 shows the spatial clustering of labels in the embedding space of the trained Contrastive Encoder for one iteration with the TalkMoves dataset. Each square in the heatmap corresponds to the median cosine similarity distance between individual dialogue act class examples. To compute this distance, we take utterances from each dialogue act class and calculate the median cosine distance between their embeddings. Each cell in the heatmap shows a scaled distance, with the blue color corresponding to closer embeddings and red corresponding to the embeddings being farther away. The heatmap on the left corresponds to the distance matrix between embeddings provided as input to Stage 2. The heatmap to the right corresponds to the distance matrix for embeddings that the Contrastive Encoder has transformed. Please note that the figure uses scaled distance values, and we show four student dialog act classes.
+
+When looking at the heatmap on the left for the input feature space, instances belonging to the dialogue act "relating to another student" (first row, label 8) are closer to the instances belonging to the dialogue act, "making a claim" (third row, label 10) than among themselves. After being transformed by the Contrastive Encoder (heatmap on the right), the smallest distances for each dialogue act fall on
+
+<table><tr><td>Classifier</td><td>LastUtterance(u_i)</td></tr><tr><td>&quot;relating to another&quot; predicted as &quot;making a claim&quot;</td><td>&quot;Reciprocal&quot; 
+&quot;I don&#x27;t agree, I measured them&#x27; 
+&quot;by five&quot;</td></tr><tr><td>&quot;relating to another&quot; predicted as &quot;relating to another&quot;</td><td>&quot;He didn&#x27;t show his work&quot; 
+&quot;You don&#x27;t know that&#x27; 
+&quot;The first one&quot;</td></tr><tr><td>&quot;making a claim&quot; predicted as &quot;making a claim&quot;</td><td>&quot;Two fifths&quot; 
+&quot;Y intercept&#x27; 
+&quot;power of three&quot;</td></tr></table>
+
+Table 2: Examples of true positive and false negative for a couple of the TalkMoves dialog acts
+
+the diagonal (blue), as one expects with clustered labels. The diagonal corresponds to utterances with the same label. The non-diagonal terms, which are utterances with different labels, are pulled farther apart (red).
+
+# 5.2 Classifier Performance by Dialogue Act
+
+Figure 4 shows a confusion matrix with counts normalized by the ground truth label counts. In the TalkMoves dataset, the classifier struggles the most with the dialogue act corresponding to "relating to another student" (label 8). It is not able to clearly distinguish between "relating to another student" and "making a claim" (label 10). Label-scarce training suggests we only cover a limited variety of examples for each dialogue act class. In Table 2, we show some examples of correct and incorrect predictions. The first row shows utterances of the type "relating to another" mislabeled as "making a claim". The other two rows are true positive examples. We can see that the example utterances are hard to distinguish between "making a claim" or "relating to another" even for a human. The difference is in the context of the dialogue, and with limited training data, such distinctions are hard to make and might lead to overfitting.
+
+# 5.3 Labeled Sample Selection
+
+The initial set of labeled samples, in essence, drives the final performance of SSCon. Figure 5 shows the performance distribution of the baseline, XGBoost, for multiple runs for different sample sizes. Two trends are evident as the number of samples increases for every dataset. The first trend is that the performance improves as more samples are available for training. The improvement in performance with sample size is an expected trend as more examples indicate more information for the classifier to model. The second trend is that the variance
+
+![](images/7763f39eb297626566414606b3937a837df4ce361efbdf81b77836fb3bd1b68f.jpg)  
+Figure 5: Drift in performance as we increase the number of labeled samples (TalkMoves dataset).
+
+in performance decreases as the number of samples increases. One can assume a certain level of variability in the labeled training sets data's pattern of utterances for each dialogue act class. The dataset can only capture some of the variances in the data when the number of labeled examples is small. Hence, the classifier's performance can vary considerably depending on included examples for smaller training sets.
+
+# 6 Related Work
+
+Kalchbrenner and Blunsom (2013) proposed a hierarchical network architecture that combines CNN and RNN models to capture discourse structure. Subsequent work on dialogue act recognition (DAR) followed a similar approach using hierarchical architectures combining CNN and LSTM (Lee and Dernoncourt, 2016; Ji et al., 2016; Liu et al., 2017). More recently, researchers enhanced the model architecture by adding a CRF layer for classification (Kumar et al., 2018; Chen et al., 2018). Raheja and Tetreault (2019) build on earlier work that solves DAR as a sequential labeling problem using deep hierarchical networks by adding context-aware self-attention with promising results on standard benchmark datasets. More recent work has built off publicly available large pre-trained language models, significantly reducing the required training data. Qin et al. (2020) showed that combining associated labels such as sentiment with DAR can improve performance. However, all the above architectures require significant training data to achieve peak performance. Our model (SS-Con) works with limited training data as it builds on dialogue context captured by a finetuned pretrained language model. We use an iterative self
+
+supervised training approach combined with a contrastive learning step to accommodate the lack of large labeled training datasets.
+
+The idea of using pre-trained models to learn a few examples has been shown to be successful for natural language processing (Miller et al., 2000; Fei-Fei et al., 2006; Brown et al., 2020), and specifically, task-oriented dialogue systems (Liu et al., 2021b; Wu et al., 2020). Using few-shot learning, Mi et al. (2022) show they can improve dialogue state tracking, intent recognition, and natural language generation with limited labeled data using talk-specific instructions through prompts. Guibon et al. (2021) propose a prototypical network for sequence labeling on conversational data. While their network is trained to support few-shot DAR, they still require significant data for episodic training. The approaches mentioned above are not suitable for label-scarce situations. We show that SS-Con outperforms their model's performance on the benchmark dataset with limited labeled data.
+
+Pretrained language models have been used for classifying texts in the K-12 math education context. Shen et al. (2021) applied a BERT-based model to classify knowledge components in descriptive math texts. Loginova and Benoit (2022) employed an LSTM model to predict math problem difficulty trained on a question-answer dataset. These models work on descriptive text samples, a distinct use case from classroom dialogues. Okur et al. (2022) developed a speech dialogue system using MathBERT for natural language understanding, trained on significant pre-labeled data, differing from our limited labeled data scenario.
+
+Self-supervision is a viable approach for DAR with limited training data. Mi et al. (2021a) use a teacher-student model iterative approach to improve performance. They use a novel text augmentation technique that adds to each iteration's training data. We show that SSCon improves upon their results on standard datasets. Our model uses contrastive learning (Hadsell et al., 2006) to cluster the different classes. A contemporary work by Tunstall et al. (2022) uses a similar contrastive learning-based approach for few-shot text classification. Liu et al.'s (2021a) trans-encoder combines two learning paradigms, cross, and bi-encoder, in a joined iterative framework to build state-of-the-art sentence similarity models. Our self-supervised contrastive learning-based (SSCon) based multiclass classifier also combines two learning paradigms
+
+like the trans-encoder by Liu et al. (2021a). An embedding-based classifier, and a twin network, trained to leverage different training goals, help each other improve. One model uses the whole dialogue context captured in a hidden layer embedding to train a classifier. At the same time, the other is a twin neural network taking the contrastive representation learning approach, clustering same dialogue acts and separating distinct dialogue acts within the embedding space.
+
+# 7 Conclusion
+
+Classroom dialogue analysis can yield significant insight into student learning. However, collecting and coding classroom dialogue datasets is very labor-intensive. To address this problem, we introduce a novel self-supervised contrastive learning approach that can automate a portion of this process and make it more efficient even when limited labeled is available. We show that our approach improves on other methods that work on limited labeled datasets. The results also show that our approach can match and exceed the performance of some models trained on the fully labeled dataset.
+
+# Limitations
+
+Selecting a representative set of examples to label becomes essential when working with limited labeled data. In this work, we use uniform sampling for our results, which might not be the best approach. We discuss these limitations in more detail in the appendix.
+
+While we evaluate our model on a multi-label dataset (DSTC2) and show improvement over standard baselines, the effectiveness of our approach on such problems needs more investigation.
+
+# Ethical Considerations
+
+While our algorithm is primarily a tool for improving classifier performance in label-scarce settings and uses publicly available, anonymized datasets, we acknowledge the potential ethical implications it may carry. Despite the neutral nature of our tool, it could unintentionally propagate or amplify biases favoring certain styles of communication. If used in real-time settings or without proper checks in place, it could inadvertently alter the natural dynamics of the classroom as teachers or students might modify their behavior based on how they believe the classifier categorizes their utterances. Like any other, we recognize that our tool could make
+
+mistakes or be misused by over-relying on quantitative aspects over qualitative aspects of instruction. Therefore, real-world application requires continuous vigilance and open dialogue with practitioners and stakeholders to ensure its use benefits teaching and learning.
+
+# Acknowledgements
+
+This work is supported by the National Science Foundation under award DRL-2112635. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.
+
+# References
+
+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.  
+Maximillian Chen, Alexandros Papangelis, Chenyang Tao, Andy Rosenbaum, Seokhwan Kim, Yang Liu, Zhou Yu, and Dilek Hakkani-Tur. 2022. Weakly supervised data augmentation through prompting for dialogue understanding. arXiv preprint arXiv:2210.14169.  
+Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, pages 785-794, New York, NY, USA. ACM.  
+Zheqian Chen, Rongqin Yang, Zhou Zhao, Deng Cai, and Xiaofei He. 2018. Dialogue act recognition via crf-attentive structured network. In The 41st international acm SIGIR conference on research & development in information retrieval, pages 225-234.  
+Li Fei-Fei, Robert Fergus, and Pietro Perona. 2006. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 28(4):594-611.  
+Chih-Wen Goo and Yun-Nung Chen. 2018. Abstractive dialogue summarization with sentence-gated modeling optimized by dialogue acts. In 2018 IEEE Spoken Language Technology Workshop (SLT), pages 735-742. IEEE.  
+Gael Guibon, Matthieu Labeau, Hélène Flamein, Luce Lefeuvre, and Chloe Clavel. 2021. Few-shot emotion recognition in conversation with sequential prototypical networks. In *The 2021 Conference on Empirical Methods in Natural Language Processing* (EMNLP 2021).
+
+Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE.  
+Tianyong Hao, Xieling Chen, and Yu Song. 2020. A topic-based bibliometric analysis of two decades of research on the application of technology in classroom dialogue. Journal of educational computing research, 58(7):1311-1341.  
+Simon Haykin. 1994. Neural networks: a comprehensive foundation. Prentice Hall PTR.  
+Matthew Henderson, Blaise Thomson, and Jason D Williams. 2014. The second dialog state tracking challenge. In Proceedings of the 15th annual meeting of the special interest group on discourse and dialogue (SIGDIAL), pages 263-272.  
+Cindy E Hmelo-Silver. 2004. Problem-based learning: What and how do students learn? Educational psychology review, 16(3):235-266.  
+Yangfeng Ji, Gholamreza Haffari, and Jacob Eisenstein. 2016. A latent variable recurrent neural network for discourse relation language models. In Proceedings of NAACL-HLT, pages 332-342.  
+Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent convolutional neural networks for discourse compositionality. In Proceedings of the 2013 Workshop on Continuous Vector Space Models and their Compositionality.  
+Harshit Kumar, Arvind Agarwal, Riddhiman Dasgupta, and Sachindra Joshi. 2018. Dialogue act sequence labeling using hierarchical encoder with crf. In Proceedings of the aai conference on artificial intelligence, volume 32.  
+Ji Young Lee and Franck Dernoncourt. 2016. Sequential short-text classification with recurrent and convolutional neural networks. In Proceedings of NAACL-HLT, pages 515-520.  
+Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986-995.  
+Zhaojiang Lin, Bing Liu, Seungwhan Moon, Paul Crook, Zhenpeng Zhou, Zhiguang Wang, Zhou Yu, Andrea Madotto, Eunjoon Cho, and Rajen Subba. 2021. 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, Online. Association for Computational Linguistics.
+
+Fangyu Liu, Yunlong Jiao, Jordan Massiah, Emine Yilmaz, and Serhii Havrylov. 2021a. Trans-encoder: Unsupervised sentence-pair modelling through self- and mutual-distillations. In International Conference on Learning Representations.  
+Jiexi Liu, Ryuichi Takanobu, Jiaxin Wen, Dazhen Wan, Hongguang Li, Weiran Nie, Cheng Li, Wei Peng, and Minlie Huang. 2021b. Robustness testing of language understanding in task-oriented dialog. 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 2467-2480.  
+Yang Liu, Kun Han, Zhao Tan, and Yun Lei. 2017. Using context information for dialog act classification in dnn framework. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2170-2178.  
+Ekaterina Loginova and Dries Benoit. 2022. Structural information in mathematical formulas for exercise difficulty prediction: a comparison of NLP representations. In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022), pages 101-106, Seattle, Washington. Association for Computational Linguistics.  
+Neil Mercer. 2010. The analysis of classroom talk: Methods and methodologies. British journal of educational psychology, 80(1):1-14.  
+Neil Mercer, Sara Hennessy, and Paul Warwick. 2019. Dialogue, thinking together and digital technology in the classroom: Some educational implications of a continuing line of inquiry. International Journal of Educational Research, 97:187-199.  
+Fei Mi, Yasheng Wang, and Yitong Li. 2022. Cins: Comprehensive instruction for few-shot learning in task-oriented dialog systems. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11076-11084.  
+Fei Mi, Wanhao Zhou, Lingjing Kong, Fengyu Cai, Minlie Huang, and Boi Faltings. 2021a. Self-training improves pre-training for few-shot learning in task-oriented dialog systems. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1887-1898.  
+Fei Mi, Wanhao Zhou, Lingjing Kong, Fengyu Cai, Minlie Huang, and Boi Faltings. 2021b. Self-training improves pre-training for few-shot learning in task-oriented dialog systems. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1887-1898, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Erik G Miller, Nicholas E Matsakis, and Paul A Viola. 2000. Learning from one example through shared
+
+densities on transforms. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), volume 1, pages 464-471. IEEE.  
+Eda Okur, Saurav Sahay, Roddy Fuentes Alba, and Lama Nachman. 2022. End-to-end evaluation of a spoken dialogue system for learning basic mathematics. In Proceedings of the 1st Workshop on Mathematical Natural Language Processing (MathNLP), pages 51-64, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.  
+Libo Qin, Wanxiang Che, Yangming Li, Mingheng Ni, and Ting Liu. 2020. Dcr-net: A deep co-interactive relation network for joint dialog act recognition and sentiment classification. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8665–8672.  
+Libo Qin, Zhouyang Li, Wanxiang Che, Minheng Ni, and Ting Liu. 2021. Co-gat: A co-interactive graph attention network for joint dialog act recognition and sentiment classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13709-13717.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners.  
+Vipul Raheja and Joel Tetreault. 2019. Dialogue act classification with context-aware self-attention. 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 3727-3733.  
+Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. 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 3982-3992.  
+Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. 2021. A survey of deep active learning. ACM computing surveys (CSUR), 54(9):1-40.  
+Lauren B Resnick, Sarah Michaels, and Catherine O'Connor. 2010. How (well structured) talk builds the mind. Innovations in educational psychology: Perspectives on learning, teaching and human development, pages 163-194.  
+Jia Tracy Shen, Michiharu Yamashita, Ethan Prihar, Neil Heffernan, Xintao Wu, Sean McGrew, and Dongwon Lee. 2021. Classifying math knowledge components via task-adaptive pre-trained bert. In Artificial Intelligence in Education: 22nd International Conference, AIED 2021, Utrecht, The Netherlands, June 14-18, 2021, Proceedings, Part I 22, pages 408-419. Springer.
+
+Yu Song, Tianyong Hao, Zhinan Liu, and Zixin Lan. 2019. A systematic review of frameworks for coding towards classroom dialogue. In International Symposium on Emerging Technologies for Education, pages 226-236. Springer.  
+Yu Song, Shunwei Lei, Tianyong Hao, Zixin Lan, and Ying Ding. 2021. Automatic classification of semantic content of classroom dialogue. Journal of Educational Computing Research, 59(3):496-521.  
+Abhijit Suresh, Jennifer Jacobs, Charis Harty, Margaret Perkoff, James H. Martin, and Tamara Sumner. 2022a. The TalkMoves dataset: K-12 mathematics lesson transcripts annotated for teacher and student discursive moves. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4654-4662, Marseille, France. European Language Resources Association.  
+Abhijit Suresh, Jennifer Jacobs, Margaret Perkoff, James H. Martin, and Tamara Sumner. 2022b. Finetuning transformers with additional context to classify discursive moves in mathematics classrooms. In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022), pages 71-81, Seattle, Washington. Association for Computational Linguistics.  
+Lewis Tunstall, Nils Reimers, Unso Eun Seo Jo, Luke Bates, Daniel Korat, Moshe Wasserblat, and Oren Pereg. 2022. Efficient few-shot learning without prompts. arXiv preprint arXiv:2209.11055.  
+Dong Wang, Ziran Li, Haitao Zheng, and Ying Shen. 2020. Integrating user history into heterogeneous graph for dialogue act recognition. In Proceedings of the 28th International Conference on Computational Linguistics, pages 4211-4221.  
+Viktor Wendel and Johannes Konert. 2016. Multiplayer serious games. In Serious Games, pages 211-241. Springer.  
+Chien-Sheng Wu, Steven CH 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.  
+Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and William B Dolan. 2020. Dialogpt: Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270-278.
+
+# A Implementation Details
+
+In the paper, we describe the iterative approach of our model, where we alternate between a contrastive representation learning model and a standard classification model to improve performance.
+
+![](images/991cab92591ba3b956b3cbaa88c6eefe49f284c2b61cf17391336ed08edbd29d.jpg)  
+Figure 6: The model's performance changes with each iteration for multiple runs. Each run starts with different labeled data sets of 10 examples per class (TalkMoves dataset).
+
+Two aspects determine the decision to stop the iteration. Figure 6 shows the model's performance in subsequent iterations for the TalkMoves dataset, starting from a label set of size ten examples per dialog act. A held-out dataset was used to calibrate the model's performance. One can see that the performance improvement is non-monotonic. In most cases, stopping after just one or two iterations is necessary to get maximum performance. The system is trying to improve based on minimal information in the labeled set, so as the size of the label set is small, there is a potential for over-fitting to the information contained in the label set. Another implementation detail is the selection of threshold confidence levels to distill soft labels as we expand the training set of the next model at the end of an iteration. Figure 7 shows the performance of our model for various prediction probability distillation thresholds. As mentioned earlier in the main paper and observed in Figure 7,  $85\%$  is the distillation threshold with the best performance distribution.
+
+# B Sample Selection by Active Learning
+
+We experimented with using active learning techniques to pick labeled examples (Ren et al., 2021). We first trained a model with just ten initial labeled examples per class. We used the model to find a new set of ten samples for each class with the least confidence and added them to the labeled set using human labels. Figure 8 compares our active learning with uniform random selection approaches as a baseline. We can see that the random approach has much more variance for every label set size. However, we also note that the best-performing model was a random set for every label size. This difference in performance might be because our ap
+
+![](images/86966fc5971adbb39b44d3867f8765f100f2a30f13fdef1a662094191d2a4ca8.jpg)  
+Figure 7: The figure shows our model's performance distribution for different soft label distillation thresholds. The runs were for 70 labeled samples per class as the initial training set from the TalkMoves dataset.
+
+![](images/8b40287e4cbfc4d9f28b4391c8db8bdd186a4275c933040c162253d6019cc17b.jpg)  
+Figure 8: Comparison between active learning selection vs uniform random selection of labeled samples.
+
+proach for active sample selection was naive. When we add a new sample based on low prediction confidence, there is no guarantee that we will improve the model. The new sample might be predicted with low confidence either because our model has limited information to decide or because the sample itself is an outlier. If it is the former adding the sample might help, but if it is the latter, adding the sample increases the number of outliers in our training set. As our training data is limited, outliers in the input data disrupt overall performance. We plan to experiment with more sophisticated active learning techniques for sample selection in the future.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? It is in a separate section after section 7 (Conclusion)  
+A2. Did you discuss any potential risks of your work? It is in the Ethical Considerations section  
+A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+2, 3.2
+
+B1. Did you cite the creators of artifacts you used? 2,3.2  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? 2.4, 3.2  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? 3.2  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3.2  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3.2
+
+C Did you run computational experiments?
+
+3
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 3  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank.

2023/Improving Classroom Dialogue Act Recognition from Limited Labeled Data with Self-Supervised Contrastive Learning Classifiers/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0716a5d9e58b8b05ebcaa22cf94fcde0ff564eeb2ba88cab5e79469d2534511
+size 385543

2023/Improving Contrastive Learning of Sentence Embeddings from AI Feedback/8930019d-cd04-44ea-a08a-812fc0aef694_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d89606a35d493b7d260665989e8018a7b1321d7a33d36470738358ec429df2be
+size 510862

2023/Improving Contrastive Learning of Sentence Embeddings from AI Feedback/full.md

@@ -0,0 +1,454 @@
+# Improving Contrastive Learning of Sentence Embeddings from AI Feedback
+
+Qinyuan Cheng, Xiaogui Yang, Tianxiang Sun, Linyang Li, Xipeng Qiu†
+
+School of Computer Science, Fudan University
+
+chengqy21@m.fudan.edu.cn
+
+# Abstract
+
+Contrastive learning has become a popular approach in natural language processing, particularly for the learning of sentence embeddings. However, the discrete nature of natural language makes it difficult to ensure the quality of positive and negative sample pairs generated through data augmentation methods. Although supervised contrastive learning can produce more accurate sample pairs with human feedback labels, it still lacks fine-grained training signals. In this paper, we propose to improve Contrastive Learning of sentence embeddings from AI Feedback (CLAIF). Our method utilizes AI feedback from large pre-trained language models (LLMs) to construct sample pairs with fine-grained sample similarity scores to improve contrastive learning. Besides, we combine human feedback and AI feedback to provide better supervision signals for supervised contrastive learning of sentence embeddings. Experimental results show that our method achieves state-of-the-art performance on several semantic textual similarity (STS) and transfer learning tasks compared to other unsupervised and supervised contrastive learning methods.
+
+# 1 Introduction
+
+Learning sentence embeddings with rich semantics is very important for many natural language processing tasks, such as semantic matching and information retrieval. Recently, pre-trained language models (Devlin et al., 2019; Liu et al., 2019; Qiu et al., 2020) provide a convenient way to get sentence embeddings. However, sentence embeddings directly generated by pre-trained language models show poor performance on semantic textual similarity (STS) tasks due to the representation degeneration problem (Gao et al., 2019). Therefore,
+
+finding ways to further improve pre-trained models to produce better sentence embeddings becomes an crucial and fundamental challenge in natural language processing.
+
+Given the shortage of labeled data for sentence embedding learning, recent studies mainly focus on unsupervised methods, such as utilizing contrastive learning methods(Yan et al., 2021; Gao et al., 2021; Chuang et al., 2022). Contrastive learning can be classified into two categories (Khosla et al., 2020): supervised contrastive learning and unsupervised contrastive learning, depending on whether additional label information is utilized to construct positive and negative sample pairs. However, the quality of positive and negative sample pairs in unsupervised contrastive learning can be difficult to ensure. Recent studies also show that data augmentation strategies in unsupervised contrastive learning may introduce some bias like length information (Wu et al., 2022) and improper negatives (Zhou et al., 2022a). While supervised contrastive learning methods can produce more accurate sample pairs by utilizing label information, such as using supervised datasets from natural language inference (Gao et al., 2021), it can only provide coarse-grained labels and lack fine-grained supervision signals. We aruge that these limitations of current contrastive learning methods restrict further performance enhancement of sentence embeddings.
+
+With the emergence of large pre-trained language models (LLMs) (Brown et al., 2020; Sun et al., 2021; Ouyang et al., 2022; Zhang et al., 2022), researchers hope powerful LLMs can help human train other AI models (Bai et al., 2022). One way is to use LLMs to generate datasets using for zero-shot learning (Schick and Schütze, 2021; Ye et al., 2022; Meng et al., 2022). These methods all use predefined labels and task descriptions to generate training inputs, instead of utilizing AI feedback as supervision signals. Therefore, these
+
+![](images/84602c63951eecb263a5a4fdccb3c6b531e3e19d61f0b94f897f863c4ca4428a.jpg)  
+Step 1: Sentence Pair Generation
+
+![](images/80a904c9427ad65b689138e787836c37635accc6a741550de9cba08ead84a3d8.jpg)  
+Step 2: Semantic Similarity Labeling  
+Figure 1: Illustration of the sample pair generation process. The darker the color, the more information the sentence shares with the original sentence.
+
+method are not suitable for tasks whose labels are continuous values and may lead to lack of diversity in training samples. Inspired by these studies, we hope to exploit the capability of LLMs to address shortcomings in contrastive learning of sentence embeddings.
+
+We propose to improve Contrastive Learning of sentence embeddings from AI Feedback (CLAIF). Specifically, we design a two-step sample pair generation method to produce high quality sentence pairs and fine-grained semantic similarity scores using AI feedback from GPT-3, as shown in Figure 1. In the first step, we mask some words in a sentence with different mask rates and then use GPT-3 to generate new sentences based on the remaining information in the masked sentence. Then we combine the generated sentences and the original sentence to construct sentence pairs. In this way, we can use the mask rate to control the amount of sharing information between two sentences in a pair, which will produce sentence pairs with different semantic similarities. In the second step, we utilize GPT-3 to generate semantic similarity scores for sentence pairs. These scores are the AI feedback on sample similarity. Since the semantic change caused by reconstructing a masked sentence is difficult to measure, we leverage the
+
+linguistic knowledge of LLMs to generate the semantic similarity score. The diversity of AI feedback similarity scores ensured by the sentence pair generation process in the first step. At last we use our generated sample pairs and similarity scores to train the model for sentence embeddings.
+
+In addition to using AI feedback alone, we also combine human feedback and AI feedback by introducing AI feedback into supervised contrastive learning of sentence embeddings which needs human feedback labels to generate positive sample pairs. We use the AI feedback similarity score for the positive sample pair as a soft label to replace the one-hot label in InfoNCE loss (He et al., 2020). We term our loss Soft InfoNCE. This process can be referred to as contrastive learning of sentence embeddings from human and AI feedback (CLHAIF).
+
+We conduct extensive experiments to show the effectiveness of our method. Sentence embeddings learned with CLAIF and CLHAIF achieve state-of-the-art performance on standard semantic textual similarity tasks and outperform strong baselines on transfer learning tasks. We also find that CLAIF results in significant improvements to the cross-encoder architecture for the sentence-pair modeling task.
+
+Our main contributions are as follows:
+
+<table><tr><td>Feedback Source</td><td>Positive Pair</td><td>Negative Pair</td><td>Loss Function</td></tr><tr><td>Zero Feedback (CLZF)</td><td>(xi,xi&#x27;)</td><td>{(xi, xj) | xj ∈ X, i ≠ j}</td><td>InfoNCE (van den Oord et al., 2018; He et al., 2020; Gao et al., 2021), NT-Xent (Chen et al., 2020)</td></tr><tr><td>Human Feedback (CLHF)</td><td>(xi, xi+)</td><td>{(xi, xi-), (xi, xj) | xj ∈ X, i ≠ j}</td><td>SupCon (Khosla et al., 2020), InfoNCE (Gao et al., 2021), KNN-Contrastive (Zhou et al., 2022b)</td></tr><tr><td>AI Feedback (CLAIF)</td><td>(xi, xi&#x27;, yi)</td><td>(xi, xi&#x27;, yi)*</td><td>Mean Squared Error</td></tr><tr><td>Human and AI Feedback (CLHAIF)</td><td>(xi, xi+, yi)</td><td>{(xi, xi-), (xi, xj) | xj ∈ X, i ≠ j}</td><td>Soft InfoNCE</td></tr></table>
+
+Table 1: The details of contrastive learning from different feedback.  $X$  is the full set containing all samples and  $x_{i}$  is the i-th sample of  $X$ , such as a sentence or an image.  $x_{i}^{\prime}$  is an augmented sample obtained by using some data augmentation strategies to  $x_{i}$ .  $x_{i}^{+}$ and  $x_{i}^{-}$ are positive sample and negative sample of  $x_{i}$  picked by human feedback information, such as class label information.  $y_{i}$  is the AI feedback sample similarity score for the i-th sample pair.*: CLAIF does not explicitly construct positive and negative pairs, sample pairs with high similarity scores can be seen as positive pairs and those with low scores can be seen as negative pairs.
+
+- We propose to improve contrastive learning of sentence embeddings from AI feedback (CLAIF) and achieve state-of-the-art performance on several semantic textual similarity tasks and transfer learning tasks.  
+- We construct a semantic textual similarity dataset with high quality sentence pairs and fine-grained AI feedback similarity scores using large pre-trained language models.  
+- We propose a method to incorporate human feedback and AI feedback to provide better supervision for contrastive learning of sentence embeddings.  
+- Experimental results show the scalability of CLAIF, which is cheaper and more efficient than collecting data from human feedback.
+
+# 2 Understanding Contrastive Learning from Different Feedback
+
+In this section, we categorize contrastive learning methods into four categories according to their feedback sources. We summarize the details of contrastive learning from different feedback in Table 1, including their feedback types, sample pairs construction methods and representative loss functions.
+
+# 2.1 Contrastive Learning from Zero Feedback
+
+Traditional contrastive learning is used for self-supervised representation learning (Hadsell et al.,
+
+2006; He et al., 2020). These methods construct positive and negative sample pairs using data augmentation strategies without any human feedback. For example, in natural language processing, Gao et al. (2021) construct positive sample pairs by doing the dropout operation twice for the same sentence and negative pairs by combining with another sentences. We refer to these methods as Contrastive Learning from Zero Feedback (CLZF). The most common loss function for CLZF is InfoNCE (van den Oord et al., 2018). Chen et al. (2020) propose NT-Xent loss, which can be seen as a variant of InfoNCE. However, due to the discrete nature of natural language, it is hard to find effective and unbiased data augmentation strategies to construct high quality sample pairs.
+
+# 2.2 Contrastive Learning from Human Feedback
+
+Recently, Khosla et al. (2020) propose to use label information to construct positive sample pairs. In sentence embeddings, Gao et al. (2021) use premise-hypothesis pairs with entailment relationship from natural language inference (NLI) datasets as positive sample pairs and still use InfoNCE for training. Since these methods leverage label information from human, we refer to them as Contrastive Learning from Human Feedback (CLHF). With the help of label information, some new losses can be used in CLHF, like SupCon (Khosla et al., 2020) and KNN-Contrastive (Zhou et al., 2022b). Although CLHF can construct more accurate sam
+
+ple pairs, it still lacks fine-grained supervision signals. For example, in InfoNCE, all positive pairs have a label of 1. But there are also differences in the similarity between different positive sample pairs.
+
+# 2.3 Contrastive Learning from AI Feedback
+
+Measuring the similarity of sample pairs in contrastive learning is a laborious task. However, thanks to emergence of LLMs, we can use LLMs to measure the similarity of sample pairs and use the AI feedback as our training signals. We refer to this approach as Contrastive Learning from AI Feedback (CLAIF). CLAIF does not need to explicitly construct positive and negative sample pairs because each sample pair has a fine-grained label. We use mean squared error (MSE) loss for the training of CLAIF in this work.
+
+# 2.4 Contrastive Learning from Human and AI Feedback
+
+Besides contrastive learning from AI feedback, we propose to combine human and AI feedback to produce better supervision signals when they are both available. We call this category contrastive learning from human and AI feedback (CLHAIF) and we propose a soft InfoNCE loss for the training of CLHAIF. We hope to use fine-grained AI feedback to refine the coarse-grained signals in current CLHF methods.
+
+# 3 Methodology
+
+In this section, we first introduce our method to generate sample pairs and the training process of CLAIF. In order to obtain high quality sentence pairs with diverse and fine-grained similarity scores, we propose a two-step sample pair generation method: Sentence Pair Generation and Semantic Similarity Labeling. The generation process is shown in Figure 1. We use these sample pairs to train language models like BERT and RoBERTa. Then we introduce CLHAIF, which combines human and AI feedback in contrastive learning of sentences embeddings.
+
+# 3.1 Sentence Pair Generation
+
+We use unpaired sentences from the training set of STS Benchmark (Cer et al., 2017) as our original sentences. As shown in Figure 1, we first mask some words of the original sentence "a man is playing a flute." with different mask rates using the
+
+![](images/279347a5afc4e63209ccf7e7ba5224262665347014feae1d433ad52b2bc055ba.jpg)  
+Figure 2: The score distribution of our generated sample pairs. The x-axis is the similarity score and the y-axis is the percentage of the score.
+
+<mask> token, in order to delete some information in the original sentence. The more words that are masked, the less information is left. We use the depth of color to indicate the degree of information sharing between two sentences in Figure 1. Then we write a task description prompt to steer GPT-3 to generate new sentences based on masked sentences. We provide our task descriptions in Appendix B. To increase the diversity of generated sentences, we merge adjacent <mask> tokens in  $50\%$  of masked sentences into one <mask> token. Then we combine the original sentence with each generated sentence to construct sentence pairs.
+
+# 3.2 Semantic Similarity Labeling
+
+In this step, we label the semantic similarity score for each sentence pair using AI feedback from GPT-3. The similarity score ranges from 0 to 1, where a score of 1 means that the semantic of the two sentences are exactly the same, and a score of 0 means that the semantic of the two sentences are completely different. We write a task description prompt to steer GPT-3 to generate a similarity score between 0 and 1 for each sample pair generated in step 1. The first step ensures the diversity of semantic similarity scores. As illustrated in Figure 2, the generated scores are diverse and distributed in the value range from 0 to 1.
+
+# 3.3 Training on Generated Pairs
+
+With the generated sample pairs, we train a language model as the sentence encoder to get better sentence embeddings. Given diverse sentence pairs which have fine-grained similarity scores, we do not need to explicitly construct positive and negative sample pairs. Therefore, we directly use the mean squared error (MSE) loss to fit the cosine similarity of each sentence pair to its AI feedback
+
+similarity score:
+
+$$
+\mathcal {L} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ \cos \left(\mathbf {h} _ {i}, \mathbf {h} _ {i} ^ {\prime}\right) - y _ {i} \right] ^ {2} \tag {1}
+$$
+
+where  $N$  is the batch size,  $\mathbf{h}_i$  and  $\mathbf{h}_i^{\prime}$  are two sentence embeddings of the i-th sentence pair  $(x_i,x_i^{\prime})$  encoded by the model,  $y_{i}$  is the corresponding similarity score and cos means the calculation of cosine similarity. During inference, we use the cosine similarity of two sentence embeddings as their semantic similarity score.
+
+# 3.4 Combining Human Feedback and AI Feedback
+
+In this section, we mainly study the cooperation of human and AI models to provide better training signals for contrastive learning, which we called CLHAIF. Reimers and Gurevych (2019) use supervised NLI datasets to learn sentence embeddings. Gao et al. (2021) construct positive and hard negative sample pairs for contrastive learning leveraging label information of NLI datasets, achieving significant improvements. However, as we mentioned in Section 2.2, CLHF does not distinguish between different positive sample pairs and assigns label of 1 for all positive pairs. In this way, all positive sample pairs are pulled together with the same extent in contrastive learning, ignoring differences in similarity between different positive pairs. Therefore, we use AI feedback to refine these coarse-grained supervision signals.
+
+At first, we use the semantic similarity labeling step in Section 3.2 to generate AI feedback similarity scores for sentence pairs constructed from supervised NLI datasets: SNLI (Bowman et al., 2015) and MNLI (Williams et al., 2018). Following Gao et al. (2021), we construct sample pairs using the label information. For the i-th sample of the NLI dataset, we can obtain two sentence pairs  $(x_{i}, x_{i}^{+})$  and  $(x_{i}, x_{i}^{-})$ , where  $x_{i}$  is the premise,  $x_{i}^{+}$  and  $x_{i}^{-}$  are entailment and contradiction hypothesis.  $(x_{i}, x_{i}^{+})$  is the positive pair and  $(x_{i}, x_{i}^{-})$  is the hard negative pair.
+
+In order to incorporate AI feedback, we propose soft InfoNCE loss by replacing the one-hot label
+
+with the AI feedback score as the soft label:
+
+$$
+\mathcal {L} = - \frac {1}{N} \sum_ {i = 1} ^ {N} l _ {i} \tag {2}
+$$
+
+$$
+l _ {i} = y _ {i} \log \frac {e ^ {\cos \left(\mathbf {h} _ {i} , \mathbf {h} _ {i} ^ {+}\right) / \tau}}{\sum_ {j = 1} ^ {N} \left(e ^ {\cos \left(\mathbf {h} _ {i} , \mathbf {h} _ {j} ^ {+}\right) / \tau} + e ^ {\cos \left(\mathbf {h} _ {i} , \mathbf {h} _ {j} ^ {-}\right) / \tau}\right)}
+$$
+
+where  $\mathbf{N}$  is the batch size,  $\mathbf{h}_i$ ,  $\mathbf{h}_i^+$  and  $\mathbf{h}_i^-$  are sentence embeddings of  $x_i$ ,  $x_i^+$  and  $x_i^-$ ,  $y_i$  is the AI feedback similarity score for the positive pair  $(x_i, x_i^+)$  and  $\tau$  is the temperature parameter.
+
+# 4 Experiments
+
+# 4.1 Evaluation Datasets
+
+We conduct extensive experiments on seven semantic textual similarity (STS) tasks and seven transfer learning tasks. The STS tasks include STS 2012-2016 (Agirre et al., 2012, 2013, 2014, 2015, 2016), STS Benchmark (Cer et al., 2017) and SICK-Relatedness (Marelli et al., 2014). The transfer learning tasks include MR (Pang and Lee, 2005), CR (Hu and Liu, 2004), SUBJ (Pang and Lee, 2004), MPQA (Wiebe et al., 2005), SST-2 (Socher et al., 2013), TREC (Voorhees and Tice, 2000) and MRPC (Dolan and Brockett, 2005).
+
+Following Gao et al. (2021), for STS tasks, we calculate the Spearman's correlation between the cosine similarity of sentence embeddings and the golden similarity scores from STS datasets. For transfer learning tasks, we train a logistic regression classifier based on fixed sentence embeddings and follow the default settings of SentEval (Connseau and Kiela, 2018). We use the same evaluation script as Gao et al. (2021) to calculate metrics.
+
+# 4.2 Baselines
+
+We compare our method with some strong baselines among three types of sentence embedding methods:
+
+Post-processing methods: These methods adopt some post-processing operations to enhance sentence embeddings which do not need to further train the backbone model. We use BERT-whitening (Su et al., 2021), BERT-flow (Li et al., 2020) and prompt based BERT (Jiang et al., 2022) as baselines.
+
+Training methods: These methods use additional data to further train the backbone model for better sentence embeddings. We use SBERT (Reimers and Gurevych, 2019), ConSERT (Yan et al., 2021),
+
+<table><tr><td>Dataset</td><td>Sample Number</td><td>Sample Type</td></tr><tr><td>Wiki-1M</td><td>1,000,000</td><td>sentence</td></tr><tr><td>NLI</td><td>275,601</td><td>sentence triplet</td></tr><tr><td>Dino</td><td>83,497</td><td>sentence pair</td></tr><tr><td>CLAIF</td><td>113,773</td><td>sentence pair</td></tr><tr><td>CLAIFscaled</td><td>1,215,618</td><td>sentence pair</td></tr></table>
+
+SimCSE (Gao et al., 2021), DiffCSE (Chuang et al., 2022) and PromptBERT (Jiang et al., 2022) as baselines.
+
+Dataset-generation based methods: Some studies generate datasets from LLMs for sentence embedding learning. We use Dino (Schick and Schütze, 2021) as our baseline. Dino generates sentence pairs based on three discrete similarity labels using GPT2-XL. For a fair comparison, we re-implement Dino using GPT-3 in our experiments.
+
+# 4.3 Implementation Details
+
+Choice of large pre-trained language models: In our experiments, we get all AI feedback from text-davinci-003, which is the latest version of GPT-3. We access text-davinci-003 through the OpenAI API.
+
+Sample pair generation: We use nine mask rates for each original sentence in sentence pair generation: 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8. For CLAIF, we use unpaired sentences from the training set of STS-B as original sentences to construct sentence pairs from scratch and randomly sample two other sentences for each original sentence to construct two sentence pairs with a similarity score of 0. For CLHAIF, following previous studies (Gao et al., 2021; Jiang et al., 2022), we use the SNLI and MNLI datasets to construct sentence pairs and add a AI feedback similarity score for each sentence pair. We only use the AI feedback scores for positive pairs in our experiments of CLHAIF. Besides, to demonstrate the scalability of CLAIF, we use sentence pairs constructed from STS-B and from NLI datasets for the training of CLAIF, which we called  $\mathrm{CLAIF}_{\mathrm{scaled}}$ . We list statistics of some datasets used for different methods in Table 2.
+
+Training details: We use the base version of the pre-trained language model BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019) as our backbone models. We use the development set of STS-B as our validation set. In CLAIF, we use the mean
+
+Table 2: Statistics of datasets for different settings. Wiki-1M is used by CLZF methods. NLI is used by CLHF methods. We use CLAIF and  $\mathrm{CLAIF}_{\mathrm{scaled}}$  to refer to our generated datasets here.  
+
+<table><tr><td>Model</td><td>SentEval Avg.</td></tr><tr><td>SimCSEBERT</td><td>85.81</td></tr><tr><td>PromptBERT</td><td>85.49</td></tr><tr><td>DiffCSEBERT</td><td>86.86</td></tr><tr><td>CLAIFBERT</td><td>86.62</td></tr><tr><td>SimCSERoBERTa</td><td>84.84</td></tr><tr><td>PromptRoBERTa</td><td>87.36</td></tr><tr><td>DiffCSERoBERTa</td><td>87.04</td></tr><tr><td>CLAIFRoBERTa</td><td>87.99</td></tr><tr><td>SimCSEBERT-supervised</td><td>86.51</td></tr><tr><td>w/ CLHAIF</td><td>86.73</td></tr><tr><td>PromptBERTsupervised</td><td>86.98</td></tr><tr><td>w/ CLHAIF</td><td>87.09</td></tr><tr><td>SimCSERoBERTa-supervised</td><td>88.08</td></tr><tr><td>w/ CLHAIF</td><td>88.82</td></tr><tr><td>PromptRoBERTasupervised</td><td>89.11</td></tr><tr><td>w/ CLHAIF</td><td>89.27</td></tr><tr><td>CLAFScaled-BERT</td><td>87.15</td></tr><tr><td>CLAFScaled-RoBERTa</td><td>89.44</td></tr></table>
+
+Table 3: The performance comparison of CLAIF and CLHAIF on transfer learning tasks. SentEval Avg is the average accuracy on seven transfer learning datasets from SentEval.
+
+pooling strategy to get sentence embeddings for BERT and RoBERTa. For CLHAIF, we take the same pooling strategy as the corresponding baseline. Other implementation details are in Appendix A.
+
+# 4.4 Main Results
+
+Semantic Textual Similarity We compare CLAIF with methods which do not use additional labeled datasets for training, including CLZF methods and dataset generation methods. The results of CLAIF on STS tasks are shown in Table 4. We observe that CLAIF achieves the best performance on the four datasets STS15, STS16, STS-B, SICK-R and get the highest averaged Spearman's correlation on seven STS datasets. And in the comparison with dataset generation methods, CLAIF outperforms Dino by 3.37 and 2.75 points on BERT and RoBERTa. Therefore, we believe that CLAIF is more effective for the learning of sentence embeddings than CLZF methods.
+
+We implement CLHAIF by incorporating AI feedback into supervised SimCSE and supervised PromptBERT/PromptRoBERTa. We compare CLHAIF with other methods that use additional labeled datasets for training. As shown in Table 5, incorporating AI feedback improves results of
+
+<table><tr><td>Model</td><td>STS12</td><td>STS13</td><td>STS14</td><td>STS15</td><td>STS16</td><td>STS-B</td><td>SICK-R</td><td>Avg.</td></tr><tr><td colspan="9">BERT-base</td></tr><tr><td>BERT-flow†</td><td>58.40</td><td>67.10</td><td>60.85</td><td>75.16</td><td>71.22</td><td>68.66</td><td>64.47</td><td>66.55</td></tr><tr><td>BERT-whitening†</td><td>57.83</td><td>66.90</td><td>60.90</td><td>75.08</td><td>71.31</td><td>68.24</td><td>63.73</td><td>66.28</td></tr><tr><td>Prompt based BERT†</td><td>60.96</td><td>73.83</td><td>62.18</td><td>71.54</td><td>68.68</td><td>70.60</td><td>67.16</td><td>67.85</td></tr><tr><td>ConBERT†</td><td>64.64</td><td>78.49</td><td>69.07</td><td>79.72</td><td>75.95</td><td>73.97</td><td>67.31</td><td>72.74</td></tr><tr><td>SimCSE†</td><td>68.40</td><td>82.41</td><td>74.38</td><td>80.91</td><td>78.56</td><td>76.85</td><td>72.23</td><td>76.25</td></tr><tr><td>DiffCSE‡</td><td>72.28</td><td>84.43</td><td>76.47</td><td>83.90</td><td>80.54</td><td>80.59</td><td>71.23</td><td>78.49</td></tr><tr><td>PromptBERT†</td><td>71.56</td><td>84.58</td><td>76.98</td><td>84.47</td><td>80.60</td><td>81.60</td><td>69.87</td><td>78.54</td></tr><tr><td>DinogPT-3</td><td>72.61</td><td>81.92</td><td>75.09</td><td>80.42</td><td>76.26</td><td>77.10</td><td>70.43</td><td>76.26</td></tr><tr><td>CLAIF</td><td>70.62</td><td>81.51</td><td>76.29</td><td>85.05</td><td>81.36</td><td>84.34</td><td>78.22</td><td>79.63</td></tr><tr><td colspan="9">RoBERTa-base</td></tr><tr><td>RoBERTa-whitening†</td><td>46.99</td><td>63.24</td><td>57.23</td><td>71.36</td><td>68.99</td><td>61.36</td><td>62.91</td><td>61.73</td></tr><tr><td>SimCSE†</td><td>70.16</td><td>81.77</td><td>73.24</td><td>81.36</td><td>80.65</td><td>80.22</td><td>68.56</td><td>76.57</td></tr><tr><td>DiffCSE‡</td><td>70.05</td><td>83.43</td><td>75.49</td><td>82.81</td><td>82.12</td><td>82.38</td><td>71.19</td><td>78.21</td></tr><tr><td>PromptRoBERTa†</td><td>73.94</td><td>84.74</td><td>77.28</td><td>84.99</td><td>81.74</td><td>81.88</td><td>69.50</td><td>79.15</td></tr><tr><td>Dino§</td><td>70.27</td><td>81.26</td><td>71.25</td><td>80.49</td><td>77.18</td><td>77.82</td><td>68.09</td><td>75.20</td></tr><tr><td>DinogPT-3</td><td>71.24</td><td>81.55</td><td>75.67</td><td>81.42</td><td>78.77</td><td>80.10</td><td>71.31</td><td>77.15</td></tr><tr><td>CLAIF</td><td>68.33</td><td>82.26</td><td>77.00</td><td>85.18</td><td>83.43</td><td>85.05</td><td>78.02</td><td>79.90</td></tr></table>
+
+Table 4: The performance comparison of CLAIF on STS tasks. †: results from (Jiang et al., 2022). ‡: results from (Chuang et al., 2022). §: results from (Schick and Schütze, 2021). Other results are from our experiments. We bold the highest results among models with the same backbone.  
+
+<table><tr><td>Model</td><td>STS12</td><td>STS13</td><td>STS14</td><td>STS15</td><td>STS16</td><td>STS-B</td><td>SICK-R</td><td>Avg.</td></tr><tr><td colspan="9">BERT-base</td></tr><tr><td>SBERT†</td><td>70.97</td><td>76.53</td><td>73.19</td><td>79.09</td><td>74.30</td><td>77.03</td><td>72.91</td><td>74.89</td></tr><tr><td>SBERT-flow†</td><td>69.78</td><td>77.27</td><td>74.35</td><td>82.01</td><td>77.46</td><td>79.12</td><td>76.21</td><td>76.60</td></tr><tr><td>SBERT-whitening†</td><td>69.65</td><td>77.57</td><td>74.66</td><td>82.27</td><td>78.39</td><td>79.52</td><td>76.91</td><td>77.00</td></tr><tr><td>ConBERT†</td><td>74.07</td><td>83.93</td><td>77.05</td><td>83.66</td><td>78.76</td><td>81.36</td><td>76.77</td><td>79.37</td></tr><tr><td>SimCSE†</td><td>75.30</td><td>84.67</td><td>80.19</td><td>85.40</td><td>80.82</td><td>84.25</td><td>80.39</td><td>81.57</td></tr><tr><td>w/ CLHAIF</td><td>74.86↓0.44</td><td>85.09↑0.42</td><td>81.24↑1.05</td><td>85.96↑0.56</td><td>81.33↑0.51</td><td>84.96↑0.71</td><td>81.36↑0.97</td><td>82.08↑0.51</td></tr><tr><td>PromptBERT*</td><td>75.10</td><td>85.54</td><td>80.58</td><td>86.00</td><td>81.24</td><td>84.57</td><td>80.36</td><td>81.91</td></tr><tr><td>w/ CLHAIF</td><td>75.03↓0.07</td><td>85.88↑0.34</td><td>81.48↑0.90</td><td>86.33↑0.33</td><td>81.40↑0.16</td><td>84.93↑0.36</td><td>80.98↑0.62</td><td>82.29↑0.38</td></tr><tr><td>CLAIFScaled</td><td>74.36</td><td>85.07</td><td>80.64</td><td>87.21</td><td>83.36</td><td>86.26</td><td>79.68</td><td>82.37</td></tr><tr><td colspan="9">RoBERTa-base</td></tr><tr><td>SRoBERTa†</td><td>71.54</td><td>72.49</td><td>70.80</td><td>78.74</td><td>73.69</td><td>77.77</td><td>74.46</td><td>74.21</td></tr><tr><td>SRoBERTa-whitening†</td><td>70.46</td><td>77.07</td><td>74.46</td><td>81.64</td><td>76.43</td><td>79.49</td><td>76.65</td><td>76.60</td></tr><tr><td>SimCSE†</td><td>76.53</td><td>85.21</td><td>80.95</td><td>86.03</td><td>82.57</td><td>85.83</td><td>80.50</td><td>82.52</td></tr><tr><td>w/ CLHAIF</td><td>76.23↓0.30</td><td>85.46↑0.25</td><td>81.48↑0.53</td><td>86.47↑0.44</td><td>83.40↑0.83</td><td>85.93↑0.10</td><td>80.95↑0.45</td><td>82.85↑0.33</td></tr><tr><td>PromptRoBERTa*</td><td>76.41</td><td>85.64</td><td>82.11</td><td>86.18</td><td>82.71</td><td>85.74</td><td>79.95</td><td>82.68</td></tr><tr><td>w/ CLHAIF</td><td>76.26↓0.15</td><td>86.01↑0.37</td><td>82.83↑0.72</td><td>86.70↑0.52</td><td>82.94↑0.23</td><td>86.04↑0.30</td><td>80.55↑0.60</td><td>83.05↑0.37</td></tr><tr><td>CLAIFScaled</td><td>72.58</td><td>84.50</td><td>79.48</td><td>86.92</td><td>84.19</td><td>85.85</td><td>79.64</td><td>81.88</td></tr></table>
+
+Table 5: The performance comparison of CLHAIF on STS tasks. †: results from Jiang et al. (2022). Other results are from our experiments. *: The results of PromptBERT and PromptRoBERTa are obtained by running official code of Jiang et al. (2022) with recommended hyperparameters.
+
+CLHF methods like supervised SimCSE on six STS datasets except STS12.
+
+Transfer Tasks In addition to STS tasks, we also evaluate several transfer learning tasks from SentEval. Experimental results show that sentence embeddings learned with CLAIF and CLHAIF also achieve better or comparable performance compared to baselines. We present the average results for seven transfer tasks in Table 3 and detailed results in Appendix C.
+
+# 4.5 Scalability of CLAIF
+
+In this section we discuss the scalability of CLAIF. The results of  $\mathrm{CLAIF}_{\mathrm{scaled}}$  in Table 5 show that using more data to scale CLAIF can bring significant improvements.  $\mathrm{CLAIF}_{\mathrm{scaled}}$  greatly outputforms CLAIF by 2.74 points on BERT-base  $(79.63 \rightarrow 82.37)$  and even outputforms or performs on par with CLHF and CLHAIF methods. We believe that using more data can further improve the performance of CLAIF. Since collecting data from AI
+
+<table><tr><td>Model</td><td>STS12</td><td>STS13</td><td>STS14</td><td>STS15</td><td>STS16</td><td>STS-B</td><td>SICK-R</td><td>Avg.</td></tr><tr><td colspan="9">BERT-base</td></tr><tr><td>Trans-Encodercross</td><td>71.94</td><td>84.14</td><td>76.39</td><td>82.87</td><td>80.65</td><td>81.06</td><td>71.16</td><td>78.32</td></tr><tr><td>CLAIFcross</td><td>70.36</td><td>83.27</td><td>79.73</td><td>87.87</td><td>84.54</td><td>85.00</td><td>78.33</td><td>81.30</td></tr><tr><td colspan="9">RoBERTa-base</td></tr><tr><td>Trans-Encodercross</td><td>72.59</td><td>83.24</td><td>76.83</td><td>84.20</td><td>82.82</td><td>82.85</td><td>69.51</td><td>78.86</td></tr><tr><td>CLAIFcross</td><td>72.80</td><td>83.75</td><td>81.52</td><td>88.66</td><td>86.61</td><td>87.05</td><td>81.28</td><td>83.10</td></tr></table>
+
+Table 6: The performance comparison of CLAIF based on the cross-encoder architecture.
+
+feedback is more cheaper than from human feedback, we argue that CLAIF has great potential in practical applications.
+
+# 4.6 Sentence-Pair Modeling
+
+In this section, we evaluate CLAIF on the sentence-pair modeling task. Cross-encoders usually outperform bi-encoders in information retrieval. However, we observe in Liu et al. (2022) that the cross-encoder does not show its superior on sentence-pair modeling. We contribute this to the lack of fine-grained training signals. We train a cross-encoder with CLAIF. Experimental results in Table 11 show that, with the help of AI feedback,  $\mathrm{CLAIF}_{\mathrm{cross}}$  brings significant improvements for cross-encoders on the sentence-pair modeling task compared to the previous model Trans-Encoder (Liu et al., 2022). More training details are in Appendix D.
+
+# 4.7 Human Evaluation
+
+In this section, we conduct human evaluation to measure the quality of generated sentences and similarity scores. We measure whether the generated sentences are fluent and whether the similarity scores are consistent with the real semantic similarities. To help human judge the consistency, we generate a natural language explanation for each generated similarity score using GPT-3. We invite 4 experts to participate in our human evaluation. Then we random pick 100 samples from the dataset used in CLAIF and assign 25 samples to each expert. In the evaluation, 92 percent of generated sentences are considered fluent and 90 percent of generated scores are considered consistent by the expert, which means our method can generate high quality sentence pairs and similarity scores.
+
+# 5 Related Work
+
+Recent studies about sentence embeddings mainly focus on using additional data to further train pre-trained language models. Yan et al. (2021) and Gao et al. (2021) propose different data aug
+
+mentation strategies for contrastive learning and achieve significant improvements using unlabeled data. Chuang et al. (2022) use equivariant contrastive learning for learning better representations. Zhou et al. (2022a) and Wu et al. (2022) address the bias caused by construction processes of negative and positive samples. Jiang et al. (2022) use different prompt templates to produce positive pairs for contrastive learning. Opitz and Frank (2022) use various semantic sentence features to construct fine-grained labels for sentence embedding training.
+
+Impressed by the powerful capabilities of LLMs (Brown et al., 2020; Ouyang et al., 2022), researchers pay more attention to using AI feedback from LLMs for zero-shot and few-shot learning. Li et al. (2023); Li and Qiu (2023) use AI feedback from language models to enhance In-context Learning and Chain-of-Thoughts. Ye et al. (2022) and Meng et al. (2022) generate datasets by taking labels and prompts as the input of LLMs and then let LLMs generate training samples. Schick and Schütze (2021) design a dataset generation method for STS tasks. They construct three natural language instructions based on three discrete similarity scores and then use these instructions to steer LLMs to construct sentence pairs. However, it is hard to use natural language to describe various similarity scores, since the similarity score is a continuous variable with values ranging from 0 to 1.
+
+# 6 Conclusion
+
+In this paper, we first formalize four types of contrastive learning: contrastive learning from zero feedback (CLZF), contrastive learning from human feedback (CLHF), contrastive learning from AI feedback (CLAIF) and contrastive learning from human and AI feedback (CLHAIF). Then we improve contrastive learning of sentence embeddings from AI feedback and combine human feedback with AI feedback to produce better supervision
+
+signals. Experimental results show that CLAIF and CLHAIF can bring substantial improvements for sentence embedding learning. We hope that learning from AI feedback can shed new lights for representation learning and contrastive learning.
+
+# Limitations
+
+To inspire future work, we conclude some limitations of our work as follows:
+
+- While our method achieves promising performance on sentence embedding related tasks like STS, the performance on other natural language processing tasks are still need to investigate.  
+- The AI feedback in our experiments comes from GPT-3, which requires a fee to use.  
+- We do not explore the effect of different task description prompts on the quality of generated sample pairs, which may influence the performance of CLAIF.  
+- In CLHAIF, we only use the AI feedback for positive sample pairs. How to utilize AI feedback for negative sample pairs remains to be studied.
+
+# Acknowledgement
+
+We would like to thank all anonymous reviewers for their valuable advice. This work was supported by the National Natural Science Foundation of China (No. 62236004 and No. 62022027).
+
+# References
+
+Eneko Agirre, Carmen Banea, Claire Cardie, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Inigo Lopez-Gazpio, Montse Maritxalar, Rada Mihalcea, German Rigau, Larraitz Uria, and Janyce Wiebe. 2015. Semeval-2015 task 2: Semantic textual similarity, english, spanish and pilot on interpretability. In Proceedings of the 9th International Workshop on Semantic Evaluation, SemEval@NAACL-HLT 2015, Denver, Colorado, USA, June 4-5, 2015, pages 252-263. The Association for Computer Linguistics.  
+Eneko Agirre, Carmen Banea, Claire Cardie, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Rada Mihalcea, German Rigau, and Janyce Wiebe. 2014. Semeval-2014 task 10: Multilingual semantic textual similarity. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August
+
+23-24, 2014, pages 81-91. The Association for Computer Linguistics.  
+Eneko Agirre, Carmen Banea, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez-Agirre, Rada Mihalcea, German Rigau, and Janyce Wiebe. 2016. Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In Proceedings of the 10th International Workshop on Semantic Evaluation, SemEval@NAACL-HLT 2016, San Diego, CA, USA, June 16-17, 2016, pages 497-511. The Association for Computer Linguistics.  
+Eneko Agirre, Daniel M. Cer, Mona T. Diab, and Aitor Gonzalez-Agirre. 2012. Semeval-2012 task 6: A pilot on semantic textual similarity. In Proceedings of the 6th International Workshop on Semantic Evaluation, SemEval@NAACL-HLT 2012, Montreal, Canada, June 7-8, 2012, pages 385-393. The Association for Computer Linguistics.  
+Eneko Agirre, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez-Agirre, and Weiwei Guo. 2013. *sem 2013 shared task: Semantic textual similarity. In Proceedings of the Second Joint Conference on Lexical and Computational Semantics*, *SEM* 2013, June 13-14, 2013, Atlanta, Georgia, USA, pages 32-43. Association for Computational Linguistics.  
+Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukosiute, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan. 2022. Constitutional AI: harmlessness from AI feedback. CoRR, abs/2212.08073.  
+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 2015, Lisbon, Portugal, September 17-21, 2015, pages 632-642. The Association for Computational Linguistics.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess,
+
+Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.  
+Daniel M. Cer, Mona T. Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. 2017. Semeval-2017 task 1: Semantic textual similarity - multilingual and cross-lingual focused evaluation. CoRR, abs/1708.00055.  
+Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. 2020. A simple framework for contrastive learning of visual representations. 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 1597-1607. PMLR.  
+Yung-Sung Chuang, Rumen Dangovski, Hongyin Luo, Yang Zhang, Shiyu Chang, Marin Soljacic, Shang-Wen Li, Scott Yih, Yoon Kim, and James R. 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, NAACL 2022, Seattle, WA, United States, July 10-15, 2022, pages 4207-4218. Association for Computational Linguistics.  
+Alexis Conneau and Douwe Kiela. 2018. Senteval: An evaluation toolkit for universal sentence representations. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, LREC 2018, Miyazaki, Japan, May 7-12, 2018. European Language Resources Association (ELRA).  
+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.  
+William B. Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In Proceedings of the Third International Workshop on Paraphrasing, IWP@IJCNLP 2005, Jeju Island, Korea, October 2005, 2005. Asian Federation of Natural Language Processing.  
+Jun Gao, Di He, Xu Tan, Tao Qin, Liwei Wang, and TieYan Liu. 2019. Representation degeneration problem in training natural language generation models. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.
+
+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, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 6894-6910. Association for Computational Linguistics.  
+Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2006), 17-22 June 2006, New York, NY, USA, pages 1735-1742. IEEE Computer Society.  
+Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. 2020. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 9726-9735. Computer Vision Foundation / IEEE.  
+Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Seattle, Washington, USA, August 22-25, 2004, pages 168-177. ACM.  
+Ting Jiang, Shaohan Huang, Zihan Zhang, Deqing Wang, Fuzhen Zhuang, Furu Wei, Haizhen Huang, Liangjie Zhang, and Qi Zhang. 2022. Promptbert: Improving BERT sentence embeddings with prompts. CoRR, abs/2201.04337.  
+Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.  
+Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. 2020. On the sentence embeddings from pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 9119-9130. Association for Computational Linguistics.  
+Xiaonan Li, Kai Lv, Hang Yan, Tianyang Lin, Wei Zhu, Yuan Ni, Guotong Xie, Xiaoling Wang, and Xipeng Qiu. 2023. Unified demonstration retriever for in-context learning.  
+Xiaonan Li and Xipeng Qiu. 2023. Mot: Pre-thinking and recalling enable chatgpt to self-improve with memory-of-thoughts.  
+Fangyu Liu, Yunlong Jiao, Jordan Massiah, Emine Yilmaz, and Serhii Havrylov. 2022. Trans-encoder: Unsupervised sentence-pair modelling through self- and mutual-distillations. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.
+
+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. CoRR, abs/1907.11692.  
+Yuxiang Lu, Yiding Liu, Jiaxiang Liu, Yunsheng Shi, Zhengjie Huang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, Shuaiqiang Wang, Dawei Yin, and Haifeng Wang. 2022. Ernie-search: Bridging cross-encoder with dual-encoder via self-on-the-fly distillation for dense passage retrieval. CoRR, abs/2205.09153.  
+Marco Marelli, Stefano Menini, Marco Baroni, Luisa Bentivogli, Raffaella Bernardi, and Roberto Zamparelli. 2014. A SICK cure for the evaluation of compositional distributional semantic models. In Proceedings of the Ninth International Conference on Language Resources and Evaluation, LREC 2014, Reykjavik, Iceland, May 26-31, 2014, pages 216-223. European Language Resources Association (ELRA).  
+Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. 2022. Generating training data with language models: Towards zero-shot language understanding. CoRR, abs/2202.04538.  
+Juri Opitz and Anette Frank. 2022. SBERT studies meaning representations: Decomposing sentence embeddings into explainable semantic features. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing, AACL/IJCNLP 2022 - Volume 1: Long Papers, Online Only, November 20-23, 2022, pages 625-638. Association for Computational Linguistics.  
+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 F. Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. CoRR, abs/2203.02155.  
+Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, 21-26 July, 2004, Barcelona, Spain, pages 271-278. ACL.  
+Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In ACL 2005, 43rd Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference, 25-30 June 2005, University of Michigan, USA, pages 115-124. The Association for Computer Linguistics.  
+Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang. 2020. Pre-trained models for natural language processing: A survey. CoRR, abs/2003.08271.
+
+Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. 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 2019, Hong Kong, China, November 3-7, 2019, pages 3980-3990. Association for Computational Linguistics.  
+Timo Schick and Hinrich Schütze. 2021. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 6943-6951. Association for Computational Linguistics.  
+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, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of SIG-DAT, a Special Interest Group of the ACL, pages 1631-1642. ACL.  
+Jianlin Su, Jiarun Cao, Weijie Liu, and Yangyiwen Ou. 2021. Whitening sentence representations for better semantics and faster retrieval. CoRR, abs/2103.15316.  
+Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, Weixin Liu, Zhihua Wu, Weibao Gong, Jianzhong Liang, Zhizhou Shang, Peng Sun, Wei Liu, Xuan Ouyang, Dianhai Yu, Hao Tian, Hua Wu, and Haifeng Wang. 2021. ERNIE 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. CoRR, abs/2107.02137.  
+Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. 2021. Augmented SBERT: data augmentation method for improving bi-encoders for pairwise sentence scoring tasks. 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 296-310. Association for Computational Linguistics.  
+Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748.  
+Ellen M. Voorhees and Dawn M. Tice. 2000. Building a question answering test collection. In SIGIR 2000: Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, July 24-28, 2000, Athens, Greece, pages 200-207. ACM.  
+Janyce Wiebe, Theresa Wilson, and Claire Cardie. 2005. Annotating expressions of opinions and emotions in
+
+language. Lang. Resour. Evaluation, 39(2-3):165-210.
+
+Adina Williams, Nikita Nangia, and Samuel R. 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, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 1112-1122. Association for Computational Linguistics.
+
+Xing Wu, Chaochen Gao, Liangjun Zang, Jizhong Han, Zhongyuan Wang, and Songlin Hu. 2022. Esimcse: Enhanced sample building method for contrastive learning of unsupervised sentence embedding. In Proceedings of the 29th International Conference on Computational Linguistics, COLING 2022, Gyeongju, Republic of Korea, October 12-17, 2022, pages 3898-3907. International Committee on Computational Linguistics.
+
+Yuanmeng Yan, Rumei Li, Sirui Wang, Fuzheng Zhang, Wei Wu, and Weiran Xu. 2021. Consert: A contrastive framework for self-supervised sentence representation transfer. 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 5065-5075. Association for Computational Linguistics.
+
+Jiacheng Ye, Jiahui Gao, Qintong Li, Hang Xu, Jiangtao Feng, Zhiyong Wu, Tao Yu, and Lingpeng Kong. 2022. Zero: Efficient zero-shot learning via dataset generation. CoRR, abs/2202.07922.
+
+Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T. Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. 2022. OPT: open pre-trained transformer language models. CoRR, abs/2205.01068.
+
+Kun Zhou, Beichen Zhang, Xin Zhao, and Ji-Rong Wen. 2022a. Debiased contrastive learning of unsupervised sentence representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 6120-6130. Association for Computational Linguistics.
+
+Yunhua Zhou, Peiju Liu, and Xipeng Qiu. 2022b. KNN-contrastive learning for out-of-domain intent classification. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5129-5141, Dublin, Ireland. Association for Computational Linguistics.
+
+# A Implementation Details
+
+For CLAIF, we train our models for 3 epochs with a batch size of 32, and set the learning rate to 2e-5. Following previous work, we use the development set of STS-B as the validation set. We evaluate the model every 125 training steps on the validation set to choose the best checkpoint during training. We conduct a grid-search of learning rate  $\in$  {1e-5,2e-5} on the validation set.
+
+For CLHAIF, we use the official implementation and the default configuration of our baselines SimCSE (Gao et al., 2021) and PromptBERT (Jiang et al., 2022). We only replace the one-hot label with our soft label.
+
+We run experiments of CLAIF on a single RTX 3090 GPU with 24Ggpu memory and experiments of CLHAIF on 4 RTX 3090 GPUs. We fix the random seed to 42 for all experiments.
+
+# B Task Descriptions
+
+We use three task description prompts in our experiments. For sentence pair generation in Section 3.1, our two prompts are:
+
+"Replace all <mask> tokens in'<masked-sentence>' to make a new sentence. The new sentence is:" and "Write two sentences that mean the same thing. Sentence 1: '<sentence1>' Sentence 2:"
+
+For semantic similarity labeling in Section 3.2, our prompt is:
+
+"The similarity score for two sentences is in the range from 0.0 to 1.0, 0.0 means completely different and 1.0 means almost the same. Now given two sentences ' <sentence1> ' and ' <sentence2 >', please give a similarity score for these two sentences: The similarity score for these two sentences is".
+
+# C Transfer Learning Tasks
+
+We list the detailed performance comparison of CLAIF and CLHAIF in Table 7 and Table 8. Experimental results show that CLAIF achieves the best performance on RoBERTa-base and comparable performance on BERT-base. CLHAIF also achieves better results compared to the baselines. Using more data to scale CLAIF also brings performance improvements on transfer learning tasks as shown in Tabel 8.
+
+<table><tr><td>Model</td><td>MR</td><td>CR</td><td>SUBJ</td><td>MPQA</td><td>SST-2</td><td>TREC</td><td>MRPC</td><td>Avg.</td></tr><tr><td colspan="9">BERT-base</td></tr><tr><td>Avg. BERT \( embeddings^† \)</td><td>78.66</td><td>86.25</td><td>94.37</td><td>88.66</td><td>84.40</td><td>92.80</td><td>69.54</td><td>84.94</td></tr><tr><td>BERT- [CLS] \( embedding^† \)</td><td>78.68</td><td>84.85</td><td>94.21</td><td>88.23</td><td>84.13</td><td>91.40</td><td>71.13</td><td>84.66</td></tr><tr><td>\( SimCSE^‡ \)</td><td>81.18</td><td>86.46</td><td>94.45</td><td>88.88</td><td>85.50</td><td>89.80</td><td>74.43</td><td>85.81</td></tr><tr><td>SimCSE w/MLM‡</td><td>82.92</td><td>87.23</td><td>95.71</td><td>88.73</td><td>86.81</td><td>87.01</td><td>78.07</td><td>86.64</td></tr><tr><td>\( DiffCSE^‡ \)</td><td>82.69</td><td>87.23</td><td>95.23</td><td>89.28</td><td>86.60</td><td>90.40</td><td>76.58</td><td>86.86</td></tr><tr><td>\( PromptBERT^† \)</td><td>80.74</td><td>85.49</td><td>93.65</td><td>89.32</td><td>84.95</td><td>88.20</td><td>76.06</td><td>85.49</td></tr><tr><td>DinoGPT-3</td><td>79.96</td><td>85.27</td><td>93.67</td><td>88.87</td><td>84.29</td><td>88.60</td><td>69.62</td><td>84.33</td></tr><tr><td>CLAIF</td><td>81.64</td><td>87.98</td><td>94.24</td><td>89.34</td><td>86.16</td><td>89.80</td><td>77.16</td><td>86.62</td></tr><tr><td colspan="9">RoBERTa-base</td></tr><tr><td>Avg. RoBERTa \( embeddings \)</td><td>84.35</td><td>88.34</td><td>95.28</td><td>86.13</td><td>89.46</td><td>93.20</td><td>74.20</td><td>87.28</td></tr><tr><td>\( SimCSE^‡ \)</td><td>81.04</td><td>87.74</td><td>93.28</td><td>86.94</td><td>86.60</td><td>84.60</td><td>73.68</td><td>84.84</td></tr><tr><td>SimCSE w/MLM‡</td><td>83.37</td><td>87.76</td><td>95.05</td><td>87.16</td><td>89.02</td><td>90.80</td><td>75.13</td><td>86.90</td></tr><tr><td>\( DiffCSE^‡ \)</td><td>82.82</td><td>88.61</td><td>94.32</td><td>87.71</td><td>88.63</td><td>90.40</td><td>76.81</td><td>87.04</td></tr><tr><td>\( PromptRoBERTa^† \)</td><td>83.82</td><td>88.72</td><td>93.19</td><td>90.36</td><td>88.08</td><td>90.60</td><td>76.75</td><td>87.36</td></tr><tr><td>DinoGPT-3</td><td>82.31</td><td>88.66</td><td>93.95</td><td>88.72</td><td>87.53</td><td>88.20</td><td>73.74</td><td>86.16</td></tr><tr><td>CLAIF</td><td>84.11</td><td>90.62</td><td>94.29</td><td>89.13</td><td>89.57</td><td>91.00</td><td>77.22</td><td>87.99</td></tr></table>
+
+Table 7: The performance comparison of CLAIF on transfer learning tasks. †: results from (Jiang et al., 2022). ‡: results from (Chuang et al., 2022). Other results are from our experiments.
+
+# D Sentence-Pair Modeling
+
+In sentence-pair modeling task, cross-encoders can be used to directly encode the sequence of two sentences and then predict a similarity score. Previous studies (Thakur et al., 2021; Liu et al., 2022; Lu et al., 2022) show that cross-encoders usually outperform bi-encoders. We find that CLAIF can significantly improve the performance of cross-encoders on sentence-pair modeling task, with the help of fine-grained AI feedback scores.
+
+We use the binary cross-entropy (BCE) loss to train cross-encoders initialized from BERT and RoBERTa:
+
+$$
+\mathcal {L} = - \frac {1}{N} \sum_ {i = 1} ^ {N} l _ {i} \tag {3}
+$$
+
+$$
+l _ {i} = y _ {i} \log \sigma (\hat {y _ {i}}) + (1 - y _ {i}) \log (1 - \sigma (\hat {y _ {i}}))
+$$
+
+where  $N$  is the batch size,  $\hat{y}_i$  is the predicted score of the i-th sentence pair,  $y_i$  is the AI feedback similarity score and  $\sigma$  is the sigmoid function.
+
+# E Cost for Data Generation
+
+According to our billings, we spent about  $100 to generate data for CLAIF and about$ 720 for the scaled dataset.
+
+# F Generated Examples
+
+We present some generated sample pairs used in CLAIF in Table 9 and some generated similarity
+
+scores for sample pairs constructed from NLI in Table 10.
+
+# G Comparison with Text-Ada-Embedding-002
+
+Recently, OpenAI has released a powerful embedding model named text-ada-embedding-002, we compare the performance of it on STS tasks with CLAIF here. The results show that CLAIF-scaled achieves better performance on STS tasks than text-ada-embedding-002.
+
+<table><tr><td>Model</td><td>MR</td><td>CR</td><td>SUBJ</td><td>MPQA</td><td>SST-2</td><td>TREC</td><td>MRPC</td><td>Avg.</td></tr><tr><td colspan="9">BERT-base</td></tr><tr><td>SBERT†</td><td>83.64</td><td>89.43</td><td>94.39</td><td>89.86</td><td>88.96</td><td>89.60</td><td>76.00</td><td>87.41</td></tr><tr><td>SimCSE†</td><td>82.69</td><td>89.25</td><td>94.81</td><td>89.59</td><td>87.31</td><td>88.40</td><td>73.51</td><td>86.51</td></tr><tr><td>w/ CLHAIF</td><td>83.11↑0.42</td><td>88.98↓0.27</td><td>94.47↓0.34</td><td>89.95↑0.36</td><td>88.58↑1.27</td><td>86.40↓2.00</td><td>75.65↑2.14</td><td>86.73↑0.22</td></tr><tr><td>PromptBERT*</td><td>83.05</td><td>88.96</td><td>94.68</td><td>89.86</td><td>88.19</td><td>87.80</td><td>76.29</td><td>86.98</td></tr><tr><td>w/ CLHAIF</td><td>83.14↑0.09</td><td>89.12↑0.16</td><td>94.65↓0.03</td><td>89.97↑0.11</td><td>87.86↓0.33</td><td>88.80↑1.00</td><td>76.06↓0.23</td><td>87.09↑0.11</td></tr><tr><td>CLAIFscaled</td><td>82.08</td><td>89.12</td><td>94.48</td><td>90.22</td><td>87.53</td><td>90.20</td><td>76.41</td><td>87.15</td></tr><tr><td colspan="9">RoBERTa-base</td></tr><tr><td>SRoBERTa†</td><td>84.91</td><td>90.83</td><td>92.56</td><td>88.75</td><td>90.50</td><td>88.60</td><td>78.14</td><td>87.76</td></tr><tr><td>SimCSE†</td><td>84.92</td><td>92.00</td><td>94.11</td><td>89.82</td><td>91.27</td><td>88.80</td><td>75.65</td><td>88.08</td></tr><tr><td>w/ CLHAIF</td><td>86.10↑1.18</td><td>91.76↓0.24</td><td>94.66↑0.55</td><td>90.07↑0.25</td><td>91.93↑0.66</td><td>91.60↑2.80</td><td>75.59↓0.06</td><td>88.82↑0.74</td></tr><tr><td>PromptRoBERTa*</td><td>86.22</td><td>91.55</td><td>95.08</td><td>90.97</td><td>91.82</td><td>91.40</td><td>76.70</td><td>89.11</td></tr><tr><td>w/ CLHAIF</td><td>86.41↑0.19</td><td>91.76↑0.21</td><td>94.90↓0.18</td><td>91.01↑0.04</td><td>92.04↑0.22</td><td>92.40↑1.00</td><td>76.35↓0.35</td><td>89.27↑0.16</td></tr><tr><td>CLAIFscaled</td><td>85.05</td><td>91.71</td><td>94.39</td><td>90.03</td><td>91.87</td><td>94.00</td><td>79.01</td><td>89.44</td></tr></table>
+
+Table 8: The performance comparison of CLHAIF on transfer learning tasks. †: results from Jiang et al. (2022). *: The results of PromptBERT and PromptRoBERTa are obtained by running official code of Jiang et al. (2022) with recommended hyperparameters.  
+
+<table><tr><td>Original Sentence</td><td>Generated Sentence</td><td>Similarity Score</td></tr><tr><td rowspan="9">a plane is taking off.</td><td>an aircraft is departing.</td><td>0.80</td></tr><tr><td>The airplane is taking off.</td><td>0.80</td></tr><tr><td>A plane is taking off swiftly</td><td>0.90</td></tr><tr><td>The blue plane is taking off.</td><td>0.75</td></tr><tr><td>Airplane is flying.</td><td>0.67</td></tr><tr><td>Bob and Joe are taking a walk.</td><td>0.00</td></tr><tr><td>Aeroplane is flying</td><td>0.67</td></tr><tr><td>Put off steam</td><td>0.00</td></tr><tr><td>Turn off lights</td><td>0.00</td></tr><tr><td rowspan="9">a man is playing a large flute.</td><td>A male individual is performing on a big flute.</td><td>0.86</td></tr><tr><td>a man is playing a large flute.</td><td>1.00</td></tr><tr><td>He she is playing a large flute.</td><td>0.78</td></tr><tr><td>a man played a wooden flute.</td><td>0.71</td></tr><tr><td>a flute is not a wooden flute</td><td>0.20</td></tr><tr><td>a boy playing a large drum</td><td>0.33</td></tr><tr><td>a man is wise.</td><td>0.00</td></tr><tr><td>The old man stood.</td><td>0.00</td></tr><tr><td>The quick brown fox jumps over the lazy dog</td><td>0.00</td></tr><tr><td rowspan="9">three men are playing chess.</td><td>There are three men playing chess.</td><td>0.94</td></tr><tr><td>Three children are playing chess.</td><td>0.80</td></tr><tr><td>Three kings are playing chess.</td><td>0.87</td></tr><tr><td>They are playing chess.</td><td>0.80</td></tr><tr><td>three men played chess together</td><td>0.78</td></tr><tr><td>three men are walking</td><td>0.00</td></tr><tr><td>John and Mary were playing chess together</td><td>0.50</td></tr><tr><td>I play blitz chess online</td><td>0.20</td></tr><tr><td>I like to play soccer and tennis.</td><td>0.00</td></tr></table>
+
+Table 9: Generated examples of sample pairs used in CLAIF.
+
+<table><tr><td>Premise</td><td>Entailment Hypothesis</td><td>Similarity Score</td></tr><tr><td>The other men shuffled.</td><td>The other men were shuffled around.</td><td>0.78</td></tr><tr><td>well it&#x27;s been very interesting</td><td>It has been very intriguing.</td><td>0.90</td></tr><tr><td>He started slowly back to the bunkhouse.</td><td>He returned slowly to the bunkhouse.</td><td>0.91</td></tr><tr><td>well what the market can bear and</td><td>The market can bear some.</td><td>0.71</td></tr><tr><td>She smiled back.</td><td>She was happy.</td><td>0.25</td></tr><tr><td>The economy could be still better.</td><td>It still have room for improvement.</td><td>0.55</td></tr><tr><td>The man should have died instantly.</td><td>The man should not have been alive.</td><td>0.14</td></tr><tr><td>Turned out, I wasn&#x27;t completely wrong.</td><td>I was not totally wrong.</td><td>0.8</td></tr></table>
+
+Table 10: Generated examples of similarity scores used in CLHAIF.  
+
+<table><tr><td>Model</td><td>STS12</td><td>STS13</td><td>STS14</td><td>STS15</td><td>STS16</td><td>STS-B</td><td>SICK-R</td><td>Avg.</td></tr><tr><td>Ada-Embedding-002</td><td>69.80</td><td>83.26</td><td>76.08</td><td>86.12</td><td>85.96</td><td>84.30</td><td>80.25</td><td>80.82</td></tr><tr><td>CLAIF-BERT</td><td>70.62</td><td>81.51</td><td>76.29</td><td>85.05</td><td>81.36</td><td>84.34</td><td>78.22</td><td>79.63</td></tr><tr><td>CLAIF-BERTscaled</td><td>74.36</td><td>85.07</td><td>80.64</td><td>87.21</td><td>83.36</td><td>86.26</td><td>79.68</td><td>82.37</td></tr></table>
+
+Table 11: The performance comparison between CLAIF and OpenAI's text-ada-embedding-002.
+
+# A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Section Limitations.
+
+A2. Did you discuss any potential risks of your work?
+
+Our work is about representation learning and contrastive learning, which are general methods and do not have potential risks.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Abstract, 1 Introduction section.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+# B Did you use or create scientific artifacts?
+
+Implementation Details section in Appendix.
+
+B1. Did you cite the creators of artifacts you used?
+
+Implementation Details section in Appendix.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Because we only use the public datasets and open source code in this work.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Our use of these public datasets and open source code is exactly what it was intended to be.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+We use publicly available datasets that are commonly used by researchers. And our work mainly focuses on representation learning.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+We use publicly available datasets that are commonly used by researchers and we cite the paper of the open code and datasets we used, where the detailed documentation can be found.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+4 Experiments section.
+
+# C Did you run computational experiments?
+
+4 Experiments section.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+We use well-known language models BERT-base and RoBERTa-base.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+4 Experiments Section and Implementation Details section in Appendix.
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+We fix the random seed in all our experiments.
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+4 Experiments Section and Implementation Details section in Appendix.
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+No response.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+No response.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+No response.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+No response.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+No response.

2023/Improving Contrastive Learning of Sentence Embeddings from AI Feedback/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b64d21711c4ccb9d5497b201d62ae08b691a8c9908f64b0b626ee00e13bdc383
+size 998296

2023/Improving Cross-task Generalization of Unified Table-to-text Models with Compositional Task Configurations/c4fe3aff-0902-4948-b26d-fcae326e24e5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66cf587e7ec68b7a76508eb70c3d692a24791e91ea45d476d6f71136f5ff9c79
+size 672966

2023/Improving Cross-task Generalization of Unified Table-to-text Models with Compositional Task Configurations/full.md

@@ -0,0 +1,544 @@
+# Improving Cross-task Generalization of Unified Table-to-text Models with Compositional Task Configurations
+
+Jifan Chen\*
+
+Yuhao Zhang
+
+Lan Liu²
+
+Rui Dong²
+
+Xinchi Chen
+
+Patrick  $\mathbf{Ng}^2$
+
+William Yang Wang
+
+Zhiheng Huang
+
+<sup>1</sup>The University of Texas at Austin
+
+$^{2}$ AWS AI Labs
+
+jfchen@cs.utexas.edu {yhzhang, liuall, ruidong}@amazon.com
+
+{xcc, patricng, wyw, zhiheng}@amazon.com
+
+# Abstract
+
+There has been great progress in unifying various table-to-text tasks using a single encoder-decoder model trained via multi-task learning (Xie et al., 2022). However, existing methods typically encode task information with a simple dataset name as a prefix to the encoder. This not only limits the effectiveness of multitask learning, but also hinders the model's ability to generalize to new domains or tasks that were not seen during training, which is crucial for real-world applications. In this paper, we propose compositional task configurations, a set of prompts presupposed to the encoder to improve cross-task generalization of unified models. We design the task configurations to explicitly specify the task type, dataset name, as well as its input and output types. We show that this not only allows the model to better learn shared knowledge across different tasks at training, but also allows us to control the model by composing new configurations that apply novel input-output combinations in a zero-shot manner. We demonstrate via experiments over ten table-to-text tasks that our method outperforms the UnifiedSKG baseline by noticeable margins in both in-domain and zero-shot settings, with average improvements of  $+0.5$  and  $+12.6$  from using a T5-large backbone, respectively.
+
+# 1 Introduction
+
+Table-to-text tasks, such as table-based question answering (Pasupat and Liang, 2015; Herzig et al., 2020), summarization (Parikh et al., 2020), or fact verification (Chen et al., 2019), are of high interest to the NLP community and have been applied in many real-world applications. Traditionally, these tasks have been studied individually, with methods commonly optimized for one or a few tasks (Liu et al., 2021; Shi et al., 2021). However, with the recent popularity of pre-trained transformer models (Raffel et al., 2020; Lewis et al., 2020; Xue et al.,
+
+![](images/57d4844e9bc2bb04441fc62e6758912d24d4452c7ea047178c1c29f4ce402e20.jpg)  
+Figure 1: An example of unifying different tasks with a single encoder-decoder model with dataset name as a prefix. The model is trained on short-form table QA and table summarization tasks, and tested on a new long-form table QA task. As there is a mismatch between the training and test tasks, the model is unable to generalize.
+
+2021), there has been a paradigm towards unifying multiple NLP tasks with a single encoder-decoder model (Khashabi et al., 2020; Sanh et al., 2022). More recently, UnifiedSKG (Xie et al., 2022) extended this paradigm to table-to-text tasks by flattening the structured input (e.g., tables) into text format, and unifying all tasks with a T5 model (Raffel et al., 2020). By training the model over 21 datasets with structured input, it has established new state-of-the-art results for most of these tasks.
+
+Despite the success, existing work often rely on a simple trick to encode task information: the name of the dataset is often used as a prefix to the encoder at both training and test time. We argue that this overly simplified design has at least two major limitations. First, since no detailed information about the task is provided, any sharable knowledge between tasks is learned in a latent manner. Second, with this design, models are trained and evaluated on their abilities to solve specific datasets, rather than tasks. As a result, we may see substantial performance degradation when we apply the model to an unseen task at test time.
+
+Figure 1 illustrates the aforementioned limitations: a unified model (such as UnifiedSKG) is trained on short-form table QA (Zhong et al., 2017) and table-based summarization (Parikh et al., 2020), and we want to test the trained model on long-form table QA (Nan et al., 2022a), where the model should take a question and a table as input and output an abstractive sentence as the answer. As there is no way to instruct the model about the information of the new task, the model can only make an educated guess by generating the most plausible text "Best Vocals" according to the training datasets, which fails to serve as a good long-form answer. We therefore argue that it is critical to test a table-to-text model's cross-task generalizability, which is captured in neither the training methods nor the evaluation setup in existing work.
+
+In this paper, we propose the use of compositional task configurations, a set of text prompts prepended to the encoder to improve the cross-task generalizability of unified table-to-text models. For a given task, we design its configuration prompt to be compositional, describing the task type, dataset name, input type, and output type. This design offers at least two key advantages. First, the task configurations explicitly inform the model what is shared between different tasks. For example, the model is able to learn from the configurations that table-based fact checking and table-based QA share the same inputs but different outputs. Second and more importantly, using task configurations allows us to have explicit control over the model's behaviors. For the example in Figure 1, we can now compose a new configuration for long-form table QA at test time to instruct the model to first produce a set of relevant cells and synthesize them to produce a long-form answer, which is within the capabilities of the two training tasks. We discuss this further in the next section.
+
+Our evaluation focuses on model's cross-task generalizability. Specifically, we train our model on 5 table-to-text datasets and test it on an additional set of 5 new datasets that cover either a new domain of an existing task or a new task of which the capabilities can be composed by the ones learned through the 5 training datasets. Our main findings can be summarized as follows:
+
+- Our method not only outperforms the strong UnifiedSKG baseline consistently on the 5 indomain datasets, but also demonstrates much stronger cross-task generalization.
+
+- In zero-shot evaluation on the 5 test-only tasks, our model outperforms UnifiedSKG by a substantial margin of  $+6.5$  and  $+12.6$  average scores from using T5-base and T5-large, respectively. Notably, we find that in zero-shot evaluation on FETAQA (Nan et al., 2022a), a long-form table QA task, while the baseline completely fails with a 0.6 F1 score, our method leads to much better generalization, achieving a 21.2 F1 score.
+
+- We also show that using the compositional task configurations allows the model to output supporting table cells that supplement its final prediction in a zero-shot manner. Human evaluation of the generated supporting cells for the TABFACT dataset reveals that more than  $80\%$  of the generated cells have high relevance to the task.
+
+# 2 Method & Tasks
+
+Prompting is a natural and feasible way to impose explicit control over the behaviors of pre-trained language models (Wei et al., 2021; Chung et al., 2022; Sanh et al., 2022). In this work, we implement the task configurations as prompts of an encoder-decoder model. Each task configuration contains the following four aspects: task type, input type, output type, and dataset name. The task type is the end goal of a task, e.g., QA and summarization, as shown in Figure 2. Input and output types specify the inputs of the encoder and the outputs of the decoder of table-to-text models, respectively. These types can be compositional, for example, both long-form and short-form table QA in Figure 2 require the decoder to output a set of relevant cells and the final answer. The dataset name specifies the dataset used for training. As different datasets can share the same task type, input and output types, we assume that the model is able to learn the shared and the unique knowledge across different datasets by adding the dataset names as configurations. When testing the model on a new dataset, we can simply omit the dataset name since it is not trained.
+
+One of the major advantages of having explicit task configurations is that it enables the model to learn the mapping between a configuration and its behavior. At test time, we can compose a new set of configurations which suits best for an unseen task using the trained configurations. Figure 2 demon
+
+![](images/46e03153c485a601e743b182866970e206f319f76c17473df665722323ef6d12.jpg)  
+Figure 2: Our method allows us to reformulate the unseen task of long-form table QA as a query-based summarization task by composing a new configuration using the existing ones from the two training tasks. At test time, this composed configuration allows the model to first identify relevant cells in the table based on the input question using the cell generation skills learned from short-form table QA, and then generates a long-form answer by utilizing the knowledge acquired from table summarization. For all three tasks, the model also takes the linearized table as input; for simplicity, we omit the linearized table and dataset name configuration in the figure.
+
+strates that by training on short-form table QA, the model learns the ability to generate a set of relevant table cells according to the question and then derive the answer based on those cells. By training on table summarization, the model learns to produce a summary based on a set of table cells.<sup>1</sup> At test time, by reformulating the long-form table QA as a query-based summarization task, our model is able to first generate a set of relevant cells (learned through short-form table QA) and then synthesize those cells to yield a long-form answer (learned from table summarization).
+
+Note that our method is an efficient extension to the original UnifiedSKG model. It only requires a small input prefix, comprising less than  $5\%$  of the total sequence length, making it flexible for generalization to more tasks and datasets.
+
+# 2.1 Datasets and Task Configurations
+
+A detailed list of our datasets with their task information is shown in Table 1. We consider 5 datasets as in-domain datasets for both train-
+
+ing and testing: WIKISQL (Zhong et al., 2017), WIKITQ (Pasupat and Liang, 2015), SQUAD (Rajpurkar et al., 2016), TOTTO (Parikh et al., 2020) and TABFACT (Chen et al., 2019). For SQUAD and TOTTO, since no official test set is released, we follow UnifiedSKG (Xie et al., 2022) and report results on the official development sets. In addition, we consider 5 datasets for test only: NQ-TABLES (Kwiatkowski et al., 2019; Herzig et al., 2021), HYBRIDQA (Chen et al., 2020), TATQA (Zhu et al., 2021), FETAQA (Nan et al., 2022b) and FEVERIOUS (Aly et al., 2021).
+
+The test-only evaluation setup aims to assess the effectiveness of our method in enabling the model to generalize to unseen tasks with new compositional configurations, as well as to a new dataset with existing configurations. Specifically, we test if the model can benefit from a combination of input configurations for tasks such as HYBRIDQA, TAT-QA, and FEVROUS, which involve both passages and tables as inputs, despite the model is only trained on one or the other during training. Similarly, we examine if the configuration for FETAQA, a combination of WIKISQL and ToTTO as shown in Figure 2, allows for explicit control over the model's behaviors, resulting in improved generalization. Finally, we assess the
+
+<table><tr><td></td><td>Dataset</td><td>Task Type</td><td>Input</td><td>Output</td><td>Unseen?</td></tr><tr><td rowspan="5">Train + Test</td><td>WIKISQL</td><td>QA (table)</td><td>query + table</td><td>cells + short-form answer</td><td>-</td></tr><tr><td>WIKITQ</td><td>QA (table)</td><td>query + table</td><td>short-form answer</td><td>-</td></tr><tr><td>SQUAD</td><td>QA (text)</td><td>query + passage</td><td>short-form answer</td><td>-</td></tr><tr><td>TOTTO</td><td>Summarization</td><td>table</td><td>cells + summary</td><td>-</td></tr><tr><td>TABFACT</td><td>Fact-check</td><td>query + table</td><td>binary answer</td><td>-</td></tr><tr><td rowspan="5">Test-only</td><td>NQ-TABLES</td><td>QA (table)</td><td>query + table</td><td>short-form answer</td><td>-</td></tr><tr><td>HYBRIDQA</td><td>QA (hybrid)</td><td>query + passage + table</td><td>short-form answer</td><td>✓</td></tr><tr><td>TAT-QA</td><td>QA (hybrid)</td><td>query + passage + table</td><td>short-form answer</td><td>✓</td></tr><tr><td>FETAQA</td><td>QA (abstractive)</td><td>query + table</td><td>cells + long-form answer</td><td>✓</td></tr><tr><td>FEVEROUS</td><td>Fact-check</td><td>query + passage + table</td><td>binary answer</td><td>✓</td></tr></table>
+
+Table 1: Datasets and tasks considered in our experiments. Tasks that have an input/output combination unseen at training time are marked with  $\sqrt{}$  in the "Unseen?" column. We include detailed statistics and task configurations applied for each dataset in Appendix A & B due to space limitation.
+
+model's ability to generalize to a new dataset, NQ-TABLES, which has the same configurations as WIKISQL and WIKITQ.
+
+For all of these datasets, we linearize the tables following the strategy used in UnifiedSKG (Xie et al., 2022). By inserting several special tokens like vertical bars to indicate the boundaries between cells and rows, a table can be linearized as: "Headers:  $h_1|...|h_m$ , Row 1:  $c_{11}|...|c_{1m}$  ... Row n:  $c_{n1}|...|c_{nm}$ ". Here,  $h_i$  denotes the  $i$ th header of a table and  $c_{ij}$  denotes the cell content in the  $i$ th row and the  $j$ th column. For simplicity, we fix the order of the task configurations to be task type, dataset name, input type, and output type. We preprocess the task configuration to the original input of a dataset and feed it to the model. To make our input and output better aligned with the configurations, we also introduce some special markings to separate different parts of inputs and outputs: Figure 3 illustrates the actual model's input and output of the short-form table QA example from Figure 2. See Appendix B for inputs and outputs constructed for all of the datasets and Appendix A for preprocessing details of each dataset.
+
+# 3 Experiments
+
+Experimental settings We evaluate our method by following the experimental setup shown in Table 1. We follow the experimental settings of UnifiedSKG (Xie et al., 2022) and use T5 (Raffel et al., 2020) as the backbone of our table-to-text model. Our implementation is based on the publicly released code of UnifiedSKG which is developed based on the transformer library (Wolf et al., 2019). To balance the size of different datasets during training, we use the temperature upsampling method proposed in the original T5 paper
+
+![](images/510d7d04054de029f5dbfa7f026a6a16dbb6089d7a1f9a0a4956de5280893fc4.jpg)
+
+![](images/dc3028c5a9a024d2035fb285f2eec2864f626c8d06fdb6ca16abdb80adc78a99.jpg)  
+Figure 3: An example of the input and output given to the model. After training, the model is able to establish a correspondence between the task configurations and the input/output format. Dataset name is omitted here for simplicity.
+
+and set the temperature to 2. For all experiments, we use a batch size of 128 and AdamW (Loshchilov and Hutter, 2018) as the optimizer with the initial learning rate set to 5e-5. We limit the maximum length of the input, including task configuration and the actual inputs, to be 1080 sentence-piece tokens. We train both the T5-base and T5-large models on the training set for 20 epochs and we use early stopping with the patience set to 2. We use deepspeed (Rasley et al., 2020) to reduce the GPU memory loads when training the T5-large model. The approximate GPU hours for T5-base and T5-large are 250 and 650 respectively on A100 GPUs with 40G memory.
+
+Baseline We mainly compare our method against UnifiedSKG (Xie et al., 2022), a strong baseline that was shown to achieve state-of-the-art results on many table-to-text tasks via multi-task training. In UnifiedSKG, for each task, a dataset name is prepended to the encoder during multi-task fine
+
+<table><tr><td colspan="8">Zero-shot Test-only Tasks</td></tr><tr><td></td><td>Models</td><td>NQ-TABLES BLEU</td><td>FETAQAEM</td><td>HYBRIDQAEM</td><td>TAT-QAEM</td><td>FEVEROUSAcc.</td><td>Avg.-</td></tr><tr><td rowspan="3">T5-base</td><td>Single Task</td><td>51.6</td><td>29.9</td><td>54.3</td><td>34.5</td><td>81.3</td><td>50.3</td></tr><tr><td>UnifiedSKG</td><td>37.8</td><td>0.6</td><td>22.5</td><td>18.2</td><td>67.5</td><td>29.3</td></tr><tr><td>Task Configs (Ours)</td><td>39.4</td><td>21.0</td><td>28.9</td><td>20.8</td><td>68.9</td><td>35.8</td></tr><tr><td rowspan="3">T5-large</td><td>Single Task</td><td>52.2</td><td>33.0</td><td>56.6</td><td>36.2</td><td>82.1</td><td>52.0</td></tr><tr><td>UnifiedSKG</td><td>42.6</td><td>0.7</td><td>34.1</td><td>20.4</td><td>41.4</td><td>27.8</td></tr><tr><td>Task Configs (Ours)</td><td>43.0</td><td>25.2</td><td>38.0</td><td>20.8</td><td>75.0</td><td>40.4</td></tr></table>
+
+<table><tr><td colspan="8">In-domain Tasks</td></tr><tr><td></td><td>Models</td><td>WIKISQL EM</td><td>WIKITQ EM</td><td>ToTTO BLEU (dev.)</td><td>SQUAD EM (dev.)</td><td>TABFACT Acc.</td><td>Avg. -</td></tr><tr><td rowspan="3">T5-base</td><td>Single Task</td><td>81.6</td><td>35.8</td><td>36.7</td><td>83.6</td><td>76.1</td><td>62.8</td></tr><tr><td>UnifiedSKG</td><td>82.9</td><td>41.1</td><td>37.2</td><td>82.5</td><td>77.1</td><td>64.2</td></tr><tr><td>Task Configs (Ours)</td><td>83.5</td><td>42.5</td><td>37.4</td><td>83.0</td><td>77.5</td><td>64.8</td></tr><tr><td rowspan="3">T5-large</td><td>Single Task</td><td>85.5</td><td>43.4</td><td>37.8</td><td>86.0</td><td>81.0</td><td>66.7</td></tr><tr><td>UnifiedSKG</td><td>86.0</td><td>48.5</td><td>38.7</td><td>86.1</td><td>83.0</td><td>68.5</td></tr><tr><td>Task Configs (Ours)</td><td>86.7</td><td>50.0</td><td>38.7</td><td>86.2</td><td>83.3</td><td>69.0</td></tr></table>
+
+Table 2: Zero-shot and in-domain performance of our proposed method (Task Configs) vs. baselines for both T5-base and T5-large. Here, "EM" denotes exact match accuracy. For all tasks we also include the results from single-task finetuning as references. Note that a direct comparison with single-task result is not fair as the latter has access to training data in each task. Higher numbers among our method and UnifiedSKG are highlighted in bold.
+
+tuning as a pseudo-task configuration. For fair comparisons, we re-trained UnifiedSKG models on the five in-domain datasets by using the authors' implementation. $^2$
+
+Evaluation For the in-domain tasks, we simply train on their training sets and evaluate on their test sets. For the test-only tasks, we evaluate our method in two settings: 1) a zero-shot setting, where we directly apply the model trained on indomain datasets and use a new set of task config design for each test dataset; 2) a few-shot setting, where for each test dataset, we further fine-tune the model using  $n$  randomly sampled training examples (where  $n$  is small). Since we observed that the few-shot training is unstable and heavily depends on the sampled examples, we report average performance from 5 different random seeds (each with a different set of few-shot examples).
+
+# 4 Results
+
+# 4.1 Main Results
+
+We present the in-domain and zero-shot evaluation results for all datasets in Table 2 and the few-shot evaluation results for OOD datasets in Figure 4. We have the following observations:
+
+First, using compositional task config shows much stronger performance on zero-shot datasets unseen at training time (Table 2). For example, the UnifiedSKG baseline fails to generalize at test time to FETAQA, a long-form table QA task where the input is a question-table pair and the output is a long-form abstractive answer. This is due to the baseline model having no clue of what format of output should be produced and what knowledge learned through the training datasets should be leveraged for this task. In contrast, by reformulating the long-form table QA as a query-based summarization task and composing the input configurations to be table and query as well as the output task config to be relevant cell and summary, our method notably improves the zero-shot performance and closes the gap between zero-shot and single-task finetuning results. Note that among the zero-shot datasets, NQ-TABLES represents a new dataset for an existing task (short-form table QA), whereas others represent new tasks unseen at train-
+
+![](images/5a5e8131a3b7bd5a6317e333a5a60d6f55b20885ddd9df2aae5ebfcb664716ed.jpg)  
+Figure 4: Few-shot performance on out-of-domain tasks, with results aggregated from 5 random training runs. Our method demonstrates better few-shot performance than the baseline in most settings, with the gap reduced as more supervised examples are added.
+
+ing. Nevertheless, we found the improvements to be consistent for all zero-shot datasets, with average improvements of  $+6.5$  and  $+12.6$  for base and large models, respectively.
+
+Second, in most cases, using compositional task config consistently improves the in-domain performance over the UnifiedSKG baseline and single-task training (Table 2). The observation is consistent for both base and large model sizes, with average improvements of  $+0.6$  and  $+0.5$  over UnifiedSKG, respectively. The improvement over single-task fine-tuning is even greater for all datasets. One explanation to this improvement is that by adding task config we explicitly encourage the model to learn the shared knowledge between different tasks and datasets.
+
+Last but not the least, in few-shot evaluation (as depicted in Figure 4), we find that using task configurations has improved few-shot learning performance for most test-time tasks. Overall the difference between our method and the UnifiedSKG baseline is particularly notable when the number of supervised examples  $(n)$  is small; and the performance gap diminishes for HYBRIDQA, TAT-QA, and FEVERIOUS as  $n$  gets larger. One possible explanation is that the prior captured by the task configurations during training is not closely aligned with these three datasets, when  $n$  getting larger, the prior introduced by the task configurations is gradually overridden by knowledge learned from the supervised data.
+
+# 4.2 Ablation of Task Configs at Training time
+
+The impact of individual configurations on model performance was evaluated by removing one configuration at a time during training. The results, presented in Table 3, indicate that the removal of the output type resulted in the largest performance drop, as the model was only able to guess the de
+
+<table><tr><td>Models</td><td>FETAQA</td><td>NQ</td><td>HYBRID</td><td>TAT</td><td>FEVR</td><td>Avg.</td></tr><tr><td>Full config</td><td>21.0</td><td>39.4</td><td>28.9</td><td>20.8</td><td>68.9</td><td>35.8</td></tr><tr><td>- dataset</td><td>21.2</td><td>36.1</td><td>25.3</td><td>19.8</td><td>68.1</td><td>34.1</td></tr><tr><td>- task type</td><td>0.4</td><td>38.4</td><td>28.3</td><td>21.3</td><td>68.5</td><td>31.4</td></tr><tr><td>- input</td><td>20.3</td><td>39.5</td><td>29.1</td><td>19.4</td><td>68.3</td><td>35.3</td></tr><tr><td>- output</td><td>17.3</td><td>34.8</td><td>17.9</td><td>16.1</td><td>68.2</td><td>30.9</td></tr></table>
+
+Table 3: Ablation of task configurations during training. We only report zero-shot performance here. We see removing any one of the configurations causes a performance drop on average.
+
+sired output type based on learned parameters. The removal of the input type had the least impact on performance. This is likely due to the fact that learning the representation of the two input types was not difficult for the model, and explicitly informing the model about the input type does not provide significant benefit, as observed in the previous section. The removal of the dataset name also results in a performance drop, particularly on the NQ-TABLES dataset, indicating that even when the task type, input, and output are the same, including the dataset name helps the model learn dataset-independent knowledge more effectively. The removal of the task type results in a complete failure on the FETAQA dataset, demonstrating that all configurations are needed to produce the correct form of output. A more detailed discussion of these findings can be found in Section 7.
+
+# 4.3 Ablation of Task Configs at Test time
+
+While our method demonstrates much stronger zero-shot task performance, it is crucial to understand the extent to which input and output configurations contribute to this success, particularly for tasks involving hybrid input or output types that are not present during training. To examine the contributions of input configurations, we remove each configuration from the hybrid tasks (HYBRIDQA,
+
+TAT-QA and FEVEROUS) at test time, with results shown in Table 4. We found that deleting either of the input configurations results in a performance drop in most cases, and the drop is quite notable when the table and passage input configurations are removed together. This suggests that the input configuration captures useful priors about the input during training, and different configurations can be combined to yield better performance in the zero-shot transfer to hybrid tasks. We also observe a similar trend in Figure 5 where we test the model performance by removing the cell output configurations for FETAQA (thereby skipping cell generation). We see that in both zero-shot and few-shot settings, model performance drops by a large margin. This shows not only that the model can generate different outputs by combining the output configurations, but also that it can better utilize the prior captured by the configurations to improve task performance.
+
+# 4.4 Human Evaluation of Generated Cells
+
+In addition to the strong task generalizability, a key advantage of applying the proposed task configurations to table-to-text tasks is that we can modify the task configurations to output more results for improved explainability, even when such a configuration combination is never seen at training time. An example of this is for the table-based fact verification task, TABFACT, instead of only generating a binary label, we can extend the output configuration to include a cell component that can serve as supporting evidence of the binary prediction. We include two examples of this setting in Figure 6.
+
+To understand how well our model can generate supporting cells without ever being trained for it, we conduct a human evaluation over 50 randomly sampled outputs from the TABFACT dataset. We ask annotators to manually evaluate the generated cells based on their level of relevance and completeness. Relevance denotes the usefulness of the generated cells in verifying a claim (precision) and completeness refers to the extent to which all of the relevant cells are generated (recall). Detailed annotation instructions are shown in Appendix C. For each aspect, we ask the annotators to select between three labels that characterize its degree: "full", "partial" or "none". Three of the authors conduct the annotations, achieving 0.72 and 0.80 Fleiss Kappa (Fleiss, 1971) for relevance and completeness, respectively. We conduct majority vote to get
+
+<table><tr><td>Models</td><td>HYBRIDQA</td><td>TAT-QA</td><td>FEVR</td><td>Avg.</td></tr><tr><td>Full Configs</td><td>28.9</td><td>20.8</td><td>68.9</td><td>39.5</td></tr><tr><td>- input:passage</td><td>28.8</td><td>19.8</td><td>68.6</td><td>39.1</td></tr><tr><td>- input:table</td><td>29.4</td><td>20.3</td><td>66.7</td><td>38.8</td></tr><tr><td>- input:all</td><td>27.3</td><td>19.3</td><td>66.1</td><td>37.6</td></tr></table>
+
+Table 4: Ablation of the input configurations at test time. The inputs of the three datasets include both table and passage. We show the model can achieve better performance by combining two input configurations.
+
+![](images/c99da7233e674f217837f2643a09811c25613e2032f64ce67ffbf15ef3a50846.jpg)  
+Figure 5: Ablation of the output configurations on FETQA. The model has better cross-task performance by generating a set of relevant cells in both zero-shot and few-shot settings. "n=0" denotes zero-shot setting.
+
+the consensus label and the results are shown in Table 5. Overall we found that the model is able to generate cells with high relevance (with  $72\%$  examples being fully relevant generations), but struggle with full completeness (with  $34\%$  fully complete).
+
+# 5 Related Work
+
+Table-to-text tasks Table-based tasks, including table-based question answering (Pasupat and Liang, 2015; Zhong et al., 2017; Chen et al., 2020; Cheng et al., 2022; Zhao et al., 2022), table-based fact-checking (Chen et al., 2019; Aly et al., 2021), table summarization (Parikh et al., 2020; Suadaa et al., 2021; Moosavi et al., 2021), have gained increasing attention in recent years. A flurry of work using transformer-based structure explored modeling table structure via pretraining, for example, TabTransformer (Huang et al., 2020), VIME (Yoon et al., 2020), TABBIE (Iida et al., 2021), TaBERT (Yin et al., 2020), TUTA (Wang et al., 2021), TabT5 (Andrejczuk et al., 2022), and TableFormer (Yang et al., 2022).
+
+Our work mainly focuses on table-to-text tasks but the ideal neural architecture for encoding table structures is not our focus. Instead, we emphasize multi-task knowledge-sharing (similar to UnifiedSKG (Xie et al., 2022)) and cross-task gen
+
+<table><tr><td colspan="5">Caption: 1976 world junior figure skating championships</td></tr><tr><td>rank</td><td>name</td><td>nation</td><td>points</td><td>places</td></tr><tr><td>1</td><td>sherri baier / robin cowan</td><td>canada</td><td>128.39</td><td>9</td></tr><tr><td>2</td><td>lorene mitchell / donald mitchell</td><td>united states</td><td>124.94</td><td>16</td></tr><tr><td>3</td><td>elizabeth cain / peter cain</td><td>australia</td><td>116.67</td><td>33</td></tr><tr><td>4</td><td>jana bláhová / ludek</td><td>czechoslovakia</td><td>113.74</td><td>36</td></tr></table>
+
+# Task config:
+
+[Task: Fact-checking] [Input: table] [Output: cells] [Output: binary answer]
+
+# Input claim:
+
+the top 2 ranked team be in north America
+
+# Output:
+
+[cell] 1 | sherri baier / robin cowan | canada | 2 | lorene mitchell / donald mitchell | united states [/cell] [answer] entailed [/answer]
+
+<table><tr><td colspan="6">Caption: Charl Schwartzel</td></tr><tr><td>tournament</td><td>wins</td><td>top - 5</td><td rowspan="6">…</td><td>top - 25</td><td>events</td></tr><tr><td>masters tournament</td><td>1</td><td>1</td><td>2</td><td>4</td></tr><tr><td>us open</td><td>0</td><td>0</td><td>3</td><td>7</td></tr><tr><td>the open championship</td><td>0</td><td>0</td><td>4</td><td>9</td></tr><tr><td>pga championship</td><td>0</td><td>0</td><td>2</td><td>8</td></tr><tr><td>totals</td><td>1</td><td>1</td><td>11</td><td>28</td></tr></table>
+
+# Task config:
+
+[Task: Fact-checking] [Input: table] [Output: cells] [Output: binary answer]
+
+# Input claim:
+
+there be a total of 28 event that make up the 4 tournament
+
+# Output:
+
+[cell] 4 [/cell] [answer] refuted [/answer]
+
+Figure 6: Two examples from the development set of TABFACT where we force the model to produce relevant cells by changing the output configurations. For the left example, the model produces a set of complete and relevant cells that help understand its "entailed" prediction. For the right example, the model is misled by the number "4" when generating the cells, and we are able to check that the final answer "refuted" is wrong.
+
+<table><tr><td></td><td>Full</td><td>Partial</td><td>None</td></tr><tr><td>Relevance</td><td>72%</td><td>10%</td><td>18%</td></tr><tr><td>Completeness</td><td>34%</td><td>48%</td><td>18%</td></tr></table>
+
+Table 5: Human evaluation results of the zero-shot cell generation quality for the TABFACT task.
+
+eralization in table-to-text tasks. Also, the proposed framework is capable of generalizing to a broader range of tasks and datasets.
+
+Task unification There have been a vein of work that tries to solve various NLP tasks using a single model. This includes encoder-decoder models like T5 (Raffel et al., 2020), UnifiedQA (Khashabi et al., 2020), UnifiedQA2 (Khashabi et al., 2022), UnifiedSKG (Xie et al., 2022); decoder-only models driven by prompts, for example, GPT3 (Brown et al., 2020), Codex (Chen et al., 2021), PaLM (Chowdhery et al., 2022). Our work extends UnifiedSKG by using an encoder-decoder model as the backbone and designing prompts to encourage better knowledge sharing between different tasks and enable control over the model's behaviors.
+
+Cross-task generalization with pretrained models Various efforts have been made to improve the ability of unified models to generalize to new tasks and datasets, including instruction-tuning using a wide range of natural language instructions (Chung et al., 2022; Sanh et al., 2022; Wei et al., 2021;
+
+Zhong et al., 2021), better design of prompts in zero-shot and few-shot setting (Wei et al., 2022; Zhou et al., 2022; Kojima et al., 2022). Our proposed method differs from instruction-tuning models like FLAN (Wei et al., 2021) in that we use a more symbolized prompt structure and it is possible to attribute cross-task generalizability to specific tasks and configurations. Also, instruction-tuning models like FLAN achieve behavior control and cross-task generalizability through costly large-scale instruction tuning. In contrast, our approach demonstrates that within a specific task domain with limited datasets like table-to-text, this can be achieved by utilizing a compositional prompt structure.
+
+Our method is also relevant to Macaw (Tafjord and Clark, 2021), ProQA (Zhong et al., 2022b), and SchemaPro (Zhong et al., 2022a), which also utilize explicit task descriptions to facilitate knowledge sharing between various NLP tasks. Our work differs in two main aspects: (1) Our work focuses on compositional generalization at test time, examining whether the model can combine different configurations from multiple tasks during training to generalize to unseen tasks at test time. (2) Our work focuses on table-to-text tasks.
+
+# 6 Conclusion
+
+We introduced compositional task configurations for unified table-to-text models. Compared to existing unified encoder-decoder models that sim
+
+ply use dataset names as input prefix, compositional task configurations allow us to specify the task type, input, and output types at a finer level, which improve multi-task learning effectiveness and cross-task generalization. Further, we showed that our method allows fine-grained control over the model's generation at test time, enabling the model to generalize to unseen tasks and improving explainability via generating high-quality supporting table cells.
+
+# 7 Limitations
+
+Task configurations are entangled with the full model parameters. In our ablation study of task configurations at training time (Table 3), we see that when training without task type, the model fails to generalize to FETAQA. Upon examining the model output, we find that although we change the output configuration to "long answer", the model still produces a short-form answer. This indicates that model behaviors are not always aligned with a single configuration, leading us to question the extent to which each individual configuration influences the model. In order to have better and more interpretable control over the models, one potential avenue for future research is to develop pluggable task configurations, where each configuration controls a more atomic function of the model and can be plugged, unplugged, and combined to yield different model behaviors.
+
+Our exploration scope is limited to table-to-text tasks. Due to the constraints of the computational resources, we haven't explored joint training with a broader range of other NLP tasks. We think with some modifications, such as the inclusion of dataset domains in the configuration set, it would be possible to extend our approach to additional datasets and tasks.
+
+# 8 Ethics Statement
+
+The authors of this paper are committed to conducting research ethically. Data used in this work has been collected from public sources and used in accordance with all applicable laws and regulations. The only area of work that involves human annotation of data is described in Section 4.4, where authors of this paper annotated a group of samples for analyzing models' behaviors. We ensure that no external human subject was involved or harmed. In addition, this work uses language models, for
+
+which the risks and potential harms are discussed in numerous previous works (Bender et al., 2021; Weidinger et al., 2021). The authors strive to ensure that the research and its results do not cause harm.
+
+# 9 Acknowledgement
+
+We would like to thanks the anonymous reviewers for their insightful comments and suggestions.
+
+# References
+
+Rami Aly, Zhijiang Guo, Michael Sejr Schlichtkrull, James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Oana Cocarascu, and Arpit Mittal. 2021. The fact extraction and VERIFICATION over unstructured and structured information (FEVEROUS) shared task. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 1-13, Dominican Republic. Association for Computational Linguistics.  
+Ewa Andrejczuk, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, and Yasemin Altun. 2022. Table-to-text generation and pre-training with tabt5. arXiv preprint arXiv:2210.09162.  
+Emily M Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.  
+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.  
+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.  
+Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. 2019. Tabfact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations.
+
+Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Yang Wang. 2020. HybridQA: A dataset of multi-hop question answering over tabular and textual data. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1026-1036, Online. Association for Computational Linguistics.  
+Zhoujun Cheng, Haoyu Dong, Zhiruo Wang, Ran Jia, Jiaqi Guo, Yan Gao, Shi Han, Jian-Guang Lou, and Dongmei Zhang. 2022. HiTab: A hierarchical table dataset for question answering and natural language generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1094-1110, Dublin, Ireland. Association for Computational Linguistics.  
+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.  
+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.  
+Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378.  
+Jonathan Herzig, Thomas Mueller, Syrine Krichene, and Julian Eisenschlos. 2021. Open domain question answering over tables via dense retrieval. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.  
+Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno, and Julian Martin Eisenschlos. 2020. TAPAS: Weakly supervised table parsing via pre-training. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Xin Huang, Ashish Khetan, Milan Cvitkovic, and Zohar Karnin. 2020. Tabtransformer: Tabular data modeling using contextual embeddings. arXiv preprint arXiv:2012.06678.  
+Hiroshi Iida, Dung Thai, Varun Manjunatha, and Mohit Iyyer. 2021. TABBIE: Pretrained representations of tabular data. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3446-3456, Online. Association for Computational Linguistics.  
+Daniel Khashabi, Yeganeh Kordi, and Hannaneh Hajishirzi. 2022. Unifiedqa-v2: Stronger generalization via broader cross-format training. arXiv preprint arXiv:2202.12359.
+
+Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UNIFIEDQA: Crossing format boundaries with a single QA system. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1896-1907, Online. Association for Computational Linguistics.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466.  
+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.  
+Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, and Jian-Guang Lou. 2021. Tapex: Table pre-training via learning a neural sql executor. In International Conference on Learning Representations.  
+Ilya Loshchilov and Frank Hutter. 2018. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Nafise Sadat Moosavi, Andreas Rücklé, Dan Roth, and Iryna Gurevych. 2021. Learning to reason for text generation from scientific tables. arXiv preprint arXiv:2104.08296.  
+Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Lin, Neha Verma, Rui Zhang, Wojciech Krysciński, Nick Schoelkopf, Riley Kong, Xiangru Tang, et al. 2022a. Fetaqa: Free-form table question answering. Transactions of the Association for Computational Linguistics, 10:35-49.  
+Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Krysciński, Hailey Schoelkopf, Riley Kong, Xiangru Tang, Mutethia Mutuma, Ben Rosand, Isabel Trindade, Renusree Bandaru, Jacob Cunningham, Caiming Xiong, Dragomir Radev, and Dragomir Radev. 2022b. FeTaQA: Free-form table question answering. Transactions of the Association for Computational Linguistics, 10:35-49.
+
+Ankur Parikh, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipanjan Das. 2020. ToTTo: A controlled table-to-text generation dataset. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1173-1186, Online. Association for Computational Linguistics.  
+Panupong Pasupat and Percy Liang. 2015. Compositional semantic parsing on semi-structured tables. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1470–1480, Beijing, China. Association for Computational Linguistics.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.  
+Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 3505-3506.  
+Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2022. Multitask prompted training enables zero-shot task generalization. In The Tenth International Conference on Learning Representations.  
+Peng Shi, Patrick Ng, Zhiguo Wang, Henghui Zhu, Alexander Hanbo Li, Jun Wang, Cicero Nogueira dos Santos, and Bing Xiang. 2021. Learning contextual representations for semantic parsing with generation-augmented pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13806-13814.  
+Lya Hulliyyatus Suadaa, Hidetaka Kamigaito, Kotaro Funakoshi, Manabu Okumura, and Hiroya Takamura. 2021. Towards table-to-text generation with numerical reasoning. 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 1451-1465, Online. Association for Computational Linguistics.  
+Oyvind Tafjord and Peter Clark. 2021. General-purpose question-answering with macaw. arXiv preprint arXiv:2109.02593.
+
+Zhiruo Wang, Haoyu Dong, Ran Jia, Jia Li, Zhiyi Fu, Shi Han, and Dongmei Zhang. 2021. Tuta: tree-based transformers for generally structured table pretraining. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 1780-1790.  
+Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. In International Conference on Learning Representations.  
+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.  
+Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, et al. 2021. Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359.  
+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.  
+Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I Wang, et al. 2022. Unifiedsgk: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. arXiv preprint arXiv:2201.05966.  
+Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics.  
+Jingfeng Yang, Aditya Gupta, Shyam Upadhyay, Luheng He, Rahul Goel, and Shachi Paul. 2022. TableFormer: Robust transformer modeling for table-text encoding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 528-537, Dublin, Ireland. Association for Computational Linguistics.  
+Pengcheng Yin, Graham Neubig, Wen-tau Yih, and Sebastian Riedel. 2020. TaBERT: Pretraining for joint understanding of textual and tabular data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8413-8426, Online. Association for Computational Linguistics.  
+Jinsung Yoon, Yao Zhang, James Jordon, and Mihaela van der Schaar. 2020. Vime: Extending the success
+
+of self-and semi-supervised learning to tabular domain. Advances in Neural Information Processing Systems, 33:11033-11043.
+
+Yilun Zhao, Yunxiang Li, Chenying Li, and Rui Zhang. 2022. MultiHiertt: Numerical reasoning over multi hierarchical tabular and textual data. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588-6600, Dublin, Ireland. Association for Computational Linguistics.
+
+Ruiqi Zhong, Kristy Lee, Zheng Zhang, and Dan Klein. 2021. Adapting language models for zero-shot learning by meta-tuning on dataset and prompt collections. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2856-2878, Punta Cana, Dominican Republic. Association for Computational Linguistics.
+
+Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2sql: Generating structured queries from natural language using reinforcement learning. CoRR, abs/1709.00103.
+
+Wanjun Zhong, Yifan Gao, Ning Ding, Zhiyuan Liu, Ming Zhou, Jiahai Wang, Jian Yin, and Nan Duan. 2022a. Improving task generalization via unified schema prompt. arXiv preprint arXiv:2208.03229.
+
+Wanjun Zhong, Yifan Gao, Ning Ding, Yujia Qin, Zhiyuan Liu, Ming Zhou, Jiahai Wang, Jian Yin, and Nan Duan. 2022b. ProQA: Structural prompt-based pre-training for unified question answering. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4230-4243, Seattle, United States. Association for Computational Linguistics.
+
+Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Olivier Bousquet, Quoc Le, and Ed Chi. 2022. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625.
+
+Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and TatSeng Chua. 2021. TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance. 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 3277-3287, Online. Association for Computational Linguistics.
+
+# A Dataset Statistics and Preprocessing
+
+The statistics and the licenses of the datasets we used in this paper are shown in Table 6. All datasets are English-based and all of them are based on Wikipedia except for TAT-QA which is based on
+
+financial reports. To fit the data into the encoder, for all datasets, we limit the max length of each cell to be 15 sentence-piece tokens. If the length (measured by sentence-piece tokens) of the linearized table is longer than 1024, we truncate random rows to reduce the table size.
+
+The original annotation of WIKISQL dataset does not include the relevant cells. We extract the relevant cells by executing the accompanied SQL query annotations. In most cases, the relevant cells equal to the final answer annotations; for the rest of the cases, aggregations or numerical operations need to be run to obtain the final answer. During training, we also create another version of the WIKISQL dataset, in which we exclude the relevant cells and only use the final answer as supervision to improve output diversity. We use both versions at training time.
+
+NQ-TABLES is a table-based QA dataset derived from the NaturalQuestions dataset (Kwiatkowski et al., 2019) and was originally released by Herzig et al. (2021). The original test set of NQ-TABLES contains 966 unique examples. In our experiments, to make the dataset more compatible with other table-based QA tasks, we evaluate on a customized version of NQ-TABLES where we only include an example if the answer is uniquely locatable as one or more table cells. This filtering step results in 549 unique triples of table, question and answers.
+
+To make ToTTO more compatible with other table-to-text tasks, we feed the selected cells as inputs to the decoder, as we mentioned in Section 2. Also, we find it helpful to create a reversed version of the ToTTO dataset, where we treat the annotated summary and the table as input and let the model predict the relevant cells. We add both versions of ToTTO to the training of all models, including the baseline.
+
+# B Task Configurations Applied for Each Dataset
+
+Below we list the task configurations applied to all datasets. For each dataset, we present input to the encoder and output from the decoder separately. For encoder, we include the template of the full input, including task configurations as well as how the dataset input is structured (with actual data replaced by "..."). For decoder, we include how we structure the annotated output during training and how we parse the output during testing.
+
+<table><tr><td></td><td>Dataset</td><td>Train</td><td>Dev</td><td>Test</td><td>License</td><td>Domain</td><td>Language</td></tr><tr><td rowspan="5">Train+Test</td><td>WIKISQL</td><td>56,355</td><td>8,421</td><td>15,878</td><td>BSD 3-Clause</td><td>Wikipedia</td><td>English</td></tr><tr><td>WIKITQ</td><td>11,321</td><td>2,810</td><td>4,344</td><td>CC BY-SA 4.0</td><td>Wikipedia</td><td>English</td></tr><tr><td>SQUAD</td><td>87,599</td><td>10,570</td><td>-</td><td>CC BY-SA 4.0</td><td>Wikipedia</td><td>English</td></tr><tr><td>TOTTO</td><td>120,761</td><td>7,700</td><td>-</td><td>CC BY-SA 3.0</td><td>Wikipedia</td><td>English</td></tr><tr><td>TABFACT</td><td>92,283</td><td>12,792</td><td>9,750</td><td>CC BY 4.0</td><td>Wikipedia</td><td>English</td></tr><tr><td rowspan="5">Test-only</td><td>NQ-TABLES</td><td>-</td><td>-</td><td>549</td><td>Apache License 2.0</td><td>Wikipedia</td><td>English</td></tr><tr><td>HYBRIDQA</td><td>-</td><td>3,466</td><td>-</td><td>CC BY 4.0</td><td>Wikipedia</td><td>English</td></tr><tr><td>TAT-QA</td><td>-</td><td>-</td><td>1,669</td><td>MIT License</td><td>Finance</td><td>English</td></tr><tr><td>FETAQA</td><td>-</td><td>-</td><td>2,003</td><td>CC BY-SA 4.0</td><td>Wikipedia</td><td>English</td></tr><tr><td>FEVEROUS</td><td>-</td><td>7,890</td><td>-</td><td>CC BY-SA 3.0</td><td>Wikipedia</td><td>English</td></tr></table>
+
+Table 6: Dataset statistics of the datasets we used in the paper. Note that for the test-only datasets, except for few-shot experiments, only the test splits of the original datasets are used. For SQUAD, HYBRIDQA, FEVERIOUS, and ToTTo, as no public test set is offered, we evaluate the model on the original development sets following UnifiedSKG (Xie et al., 2022). For NQ-TABLES, we use a modified version of it in our experiments as described in Appendix A.
+
+# B.1 WikiSQL
+
+# Encoder:
+
+[Task: QA] [Dataset: WikiSQL] [Input: query] [Input: table] [Output: cells] [Output: short answer] [query] ... [/query] [table] ... [/table]
+
+# Decoder:
+
+[cell] ... [/cell] [answer] ... [/answer]
+
+# B.2 WikiTQ
+
+# Encoder:
+
+[Task: QA] [Dataset: WikiTQ] [Input: query] [Input: table] [Output: short answer] [query] ... [/query] [table] ... [/table]
+
+# Decoder:
+
+[answer] ... [/answer]
+
+# B.3 SQuAD
+
+# Encoder:
+
+[Task: QA] [Dataset: SQuAD] [Input: query] [Input: passage] [Output: short answer] [query] ... [/query] [passage] ... [/passage]
+
+# Decoder:
+
+[answer] ... [/answer]
+
+# B.4 ToTTo
+
+# Encoder:
+
+[Task: Summarization] [Dataset: ToTTo] [Output: cells] [Output: long answer]
+
+# Decoder:
+
+[cell] ... [/cell] [answer] ... [/answer]
+
+# B.5 TabFact
+
+# Encoder:
+
+[Task: Fact-checking] [Dataset: TabFact]  
+[Input: query] [Input: table] [Output: binary answer] [query] ... [/query] [table] ... [/table]
+
+# Decoder:
+
+[answer] ... [/answer]
+
+# B.6 NQ-Tables
+
+# Encoder:
+
+[Task: QA] [Input: query] [Input: table]  
+[Output: short answer] [query] ... [/ query] [table] ... [/table]
+
+# Decoder:
+
+[answer] ... [/answer]
+
+# B.7 HybridQA
+
+# Encoder:
+
+[Task: QA] [Input: query] [Input: table] [Input: passage] [Output: short answer] [query] ... [/query] [table] ... [/table] [passage] ... [/passage]
+
+# Decoder:
+
+[answer] ... [/answer]
+
+# B.8 TAT-QA
+
+# Encoder:
+
+[Task: QA] [Input: query] [Input: table] [Input: passage] [Output: short answer] [query] ... [/query] [table] ... [/table] [passage] ... [/passage]
+
+# Decoder:
+
+[answer] ... [/answer]
+
+# B.9 FeTaQA
+
+# Encoder:
+
+```txt
+[Task: Summarization] [Input: query] [Input: table] [Output: cells] [Output: long answer] [query] ... [/query] [table] ... [/table]
+```
+
+# Decoder:
+
+```txt
+[cell] ... [/cell] [answer] ... [/answer]
+```
+
+# B.10 FEVEROUS
+
+# Encoder:
+
+```txt
+[Task: Fact-checking] [Input: query] [Input: table] [Input: passage] [Output: binary answer] [query] ...[/query] [table] ...[/table] [passage] ...[/ passage]
+```
+
+# Decoder:
+
+```ini
+[answer] ... [/answer]
+```
+
+# B.11 WikisQL-Answer-only
+
+# Encoder:
+
+```txt
+[Task: QA] [Dataset: WikiSQL] [Input: query] [Input: table] [Output: short answer] [query] ... [/query] [table] ... [/table]
+```
+
+# Decoder:
+
+```ini
+[answer] ... [/answer]
+```
+
+# B.12 ToTTo-Reverse
+
+# Encoder:
+
+```txt
+[Task: Cell-generation] [Input: query] [Input: table] [Output: cell] [query] [... [/query] [table] ... [/table]
+```
+
+# Decoder:
+
+```json
+[cell] ... [/cell]
+```
+
+# C Annotation Interface
+
+The annotation interface we used for our human study in this paper is shown in Figure 7
+
+# Zero-shot cell generation (TabFact)
+
+# Annotation guideline:
+
+For each claim, the corresponding table, and the generated cells, our main goal is to evaluate the generated cells under the following two aspects:
+
+1. cell relevance: whether the predicted cells are relevant to check the claim (precision). We have the following three labels: (1) relevant: all cells are relevant (precision == 1) (2) partially relevant: some cells are relevant  $(0 < \text{precision} < 1)$  (3) irrelevant: none of the cells are relevant (precision == 0)  
+2. cell completeness: whether the predicted cells contain all information needed. For example, claim mentions two entities but only one is predicted. We have the following three labels: (1) complete: all necessary information is contained by the cells. (2) partially complete: only part of the information is covered by the cells. (3) incomplete: none of the information is covered by the cells.
+
+Example ID: 17
+
+Submit
+
+# Input claim:
+
+the top 2 ranked team be in north america
+
+# Output:
+
+[cell] 1 | sherri baier / robin cowan | canada | 2 | lorene mitchell / donald mitchell | united states [/cell] [answer] entailed [/answer]
+
+<table><tr><td>rank</td><td>name</td><td>nation</td><td>points</td><td>places</td></tr><tr><td>1</td><td>sherri baier / robin cowan</td><td>canada</td><td>128.39</td><td>9</td></tr><tr><td>2</td><td>lorene mitchell / donald mitchell</td><td>united states</td><td>124.94</td><td>16</td></tr><tr><td>3</td><td>elizabeth cain / peter cain</td><td>australia</td><td>116.67</td><td>33</td></tr><tr><td>4</td><td>jana bláhová / ludek</td><td>czechoslovakia</td><td>113.74</td><td>36</td></tr><tr><td>5</td><td>sabine fuchs / xavier vide</td><td>france</td><td>114.12</td><td>39</td></tr><tr><td>6</td><td>karen wood / stephen baker</td><td>united kingdom</td><td>100.33</td><td>55</td></tr><tr><td>7</td><td>catherine brunet / philippe brunet</td><td>france</td><td>94.27</td><td>62</td></tr></table>
+
+# table caption:
+
+1976 world junior figure skating championships
+
+Figure 7: Instruction and the annotation interface we used for the human study in section 4.4.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? Section 7  
+A2. Did you discuss any potential risks of your work? Section 8  
+A3. Do the abstract and introduction summarize the paper's main claims? Last paragraph of Section 1  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Section 2.1
+
+B1. Did you cite the creators of artifacts you used? Section 2.1  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix A  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Appendix A  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We checked the data-collecting procedure of the datasets we used in this paper and found no specific offensive content is included. However, we did not take further action to conduct a more comprehensive review of each dataset, as they are widely used within the community and the large volume of datasets we employed made this impractical.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Appendix A  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix A
+
+# C Did you run computational experiments?
+
+Section 3
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Section 3
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Section 3
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+Section 3
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+Section 3
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Section 4.4
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+Section 4.4 Appendix C
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+Not applicable. Left blank.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+Not applicable. Left blank.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+Not applicable. Left blank.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+Not applicable. Left blank.

2023/Improving Cross-task Generalization of Unified Table-to-text Models with Compositional Task Configurations/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a0844b121e2d0b7642fa434fa697b3e7cccfe3c60af9597084e989b6cff9f02
+size 599414

2023/Improving Diachronic Word Sense Induction with a Nonparametric Bayesian method/650dc53b-8e23-4a7c-9f4e-10f8ef2f245a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a09354483c9774cb39981e2029827b4719ecf6a3cd65ace2ffc84da9ec7c1a86
+size 1293112

2023/Improving Diachronic Word Sense Induction with a Nonparametric Bayesian method/full.md

@@ -0,0 +1,434 @@
+# Improving Diachronic Word Sense Induction with a Nonparametric Bayesian method
+
+Ashjan Alsulaimani
+
+School of Computer Science and Statistics
+
+Trinity College of Dublin
+
+alsulaia@td.ie
+
+Erwan Moreau
+
+School of Computer Science
+
+and Statistics
+
+Trinity College of Dublin
+
+moreaue@tcd.ie
+
+# Abstract
+
+Diachronic Word Sense Induction (DWSI) is the task of inducing the temporal representations of a word meaning from the context, as a set of senses and their prevalence over time. We introduce two new models for DWSI, based on topic modelling techniques: one is based on Hierarchical Dirichlet Processes (HDP), a nonparametric model; the other is based on the Dynamic Embedded Topic Model (DETM), a recent dynamic neural model. We evaluate these models against two state of the art DWSI models, using a time-stamped labelled dataset from the biomedical domain. We demonstrate that the two proposed models perform better than the state of the art. In particular, the HDP-based model drastically outperforms all the other models, including the dynamic neural models.
+
+# 1 Introduction
+
+Word meanings evolve over time. Recent research works have focused on how to model such dynamic behaviour. The unsupervised task of Diachronic Word Sense Induction (DWSI) aims to capture how the meaning of a word varies continuously over time, in particular when new senses appear or old senses disappear. DWSI takes the time dimension into account and assumes that the data spans over a long continuous period of time in order to model the progressive evolution of senses across time.
+
+The dynamic behaviour of words contributes to semantic ambiguity, which is a challenge in many NLP tasks. DWSI can serve as an analytical tool to help building terminology resources and indexing documents more accurately and therefore can be beneficial for information retrieval tasks.
+
+DWSI follows the probabilistic graphical modelling approach to approximate the true meanings from the observed data. Thus, in this paper, we explore the relation of DWSI with topic modelling in general and to the dynamic topic modelling techniques in particular: they both aim to discover a latent variable (sense or topic respectively) from a sequential collection of documents. Despite a close relation between the tasks, topic modelling techniques are not fully explored or compared against in the current state of the art of DWSI.
+
+The state of the art of DWSI consists of only two models: (Emms and Kumar Jayapal, 2016) and (Frermann and Lapata, 2016). They are both designed specifically for DWSI; both are parametric; and both are dynamic, in the sense that they both introduce a time variable into the model in order to capture the evolution of the meaning over time. Emms and Kumar Jayapal (2016) propose a parametric generative model (NEO) where each sense is represented as a  $|V|$ -dimensional multinomial distribution over the vocabulary  $V$ , each document is represented as a mixture of senses, and the dependency of the sense proportions on time is represented as a  $K$ -dimensional multinomial distribution over the  $K$  senses. The parameters of the model have finite Dirichlet priors. A more complex model called SCAN (Frermann and Lapata, 2016) allows each sense distribution over the vocabulary to evolve sequentially from adjacent time slices, as well as the senses proportion. The multinomial parameters of words and senses have logistic normal priors.
+
+The two above-mentioned models are parametric, in the sense that the number of senses (which reflects the structure of the hidden meanings in the data) is a hyper-parameter which has to be known a priori. This is not ideal given the nature of the DWSI task, which is meant to infer senses from the
+
+data. The same issue has been studied for the tasks of topic modelling and WSI; Hierarchical Dirichlet Processes (HDP), a nonparametric hierarchical model introduced by Teh et al. (2006), offer an powerful solution to this problem. HDP extends Latent Dirichlet Allocation (LDA) (Blei et al., 2003) by placing Dirichlet processes priors (DPs) (Ferguson, 1973) on the infinite-dimensional space of multinomial probability distributions. Thus the number of mixture components is infinite a priori and to be inferred from the data. In contrast, LDA posits a predefined number  $K$  of topics, each of which is a multinomial distribution over the vocabulary. Each document has specific topic proportions from a Dirichlet prior, and the topics are shared among the documents. Additionally, the HDP model allows sharing topics not only among documents but also across hierarchical levels by the use of multiple DPs.
+
+The intuition behind our approach relies on the fact that the hierarchical DPs allow "new" senses to appear as needed, thanks to the theoretically infinite number of possible senses. Therefore, the hierarchical design of Dirichlet processes can capture the dynamic behaviour of the words, while inferring the optimal number of clusters directly from the data across time.
+
+Word embeddings are another natural direction of potential improvement for DWSI. Introduced by Rumelhart and Abrahamson (1973); Bengio et al. (2003, 2006), they provide a distributed representation where words with similar meanings are close in a lower-dimensional vector space. Recently, various models have been proposed which integrate word embeddings for topic modelling, however these models do not necessarily represent both words and topics using embeddings. Dieng et al. (2019) provide an elegant solution to this problem: Dynamic Embedded Topic Model (DETM) is a parametric generative model inspired by D-LDA (Dynamic LDA) Blei and Lafferty (2006), in which each word is represented with a word embedding, and per-time topics are represented as embeddings as well. Topics and topic proportions evolve sequentially from adjacent time slices. DETM also directly models per-topic conditional probability of a word as the exponentiated inner product between the word embeddings and per-time topic embeddings. This results in a closer semantic correspondence between words and topics, and thus
+
+leads to better topics quality.
+
+By contrast to previous contributions in DWSI which were mostly theoretical, this paper is an empirical contribution focusing on adapting different existing topic modelling techniques to DWSI. The aim is to set the state of the art DWSI models up against two serious competitors, in order to check whether they actually fit the task of DWSI optimally. In this perspective, we adapt HDP and DETM to the task of DWSI, describing our approach in §3. We test the ability of these models to detect meaning change over time using the evaluation framework proposed by (Alsulaimani et al., 2020), described in §4: using a large corpus of biomedical time-stamped data, including 188 ambiguous target words, we compare the proposed models with the current state of the art models NEO and SCAN. The results, presented in §5, show that HDP-based models achieve the best results over the dataset, establishing a new state of the art for DWSI.
+
+# 2 Related Work
+
+Topic modelling techniques are hierarchical probabilistic Bayesian models used originally for discovering topics in a collection of documents (Blei et al., 2010). Topic models have also been adopted for the Word Sense Induction (WSI) task, as introduced by Brody and Lapata, 2009; Yao and Van Durme, 2011): word senses are treated as topics, and a short window around the target word (context) is considered instead of a full document. Topic modelling techniques have been extended further to similar tasks, such as Novel Sense Detection.
+
+Novel Sense Detection (NSD; also called Novel Sense Identification), introduced by Lau et al. (2012), consists of determining whether a target word acquires a new sense over two independent periods of time, separated by a large gap. Several authors have used Hierarchical Dirichlet Processes (HDP) for this task over a small set of target words and/or small set of data (Lau et al., 2012, 2014; Cook et al., 2014). Yao and Van Durme (2011); Lau et al. (2012) show in a preliminary study that HDP is also superior to LDA for WSI, due to its ability to adapt to varying degrees of granularity. Lau et al. (2012) extend this study using an oracle-based method to identify new senses
+
+from HDP predictions for the task of NSD, and for only five target words. Sarsfield and Tayyar Madabushi (2020) used HDP for NSD on a larger dataset (Schlechtweg et al., 2020), which was proposed in a recent shared task about Lexical Semantic Change Detection (LsCD), a refined version of NSD: LsCD intends to answer the question of whether the meaning of a target word has changed between two independent periods of time (also separated by a large time gap). In the LsCD task, methods based on static word embeddings (where the meaning of the word is represented by a single vector) achieved the highest performance.
+
+In contrast to NSD/LSCD, DWSI takes the time dimension into account and thus the task of DWSI is technically broader: it aims to discriminate senses and also models the temporal dynamics of word meaning across a long continuous period of time, e.g. year by year. As a result, DWSI can track the evolution of senses, the emergence of new senses and detect the year where a new sense appears. The DWSI task is introduced independently by Emms and Kumar Jayapal (2016) and Frermann and Lapata (2016); given a target word and a time-stamped corpus, both models estimate two main parameters: the senses as distributions over words, and the senses proportions over time. Frermann and Lapata (2016) extend this by also inferring the subtle meaning changes within a single sense over time, i.e. by allowing different word distributions over time for the same sense.
+
+However, these models are parametric and require the number of senses to be chosen in advance. Previous approaches dealt with this issue by increasing the number of senses. For example, Emms and Kumar Jayapal (2016) vary the number of senses manually for every target word, while Frermann and Lapata (2016) choose an arbitrary fixed large number of senses for all the target words.
+
+Additionally, evaluating and comparing such models on the DWSI task is difficult: the lack of large scale time-stamped and sense-annotated data hinders direct quantitative evaluation. The state of the art models, (Emms and Kumar Jayapal, 2016; Frermann and Lapata, 2016), were originally evaluated only qualitatively on a few hand-picked target words, with a manual investigation of the quality of the associated top words in each cluster; Frermann and Lapata (2016) also evaluated their model on
+
+several indirect tasks. Alsulaimani et al. (2020) demonstrate that these evaluation methods are insufficient, and consequently propose a quantitative evaluation of these DWSI models based on a large set of data. In particular, they show that the senses size distribution plays a significant role in capturing the senses representations and emergence of new senses. The number of senses is clearly a crucial hyperparameter for a DWSI model, the choice of which should in theory depend on the characteristics of the data.
+
+# 3 Approach
+
+# 3.1 Parameters Notation
+
+DWSI aims to discover the senses  $S$  across time  $Y$  for each target word in a sequential collection of documents, where senses are latent variables and the number of senses is unknown a priori. A DWSI model estimates at least two multinomial distributions:
+
+-  $P(W|S)$ , the word given sense distribution. The changes within senses across time can also be represented as  $P(W|S,Y)$ , the word given sense and year distribution. These distributions represent the sense.  
+-  $P(S|Y)$ , the sense given year distribution. This distribution represents the relative prevalence of a sense over time.
+
+# 3.2 HDP-DWSI
+
+HDP allows senses (i.e. clusters) to appear when a new context occurs, as the number of senses is determined by the data. HDP-DWSI directly relies on this property: in the first step, all the documents, independently from their year, are clustered by HDP. Appendix A provides details about the description of HDP. This means that in this step the documents are assumed to be exchangeable, as opposed to dynamic models in which documents are only exchangeable within a time period. In the second step, the year of the document (observed variable) is reintroduced and the time-related multinomial parameters  $P(S = s|Y = y)$  are estimated by marginalising across the documents of each year  $j$  independently  $\sum_{d\in y}\frac{freq(s_d)}{\sum_{s'}freq(s'_d)}$ , where  $freq(s_d)$
+
+the number of words predicted as sense  $s$  in the document  $d$ , and  $d \in y$  represents the condition that the document  $d$  belongs to year  $y$ .
+
+HDP-DWSI is intended to be used as a nonparametric method, but a parametric mode is also proposed for the purpose of evaluation and comparison against parametric models. In the nonparametric mode, the model parameters are obtained directly as described above. In the parametric mode, an additional step is required to reduce the number of senses because HDP-DWSI tends to induce a higher number of clusters than the gold number of senses, i.e. to split senses into multiple clusters. Depending on the context of the application, it can also be relevant to reduce the number of senses even in the nonparametric mode. This can also be done with the method described below for the parametric mode, called  $\mathrm{HDP - DWSI_m}$ .
+
+HDP-DWSI $_m$  consists in merging the predicted senses which are the most semantically similar. Agglomerative hierarchical clustering (Ward Jr, 1963) is used to merge senses, based on a sense cooccurrence matrix obtained from the HDP clustering output.
+
+Pointwise Mutual Information (PMI) is used to represent how strongly two predicted senses are statistically associated, under the assumption of independence:
+
+$$
+P M I \left(s _ {i}, s _ {j}\right) = \log_ {2} \frac {P \left(s _ {i} , s _ {j}\right)}{P \left(s _ {i}\right) P \left(s _ {j}\right)} \tag {1}
+$$
+
+where  $i \neq j$  and  $P(s_i, s_j)$  is the joint probability of observing both  $s_i$  and  $s_j$  in the same document.  $P(s_i)$  (resp.  $P(s_j)$ ) is the probability of a predicted sense with respect to the entire corpus, i.e., an occurrence is counted for every document in which the predicted sense  $s_i$  (resp.  $s_j$ ) independently occurs.
+
+Moreover, since a pair of predicted senses with negative PMI is uninformative for the purpose of merging similar senses, Positive Pointwise Mutual Information (PPMI), as defined in Equation 2, is used for constructing the sense cooccurrence matrix.
+
+$$
+P P M I = \left\{ \begin{array}{l l} P M I \left(s _ {i}, s _ {j}\right) & \text {i f} P M I \left(s _ {i}, s _ {j}\right) > 0 \\ 0 & \text {e l s e} \end{array} \right. \tag {2}
+$$
+
+(P)PMI is sensitive to low frequency events, particularly in the event when one of the predicted
+
+senses (or both of them) is/are less frequent with respect to the whole corpus; thus it is possible that two senses mostly cooccur together by chance, yet obtain a high (P)PMI value. In such a case, the two predicted senses are not semantically associated, so this is a potential bias in the merging process. To counter this bias, we use the linkage criterion defined in Equation 3 as the average of the PPMI values weighted by their corresponding joint probabilities. The linkage criterion for two clusters  $C_1, C_2$ :
+
+$$
+\sum_ {\substack {\forall s _ {1} \in C _ {1} \\ \forall s _ {2} \in C _ {2}}} w \left(s _ {1}, s _ {2}\right) \times P P M I \left(s _ {1}, s _ {2}\right) \tag{3}
+$$
+
+$$
+\text{where} w(s_{1},s_{2}) = \frac{P(s_{1},s_{2})}{\sum_{\substack{\forall s_{1}\in C_{1}\\ \forall s_{2}\in C_{2}}}P(s_{1},s_{2})}
+$$
+
+The evaluation method proposed by Alsulaimani et al. (2020) (see §4) relies on the gold number of senses, as it is originally intended for parametric methods. In order to compare an HDP-based model against parametric models in an equivalent setting, the HDP-DWSI<sub>m</sub> merging method is used to reduce the predicted number of senses to the gold-standard number of senses.
+
+# 3.3 DETM-DWSI
+
+DETM represents not only the observed words but also latent topics/senses as embeddings, while preserving the traditional representation of a topic/sense as a probability distribution across words. The categorical distributions over the vocabulary is time dependent, i.e.  $P(W|S,Y)$  and is derived from the corresponding word embeddings and sense embedding at a given time. DETM also places time-dependent priors over senses proportions: the use of Markov chain over the sense proportions allows smoothness of the variations between the adjacent senses at neighboring times (see Appendix A for the description of DETM). We propose two modes for DETM-DWSI as follows:
+
+- In the regular DETM-DWSI, both the word and sense embeddings are trained simultaneously. This mode does not require any additional resource but the corpus must be large enough for the embeddings to be accurate.  
+- In DETM-DWSI $_i$ , the model is trained with prefitted word embeddings. This mode leverages the external information contained in the
+
+embeddings, potentially obtaining a more accurate representation of the senses as a consequence. It also allows the application of the model to text containing words not present in the corpus, as long as their embedding is available.
+
+In the experiments described below, the DETM-DWSI models are trained using the BioWordVec pretrained word embeddings $^2$  (Zhang et al., 2019). The fastText subword embedding model (Bojanowski et al., 2017) is a variant of the continuous skip-gram model (Mikolov et al., 2013). The fastText subword embedding can learn a distinct vector for each word while exploiting subword information in a unified n-gram embedding space. BioWordVec embeddings are trained with fastText on the PubMed text and MeSH terms, combined into a unified embedding space. In the biomedical domain, the advantage of a subword embedding model is that it can handle Out of Vocabulary (OOV) words (Zhang et al., 2019). $^3$  This leads to a more precise word representation, in theory better able to capture the semantics of specialised concepts. We use the intrinsic BioWordVec embeddings (as opposed to the extrinsic type), meant to represent the semantic similarity between words (Zhang et al., 2019).
+
+# 4 Experimental Setup
+
+# 4.1 Data
+
+We use the DWSI evaluation framework proposed by Alsulaimani et al. (2020): the biomedical literature is used as a source of labelled and timestamped data which covers the years 1946 to 2019. $^{4}$  The dataset is collected from resources provided by the US National Library of Medicine (NLM): PubMed (a platform which includes the major biomedical literature databases) and MeSH (a controlled vocabulary thesaurus, created manually to index NLM databases). $^{5}$  The data is preprocessed
+
+as in (Alsulaimani et al., 2020). The data consists of 188 ambiguous target words and 379 gold-standard senses (Jimeno-Yepes et al., 2011): 75 ambiguous target words have 2 senses, 12 have 3 and one has 5 senses. The total data size is  $15.36 \times 10^{9}$  words, and the average number of documents is 61,352 by sense. The input documents for every target word consist of the occurrences of the target word which are provided with a window of 5-word context on each side as well as the year of publication. The gold-standard sense label is also available for evaluation purposes.
+
+# 4.2 Algorithms Settings
+
+- The HDP-DWSI and HDP-DWSIm models are trained using the official  $\mathrm{C} + +$  implementation of HDP. No additional preprocessing is needed.  
+- The DETM-DWSI and DETM-DWSI $_i$  models are trained using the implementation provided by Dieng et al. (2019). The preprocessing is adapted to the DWSI dataset: since the data is strongly imbalanced across time, stratified sampling is used in order to ensure a representative time distribution (with at least one instance by year) across the data partitions. The data is split into  $85\%$  of instances for training and  $15\%$  for validation. The document frequency thresholds are unused so as to include all the words. For efficiency reasons, during training the number of instances is capped at 2,000 instances per year.
+
+# 4.3 Evaluation Methodology
+
+Since DWSI is an unsupervised task (clustering) and our evaluation is based on the external sense labels, both the estimation of the model and the evaluation are performed on the full set of documents for each target word. The gold-standard number of senses of each ambiguous target word is provided for all the parametric models (excluding HDP-DWSI). The default parameters are used in all the systems, except the number of iter
+
+tions/epochs (set to 500 for all the systems), and specifically for DETM-DWSI the batch size is set to 1000 and the dimension of the embeddings is set to 200.
+
+After estimating each model for each ambiguous target word, the posterior probability is calculated for every document. The sense with the highest probability is assigned.
+
+# 4.4 Evaluation Measures
+
+We follow Alsulaimani et al. (2020) for the evaluation measures with some adjustments, detailed below.
+
+The "Global Matching" method, presented by Alsulaimani et al. (2020), consists in determining a one-to-one assignment between predicted senses and gold senses based on their joint frequency: the pair with the highest frequency is matched first, and this process is iterated until all the senses are matched. In the case of HDP-DWSI, the number of predicted senses may be higher than the gold number of senses, and the instances of the predicted senses which remain unmatched are considered as false negative. This allows to compare HDP-DWSI with the parametric models, assuming that in theory the ideal nonparametric model would infer exactly the true number of senses. Of course,  $\mathrm{HDP - DWSI_m}$  is by definition more appropriate for a comparison in the parametric setting of HDP-based methods.
+
+We also propose to use the V-measure as a different method of evaluation. The V-measure is introduced by Rosenberg and Hirschberg (2007), providing a different way to evaluate a clustering solution. In this case, it evaluates every cluster against every gold sense without relying on a matching method, thus providing an objective assessment even when the number of the clusters is higher than the true number of senses. The V-measure is based on entropy (entropy is a measure of the uncertainty associated with a random variable): it is defined as the harmonic mean of homogeneity and completeness, which are both based on the normalised conditional entropy.
+
+Alsulaimani et al. (2020) also propose to evalu
+
+ate the emergence of a new sense by considering whether the system predicts the true emergence year of a sense. This requires a method to determine the year from the  $P(S|Y)$  distribution, for which the original algorithm "EmergeTime" was proposed in Jayapal (2017). We introduce "LRE-mergeTime" (see Appendix B Algorithm 1), an improved version of "EmergeTime" using linear regression instead of multiple thresholds within a window. Indeed, the original algorithm is very sensitive to the noise which sometimes occurs in the emergence pattern. Linear regression handles this issue better, since it measures the global trend across the window.[10]
+
+The emergence year is evaluated as in (Alsulaimani et al., 2020): (1) with standard classification measures, considering the sense as correctly predicted if the year is within 5 years of the true emergence year; (2) with (normalized) Mean Absolute Error, representing the average difference in number of years but also penalizing the wrongly predicted presence/absence of emergence.
+
+Finally we also use the distance between the true and predicted evolution of the senses over time  $(P(S|Y))$  as an evaluation method for DWSI, again following Alsulaimani et al. (2020).
+
+# 5 Results
+
+# 5.1 Qualitative exploration
+
+We explore the temporal meanings of "SARS-associated coronavirus" over the years (2002-2018) as an example. The ambiguous word has two gold-standard senses described by UMLS concepts  $C1175175$  and  $C1175743$ : Severe Acute Respiratory Syndrome (refers to the disease caused by the virus) and SARS Virus (refers to the virus related to the Coronavirus family causing the disease) respectively. The top words represented by the inferred parameter word given sense, identified by HDP
+
+$\mathrm{DWSI}_{\mathrm{m}}$  for the first sense are  $\{$  patients, outbreak, sars, 2003, epidemic, health, case, transmission, hospital\} and for the second sense are  $\{$  cov, sars, coronavirus, patients, infection, protein, respiratory, acute, syndrome, cells\}. Figure 1 shows the relative prevalence of the two inferred and gold senses over time, and Table 1 shows the top inferred words/uses associated with sense  $C_{1175175}$  at specific times.
+
+In Figure 1, both senses data start in 2002, however the prevalence of sense  $C_{1175175}$  was decreasing progressively from 2002 to 2018 since SARS was successfully contained in 2004, while the prevalence of the sense  $C_{1175743}$  kept increasing since the research about the SARS virus became a priority for the public health around the world.
+
+The temporal changes of the top words within C1175175 are highlighted in Table 1. Historically, the first known case of SARS appears in November 2002, causing the 2002-2004 SARS outbreaks in cities and hospitals. Global attention then started and in 2016, for instance, the top words shifted to facemask, post, era, sars. Finally, the year 2018 shows the concerns about a second wave of SARS.
+
+<table><tr><td>2002</td><td>2003</td><td>2004</td><td>2016</td><td>2017</td><td>2018</td></tr><tr><td>case</td><td>patients</td><td>patients</td><td>outbreak</td><td>outbreak</td><td>second</td></tr><tr><td>outbreak</td><td>outbreak</td><td>outbreak</td><td>facemask</td><td>2003</td><td>2003</td></tr><tr><td>lessons</td><td>case</td><td>sars</td><td>post</td><td>patients</td><td>impact</td></tr><tr><td>learned</td><td>health</td><td>transmission</td><td>2003</td><td>china</td><td>epidemic</td></tr><tr><td>health</td><td>2003</td><td>hospital</td><td>era</td><td>data</td><td>wave</td></tr><tr><td>chief</td><td>sars</td><td>case</td><td>sars</td><td>outbreaks</td><td>n’t</td></tr><tr><td>falls</td><td>hospital</td><td>patient</td><td>hong</td><td>health</td><td>link</td></tr></table>
+
+![](images/966468778507235f73c7f4e6a5705e537cdf02cbb310352c0b360eaec743e589.jpg)  
+Figure 1: Dynamic representations of "SARS-associated coronavirus". On the Y-axis,  $P(S|Y)$  shows the relative prevalence of the gold senses as well as the predicted senses across time estimated by HDP-DWSI $_m$ .
+
+# 5.2 Matching-based Evaluation
+
+Table 2 shows the performance of the six models according to standard classification and regression measures using "Global Matching". In general, DWSI models based on HDP perform well compared to NEO or SCAN. In the case of HDP-DWSI, "Global Matching" causes two observable effects: it increases precision, by allowing the system to choose the best predicted clusters matched with the gold senses; but it also decreases recall by introducing a large number of false negative cases due the discarded unmatched predicted clusters. Nevertheless the macro F1 score for HDP-DWSI is much higher than both NEO and SCAN, by  $17.7\%$  and  $13.8\%$  respectively. This shows that HDP-DWSI can distinguish minority senses significantly better. This can also be seen in Table 3 which shows the mean F1-score by senses size.
+
+Table 1: Temporal evolution of the top-7 words for the sense Severe Acute Respiratory Syndrome learned by HDP-DWSIm, at specific times.  
+
+<table><tr><td rowspan="2">Systems</td><td colspan="3">Macro-average</td><td colspan="3">Micro-average</td><td></td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>MAE</td></tr><tr><td>DETM-DWSIi</td><td>0.553</td><td>0.561</td><td>0.557</td><td>0.704</td><td>0.704</td><td>0.704</td><td>0.401</td></tr><tr><td>DETM-DWSI</td><td>0.559</td><td>0.590</td><td>0.574</td><td>0.650</td><td>0.650</td><td>0.650</td><td>0.379</td></tr><tr><td>HDP-DWSI</td><td>0.726</td><td>0.599</td><td>0.657</td><td>0.739</td><td>0.424</td><td>0.539</td><td>-</td></tr><tr><td>HDP-DWSIm</td><td>0.666</td><td>0.681</td><td>0.674</td><td>0.744</td><td>0.744</td><td>0.744</td><td>0.26</td></tr><tr><td>NEO</td><td>0.548</td><td>0.569</td><td>0.558</td><td>0.595</td><td>0.595</td><td>0.595</td><td>0.425</td></tr><tr><td>SCAN</td><td>0.562</td><td>0.591</td><td>0.577</td><td>0.558</td><td>0.558</td><td>0.558</td><td>0.444</td></tr></table>
+
+Table 2: Global performance results for all systems using "Global Matching". P/R/F1 stand for Precision/Recall/F1-score (higher is better) MAE stands for Mean Absolute Error (lower is better). Best performance in bold.
+
+The superiority of HDP-DWSI $_m$  is even clearer: the macro F1 score is  $20.8\%$  higher than NEO and  $16.8\%$  higher than SCAN; the performance difference in micro F1 score is even stronger:  $21.0\%$  above DETM-DWSI,  $17.4\%$  higher than DETM-DWSI,  $25.0\%$  above NEO and  $33.3\%$  above SCAN. Contrary to the differences between NEO and SCAN, HDP-DWSI $_m$  improves performance significantly across the board: both precision and recall are drastically higher, according to both micro and macro scores. This means that HDP-based models are fundamentally much better at discriminating the different senses (with a very significant p-value  $< 0.05$ ), as opposed to strategically favouring large senses for instance. This is confirmed in Table 3. $^{11}$
+
+The two DETM-based models perform very well, in particular achieving micro F1-score much higher than NEO and SCAN. However their macro-average performance is comparable to NEO and
+
+SCAN, a clear sign that they do not separate the senses better. Table 3 confirms that the DETM-based models perform closely to NEO and SCAN.
+
+Finally the MAE scores confirm that DETM-DWSI and DETM-DWSI perform better than NEO and SCAN, but also that these four models are drastically outperformed by HDP-DWSI $_m$ .
+
+<table><tr><td rowspan="2">Number of Senses</td><td rowspan="2">Sense rank</td><td colspan="6">Mean F1 score</td></tr><tr><td>N</td><td>S</td><td>H</td><td>Hm</td><td>D</td><td>Di</td></tr><tr><td>-</td><td>first</td><td>0.299</td><td>0.321</td><td>0.532</td><td>0.438</td><td>0.314</td><td>0.283</td></tr><tr><td>-</td><td>last</td><td>0.732</td><td>0.692</td><td>0.658</td><td>0.857</td><td>0.739</td><td>0.772</td></tr><tr><td>2</td><td>first</td><td>0.315</td><td>0.335</td><td>0.557</td><td>0.462</td><td>0.330</td><td>0.294</td></tr><tr><td>2</td><td>second</td><td>0.740</td><td>0.6995</td><td>0.659</td><td>0.863</td><td>0.744</td><td>0.777</td></tr><tr><td>3</td><td>first</td><td>0.100</td><td>0.143</td><td>0.224</td><td>0.132</td><td>0.111</td><td>0.134</td></tr><tr><td>3</td><td>second</td><td>0.253</td><td>0.390</td><td>0.553</td><td>0.499</td><td>0.237</td><td>0.248</td></tr><tr><td>3</td><td>third</td><td>0.629</td><td>0.597</td><td>0.655</td><td>0.778</td><td>0.681</td><td>0.708</td></tr></table>
+
+Table 3: Comparison of the performance by sense according to the "Global Matching" method, ranked by proportion within a target. The sense rank is ordered by the size of senses (in number of instances), from the smallest sense (rank first) to the largest (rank last). "-" means any number of senses (all the data). The systems are referred to by their initials.
+
+# 5.3 Entropy-based Evaluation
+
+<table><tr><td rowspan="2">Systems</td><td colspan="2">V-measure</td><td colspan="2">homogeneity</td><td colspan="2">completeness</td></tr><tr><td>Mean</td><td>Median</td><td>Mean</td><td>Median</td><td>Mean</td><td>Median</td></tr><tr><td>DETM-DWSIi</td><td>0.093</td><td>0.021</td><td>0.106</td><td>0.059</td><td>0.089</td><td>0.016</td></tr><tr><td>DETM-DWSI</td><td>0.092</td><td>0.026</td><td>0.111</td><td>0.059</td><td>0.085</td><td>0.020</td></tr><tr><td>HDP-DWSI</td><td>0.213</td><td>0.161</td><td>0.384</td><td>0.349</td><td>0.157</td><td>0.107</td></tr><tr><td>HDP-DWSIm</td><td>0.272</td><td>0.110</td><td>0.289</td><td>0.154</td><td>0.268</td><td>0.094</td></tr><tr><td>NEO</td><td>0.046</td><td>0.018</td><td>0.053</td><td>0.026</td><td>0.043</td><td>0.014</td></tr><tr><td>SCAN</td><td>0.080</td><td>0.021</td><td>0.098</td><td>0.041</td><td>0.074</td><td>0.015</td></tr></table>
+
+Table 4: V-measure, homogeneity and completeness for all the systems. Both the mean and median across targets are reported, because the strong differences between targets in terms of size and distribution of the senses may cause a bias with the mean.
+
+Table 4 shows the results of the systems for V-measure, with details about homogeneity and completeness. HDP-DWSI and HDP-DWSIm perform the best at all three levels, with values far above the other systems. HDP-DWSI has the highest homogeneity mean, because this model produces a higher number of smaller predicted senses; these predicted senses are therefore more homogeneous in general, but also less complete since the gold senses are often split. HDP-DWSIm merges the senses predicted by HDP-DWSI, thus obtaining lower homogeneity but compensating with higher completeness, leading to higher mean V-measure.
+
+Figure 2 offers a more precise picture of the differences between systems about their V-measure distribution. It confirms that DETM-DWSI, DETMDWSI and SCAN perform very similarly. It
+
+![](images/ae87f09bf8383d712835228809ee15bf5b7d83a2ef9672761a4a33f77f0e491f.jpg)  
+Figure 2: Quantile plot of the V-measure scores by system, with the quantile rank shown on the X axis and the corresponding value on the Y axis. Example: for HDP-DWSI, the median  $(x = 0.5)$  is  $y = 0.16$ . The graph is obtained by sorting the values, then normalising their rank between 0 and 1.
+
+shows that the higher performance of DETM-DWSI, DETM-DWSi and SCAN compared to NEO is due to a minority of targets, as their  $75\%$  lowest scores are almost identical. These targets cause most of the high difference in mean between NEO and SCAN, as the smaller difference in medians shows.
+
+By contrast, HDP-DWSI and HDP-DWSIm have a much smaller proportion of low scores. Interestingly, HDP-DWSI has higher low scores than HDP-DWSIm, i.e. HDP-DWSI performs better until both systems reach the median. However HDP-DWSIm skyrockets just after the median and surpasses HDP by having much higher high scores. This explains why the median is slightly lower for HDP-DWSIm than HDP while the mean is much higher for HDP-DWSIm.12
+
+# 5.4 Comparison between Measures
+
+<table><tr><td rowspan="2">Measure</td><td colspan="6">V-measure</td></tr><tr><td>N</td><td>S</td><td>H</td><td>Hm</td><td>D</td><td>Di</td></tr><tr><td>Macro F-Score</td><td>0.730</td><td>0.799</td><td>0.856</td><td>0.901</td><td>0.795</td><td>0.794</td></tr><tr><td>Micro F-Score</td><td>0.583</td><td>0.850</td><td>0.806</td><td>0.714</td><td>0.713</td><td>0.494</td></tr></table>
+
+Table 5: Pearson correlation coefficients: the relationship between the performance according to different measures. All the results are significantly correlated with p-value  $<= 5.6\mathrm{e} - 13$ . The systems are referred to by their initials.
+
+V-measure can introduce a bias towards systems which predict a number of clusters larger than the number of gold senses. Such systems tend to have very high homogeneity scores and low completeness scores. However, this is not the case for HDP-DWSI. The HDP-DWSI performance is high not only according to the V-measure but also confirmed
+
+by the F1 scores. The number of senses predicted by HDP-DWSI in average is 8 senses, with the minimum 4 senses and the maximum 13 senses. The Pearson correlation between homogeneity and completeness is 0.853 and with very significant p-value, 2.2e-16. Also, it is found that there is virtually no correlation between the predicted number of senses and either the size of the data or V-measure by target: 0.065, 0.008 (non significant: p-value = 0.3746, 0.261). This indicates that HDP-DWSI is not biased towards generating more senses when the data is larger.
+
+Table 5 shows that all the evaluation measures are significantly correlated. The macro-F1 scores are positively correlated in all four systems. However, the micro F-score favours systems that perform well on the majority sense, whereas the V-measure explicitly evaluates every cluster, taking into account not only the majority sense but also the minority one. Therefore systems which favour the majority sense, like NEO and DETM-DWSi, have a lower correlation.
+
+# 5.5 Emergence-based Evaluation
+
+<table><tr><td>System</td><td>precision</td><td>recall</td><td>F1 score</td><td>global mean absolute error</td><td>normalised global mean absolute error</td></tr><tr><td>\( DETM-DWSI_i \)</td><td>0.500</td><td>0.009</td><td>0.019</td><td>48.713</td><td>0.812</td></tr><tr><td>\( DETM-DWSI \)</td><td>0.385</td><td>0.050</td><td>0.088</td><td>45.685</td><td>0.761</td></tr><tr><td>\( HDP-DWSI_m \)</td><td>0.371</td><td>0.254</td><td>0.301</td><td>23.148</td><td>0.403</td></tr><tr><td>NEO</td><td>0.383</td><td>0.397</td><td>0.390</td><td>23.967</td><td>0.399</td></tr><tr><td>SCAN</td><td>0.374</td><td>0.162</td><td>0.226</td><td>39.634</td><td>0.666</td></tr></table>
+
+Table 6: Sense emergence evaluation results for all the systems. The values in bold indicate the best score achieved among the systems.
+
+DWSI systems can also be evaluated based on their ability to predict the year of emergence of a new sense. Table 6 shows the performance of the systems after applying "LREmergeTime" (see §4.4) on the predictions of the systems. HDP-DWSIm and NEO perform closely to each other and much better than the other systems, according to both classification measures and MAE. NEO was designed and implemented with a focus on detecting sense emergence, this probably explains why it performs particularly well in this task (Jayapal, 2017).
+
+# 5.6 Evaluation based on the predicted evolution over time
+
+Table 7 shows for every system how well their prediction of  $P(S|Y)$  matches the true evolution of sense. Among all the systems, HDP-DWSIm predicts the closest  $P(S|Y)$  to the true evolution
+
+<table><tr><td rowspan="2">System</td><td colspan="2">Distance Global mean</td></tr><tr><td>DTW</td><td>Euclidean</td></tr><tr><td>DETM-DWSI</td><td>0.191</td><td>0.134</td></tr><tr><td>DETM-DWSIi</td><td>0.165</td><td>0.106</td></tr><tr><td>HDP-DWSIm</td><td>0.115</td><td>0.067</td></tr><tr><td>NEO</td><td>0.182</td><td>0.124</td></tr><tr><td>SCAN</td><td>0.222</td><td>0.142</td></tr></table>
+
+Table 7: Mean distance between the true and predicted sense, measured by Dynamic Time Warping (DTW) and Euclidean distance (lower is better). The results in bold indicate the best system.
+
+according to both distance measures. This confirms that not only  $\mathrm{HDP - DWSI_m}$  produces accurate predictions of the emergence year of novel senses but also predicts accurately the  $P(S|Y)$  trends in general, with significantly less errors than the other systems.
+
+# 6 Conclusion and Discussion
+
+In this paper we adapted two topic modelling methods to the task of DWSI and evaluated them against two state of art DWSI systems, NEO and SCAN, using the evaluation framework proposed by Alsulaimani et al. (2020). We also compared using the V-measure, and proposed an improved version of the emergence algorithm.
+
+The results show that HDP-based models are able to fit the data better than the parametric models. The results strongly show that merging HDP-DWSI clusters performs better than the DETM-DWSI models and LDA-like clustering, such as NEO and SCAN. The properties of HDP make it better at accurately fitting the topics/senses, in particular when there is a high imbalance between the senses proportions, i.e. with senses smaller in size (see Table 3). Furthermore, the fact that  $\mathrm{HDP - DWSI_m}$  outperforms all the other parametric models also demonstrates that these models do not find the optimal separation between the senses. It seems that the additional complexity of the time dimension together with the parametric constraints do not cope well with data imbalance across years.
+
+One could naturally assume that models designed specifically for a task would perform better on it. Implicitly, the research community encourages the creation of new models and tends to reward theoretical contribution over empirical ones. Thus there might be a bias in favor of designing sophisticated ad-hoc models (like NEO and SCAN) rather than adapting existing robust models (like HDP).
+
+# 7 Limitations
+
+# 7.1 Biomedical Domain
+
+The dataset used in these experiments belongs to the biomedical domain and it is in English language. There is no clear reason why the comparison between models would lead to different results on different domains, therefore we would expect the reported results (at least the major tendencies) to be also valid on the general domain.
+
+Nevertheless this assumption would need to be tested experimentally. To our knowledge, there is no equivalent dataset available in the general domain which satisfies the two following conditions:
+
+- Time-stamped documents spanning a relatively long period of time;  
+- Every document labelled with the sense of the target word.
+
+# 7.2 Duration of the Training Stage
+
+In the table below, we present the computational cost of training the different models presented in this paper. Most of the experiments were carried out on a computing cluster containing 20 to 30 machines with varying characteristics, thus the total duration is approximative.
+
+Computing times are reported in hours of CPU/GPU activity required to train the total of 188 target datasets. It is important to note that the two DETM models are trained on GPUs, whereas all the other models are trained on regular CPUs. Thus in overall computing power, the DETM models are the most costly to train (more than HDP, despite the higher duration).
+
+<table><tr><td>System</td><td>Duration</td><td>Notes</td></tr><tr><td>DETM-DWSIi</td><td>523.4</td><td>Trained on GPU</td></tr><tr><td>DETM-DWSI</td><td>474.2</td><td>Trained on GPU</td></tr><tr><td>HDP-DWSI</td><td>2,471.4</td><td></td></tr><tr><td>HDP-DWSIm</td><td>0.1</td><td>Only the merging process</td></tr><tr><td>NEO</td><td>25.1</td><td></td></tr><tr><td>SCAN</td><td>77.9</td><td></td></tr></table>
+
+# Acknowledgements
+
+We would like to thank the anonymous reviewers for their valuable comments. The first author is
+
+grateful to the Custodian of the Two Holy Mosques Scholarship Program from the Saudi Arabian Government as well as to Mr. Saud Alsulaimani for supporting this work. This work was conducted using high-performance clusters facilitated by the ADAPT Centre, Trinity College of Dublin.
+
+# References
+
+Eneko Agirre and Aitor Soroa. 2007. Semeval-2007 task 02: Evaluating word sense induction and discrimination systems. In Proceedings of the fourth international workshop on semantic evaluations (semeval-2007), pages 7-12.
+
+Ashjan Alsulaimani, Erwan Moreau, and Carl Vogel. 2020. An evaluation method for diachronic word sense induction. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3171-3180. Association for Computational Linguistics.
+
+Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. 2003. A neural probabilistic language model. journal of machine learning research, 3 (feb): 1137-1155, 2003. Google Scholar Google Scholar Digital Library Digital Library.
+
+Yoshua Bengio, Holger Schwenk, Jean-Sébastien Senécal, Frédéric Morin, and Jean-Luc Gauvain. 2006. Innovations in machine learning. *Neural Probabilistic Language Models*, 194:137-186.
+
+David Blei, Lawrence Carin, and David Dunson. 2010. Probabilistic topic models: A focus on graphical model design and applications to document and image analysis. IEEE signal processing magazine, 27(6):55.
+
+David M Blei and John D Lafferty. 2006. Dynamic topic models. In Proceedings of the 23rd international conference on Machine learning, pages 113-120. ACM.
+
+David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993-1022.
+
+Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5:135-146.
+
+Samuel Brody and Mirella Lapata. 2009. Bayesian word sense induction. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, pages 103-111. Association for Computational Linguistics.
+
+Paul Cook, Joy Han Lau, Diana McCarthy, and Timothy Baldwin. 2014. Novel word-sense identification. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1624-1635.
+
+Adji B Dieng, Francisco JR Ruiz, and David M Blei. 2019. The dynamic embedded topic model. arXiv preprint arXiv:1907.05545.  
+Martin Emms and Arun Kumar Jayapal. 2016. Dynamic generative model for diachronic sense emergence detection. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1362-1373.  
+Thomas S. Ferguson. 1973. A bayesian analysis of some nonparametric problems. Annals of Statistics, 1:209-230.  
+Lea Frermann and Mirella Lapata. 2016. A bayesian model of diachronic meaning change. Transactions of the Association for Computational Linguistics, 4:31-45.  
+Arun Jayapal. 2017. Finding Sense Changes by Unsupervised Methods. PhD thesis, Trinity College Dublin.  
+Antonio J Jimeno-Yepes, Bridget T McInnes, and Alan R Aronson. 2011. Exploiting mesh indexing in medline to generate a data set for word sense disambiguation. BMC bioinformatics, 12(1):223.  
+Jey Han Lau, Paul Cook, Diana McCarthy, Spandana Gella, and Timothy Baldwin. 2014. Learning word sense distributions, detecting unattested senses and identifying novel senses using topic models. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 259-270.  
+Jey Han Lau, Paul Cook, Diana McCarthy, David Newman, and Timothy Baldwin. 2012. Word sense induction for novel sense detection. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 591-601. Association for Computational Linguistics.  
+Suresh Manandhar, Ioannis Klapaftis, Dmitriy Dligach, and Sameer Pradhan. 2010. SemEval-2010 task 14: Word sense induction &disambiguation. In Proceedings of the 5th International Workshop on Semantic Evaluation, pages 63–68, Uppsala, Sweden. Association for Computational Linguistics.  
+Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111-3119.  
+Andrew Rosenberg and Julia Hirschberg. 2007. V-measure: A conditional entropy-based external cluster evaluation measure. In Proceedings of the 2007 joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP-CoNLL), pages 410-420.  
+David E Rumelhart and Adele A Abrahamson. 1973. A model for analogical reasoning. Cognitive Psychology, 5(1):1-28.
+
+Eleri Sarsfield and Harish Tayyar Madabushi. 2020. UoB at SemEval-2020 task 1: Automatic identification of novel word senses. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 239-245, Barcelona (online). International Committee for Computational Linguistics.  
+Dominik Schlechtweg, Barbara McGillivray, Simon Hengchen, Haim Dubossarsky, and Nina Tahmasebi. 2020. Semeval-2020 task 1: Unsupervised lexical semantic change detection. arXiv preprint arXiv:2007.11464.  
+Yee Whye Teh, Michael I Jordan, Matthew J Beal, and David M Blei. 2006. Hierarchical dirichlet processes. Journal of the american statistical association, 101(476):1566-1581.  
+Joe H Ward Jr. 1963. Hierarchical grouping to optimize an objective function. Journal of the American statistical association, 58(301):236-244.  
+Xuchen Yao and Benjamin Van Durme. 2011. Nonparametric bayesian word sense induction. In Proceedings of TextGraphs-6: Graph-based Methods for Natural Language Processing, pages 10-14. Association for Computational Linguistics.  
+Yijia Zhang, Qingyu Chen, Zhihao Yang, Hongfei Lin, and Zhiyong Lu. 2019. Biowordvec, improving biomedical word embeddings with subword information and mesh. Scientific data, 6(1):1-9.
+
+# A Hierarchical Bayesian Models Background
+
+# A.1 Hierarchical Dirichlet Processes
+
+Hierarchical Dirichlet Processes (HDP), introduced by Teh et al. (2006), uses Dirichlet processes priors (DPs), on the infinite-dimensional space of multinomial probability distributions and thus the number of mixture components (senses) is infinite a priori. The Hierarchical DPs allow new senses to emerge naturally at any point in time and guarantee the senses are shared within and across the documents. The DP provides a distribution on distributions over an arbitrary space.  $H$  is a symmetric Dirichlet on the word simplex and  $\gamma$  is a concentration parameter that controls the amount of variability of senses on the base distribution  $G_{0}$ , a distribution over senses drawn from a DP.  $\alpha$  is also a concentration parameter that controls the amount of variability of per-document senses on  $G_{d}$ , a multinomial probability distribution over senses drawn from a DP. Then, for each word  $w$  we draw a sense  $\beta_{d,n}$  from  $G_{d}$  and finally draw the word  $w$  from that sense  $\beta_{d,n}$ . The graphical model and the generative story of HDP are described in Figure 3.
+
+# A.2 Dynamic Embedded Topic Model
+
+Dynamic Embedded Topic Model (DETM), introduced by Dieng et al. (2019), uses embedding representations of words and topics. For each term  $v$  it considers an  $L$ -dimensional embedding representation  $p_v$ . It also considers an embedding  $\alpha_k^t \in \mathbf{R}^L$  for each topic  $k$  at a given time step  $t = 1, \dots, T$ . The topics (i.e. distributions over the vocabulary) are represented by the normalised exponentiated dot product between the embedding representation of the word and the assigned topic's embedding at every time  $t$  for each word in a document  $d$ :  $p(w_{d,n} = v|z_{d,n} = k, \alpha_k^{t_d}) \propto \exp\{p_v^T \alpha_k^{(t_d)}\}$ . The DETM uses a Markov chain over the topic embeddings  $\alpha_k^t$  and thus they evolve under Gaussian noise with variance  $\delta^2$ . Moreover, DETM posits time-varying prior, the logistic-normal distribution  $\mathcal{LN}$  over the topic proportions  $\theta_d$ , which depends on a latent variable  $\eta_{t_d}$ .
+
+# B Emergence Algorithm
+
+"LREmergeTime " algorithm in 1 is linear regression based algorithm, an improved version of "EmergeTime" proposed by (Jayapal, 2017).
+
+# ALGORITHM 1
+
+Emergence Detection algorithm based on linear regression
+
+Input  $\pi$  ..  $\pi [i]$  is the probability at time  $i$  , with  $1\leq i\leq N$    
+Input  $r$  : window size  $\triangleright$  Value used for window size: 5   
+Input  $s$  : slope threshold.  $\triangleright$  Value used for slope threshold: 0.04 function LREMERGTIME(  $\pi ,r,s)$  Surges  $=$ $\phi$  for n:=1 to (N-r+1) do if SurgeStart(n,π,s) then Surges  $=$  Surges  $\cup \{n\}$  end if end for if Surges  $\neq \phi$  then return min(Surges) else return  $\phi$  end if end function function SURGESTART(n,  $\pi ,s)$  (slope, intercept)  $=$  fit linear regression model on  $X =$ $[n,\dots ,n + r - 1]$  and  $Y = [\pi [n],\dots ,\pi [n + r - 1]]$  if slope  $<  s*\max (\pi)$  then return false end if PrevYears  $= \{n^{\prime}:1\leq n^{\prime} <   n\}$  if  $|\{n^{\prime}:n^{\prime}\in \mathrm{PrevYearand}\pi [n^{\prime}]\leq 0.1*$  max(π)}|/ |PrevYear|  $\geq 0.8$  then return true else return false end if end function
+
+# C Data: Gold Standard Dataset
+
+The table C below shows the gold standard output (senses and year of emergence), as obtained by the "LREmergeTime" emergence detection algorithm based on the original gold data in (Alsulaimani et al., 2020).
+
+The total number of targets which has emergence is 146 and which has no emergence is 42. This consists of 233 senses with emergence and 158 senses with no emergence. The table includes three type of emergence:
+
+- N: Number of senses  
+- LRET: "LREmergeTime" emergence year,
+
+![](images/33bbb43e76bfcb0ee381eb3759beb0058b7e677a275205f5b432936b85abd5f1.jpg)  
+Figure 3: Left: graphical representation of HDP for DWSI. Observed variables represented by shaded nodes and latent variables by clear nodes. Right: the corresponding generative process. Note that in DWSI the sense related variables replace the topic related variables.
+
+- Draw the base distribution over senses  $G_0 \sim DP(\gamma, H)$ ,  
+- For  $d \in 1, \dots, D$ , draw the per-document distribution over senses  $G_{d} \sim DP(\alpha, G_{0})$  
+- For each word  $w \in 1, \ldots, N_d$  in each document  $d$ ,
+
+- Draw the sense for the word  $\beta_{d,n} \sim G_d$  
+- Draw the word  $w_{d,n} \sim Mult(\beta_{d,n})$
+
+![](images/03a756f9b09ec3417ae3ce8ccddc8bffeb785625608a605a0d4781fc83ac7986.jpg)  
+Figure 4: Left: graphical representation of DETM for DWSI. Observed variables represented by shaded nodes and latent variables by clear nodes. Right: the corresponding generative process. Note that in DWSI the sense related variables replace the topic related variables.
+
+- Draw initial sense embedding  $\alpha_{k}^{(0)}\sim \mathcal{N}(0,I)$  
+- Draw initial sense proportion mean  $\eta_0 \sim \mathcal{N}(0, I)$  
+- For time step  $t = 1, \dots, T$ :
+
+- Draw sense embeddings  $\alpha_{k}^{(t)}$ $\mathcal{N}(\alpha_k^{(t - 1)},\delta^2 I)$  for  $k = 1,\dots..,K$  
+- Draw sense proportion means  $\eta_t \sim \mathcal{N}(\eta_{t-1}, \delta^2 I)$
+
+- For each document  $d$ :
+
+- Draw sense proportions  $\theta_{d}$ $\mathcal{LN}(\eta_{t_d},a^2 I)$  
+- For each word  $n$  in the document  $d$ :
+
+\*Draw sense assignment  $z_{d,n}$  \~ Cat(0d)  
+\*Draw word  $w_{d,n}$  \~ Cat(softmax  $(p^T\alpha_{z_{d,n}}^{t_d}))$
+
+- ET: "EmergeTime" emergence year,  
+- FYO: indicates the "First Year Occurrence" of a sense, determined by the start date of each sense in the data,  
+- MS: indicates the "Manual Surge", i.e. the visual manual annotations by the authors. The value "NA" indicates cases when no emergence found and "?" indicates visually ambiguous cases found during the manual annotation by the authors.
+
+<table><tr><td>Target</td><td>N</td><td>CUI</td><td>ET</td><td>LRET</td><td>FYO</td><td>MS</td></tr><tr><td>AA</td><td>2</td><td>C0001972</td><td></td><td>1945</td><td>1947</td><td>1947</td></tr><tr><td>AA</td><td>2</td><td>C0002520</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>ADA</td><td>2</td><td>C0001457</td><td></td><td></td><td>1955</td><td>?</td></tr><tr><td>ADA</td><td>2</td><td>C0002456</td><td></td><td></td><td>1955</td><td>?</td></tr><tr><td>ADH</td><td>2</td><td>C0001942</td><td>1978</td><td>1977</td><td>1976</td><td>1979</td></tr><tr><td>ADH</td><td>2</td><td>C0003779</td><td></td><td></td><td>1975</td><td></td></tr><tr><td>ADP</td><td>2</td><td>C0001459</td><td></td><td></td><td>1956</td><td></td></tr><tr><td>ADP</td><td>2</td><td>C0004374</td><td>1958</td><td>1956</td><td>1959</td><td>1959</td></tr><tr><td>Adrenal</td><td>2</td><td>C0001625</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Adrenal</td><td>2</td><td>C0014563</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Ala</td><td>3</td><td>C0001898</td><td></td><td></td><td>1947</td><td></td></tr><tr><td>Ala</td><td>3</td><td>C0002563</td><td>1954</td><td>1949</td><td>1953</td><td>1973</td></tr><tr><td>Ala</td><td>3</td><td>C0051405</td><td></td><td>1982</td><td>1947</td><td>1979</td></tr><tr><td>ALS</td><td>2</td><td>C0002736</td><td></td><td></td><td>1948</td><td></td></tr><tr><td>ALS</td><td>2</td><td>C0003372</td><td></td><td>1964</td><td>1968</td><td>1968</td></tr><tr><td>ANA</td><td>2</td><td>C0002463</td><td>1962</td><td>1962</td><td>1963</td><td>1963</td></tr><tr><td colspan="7">continued on next column or page</td></tr><tr><td>ANA</td><td>2</td><td>C0003243</td><td></td><td></td><td>1962</td><td></td></tr><tr><td>Astragalus</td><td>2</td><td>C0039277</td><td></td><td></td><td>1947</td><td>?</td></tr><tr><td>Astragalus</td><td>2</td><td>C0330845</td><td></td><td></td><td>1947</td><td>?</td></tr><tr><td>B-Cell Leukemia</td><td>2</td><td>C0023434</td><td></td><td></td><td>1986</td><td></td></tr><tr><td>B-Cell Leukemia</td><td>2</td><td>C2004493</td><td></td><td>1986</td><td>1988</td><td>1988</td></tr><tr><td>BAT</td><td>2</td><td>C0006298</td><td></td><td>1946</td><td>1949</td><td>1949</td></tr><tr><td>BAT</td><td>2</td><td>C0008139</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>BLM</td><td>2</td><td>C0005740</td><td></td><td></td><td>1971</td><td></td></tr><tr><td>BLM</td><td>2</td><td>C0005859</td><td></td><td>1978</td><td>1981</td><td>1981</td></tr><tr><td>Borrelia</td><td>2</td><td>C0006033</td><td></td><td></td><td>1979</td><td></td></tr><tr><td>Borrelia</td><td>2</td><td>C0024198</td><td></td><td>1980</td><td>1983</td><td>1983</td></tr><tr><td>BPD</td><td>2</td><td>C0006012</td><td></td><td></td><td>1980</td><td>?</td></tr><tr><td>BPD</td><td>2</td><td>C0006287</td><td>1980</td><td>1980</td><td>1981</td><td>?</td></tr><tr><td>BR</td><td>2</td><td>C0006137</td><td></td><td>1946</td><td>1946</td><td>?</td></tr><tr><td>BR</td><td>2</td><td>C0006222</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>Brucella abortus</td><td>2</td><td>C0006304</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>Brucella abortus</td><td>2</td><td>C0302363</td><td></td><td></td><td>1946</td><td>1958</td></tr><tr><td>BSA</td><td>2</td><td>C0005902</td><td></td><td></td><td>1952</td><td>?</td></tr><tr><td>BSA</td><td>2</td><td>C0036774</td><td></td><td>1952</td><td>1952</td><td>?</td></tr><tr><td>BSE</td><td>2</td><td>C0085105</td><td></td><td></td><td>1991</td><td>?</td></tr><tr><td>BSE</td><td>2</td><td>C0085209</td><td>1991</td><td></td><td>1991</td><td>?</td></tr><tr><td>Ca</td><td>3</td><td>C0006675</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Ca</td><td>3</td><td>C0006754</td><td>1945</td><td></td><td>1945</td><td>?</td></tr><tr><td>Ca</td><td>3</td><td>C0006823</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>CAD</td><td>2</td><td>C0011905</td><td></td><td></td><td>1983</td><td></td></tr><tr><td>CAD</td><td>2</td><td>C1956346</td><td>1983</td><td>1983</td><td>1985</td><td>1985</td></tr><tr><td>Callus</td><td>2</td><td>C0006767</td><td></td><td></td><td>1972</td><td>?</td></tr><tr><td>Callus</td><td>2</td><td>C0376154</td><td></td><td></td><td>1972</td><td>?</td></tr><tr><td>CAM</td><td>2</td><td>C0007578</td><td></td><td></td><td>1981</td><td></td></tr><tr><td>CAM</td><td>2</td><td>C0178551</td><td>2002</td><td>2001</td><td>2003</td><td>2003</td></tr><tr><td>CCD</td><td>2</td><td>C0008928</td><td></td><td></td><td>1965</td><td></td></tr><tr><td>CCD</td><td>2</td><td>C0751951</td><td>1997</td><td>1995</td><td>1965</td><td>1998</td></tr><tr><td>CCl4</td><td>2</td><td>C0007022</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>CCl4</td><td>2</td><td>C0209338</td><td>1994</td><td>1992</td><td>1991</td><td>1994</td></tr><tr><td>CDA</td><td>2</td><td>C0002876</td><td></td><td></td><td>1979</td><td></td></tr><tr><td>CDA</td><td>2</td><td>C0092801</td><td>1982</td><td>1979</td><td>1983</td><td>1988</td></tr><tr><td>CDR</td><td>2</td><td>C0011485</td><td></td><td></td><td>1973</td><td></td></tr><tr><td>CDR</td><td>2</td><td>C0021024</td><td></td><td>1994</td><td>1998</td><td>1998</td></tr><tr><td>Cell</td><td>2</td><td>C0007634</td><td></td><td></td><td>1969</td><td></td></tr><tr><td>Cell</td><td>2</td><td>C1136359</td><td>2010</td><td>1998</td><td>1999</td><td>2002</td></tr><tr><td>Cement</td><td>2</td><td>C0011343</td><td></td><td></td><td>1957</td><td>?</td></tr><tr><td>Cement</td><td>2</td><td>C1706094</td><td></td><td></td><td>1957</td><td>?</td></tr><tr><td>CH</td><td>2</td><td>C0008115</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>CH</td><td>2</td><td>C0039021</td><td>1946</td><td>1946</td><td>1946</td><td>?</td></tr><tr><td>Cholera</td><td>2</td><td>C0008354</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Cholera</td><td>2</td><td>C0008359</td><td></td><td>1945</td><td>1946</td><td>1961</td></tr><tr><td>CI</td><td>2</td><td>C0008107</td><td></td><td></td><td>1949</td><td></td></tr><tr><td>CI</td><td>2</td><td>C0022326</td><td></td><td></td><td>1951</td><td>1955</td></tr><tr><td>Cilia</td><td>2</td><td>C0008778</td><td></td><td>1950</td><td>1950</td><td>?</td></tr><tr><td>Cilia</td><td>2</td><td>C0015422</td><td></td><td></td><td>1950</td><td>?</td></tr><tr><td>CIS</td><td>2</td><td>C0007099</td><td></td><td></td><td>1972</td><td></td></tr><tr><td>CIS</td><td>2</td><td>C0162854</td><td>1991</td><td>1989</td><td>1992</td><td>1992</td></tr><tr><td>CLS</td><td>2</td><td>C0265252</td><td></td><td>1998</td><td>2002</td><td>2002</td></tr><tr><td>CLS</td><td>2</td><td>C0343084</td><td></td><td></td><td>1996</td><td></td></tr><tr><td>Coffee</td><td>2</td><td>C0009237</td><td></td><td></td><td>1960</td><td></td></tr><tr><td>Coffee</td><td>2</td><td>C0085952</td><td>2001</td><td>1998</td><td>1962</td><td>2002</td></tr><tr><td>Cold</td><td>3</td><td>C0009264</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Cold</td><td>3</td><td>C0009443</td><td>1945</td><td>1945</td><td>1945</td><td></td></tr><tr><td>Cold</td><td>3</td><td>C0024117</td><td>1998</td><td>1997</td><td>1959</td><td>1998</td></tr><tr><td>Compliance</td><td>2</td><td>C0009563</td><td></td><td></td><td>1974</td><td>?</td></tr><tr><td>Compliance</td><td>2</td><td>C1321605</td><td></td><td>1974</td><td>1974</td><td>?</td></tr><tr><td>Cortex</td><td>2</td><td>C0001614</td><td>1948</td><td>1947</td><td>1950</td><td>?</td></tr><tr><td>Cortex</td><td>2</td><td>C0007776</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Cortical</td><td>3</td><td>C0001613</td><td>1945</td><td>1945</td><td>1945</td><td></td></tr><tr><td>Cortical</td><td>3</td><td>C0007776</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Cortical</td><td>3</td><td>C0022655</td><td></td><td></td><td>1947</td><td>1971</td></tr><tr><td>CP</td><td>3</td><td>C0007789</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>CP</td><td>3</td><td>C0008925</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>CP</td><td>3</td><td>C0033477</td><td>1971</td><td>1969</td><td>1946</td><td>1971</td></tr><tr><td>CPDD</td><td>2</td><td>C0008838</td><td>1971</td><td>1971</td><td>1972</td><td>1972</td></tr><tr><td>CPDD</td><td>2</td><td>C0553730</td><td></td><td></td><td>1971</td><td></td></tr><tr><td>Crack</td><td>2</td><td>C0040441</td><td></td><td></td><td>1986</td><td></td></tr><tr><td>Crack</td><td>2</td><td>C0085163</td><td></td><td>1987</td><td>1990</td><td>1990</td></tr><tr><td>CRF</td><td>2</td><td>C0010132</td><td></td><td>1954</td><td>1956</td><td>1967</td></tr><tr><td>CRF</td><td>2</td><td>C0022661</td><td></td><td></td><td>1954</td><td></td></tr><tr><td>cRNA</td><td>2</td><td>C0056208</td><td>1981</td><td>1978</td><td>1982</td><td>1984</td></tr></table>
+
+continued on next column or page
+
+<table><tr><td>Target</td><td>N</td><td>CUI</td><td>ET</td><td>LRET</td><td>FYO</td><td>MS</td></tr><tr><td>cRNA</td><td>2</td><td>C1321571</td><td></td><td></td><td>1975</td><td></td></tr><tr><td>CTX</td><td>2</td><td>C0010583</td><td></td><td></td><td>1960</td><td></td></tr><tr><td>CTX</td><td>2</td><td>C0238052</td><td>1997</td><td>1992</td><td>1974</td><td>1996</td></tr><tr><td>DAT</td><td>2</td><td>C0002395</td><td></td><td></td><td>1974</td><td></td></tr><tr><td>DAT</td><td>2</td><td>C0114838</td><td>1989</td><td>1988</td><td>1989</td><td>1991</td></tr><tr><td>DBA</td><td>2</td><td>C0025923</td><td></td><td></td><td>1972</td><td></td></tr><tr><td>DBA</td><td>2</td><td>C1260899</td><td>1999</td><td>1998</td><td>2001</td><td>2001</td></tr><tr><td>dC</td><td>2</td><td>C0011485</td><td>1971</td><td>1969</td><td>1973</td><td>1973</td></tr><tr><td>dC</td><td>2</td><td>C0012764</td><td></td><td></td><td>1966</td><td></td></tr><tr><td>DDD</td><td>2</td><td>C0011037</td><td></td><td></td><td>1962</td><td></td></tr><tr><td>DDD</td><td>2</td><td>C0026256</td><td></td><td></td><td>1963</td><td>1973</td></tr><tr><td>DDS</td><td>3</td><td>C0010980</td><td></td><td></td><td>1960</td><td></td></tr><tr><td>DDS</td><td>3</td><td>C0085104</td><td>1988</td><td>1987</td><td>1990</td><td>1990</td></tr><tr><td>DDS</td><td>3</td><td>C0950121</td><td>1999</td><td>1998</td><td>2001</td><td>2001</td></tr><tr><td>DE</td><td>2</td><td>C0011198</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>DE</td><td>2</td><td>C0017480</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>DI</td><td>2</td><td>C0011848</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>DI</td><td>2</td><td>C0032246</td><td></td><td></td><td>1946</td><td>1976</td></tr><tr><td>Digestive</td><td>2</td><td>C0012238</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Digestive</td><td>2</td><td>C0012240</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>DON</td><td>2</td><td>C0012020</td><td></td><td></td><td>1975</td><td></td></tr><tr><td>DON</td><td>2</td><td>C0028652</td><td>1979</td><td>1978</td><td>1981</td><td>1981</td></tr><tr><td>drinking</td><td>2</td><td>C0001948</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>drinking</td><td>2</td><td>C0684271</td><td>1946</td><td></td><td>1946</td><td>?</td></tr><tr><td>eCG</td><td>2</td><td>C0018064</td><td>1989</td><td>1989</td><td>1945</td><td>?</td></tr><tr><td>eCG</td><td>2</td><td>C1623258</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Eels</td><td>2</td><td>C0013671</td><td></td><td></td><td>1951</td><td></td></tr><tr><td>Eels</td><td>2</td><td>C0677644</td><td>2003</td><td>2000</td><td>2004</td><td>2004</td></tr><tr><td>EGG</td><td>2</td><td>C0013710</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>EGG</td><td>2</td><td>C0029974</td><td></td><td>1945</td><td>1945</td><td>1945</td></tr><tr><td>EM</td><td>2</td><td>C0014921</td><td>1973</td><td>1972</td><td>1975</td><td>1975</td></tr><tr><td>EM</td><td>2</td><td>C0026019</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>EMS</td><td>2</td><td>C0013961</td><td></td><td></td><td>1967</td><td></td></tr><tr><td>EMS</td><td>2</td><td>C0015063</td><td>1974</td><td>1971</td><td>1975</td><td>1975</td></tr><tr><td>Epi</td><td>2</td><td>C0014563</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Epi</td><td>2</td><td>C0014582</td><td></td><td>1988</td><td>1980</td><td>1980</td></tr><tr><td>ERP</td><td>2</td><td>C0008310</td><td>1978</td><td>1977</td><td>1978</td><td>1978</td></tr><tr><td>ERP</td><td>2</td><td>C0015214</td><td></td><td></td><td>1956</td><td></td></tr><tr><td>ERUPTION</td><td>2</td><td>C0015230</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>ERUPTION</td><td>2</td><td>C1533692</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Erythrocytes</td><td>2</td><td>C0014772</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Erythrocytes</td><td>2</td><td>C0014792</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Exercises</td><td>2</td><td>C0015259</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Exercises</td><td>2</td><td>C0452240</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>FA</td><td>2</td><td>C0015625</td><td></td><td></td><td>1946</td><td>1975</td></tr><tr><td>FA</td><td>2</td><td>C0016410</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Fe</td><td>2</td><td>C0302583</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Fe</td><td>2</td><td>C0376520</td><td>1995</td><td>1992</td><td>1946</td><td>1996</td></tr><tr><td>Fish</td><td>2</td><td>C0016163</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Fish</td><td>2</td><td>C0162789</td><td>1990</td><td>1988</td><td>1953</td><td>1992</td></tr><tr><td>Follicle</td><td>2</td><td>C0018120</td><td></td><td></td><td>1949</td><td>?</td></tr><tr><td>Follicle</td><td>2</td><td>C0221971</td><td></td><td>1949</td><td>1949</td><td>?</td></tr><tr><td>Follicles</td><td>2</td><td>C0018120</td><td></td><td></td><td>1949</td><td>?</td></tr><tr><td>Follicles</td><td>2</td><td>C0221971</td><td></td><td>1949</td><td>1949</td><td>?</td></tr><tr><td>FTC</td><td>2</td><td>C0041713</td><td></td><td></td><td>1982</td><td></td></tr><tr><td>FTC</td><td>2</td><td>C0206682</td><td>1992</td><td>1989</td><td>1993</td><td>1993</td></tr><tr><td>GAG</td><td>2</td><td>C0017346</td><td>1988</td><td>1986</td><td>1982</td><td>1989</td></tr><tr><td>GAG</td><td>2</td><td>C0017973</td><td></td><td></td><td>1949</td><td></td></tr><tr><td>Ganglion</td><td>2</td><td>C0017067</td><td>1946</td><td></td><td>1946</td><td></td></tr><tr><td>Ganglion</td><td>2</td><td>C1258666</td><td>2006</td><td>1946</td><td>1946</td><td>2002</td></tr><tr><td>Gas</td><td>2</td><td>C0016204</td><td>1945</td><td></td><td>1945</td><td>?</td></tr><tr><td>Gas</td><td>2</td><td>C0017110</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Glycoside</td><td>2</td><td>C0007158</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>Glycoside</td><td>2</td><td>C0017977</td><td>1946</td><td>1946</td><td>1946</td><td>?</td></tr><tr><td>Haemophilus ducreyi</td><td>2</td><td>C0007947</td><td></td><td></td><td>1977</td><td></td></tr><tr><td>Haemophilus ducreyi</td><td>2</td><td>C0018481</td><td></td><td>1977</td><td>1978</td><td>1978</td></tr><tr><td>HCl</td><td>2</td><td>C0020259</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>HCl</td><td>2</td><td>C0023443</td><td>1975</td><td>1959</td><td>1954</td><td>1976</td></tr><tr><td>Hemlock</td><td>2</td><td>C0242872</td><td>2004</td><td>2002</td><td>2002</td><td>?</td></tr><tr><td>Hemlock</td><td>2</td><td>C0949851</td><td></td><td></td><td>2002</td><td>?</td></tr><tr><td>Heregulin</td><td>2</td><td>C0626201</td><td></td><td>1992</td><td>1994</td><td>1994</td></tr><tr><td>Heregulin</td><td>2</td><td>C0752253</td><td></td><td></td><td>1992</td><td></td></tr><tr><td>HGF</td><td>2</td><td>C0021760</td><td></td><td>1984</td><td>1984</td><td>?</td></tr><tr><td>HGF</td><td>2</td><td>C0062534</td><td></td><td></td><td>1984</td><td>?</td></tr><tr><td>Hip</td><td>2</td><td>C0019552</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>Hip</td><td>2</td><td>C0022122</td><td></td><td></td><td>1947</td><td>?</td></tr></table>
+
+continued on next column or page
+
+<table><tr><td>Target</td><td>N</td><td>CUI</td><td>ET</td><td>LRET</td><td>FYO</td><td>MS</td></tr><tr><td>HIV</td><td>2</td><td>C0019682</td><td></td><td></td><td>1985</td><td></td></tr><tr><td>HIV</td><td>2</td><td>C0019693</td><td>1987</td><td>1985</td><td>1987</td><td>1987</td></tr><tr><td>HPS</td><td>2</td><td>C0079504</td><td></td><td>1996</td><td>2000</td><td>2000</td></tr><tr><td>HPS</td><td>2</td><td>C0242994</td><td></td><td></td><td>1994</td><td></td></tr><tr><td>HR</td><td>2</td><td>C0010343</td><td></td><td>1947</td><td>1950</td><td>1992</td></tr><tr><td>HR</td><td>2</td><td>C0018810</td><td></td><td></td><td>1947</td><td></td></tr><tr><td>IA</td><td>2</td><td>C0021487</td><td>1946</td><td>1946</td><td>1946</td><td>1946</td></tr><tr><td>IA</td><td>2</td><td>C0022037</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>Ice</td><td>3</td><td>C0020746</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>Ice</td><td>3</td><td>C0025611</td><td></td><td>1946</td><td>1946</td><td>1946</td></tr><tr><td>Ice</td><td>3</td><td>C0534519</td><td>1990</td><td>1990</td><td>1991</td><td>1991</td></tr><tr><td>INDO</td><td>2</td><td>C0021246</td><td>1961</td><td>1959</td><td>1963</td><td>1963</td></tr><tr><td>INDO</td><td>2</td><td>C0021247</td><td></td><td></td><td>1949</td><td></td></tr><tr><td>Ion</td><td>2</td><td>C0022023</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Ion</td><td>2</td><td>C0022024</td><td>1945</td><td>1945</td><td>1946</td><td>1946</td></tr><tr><td>IP</td><td>2</td><td>C0021069</td><td>2000</td><td>1997</td><td>1989</td><td>2001</td></tr><tr><td>IP</td><td>2</td><td>C0021171</td><td></td><td></td><td>1986</td><td></td></tr><tr><td>Iris</td><td>2</td><td>C0022077</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Iris</td><td>2</td><td>C1001362</td><td></td><td>1945</td><td>1946</td><td>2001</td></tr><tr><td>JP</td><td>2</td><td>C0022341</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>JP</td><td>2</td><td>C0031106</td><td>1946</td><td>1946</td><td>1947</td><td>1983</td></tr><tr><td>LABOR</td><td>2</td><td>C0022864</td><td>1945</td><td>1945</td><td>1945</td><td>1945</td></tr><tr><td>LABOR</td><td>2</td><td>C0043227</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Lactation</td><td>2</td><td>C0006147</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Lactation</td><td>2</td><td>C0022925</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Language</td><td>2</td><td>C0023008</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>Language</td><td>2</td><td>C0033348</td><td>1986</td><td>1954</td><td>1958</td><td>1985</td></tr><tr><td>Laryngeal</td><td>2</td><td>C0023078</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Laryngeal</td><td>2</td><td>C0023081</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Lawsonia</td><td>2</td><td>C0752045</td><td></td><td></td><td>2000</td><td></td></tr><tr><td>Lawsonia</td><td>2</td><td>C1068388</td><td></td><td></td><td>2002</td><td>2002</td></tr><tr><td>Leishmaniasis</td><td>2</td><td>C0023281</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Leishmaniasis</td><td>2</td><td>C1548483</td><td>2005</td><td>1996</td><td>1947</td><td>2000</td></tr><tr><td>lens</td><td>3</td><td>C0023308</td><td>1951</td><td>1948</td><td>1952</td><td>1978</td></tr><tr><td>lens</td><td>3</td><td>C0023317</td><td></td><td>1945</td><td>1945</td><td></td></tr><tr><td>lens</td><td>3</td><td>C0023318</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Lupus</td><td>3</td><td>C0024131</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Lupus</td><td>3</td><td>C0024138</td><td>1945</td><td>1945</td><td>1946</td><td>1946</td></tr><tr><td>Lupus</td><td>3</td><td>C0024141</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>lymphogranulomatosis</td><td>2</td><td>C0019829</td><td></td><td>1945</td><td>1945</td><td>1945</td></tr><tr><td>lymphogranulomatosis</td><td>2</td><td>C0036202</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>MAF</td><td>2</td><td>C0079786</td><td></td><td></td><td>1980</td><td></td></tr><tr><td>MAF</td><td>2</td><td>C0919482</td><td>2001</td><td>1994</td><td>1998</td><td>1998</td></tr><tr><td>Malaria</td><td>2</td><td>C0024530</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Malaria</td><td>2</td><td>C0206255</td><td>1991</td><td>1988</td><td>1945</td><td>1992</td></tr><tr><td>MBP</td><td>2</td><td>C0014063</td><td></td><td></td><td>1973</td><td></td></tr><tr><td>MBP</td><td>2</td><td>C0065661</td><td>1999</td><td>1998</td><td>1984</td><td>2001</td></tr><tr><td>MCC</td><td>2</td><td>C0007129</td><td></td><td></td><td>1988</td><td></td></tr><tr><td>MCC</td><td>2</td><td>C0162804</td><td>1990</td><td>1989</td><td>1991</td><td>1993</td></tr><tr><td>Medullary</td><td>2</td><td>C0001629</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>Medullary</td><td>2</td><td>C0025148</td><td></td><td></td><td>1947</td><td>1947</td></tr><tr><td>MHC</td><td>2</td><td>C0024518</td><td></td><td></td><td>1978</td><td></td></tr><tr><td>MHC</td><td>2</td><td>C0027100</td><td></td><td>1991</td><td>1986</td><td>1994</td></tr><tr><td>Milk</td><td>2</td><td>C0026131</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Milk</td><td>2</td><td>C0026140</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Moles</td><td>2</td><td>C0027960</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>Moles</td><td>2</td><td>C0324740</td><td></td><td>1946</td><td>1946</td><td>1974</td></tr><tr><td>MRS</td><td>2</td><td>C0024487</td><td></td><td>1959</td><td>1961</td><td>1961</td></tr><tr><td>MRS</td><td>2</td><td>C0025235</td><td></td><td></td><td>1950</td><td></td></tr><tr><td>NBS</td><td>2</td><td>C0027819</td><td></td><td></td><td>1947</td><td></td></tr><tr><td>NBS</td><td>2</td><td>C0398791</td><td>2003</td><td>2002</td><td>2002</td><td>2006</td></tr><tr><td>NEUROFIBROMA...</td><td>2</td><td>C0085113</td><td></td><td></td><td>1990</td><td></td></tr><tr><td>NEUROFIBROMA...</td><td>2</td><td>C0162678</td><td>1990</td><td>1990</td><td>1991</td><td>1991</td></tr><tr><td>NM</td><td>2</td><td>C0025033</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>NM</td><td>2</td><td>C0027972</td><td>1963</td><td>1962</td><td>1946</td><td>1946</td></tr><tr><td>NPC</td><td>2</td><td>C0028587</td><td></td><td></td><td>1998</td><td></td></tr><tr><td>NPC</td><td>2</td><td>C0220756</td><td>2005</td><td>2002</td><td>2006</td><td>2006</td></tr><tr><td>Nurse</td><td>2</td><td>C0006147</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Nurse</td><td>2</td><td>C0028661</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Nursing</td><td>2</td><td>C0006147</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Nursing</td><td>2</td><td>C0028677</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>OCD</td><td>2</td><td>C0028768</td><td></td><td></td><td>1975</td><td></td></tr><tr><td>OCD</td><td>2</td><td>C0029421</td><td>1983</td><td>1980</td><td>1984</td><td>1984</td></tr><tr><td>OH</td><td>2</td><td>C0028905</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>OH</td><td>2</td><td>C0063146</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>Orf</td><td>2</td><td>C0013570</td><td></td><td></td><td>1980</td><td></td></tr></table>
+
+continued on next column or page
+
+<table><tr><td>Target</td><td>N</td><td>CUI</td><td>ET</td><td>LRET</td><td>FYO</td><td>MS</td></tr><tr><td>Orf</td><td>2</td><td>C0079941</td><td>1986</td><td>1985</td><td>1982</td><td>1982</td></tr><tr><td>ORI</td><td>2</td><td>C0206601</td><td></td><td></td><td>1993</td><td></td></tr><tr><td>ORI</td><td>2</td><td>C0242961</td><td>1993</td><td>1993</td><td>1993</td><td>1993</td></tr><tr><td>PAC</td><td>2</td><td>C0033036</td><td></td><td></td><td>1995</td><td></td></tr><tr><td>PAC</td><td>2</td><td>C0949780</td><td></td><td>1997</td><td>2001</td><td>2001</td></tr><tr><td>PAF</td><td>2</td><td>C0032172</td><td></td><td></td><td>1979</td><td></td></tr><tr><td>PAF</td><td>2</td><td>C0037019</td><td></td><td></td><td>1980</td><td>1980</td></tr><tr><td>Parotitis</td><td>2</td><td>C0026780</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Parotitis</td><td>2</td><td>C0030583</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>PCA</td><td>5</td><td>C0030131</td><td>1972</td><td>1971</td><td>1974</td><td>1974</td></tr><tr><td>PCA</td><td>5</td><td>C0030625</td><td></td><td></td><td>1957</td><td></td></tr><tr><td>PCA</td><td>5</td><td>C0078944</td><td>1987</td><td>1986</td><td>1989</td><td>1989</td></tr><tr><td>PCA</td><td>5</td><td>C0149576</td><td>1957</td><td>1957</td><td>1957</td><td>1957</td></tr><tr><td>PCA</td><td>5</td><td>C0429865</td><td>1999</td><td>1998</td><td>1960</td><td>2001</td></tr><tr><td>PCB</td><td>2</td><td>C0032447</td><td></td><td></td><td>1971</td><td>?</td></tr><tr><td>PCB</td><td>2</td><td>C0033223</td><td></td><td></td><td>1971</td><td>?</td></tr><tr><td>PCD</td><td>2</td><td>C0022521</td><td></td><td></td><td>1971</td><td></td></tr><tr><td>PCD</td><td>2</td><td>C0162638</td><td></td><td>1988</td><td>1991</td><td>1991</td></tr><tr><td>PCP</td><td>2</td><td>C0030855</td><td></td><td></td><td>1972</td><td>?</td></tr><tr><td>PCP</td><td>2</td><td>C0031381</td><td></td><td>1972</td><td>1972</td><td>?</td></tr><tr><td>PEP</td><td>2</td><td>C0031642</td><td></td><td></td><td>1971</td><td></td></tr><tr><td>PEP</td><td>2</td><td>C0135981</td><td>1978</td><td>1976</td><td>1980</td><td>1980</td></tr><tr><td>PHA</td><td>2</td><td>C0030779</td><td>2002</td><td>2007</td><td>1976</td><td>1976</td></tr><tr><td>PHA</td><td>2</td><td>C0031858</td><td></td><td>1975</td><td>1975</td><td></td></tr><tr><td>Pharmaceutical</td><td>2</td><td>C0013058</td><td>1963</td><td>1962</td><td>1963</td><td>1963</td></tr><tr><td>Pharmaceutical</td><td>2</td><td>C0031336</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Phosphorus</td><td>2</td><td>C0031705</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Phosphorus</td><td>2</td><td>C0080014</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Phosphorylase</td><td>2</td><td>C0017916</td><td></td><td></td><td>1971</td><td></td></tr><tr><td>Phosphorylase</td><td>2</td><td>C0917783</td><td>2005</td><td>1998</td><td>1973</td><td>2001</td></tr><tr><td>pI</td><td>2</td><td>C0022171</td><td></td><td></td><td>1975</td><td>?</td></tr><tr><td>pI</td><td>2</td><td>C0812425</td><td></td><td></td><td>1975</td><td>?</td></tr><tr><td>Plague</td><td>2</td><td>C0032064</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Plague</td><td>2</td><td>C0032066</td><td>1959</td><td>1957</td><td>1946</td><td>1960</td></tr><tr><td>Plaque</td><td>2</td><td>C0011389</td><td></td><td></td><td>1950</td><td>?</td></tr><tr><td>Plaque</td><td>2</td><td>C0333463</td><td></td><td></td><td>1950</td><td>?</td></tr><tr><td>Platelet</td><td>2</td><td>C0005821</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Platelet</td><td>2</td><td>C0032181</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Pleuropneumonia</td><td>2</td><td>C0026934</td><td>1945</td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Pleuropneumonia</td><td>2</td><td>C0032241</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>POL</td><td>2</td><td>C0017360</td><td></td><td>1986</td><td>1989</td><td>1989</td></tr><tr><td>POL</td><td>2</td><td>C0032356</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>posterior pituitary</td><td>2</td><td>C0032009</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>posterior pituitary</td><td>2</td><td>C0032017</td><td></td><td>1946</td><td>1947</td><td>1946</td></tr><tr><td>Potassium</td><td>2</td><td>C0032821</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Potassium</td><td>2</td><td>C0162800</td><td>1990</td><td>1989</td><td>1948</td><td>1992</td></tr><tr><td>PR</td><td>2</td><td>C0034044</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>PR</td><td>2</td><td>C0034833</td><td>1972</td><td>1972</td><td>1973</td><td>1973</td></tr><tr><td>Projection</td><td>2</td><td>C0016538</td><td></td><td>1970</td><td>1970</td><td>?</td></tr><tr><td>Projection</td><td>2</td><td>C0033363</td><td></td><td></td><td>1970</td><td>?</td></tr><tr><td>PVC</td><td>2</td><td>C0032624</td><td></td><td></td><td>1974</td><td></td></tr><tr><td>PVC</td><td>2</td><td>C0151636</td><td></td><td>1991</td><td>1988</td><td>1992</td></tr><tr><td>RA</td><td>3</td><td>C0002893</td><td></td><td>1945</td><td>1946</td><td>?</td></tr><tr><td>RA</td><td>3</td><td>C0003873</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>RA</td><td>3</td><td>C0034625</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Radiation</td><td>2</td><td>C0851346</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Radiation</td><td>2</td><td>C1522449</td><td></td><td></td><td>1946</td><td>1946</td></tr><tr><td>RB</td><td>2</td><td>C0035335</td><td></td><td></td><td>1947</td><td></td></tr><tr><td>RB</td><td>2</td><td>C0035930</td><td></td><td>1947</td><td>1951</td><td>1951</td></tr><tr><td>RBC</td><td>2</td><td>C0014772</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>RBC</td><td>2</td><td>C0014792</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>rDNA</td><td>2</td><td>C0012931</td><td></td><td></td><td>1976</td><td></td></tr><tr><td>rDNA</td><td>2</td><td>C0012933</td><td>1980</td><td>1978</td><td>1981</td><td>1981</td></tr><tr><td>Respiration</td><td>2</td><td>C0035203</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Respiration</td><td>2</td><td>C0282636</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Retinal</td><td>2</td><td>C0035298</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Retinal</td><td>2</td><td>C0035331</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Root</td><td>2</td><td>C0040452</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Root</td><td>2</td><td>C0242726</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>RSV</td><td>2</td><td>C0035236</td><td></td><td>1957</td><td>1960</td><td>1960</td></tr><tr><td>RSV</td><td>2</td><td>C0086943</td><td></td><td></td><td>1955</td><td></td></tr><tr><td>SARS</td><td>2</td><td>C1175175</td><td></td><td></td><td>2002</td><td></td></tr><tr><td>SARS</td><td>2</td><td>C1175743</td><td>2002</td><td>2002</td><td>2002</td><td>2002</td></tr><tr><td>SARS-assoc...</td><td>2</td><td>C1175175</td><td></td><td></td><td>2002</td><td></td></tr><tr><td>SCD</td><td>2</td><td>C1175743</td><td>2002</td><td>2002</td><td>2002</td><td>2002</td></tr></table>
+
+continued on next column or page
+
+<table><tr><td>Target</td><td>N</td><td>CUI</td><td>ET</td><td>LRET</td><td>FYO</td><td>MS</td></tr><tr><td>SCD</td><td>2</td><td>C0085298</td><td>1988</td><td>1987</td><td>1950</td><td>1989</td></tr><tr><td>Schistosoma...</td><td>2</td><td>C0036319</td><td></td><td></td><td>1971</td><td></td></tr><tr><td>Schistosoma...</td><td>2</td><td>C0036330</td><td></td><td>1981</td><td>1977</td><td>1985</td></tr><tr><td>SLS</td><td>2</td><td>C0037231</td><td></td><td>1987</td><td>1991</td><td>1991</td></tr><tr><td>SLS</td><td>2</td><td>C0037506</td><td></td><td></td><td>1971</td><td></td></tr><tr><td>Sodium</td><td>2</td><td>C0037473</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Sodium</td><td>2</td><td>C0037570</td><td>1945</td><td>1945</td><td>1945</td><td>1945</td></tr><tr><td>SPR</td><td>2</td><td>C0164209</td><td></td><td></td><td>1981</td><td></td></tr><tr><td>SPR</td><td>2</td><td>C0597731</td><td>1996</td><td>1994</td><td>1998</td><td>1998</td></tr><tr><td>SS</td><td>2</td><td>C0039101</td><td></td><td></td><td>1948</td><td></td></tr><tr><td>SS</td><td>2</td><td>C0085077</td><td>1990</td><td>1960</td><td>1964</td><td>1990</td></tr><tr><td>Staph</td><td>2</td><td>C0038160</td><td></td><td></td><td>1945</td><td>1945</td></tr><tr><td>Staph</td><td>2</td><td>C0038170</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>STEM</td><td>2</td><td>C0162731</td><td></td><td></td><td>1992</td><td></td></tr><tr><td>STEM</td><td>2</td><td>C0242767</td><td></td><td>1992</td><td>1994</td><td>1994</td></tr><tr><td>Sterilization</td><td>2</td><td>C0038280</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Sterilization</td><td>2</td><td>C0038288</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Strep</td><td>2</td><td>C0038395</td><td></td><td>1945</td><td>1945</td><td>1945</td></tr><tr><td>Strep</td><td>2</td><td>C0038402</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Synapsis</td><td>2</td><td>C0039062</td><td></td><td></td><td>1950</td><td></td></tr><tr><td>Synapsis</td><td>2</td><td>C0598501</td><td>1998</td><td>1950</td><td>1951</td><td>1951</td></tr><tr><td>TAT</td><td>3</td><td>C0017375</td><td>1988</td><td>1985</td><td>1989</td><td>1989</td></tr><tr><td>TAT</td><td>3</td><td>C0039341</td><td>1983</td><td>1982</td><td>1985</td><td>1985</td></tr><tr><td>TAT</td><td>3</td><td>C0039756</td><td></td><td></td><td>1975</td><td></td></tr><tr><td>Tax</td><td>2</td><td>C0039371</td><td></td><td></td><td>1975</td><td></td></tr><tr><td>Tax</td><td>2</td><td>C0144576</td><td>1992</td><td>1989</td><td>1983</td><td>1993</td></tr><tr><td>TEM</td><td>2</td><td>C0040975</td><td></td><td></td><td>2004</td><td></td></tr><tr><td>TEM</td><td>2</td><td>C0678118</td><td></td><td></td><td>2002</td><td></td></tr><tr><td>THYMUS</td><td>3</td><td>C0040112</td><td>1948</td><td>1946</td><td>1949</td><td>1949</td></tr><tr><td>THYMUS</td><td>3</td><td>C0040113</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>THYMUS</td><td>3</td><td>C1015036</td><td></td><td>1946</td><td>1946</td><td></td></tr><tr><td>TLC</td><td>2</td><td>C0008569</td><td></td><td>1959</td><td>1959</td><td>?</td></tr><tr><td>TLC</td><td>2</td><td>C0040509</td><td>1974</td><td>1972</td><td>1959</td><td>?</td></tr><tr><td>TMJ</td><td>2</td><td>C0039493</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>TMJ</td><td>2</td><td>C0039496</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>TMP</td><td>2</td><td>C0040079</td><td></td><td>1972</td><td>1975</td><td>1975</td></tr><tr><td>TMP</td><td>2</td><td>C0041041</td><td></td><td></td><td>1970</td><td></td></tr><tr><td>TNC</td><td>2</td><td>C0076088</td><td>1983</td><td>1982</td><td>1985</td><td>1985</td></tr><tr><td>TNC</td><td>2</td><td>C0077400</td><td></td><td></td><td>1980</td><td></td></tr><tr><td>TNT</td><td>2</td><td>C0041070</td><td></td><td></td><td>1982</td><td>1982</td></tr><tr><td>TNT</td><td>2</td><td>C0077404</td><td></td><td></td><td>1981</td><td></td></tr><tr><td>Tolerance</td><td>2</td><td>C0013220</td><td></td><td></td><td>1946</td><td>?</td></tr><tr><td>Tolerance</td><td>2</td><td>C0020963</td><td></td><td>1946</td><td>1946</td><td>?</td></tr><tr><td>tomography</td><td>2</td><td>C0040395</td><td></td><td></td><td>1947</td><td>?</td></tr><tr><td>tomography</td><td>2</td><td>C0040405</td><td></td><td></td><td>1947</td><td>?</td></tr><tr><td>Torula</td><td>2</td><td>C0010414</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Torula</td><td>2</td><td>C0010415</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>TPA</td><td>2</td><td>C0032143</td><td>1983</td><td>1982</td><td>1982</td><td>1985</td></tr><tr><td>TPA</td><td>2</td><td>C0039654</td><td></td><td></td><td>1975</td><td></td></tr><tr><td>TPO</td><td>2</td><td>C0021965</td><td>1974</td><td>1974</td><td>1975</td><td>1975</td></tr><tr><td>TPO</td><td>2</td><td>C0040052</td><td></td><td></td><td>1974</td><td></td></tr><tr><td>TRF</td><td>2</td><td>C0021759</td><td></td><td></td><td>1980</td><td>1980</td></tr><tr><td>TRF</td><td>2</td><td>C0040162</td><td></td><td></td><td>1968</td><td></td></tr><tr><td>TSF</td><td>2</td><td>C0021756</td><td>1976</td><td>1974</td><td>1977</td><td>1977</td></tr><tr><td>TSF</td><td>2</td><td>C0040052</td><td></td><td></td><td>1974</td><td></td></tr><tr><td>TYR</td><td>2</td><td>C0041484</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>TYR</td><td>2</td><td>C0041485</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>US</td><td>2</td><td>C0041618</td><td>1971</td><td>1964</td><td>1945</td><td>1966</td></tr><tr><td>US</td><td>2</td><td>C0041703</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>Ventricles</td><td>2</td><td>C0007799</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>Ventricles</td><td>2</td><td>C0018827</td><td></td><td></td><td>1945</td><td>?</td></tr><tr><td>veterinary</td><td>2</td><td>C0042615</td><td></td><td></td><td>1945</td><td></td></tr><tr><td>veterinary</td><td>2</td><td>C0206212</td><td></td><td>1959</td><td>1963</td><td>1993</td></tr><tr><td>Wasp</td><td>2</td><td>C0043041</td><td></td><td></td><td>1975</td><td></td></tr><tr><td>Wasp</td><td>2</td><td>C0258432</td><td>1993</td><td>1991</td><td>1994</td><td>1994</td></tr><tr><td>WBS</td><td>2</td><td>C0004903</td><td></td><td></td><td>1982</td><td></td></tr><tr><td>WBS</td><td>2</td><td>C0175702</td><td>1994</td><td>1991</td><td>1995</td><td>1995</td></tr><tr><td>WT1</td><td>2</td><td>C0027708</td><td></td><td></td><td>1946</td><td></td></tr><tr><td>WT1</td><td>2</td><td>C0148873</td><td>1991</td><td>1989</td><td>1991</td><td>1991</td></tr><tr><td>Yellow Fever</td><td>2</td><td>C0043395</td><td></td><td>1945</td><td>1945</td><td>?</td></tr><tr><td>Yellow Fever</td><td>2</td><td>C0301508</td><td></td><td></td><td>1945</td><td>?</td></tr></table>
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? Section 7  
+A2. Did you discuss any potential risks of your work? Not applicable. Left blank.  
+A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Section 1 and 4 and 5
+
+B1. Did you cite the creators of artifacts you used? Section 4 and 5  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 4  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 4 and 5  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. The data used in this research is a secondary data which was previously published. The data source files were taken from NML and is made of biomedical scientific publications.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 7  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 4
+
+C Did you run computational experiments?
+
+Section 4 and 5
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4 and 7
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 5  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 4 and 5
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank.

2023/Improving Diachronic Word Sense Induction with a Nonparametric Bayesian method/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0619f8c45bf11ece3eb649368398c219a210a5af46cdb250e771a3ac2574b64
+size 1259773

2023/Improving Embedding-based Unsupervised Keyphrase Extraction by Incorporating Structural Information/a6ac9ba6-311e-4129-8f7e-5207177224bf_content_list.json

@@ -0,0 +1,1320 @@
+[
+    {
+        "type": "text",
+        "text": "Improving Embedding-based Unsupervised Keyphrase Extraction by Incorporating Structural Information",
+        "text_level": 1,
+        "bbox": [
+            142,
+            89,
+            855,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Mingyang Song, Huafeng Liu*, Yi Feng, Liping Jing*",
+        "bbox": [
+            273,
+            149,
+            732,
+            167
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Beijing Key Lab of Traffic Data Analysis and Mining",
+        "bbox": [
+            285,
+            168,
+            717,
+            183
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Beijing Jiaotong University, Beijing, China",
+        "bbox": [
+            327,
+            184,
+            675,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "mingyang.song@bjtu.edu.cn",
+        "bbox": [
+            374,
+            200,
+            626,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Keyphrase extraction aims to extract a set of keyphrases with the central idea of the document. In a structured document, there are certain locations (e.g., the title or the first sentence) where a keyphrase is most likely to appear. However, when extracting keyphrases from the document, most existing embedding-based unsupervised keyphrase extraction models ignore the indicative role of the highlights in certain locations, leading to wrong keyphrases extraction. In this paper, we propose a new Highlight-Guided Unsupervised Keyphrase Extraction model (HGUKE) to address the above issue. Specifically, HGUKE first models the phrase-document relevance via the highlights of the documents. Next, HGUKE calculates the cross-phrase relevance between all candidate phrases. Finally, HGUKE aggregates the above two relevance as the importance score of each candidate to rank and extract keyphrases. The experimental results on three benchmarks demonstrate that HGUKE outperforms the state-of-the-art unsupervised keyphrase extraction baselines.",
+        "bbox": [
+            141,
+            278,
+            460,
+            605
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            615,
+            258,
+            631
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Keyphrase extraction is the fundamental task of automatically extracting a set of salient phrases from a document that concisely describes its primary content (Hasan and Ng, 2014; Song et al., 2023a). Figure 1 shows an example of the source document and its corresponding keyphrases.",
+        "bbox": [
+            112,
+            640,
+            487,
+            736
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent developments in pre-trained language models (Devlin et al., 2019) have heightened the need for utilizing pre-trained embeddings on natural language processing tasks, which significantly improves the performance of embedding-based unsupervised keyphrase extraction models (Sun et al., 2020; Liang et al., 2021; Zhang et al., 2022). Existing embedding-based models mainly consist of two components: candidate keyphrase extraction and keyphrase importance estimation (Hasan and",
+        "bbox": [
+            112,
+            738,
+            487,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Title:",
+        "bbox": [
+            514,
+            254,
+            544,
+            263
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Measuring keyboard response delays by comparing keyboard and joystick inputs Abstract:",
+        "bbox": [
+            514,
+            265,
+            858,
+            282
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The response characteristics of PC keyboards have to be identified when they are used as response devices in psychological experiments. In the past, the proposed method has been to check the characteristics independently by means of external measurement equipment. However, with the availability of different PC models and the rapid pace of model change, there is an urgent need for the development of convenient and accurate methods of checking. The method proposed consists of raising the precision of the PC's clock to the microsecond level and using a joystick connected to the MIDI terminal of a sound board to give the PC an independent timing function. Statistical processing of the data provided by this method makes it possible to estimate accurately the keyboard scanning interval time and the average keyboard delay time...",
+        "bbox": [
+            512,
+            284,
+            870,
+            382
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Keyphrases:",
+        "text_level": 1,
+        "bbox": [
+            515,
+            382,
+            579,
+            391
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "keyboard response delay measurement, joystick input, keyboard input, pc keyboard, psychological experiment, model change, check, pc clock precision, midi terminal, sound board, independent timing function, statistical data process, keyboard scan interval time, average keyboard delay time",
+        "bbox": [
+            514,
+            392,
+            873,
+            428
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: Randomly sampled document with its corresponding keyphrases from the benchmark keyphrase extraction dataset Inspec. Bold orange represents the content related to the title, and underlined indicates the content related to the first sentence.",
+        "bbox": [
+            507,
+            445,
+            882,
+            517
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ng, 2014; Song et al., 2021, 2022a). The former extracts continuous words from the document as candidate keyphrases through heuristic rules, and the latter estimates the importance of candidate phrases by matching similarity with their corresponding document.",
+        "bbox": [
+            507,
+            546,
+            882,
+            642
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Generally, the source document has both salient information and noises (redundant content). Hence, there may be a deviation when directly using the phrase-document relevance as the importance score of each candidate to select keyphrases. For many specific-domain documents (e.g., news or scientific articles), the highlights (the title or the first sentence) typically contains the central information of the source document (as shown in Figure 1), which has more significant guidance for extracting keyphrases. However, the recent embedding-based unsupervised keyphrase extraction models ignore the effect of the highlight information, leading to extract wrong keyphrases.",
+        "bbox": [
+            507,
+            644,
+            882,
+            869
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Motivated by the above issues, we propose a new Highlight-Guided Unsupervised Keyphrase Extraction model (HGUKE), which estimates the impor",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding Author",
+        "bbox": [
+            141,
+            904,
+            290,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1041",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 1041-1048",
+        "bbox": [
+            228,
+            945,
+            768,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            959,
+            699,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/0b5f62ff3c2da3709a5857f174293e861e38b41654e20fac1235f2730652b7ba.jpg",
+        "image_caption": [
+            "Figure 2: The model architecture of HGUKE."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            82,
+            480,
+            198
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tance score of each candidate phrase by jointly considering the global and local relevance between candidate phrases and their corresponding document. Concretely, HGUKE first calculates the global relevance by leveraging the highlights rather than the whole document and then locally computes the cross-phrase relevance between all candidate keyphrases, as illustrated in Figure 2. Finally, HGUKE aggregates the global and local relevance as the importance score of each candidate keyphrase to rank and extract keyphrases. Experimental results demonstrate that the proposed model HGUKE outperforms the recent state-of-the-art embedding-based unsupervised keyphrase extraction baselines on three benchmark keyphrase extraction datasets.",
+        "bbox": [
+            112,
+            252,
+            489,
+            508
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            523,
+            265,
+            539
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Candidate Keyphrase Extraction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            551,
+            420,
+            567
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To extract candidate keyphrases from the source document, we follow the previous studies (Liang et al., 2021; Song et al., 2022b; Ding and Luo, 2021) and leverage Stanford CoreNLP Tools<sup>1</sup> for tokenizing, part-of-speech tagging and noun phrase chunking. Concretely, in our model, the regular expression  $\\{<NN.*|JJ> * <NN.*>\\}$  is designed to extract noun phrases as the candidate keyphrases via the python package NLTK<sup>2</sup>.",
+        "bbox": [
+            112,
+            573,
+            489,
+            720
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Phrase and Document Encoding",
+        "text_level": 1,
+        "bbox": [
+            112,
+            732,
+            415,
+            747
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "After constructing a set of candidate keyphrases  $\\mathbf{P} = \\{p_1,\\dots ,p_i,\\dots ,p_{|\\mathbf{P}|}\\}$  for the source document via the above method, we adopt the pre-trained language model BERT (Devlin et al., 2019) as the embedding layer to obtain pre-trained word embeddings  $\\mathbf{H} = \\{h_1,\\dots ,h_m,\\dots ,h_{|\\mathbf{D}|}\\}$  for the source document  $\\mathbf{D} = \\{w_{1},\\ldots ,w_{m},\\ldots ,w_{|\\mathbf{D}|}\\}$  where  $h_m$  indicates the  $m$ -th word in the document.",
+        "bbox": [
+            112,
+            753,
+            489,
+            882
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Next, we leverage word embeddings to obtain candidate keyphrase representations. To capture the central semantics of the candidate keyphrases, we obtain candidate keyphrase representations by leveraging the max pooling operation, which is a simple and effective parameter-free approach and can be calculated as follows,",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh _ {p _ {i}} = \\operatorname {M a x} - \\operatorname {P o o l i n g} \\left(\\left\\{h _ {1}, \\dots , h _ {k}, \\dots , h _ {\\left| p _ {i} \\right|} \\right\\}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            522,
+            210,
+            882,
+            229
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $h_{p_i}$  is the representation of the  $i$ -th candidate keyphrase and  $|p_i|$  indicates the length of  $p_i$ . Specifically,  $h_k$  represents the word in the document associated with the candidate keyphrase  $p_i$ . At the same time, we use the mean pooling operation to obtain the highlight representation  $h_s$  of the document.",
+        "bbox": [
+            507,
+            239,
+            884,
+            351
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 Phrase-Document Relevance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            363,
+            781,
+            378
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To obtain more relevant candidates, we model the similarity between candidate phrases and the corresponding document as follows,",
+        "bbox": [
+            507,
+            384,
+            882,
+            432
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {i} ^ {h} = \\frac {1}{\\left| \\left| h _ {s} - h _ {p _ {i}} \\right| \\right| _ {1}}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            618,
+            443,
+            880,
+            478
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $p_i^h$  denotes the phrase-document relevance of  $i$ -th candidate keyphrases and  $||\\cdot||_1$  indicates the Manhattan Distance.",
+        "bbox": [
+            507,
+            489,
+            882,
+            536
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For news and scientific articles, keyphrases often appear at the beginning or front position (Florescu and Caragea, 2017a,b), which means that the position information is important and indicative for extracting keyphrases. For example, the word appearing at 2-th, 5-th and 10-th, has a weight  $\\rho_{i} = 1 / 2 + 1 / 5 + 1 / 10 = 0.8$ . Inspired by the previous work (Florescu and Caragea, 2017b; Liang et al., 2021), we adopt a position regularization as follows,  $\\rho_{i} = \\mathrm{softmax}(e^{1 / i})$ , where  $\\rho_{i}$  is the position regularization factor of the  $i$ -th candidate phrase. Then, the weighted phrase-document relevance  $\\hat{p}_i^h$  can be re-calculated as follows,",
+        "bbox": [
+            507,
+            538,
+            884,
+            747
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {p} _ {i} ^ {h} = p _ {i} ^ {h} \\cdot \\rho_ {i}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            643,
+            758,
+            880,
+            778
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Here, we finally employ  $\\hat{p}_i^h$  to estimate the phrase-document relevance of the  $i$ -th candidate phrase.",
+        "bbox": [
+            507,
+            790,
+            882,
+            822
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.4 Cross-Phrase Relevance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            834,
+            744,
+            848
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Generally, the phrase-document relevance is calculated between the highlight information and each candidate independently, and consequently, it cannot determine which candidates are better than the",
+        "bbox": [
+            507,
+            854,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ https://stanfordnlp.github.io/CoreNLP/",
+        "bbox": [
+            134,
+            890,
+            381,
+            904
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://github.com/ntlk",
+        "bbox": [
+            136,
+            904,
+            282,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1042",
+        "bbox": [
+            482,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/c71e558f80b1bb4bc92ea9aa0871fd3695b8948de2792a93a971f0c406ada171.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">DUC2001</td><td colspan=\"3\">Inspec</td><td colspan=\"3\">SemEval2010</td></tr><tr><td>F1@5</td><td>F1@10</td><td>F1@15</td><td>F1@5</td><td>F1@10</td><td>F1@15</td><td>F1@5</td><td>F1@10</td><td>F1@15</td></tr><tr><td colspan=\"10\">Statistical Keyphrase Extraction Models</td></tr><tr><td>TF-IDF (Jones, 2004)</td><td>9.21</td><td>10.63</td><td>11.06</td><td>11.28</td><td>13.88</td><td>13.83</td><td>2.81</td><td>3.48</td><td>3.91</td></tr><tr><td>YAKE (Campos et al., 2018)</td><td>12.27</td><td>14.37</td><td>14.76</td><td>18.08</td><td>19.62</td><td>20.11</td><td>11.76</td><td>14.4</td><td>15.19</td></tr><tr><td colspan=\"10\">Graph-based Keyphrase Extraction Models</td></tr><tr><td>TextRank (Mihalcea and Tarau, 2004)</td><td>11.80</td><td>18.28</td><td>20.22</td><td>27.04</td><td>25.08</td><td>36.65</td><td>3.80</td><td>5.38</td><td>7.65</td></tr><tr><td>SingleRank (Wan and Xiao, 2008)</td><td>20.43</td><td>25.59</td><td>25.70</td><td>27.79</td><td>34.46</td><td>36.05</td><td>5.90</td><td>9.02</td><td>10.58</td></tr><tr><td>TopicRank (Bougouin et al., 2013)</td><td>21.56</td><td>23.12</td><td>20.87</td><td>25.38</td><td>28.46</td><td>29.49</td><td>12.12</td><td>12.90</td><td>13.54</td></tr><tr><td>PositionRank (Florescu and Caragea, 2017b)</td><td>23.35</td><td>28.57</td><td>28.60</td><td>28.12</td><td>32.87</td><td>33.32</td><td>9.84</td><td>13.34</td><td>14.33</td></tr><tr><td>MultipartiteRank (Boudin, 2018)</td><td>23.20</td><td>25.00</td><td>25.24</td><td>25.96</td><td>29.57</td><td>30.85</td><td>12.13</td><td>13.79</td><td>14.92</td></tr><tr><td colspan=\"10\">Embedding-based Keyphrase Extraction Models</td></tr><tr><td>EmbedRankd2v (Bennani-Smires et al., 2018)</td><td>24.02</td><td>28.12</td><td>28.82</td><td>31.51</td><td>37.94</td><td>37.96</td><td>3.02</td><td>5.08</td><td>7.23</td></tr><tr><td>KeyGames (Saxena et al., 2020)</td><td>24.42</td><td>28.28</td><td>29.77</td><td>32.12</td><td>40.48</td><td>40.94</td><td>11.93</td><td>14.35</td><td>14.62</td></tr><tr><td>SIFRank (Sun et al., 2020)</td><td>24.27</td><td>27.43</td><td>27.86</td><td>29.11</td><td>38.80</td><td>39.59</td><td>-</td><td>-</td><td>-</td></tr><tr><td>JointGL (Liang et al., 2021)</td><td>28.62</td><td>35.52</td><td>36.29</td><td>32.61</td><td>40.17</td><td>41.09</td><td>13.02</td><td>19.35</td><td>21.72</td></tr><tr><td>MDERank (Zhang et al., 2022)</td><td>23.31</td><td>26.65</td><td>26.42</td><td>27.85</td><td>34.36</td><td>36.40</td><td>13.05</td><td>18.27</td><td>20.35</td></tr><tr><td>HGUKE</td><td>31.31</td><td>37.24</td><td>38.31</td><td>34.18</td><td>41.05</td><td>42.16</td><td>14.07</td><td>20.52</td><td>23.10</td></tr></table>",
+        "bbox": [
+            122,
+            82,
+            877,
+            332
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Performance of the selected baselines and our model on DUC2001, Inspec and SemEval2010 test sets. F1 scores on the top 5, 10, and 15 keyphrases are reported. The best results are bolded in the table.",
+        "bbox": [
+            112,
+            348,
+            880,
+            378
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "others. To determine which candidate phrases are more salient than the others, we sum the semantic relatedness between the  $i$ -th candidate phrases and all candidates as the cross-phrase relevance. Thus, it calculates the local relevance as follows,",
+        "bbox": [
+            112,
+            401,
+            487,
+            481
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {i} ^ {p} = \\sum_ {j = 1, j \\neq i} \\left(h _ {p _ {i}} h _ {p _ {j}} ^ {\\top} - \\lambda \\delta_ {i}\\right). \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            183,
+            494,
+            485,
+            529
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\delta_{i} = \\mathrm{Mean}(\\sum_{j = 1,j\\neq i}h_{p_i}h_{p_j}^\\top)$ . Here, we treat  $\\delta_{i}$  as a de-noisy factor to filter the noises, which is far different from the  $i$ -th candidate keyphrase in the document.",
+        "bbox": [
+            112,
+            541,
+            487,
+            606
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.5 Relevance Aggregation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            618,
+            341,
+            634
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We aggregate the phrase-document relevance and the cross-phrase relevance into a whole score as the importance score of each candidate via a simple multiplication,",
+        "bbox": [
+            112,
+            639,
+            487,
+            703
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nr _ {i} = \\hat {p} _ {i} ^ {h} \\cdot p _ {i} ^ {p} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            253,
+            715,
+            485,
+            734
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $r_i$  indicates the importance score of the  $i$ -th candidate phrase. Then, we rank all candidates with their importance score  $r_i$  and extract top-ranked  $k$  phrases as keyphrases of the source document.",
+        "bbox": [
+            112,
+            746,
+            487,
+            812
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Experiments and Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            822,
+            366,
+            840
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            848,
+            334,
+            865
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This paper conducts experiments on three benchmark and popular used keyphrase datasets, which includes DUC2001 (Wan and Xiao, 2008), Inspec",
+        "bbox": [
+            112,
+            870,
+            487,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/31ff37f3be4c19f5a7edca88e73628e8cb7292dd3f8e70a429de9097f23db3a0.jpg",
+        "image_caption": [
+            "Figure 3: The results of calculating the phrase-document relevance via the whole document and the highlights."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            403,
+            873,
+            531
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(Hulth, 2003), and SemEval2010 (Kim et al., 2010). Due to page limits, please refer to the corresponding articles for the details of the three datasets.",
+        "bbox": [
+            507,
+            596,
+            882,
+            645
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following the previous work (Liang et al., 2021; Ding and Luo, 2021; Song et al., 2023b), we use the standard practice and evaluate the performance of our model in terms of f-measure at the top-K keyphrases (F1@K) and adopt stemming to both extracted keyphrases and gold truth. Concretely, we report F1@5, F1@10, and F1@15 of each model on three benchmark datasets.",
+        "bbox": [
+            507,
+            646,
+            882,
+            772
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We adopt the pre-trained language model BERT (Devlin et al., 2019) as the backbone of our model, initialized from their pre-trained weights. In our experiments,  $\\lambda$  is set to 0.9 for three benchmark datasets.",
+        "bbox": [
+            507,
+            774,
+            882,
+            853
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Overall Performance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            721,
+            881
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1 shows the performance of baselines and our model on three benchmark datasets (DUC2001, In",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1043",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/9dfd171f859cb53fbf496c54f76d6701935c1ff167f9e59764589bfc223609e0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Pooling Methods</td><td colspan=\"3\">DUC2001</td><td colspan=\"3\">Inspec</td><td colspan=\"3\">SemEval2010</td></tr><tr><td>F1@5</td><td>F1@10</td><td>F1@15</td><td>F1@5</td><td>F1@10</td><td>F1@15</td><td>F1@5</td><td>F1@10</td><td>F1@15</td></tr><tr><td>Max-Pooling</td><td>25.43</td><td>33.24</td><td>36.10</td><td>33.95</td><td>41.21</td><td>42.12</td><td>9.92</td><td>17.20</td><td>21.54</td></tr><tr><td>Mean-Pooling</td><td>31.31</td><td>37.24</td><td>38.31</td><td>34.18</td><td>41.05</td><td>42.16</td><td>14.07</td><td>20.52</td><td>23.10</td></tr></table>",
+        "bbox": [
+            134,
+            82,
+            862,
+            156
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/4753e3f875817d0f9b1edb2306ee2bff96429962f47f002fd3426cfdf3c987cc.jpg",
+        "table_caption": [
+            "Table 2: The results of different pooling methods for document embedding."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Different Similarity Measures</td><td colspan=\"3\">DUC2001</td><td colspan=\"3\">Inspec</td><td colspan=\"3\">SemEval2010</td></tr><tr><td>F1@5</td><td>F1@10</td><td>F1@15</td><td>F1@5</td><td>F1@10</td><td>F1@15</td><td>F1@5</td><td>F1@10</td><td>F1@15</td></tr><tr><td>Cosine Similarity</td><td>30.26</td><td>36.14</td><td>37.19</td><td>33.70</td><td>40.51</td><td>41.53</td><td>13.15</td><td>20.78</td><td>23.26</td></tr><tr><td>Euclidean Distance</td><td>30.67</td><td>36.65</td><td>37.95</td><td>34.04</td><td>41.02</td><td>42.24</td><td>13.43</td><td>20.18</td><td>23.44</td></tr><tr><td>Manhattan Distance</td><td>31.31</td><td>37.24</td><td>38.31</td><td>34.18</td><td>41.05</td><td>42.16</td><td>14.07</td><td>20.52</td><td>23.10</td></tr></table>",
+        "bbox": [
+            134,
+            202,
+            862,
+            293
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 3: The results of different similarity measure methods for the phrase-document relevance.",
+        "bbox": [
+            169,
+            309,
+            823,
+            324
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "spec, and SemEval2010). The results show that our method significantly improves over state-of-the-art unsupervised keyphrase extraction baselines. Compared with the current state-of-the-art models, our model achieves significantly better performance on F1@5, F1@10, and F1@15 evaluation metrics, demonstrating the effectiveness of estimating the importance of candidate phrases by leveraging the highlights to calculate the relevance.",
+        "bbox": [
+            112,
+            349,
+            487,
+            492
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Compared with EmbedRank (Bennani-Smires et al., 2018), KeyGames (Saxena et al., 2020), and SIFRank (Sun et al., 2020), HGUKE achieves significant improvement, which benefits from using the highlights to calculate the importance score of each candidate keyphrase. Compared with the best baseline JointGL, our model achieves better performance on several benchmark keyphrase extraction datasets in all evaluation metrics. The main reason for this improvement is that we use the highlights as the guidance information instead of the whole document when estimating the importance of keyphrases.",
+        "bbox": [
+            112,
+            495,
+            489,
+            705
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Ablation Test",
+        "text_level": 1,
+        "bbox": [
+            112,
+            719,
+            265,
+            733
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The ablation experiments on three benchmark keyphrase extraction datasets are shown in Figure 3. It can be seen from the results that using the highlight information can significantly improve the performance of keyphrase extraction, which benefits from estimating the importance score of each candidate by using its corresponding highlight information rather than the whole document. We consider the main reason is that the title or the first sentence of the document usually has a strong guidance for extracting keyphrases.",
+        "bbox": [
+            112,
+            741,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4 Impact of Pooling Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            349,
+            768,
+            365
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we study different pooling methods, including mean- and max-pooling operations. For all pooling methods, HGUKE using the last BERT layer achieves the best results, demonstrating that HGUKE benefits from stronger contextualized semantic representations. We can see the results in Table 2 that the document encoded via the mean-pooling operation obtains the best performance.",
+        "bbox": [
+            507,
+            370,
+            884,
+            499
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.5 Impact of Different Similarity Measures",
+        "text_level": 1,
+        "bbox": [
+            507,
+            508,
+            870,
+            524
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our model adopts Manhattan Distance to measure the textual similarity between candidate phrases and the highlight information. Furthermore, we attempt to employ different measures to estimate the phrase-document relevance. The results of different similarity measures are shown in Table 3, and we can see that the advantage of Manhattan Distance is obvious.",
+        "bbox": [
+            507,
+            529,
+            882,
+            657
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            669,
+            663,
+            684
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Most existing unsupervised keyphrase extraction methods can be mainly divided into four categories: statistics-based, topic-based, graph-based, and embedding-based models. Specifically, statistics-based models (Salton and Buckley, 1988; Witten et al., 1999) usually extract keyphrases by estimating the importance of candidate phrases with different statistic features, such as word frequency feature, phrase position feature, linguistic features of natural language, etc. Topic-based models (Liu et al., 2009, 2010) typically utilize topic information to determine whether a candidate phrase is a keyphrase. Graph-based models (Mihalcea and Tarau, 2004; Grineva et al., 2009) represent the",
+        "bbox": [
+            507,
+            694,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1044",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "document as a graph and rank candidate phrases by graph-based similarities.",
+        "bbox": [
+            112,
+            84,
+            485,
+            115
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Embedding-based models usually adopt the pretrained embeddings to obtain document and candidate phrase representations and calculate the importance score of each candidate depending on the obtained representations. Benefiting from the development of transformer-based pre-trained language models (Devlin et al., 2019) in the natural language processing field, embedding-based models (Bennani-Smires et al., 2018; Sun et al., 2020; Liang et al., 2021) have achieved outstanding performance. Concretely, embedding-based models mainly consist of two procedures: candidate keyphrase representation and keyphrase importance estimation (Hasan and Ng, 2014; Song et al., 2023a). The first procedure utilizes natural language linguistic features to construct candidate keyphrases and represents them by pre-trained embedding approaches (e.g., BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019)). The second procedure estimates the importance of candidate phrases from different perspectives to determine whether a candidate phrase is a keyphrase.",
+        "bbox": [
+            115,
+            117,
+            489,
+            470
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Unlike the existing unsupervised keyphrase extraction models, we use the highlight information of the document to calculate the phrase-document relevance instead the whole document.",
+        "bbox": [
+            112,
+            472,
+            487,
+            535
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            548,
+            401,
+            564
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we incorporate structural information to improve the performance of embedding-based unsupervised keyphrase extraction. Specifically, in this paper, we propose a new Highlight-Guided Unsupervised Keyphrase Extraction model (HGUKE), which calculates the phrase-document relevance via the highlight information instead of the whole document to select relevant candidate phrases. Extensive experiments demonstrate that HGUKE outperforms the state-of-the-art unsupervised baselines. Future research may investigate adopting different structural information of the source document to improve the performance of unsupervised keyphrase extraction.",
+        "bbox": [
+            112,
+            574,
+            489,
+            800
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            813,
+            307,
+            829
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We thank the three anonymous reviewers for carefully reading our paper and their insightful comments and suggestions. This work was partly supported by the Fundamental Research Funds for the Central Universities (2019JBZ110); the National",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Natural Science Foundation of China under Grant 62176020; the National Key Research and Development Program (2020AAA0106800); the Beijing Natural Science Foundation under Grant L211016; CAAI-Huawei MindSpore Open Fund; and Chinese Academy of Sciences (OEIP-O-202004).",
+        "bbox": [
+            507,
+            84,
+            884,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            192,
+            645,
+            208
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "There are still some limitations of our work. In the future, we plan to enhance the procedure of extracting candidate keyphrase, to improve the upper bound of the performance of keyphrase extraction. One possible way is to generate candidate phrases of the document by utilizing the high-level semantic relatedness (e.g., attention weights) instead of using the surface-or syntactic-level information.",
+        "bbox": [
+            507,
+            218,
+            884,
+            347
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            374,
+            608,
+            388
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Kamil Bennani-Smires, Claudiu Musat, Andreae Hossmann, Michael Baeriswyl, and Martin Jaggi. 2018. Simple unsupervised keyphrase extraction using sentence embeddings. In CoNLL, pages 221-229. Association for Computational Linguistics.",
+            "Florian Boudin. 2018. Unsupervised keyphrase extraction with multipartite graphs. In *NAACL-HLT* (2), pages 667-672. Association for Computational Linguistics.",
+            "Adrien Bougouin, Florian Boudin, and Beatrice Daille. 2013. Topicrank: Graph-based topic ranking for keyphrase extraction. In IJCNLP, pages 543-551. Asian Federation of Natural Language Processing / ACL.",
+            "Ricardo Campos, Vitor Mangaravite, Arian Pasquali, Alipio Mario Jorge, Celia Nunes, and Adam Jatowt. 2018. Yake! collection-independent automatic keyword extractor. In ECIR, volume 10772 of Lecture Notes in Computer Science, pages 806-810. Springer.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In *NAACL-HLT*, pages 4171–4186. Association for Computational Linguistics.",
+            "Haoran Ding and Xiao Luo. 2021. Attentionrank: Unsupervised keyphrase extraction using self and cross attentions. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1919-1928.",
+            "Corina Florescu and Cornelia Caragea. 2017a. A position-biased pagerank algorithm for keyphrase extraction. In AAAI, pages 4923-4924. AAAI Press.",
+            "Corina Florescu and Cornelia Caragea. 2017b. Positionrank: An unsupervised approach to keyphrase"
+        ],
+        "bbox": [
+            509,
+            395,
+            885,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1045",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "extraction from scholarly documents. In ACL (1), pages 1105-1115. Association for Computational Linguistics.",
+            "Maria P. Grineva, Maxim N. Grinev, and Dmitry Lizorkin. 2009. Extracting key terms from noisy and multitheme documents. In WWW, pages 661-670. ACM.",
+            "Kazi Saidul Hasan and Vincent Ng. 2014. Automatic keyphrase extraction: A survey of the state of the art. In ACL (1), pages 1262-1273. The Association for Computer Linguistics.",
+            "Anette Hulth. 2003. Improved automatic keyword extraction given more linguistic knowledge. In EMNLP.",
+            "Karen Sparck Jones. 2004. A statistical interpretation of term specificity and its application in retrieval. J. Documentation, 60(5):493-502.",
+            "Su Nam Kim, Olena Medelyan, Min-Yen Kan, and Timothy Baldwin. 2010. SemEval-2010 task 5: Automatic keyphrase extraction from scientific articles. In SemEval@ACL, pages 21-26. The Association for Computer Linguistics.",
+            "Xinnian Liang, Shuangzhi Wu, Mu Li, and Zhoujun Li. 2021. Unsupervised keyphrase extraction by jointly modeling local and global context. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 155-164, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Feifan Liu, Deana Pennell, Fei Liu, and Yang Liu. 2009. Unsupervised approaches for automatic keyword extraction using meeting transcripts. In HLT-NAACL, pages 620-628. The 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. 2019. Roberta: A robustly optimized bert pretraining approach. CoRR, abs/1907.11692.",
+            "Zhiyuan Liu, Wenyi Huang, Yabin Zheng, and Maosong Sun. 2010. Automatic keyphrase extraction via topic decomposition. In EMNLP, pages 366-376. ACL.",
+            "Rada Mihalcea and Paul Tarau. 2004. Textrank: Bringing order into text. In EMNLP, pages 404-411. ACL.",
+            "Gerard Salton and Chris Buckley. 1988. Term weighting approaches in automatic text retrieval. Information Processing and Management, 24:513-523. Also available in Sparck Jones and Willett (1997).",
+            "Arnav Saxena, Mudit Mangal, and Goonjan Jain. 2020. Keygames: A game theoretic approach to automatic keyphrase extraction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2037-2048."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Mingyang Song, Yi Feng, and Liping Jing. 2022a. Hyperbolic relevance matching for neural keyphrase extraction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2022, Seattle, WA, United States, July 10-15, 2022, pages 5710-5720. Association for Computational Linguistics.",
+            "Mingyang Song, Yi Feng, and Liping Jing. 2022b. Utilizing BERT intermediate layers for unsupervised keyphrase extraction. In Proceedings of the 5th International Conference on Natural Language and Speech Processing (ICNLSP 2022), pages 277-281, Trento, Italy. Association for Computational Linguistics.",
+            "Mingyang Song, Yi Feng, and Liping Jing. 2023a. A survey on recent advances in keyphrase extraction from pre-trained language models. In *Findings of the Association for Computational Linguistics: EACL* 2023, pages 2153-2164, Dubrovnik, Croatia. Association for Computational Linguistics.",
+            "Mingyang Song, Liping Jing, and Lin Xiao. 2021. Importance Estimation from Multiple Perspectives for Keyphrase Extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Mingyang Song, Lin Xiao, and Liping Jing. 2023b. Learning to extract from multiple perspectives for neural keyphrase extraction. Computer Speech & Language, 81:101502.",
+            "Yi Sun, Hangping Qiu, Yu Zheng, Zhongwei Wang, and Chaoran Zhang. 2020. Sifrank: A new baseline for unsupervised keyphrase extraction based on pre-trained language model. IEEE Access, 8:10896-10906.",
+            "Xiaojun Wan and Jianguo Xiao. 2008. Single document keyphrase extraction using neighborhood knowledge. In AAAI, pages 855-860. AAAI Press.",
+            "Ian H. Witten, Gordon W. Paynter, Eibe Frank, Carl Gutwin, and Craig G. Nevill-Manning. 1999. Kea: Practical automatic keyphrase extraction. In ACM DL, pages 254-255. ACM.",
+            "Linhan Zhang, Qian Chen, Wen Wang, Chong Deng, Shiliang Zhang, Bing Li, Wei Wang, and Xin Cao. 2022. Mderank: A masked document embedding rank approach for unsupervised keyphrase extraction. In *Findings of the Association for Computational Linguistics: ACL* 2022, Dublin, Ireland, May 22-27, 2022, pages 396-409. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            848
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1046",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? 7",
+            "A2. Did you discuss any potential risks of your work? 7",
+            "A3. Do the abstract and introduction summarize the paper's main claims?",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            695,
+            288
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            115,
+            299,
+            489,
+            315
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3",
+        "bbox": [
+            134,
+            321,
+            146,
+            332
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? 3",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? 3",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? 3",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3"
+        ],
+        "bbox": [
+            129,
+            346,
+            880,
+            752
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            115,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3",
+        "bbox": [
+            134,
+            788,
+            146,
+            800
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3",
+        "bbox": [
+            129,
+            812,
+            880,
+            858
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            866,
+            877,
+            889
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1047",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 3",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? 3",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?"
+        ],
+        "bbox": [
+            129,
+            83,
+            880,
+            279
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response."
+        ],
+        "bbox": [
+            127,
+            340,
+            880,
+            640
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1048",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    }
+]