Add Batch 272281d1-5277-47f6-9013-555f1a9b3810 data

.gitattributes

@@ -6937,3 +6937,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/Enabling[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Generate[[:space:]]Text[[:space:]]with[[:space:]]Citations/74c2693d-b972-4cd1-be80-444c065636c2_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/End-to-End[[:space:]]Single-Channel[[:space:]]Speaker-Turn[[:space:]]Aware[[:space:]]Conversational[[:space:]]Speech[[:space:]]Translation/38626d7d-d3ae-428e-8764-fc579b38b613_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/End-to-end[[:space:]]Task-oriented[[:space:]]Dialogue_[[:space:]]A[[:space:]]Survey[[:space:]]of[[:space:]]Tasks,[[:space:]]Methods,[[:space:]]and[[:space:]]Future[[:space:]]Directions/28309954-d2f5-4a43-8aa6-31dcf545f556_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Biomedical[[:space:]]Lay[[:space:]]Summarisation[[:space:]]with[[:space:]]External[[:space:]]Knowledge[[:space:]]Graphs/49634b26-4af9-4204-9100-ca8a3aea5c79_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Chat[[:space:]]Language[[:space:]]Models[[:space:]]by[[:space:]]Scaling[[:space:]]High-quality[[:space:]]Instructional[[:space:]]Conversations/b1f4f81f-0f2a-4ef5-813b-a9152c0d03d7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Code-Switching[[:space:]]for[[:space:]]Cross-lingual[[:space:]]SLU_[[:space:]]A[[:space:]]Unified[[:space:]]View[[:space:]]of[[:space:]]Semantic[[:space:]]and[[:space:]]Grammatical[[:space:]]Coherence/b8a9adec-48b2-48e3-8fb6-523de3c4eab2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Computation[[:space:]]Efficiency[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Weight[[:space:]]and[[:space:]]Activation[[:space:]]Quantization/811af1b7-d7f8-4044-9ba8-d543804a67ed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Generative[[:space:]]Retrieval[[:space:]]with[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]from[[:space:]]Relevance[[:space:]]Feedback/8a8868ae-16b0-40bb-b3ea-fdf45defe7f6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Low-resource[[:space:]]Fine-grained[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition[[:space:]]by[[:space:]]Leveraging[[:space:]]Coarse-grained[[:space:]]Datasets/dff73b45-756f-4d7a-a8f6-789c5d1784de_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Structured[[:space:]]Evidence[[:space:]]Extraction[[:space:]]for[[:space:]]Fact[[:space:]]Verification/ebbbd673-d241-414d-91ff-1168a052ced6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Task-oriented[[:space:]]Dialogue[[:space:]]Systems[[:space:]]with[[:space:]]Generative[[:space:]]Post-processing[[:space:]]Networks/ff4f198b-74ec-4f81-bac0-82b9a8b9efc4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Textbooks[[:space:]]with[[:space:]]Visuals[[:space:]]from[[:space:]]the[[:space:]]Web[[:space:]]for[[:space:]]Improved[[:space:]]Learning/dc473d46-c5f9-437f-83b3-208c03329f2a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]Uncertainty-Based[[:space:]]Hallucination[[:space:]]Detection[[:space:]]with[[:space:]]Stronger[[:space:]]Focus/cda02f1a-0de9-4677-aa94-fd534e0ef30f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Enhancing[[:space:]]the[[:space:]]Ranking[[:space:]]Context[[:space:]]of[[:space:]]Dense[[:space:]]Retrieval[[:space:]]through[[:space:]]Reciprocal[[:space:]]Nearest[[:space:]]Neighbors/61e28cbf-0316-4d7a-b234-c2599fc3167a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/EntSUMv2_[[:space:]]Dataset,[[:space:]]Models[[:space:]]and[[:space:]]Evaluation[[:space:]]for[[:space:]]More[[:space:]]Abstractive[[:space:]]Entity-Centric[[:space:]]Summarization/3ea6bb2d-6c35-4eb4-968f-323b9376f651_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/EpiK-Eval_[[:space:]]Evaluation[[:space:]]for[[:space:]]Language[[:space:]]Models[[:space:]]as[[:space:]]Epistemic[[:space:]]Models/acd931ec-4343-4760-9dea-715b833b13fa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Establishing[[:space:]]Trustworthiness_[[:space:]]Rethinking[[:space:]]Tasks[[:space:]]and[[:space:]]Model[[:space:]]Evaluation/a0ed954d-4f48-4c09-a0d0-b7abf4c4f376_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/EtiCor_[[:space:]]Corpus[[:space:]]for[[:space:]]Analyzing[[:space:]]LLMs[[:space:]]for[[:space:]]Etiquettes/ee5863b7-77ee-4c41-bf44-2bccc4b2a896_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Euphemistic[[:space:]]Abuse[[:space:]]–[[:space:]]A[[:space:]]New[[:space:]]Dataset[[:space:]]and[[:space:]]Classification[[:space:]]Experiments[[:space:]]for[[:space:]]Implicitly[[:space:]]Abusive[[:space:]]Language/0c031145-d747-44d7-adbb-ed27609f9a77_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluating[[:space:]]Bias[[:space:]]and[[:space:]]Fairness[[:space:]]in[[:space:]]Gender-Neutral[[:space:]]Pretrained[[:space:]]Vision-and-Language[[:space:]]Models/7fd84048-0a15-44a8-be67-8df721564750_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluating[[:space:]]Cross-Domain[[:space:]]Text-to-SQL[[:space:]]Models[[:space:]]and[[:space:]]Benchmarks/586821f9-a5c2-4b29-a96d-378c12f3a8da_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluating[[:space:]]Evaluation[[:space:]]Metrics_[[:space:]]A[[:space:]]Framework[[:space:]]for[[:space:]]Analyzing[[:space:]]NLG[[:space:]]Evaluation[[:space:]]Metrics[[:space:]]using[[:space:]]Measurement[[:space:]]Theory/2e161355-e7a4-4e04-afbd-7d80bcced85b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluating[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]on[[:space:]]Controlled[[:space:]]Generation[[:space:]]Tasks/66b1b0ad-47f1-4040-9e0d-bbd71e78f70c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluating[[:space:]]Object[[:space:]]Hallucination[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/7f273bea-6bc8-4ce4-bb05-a3226eadb7fb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluating[[:space:]]and[[:space:]]Modeling[[:space:]]Attribution[[:space:]]for[[:space:]]Cross-Lingual[[:space:]]Question[[:space:]]Answering/5ced1149-926b-47eb-8bfe-c15a656ac7aa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluating[[:space:]]the[[:space:]]Rationale[[:space:]]Understanding[[:space:]]of[[:space:]]Critical[[:space:]]Reasoning[[:space:]]in[[:space:]]Logical[[:space:]]Reading[[:space:]]Comprehension/a65c9025-63a3-4a67-86f6-a36a8f88de03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluation[[:space:]]Metrics[[:space:]]in[[:space:]]the[[:space:]]Era[[:space:]]of[[:space:]]GPT-4_[[:space:]]Reliably[[:space:]]Evaluating[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]on[[:space:]]Sequence[[:space:]]to[[:space:]]Sequence[[:space:]]Tasks/ce88d0a0-d4fc-4009-8f69-383cd3b6028f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Evaluation[[:space:]]of[[:space:]]African[[:space:]]American[[:space:]]Language[[:space:]]Bias[[:space:]]in[[:space:]]Natural[[:space:]]Language[[:space:]]Generation/e4af9ac7-45e8-4c93-a80a-08f1edd00446_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Event[[:space:]]Causality[[:space:]]Extraction[[:space:]]via[[:space:]]Implicit[[:space:]]Cause-Effect[[:space:]]Interactions/53824f97-c68a-4af7-83f9-be42aba637b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Event[[:space:]]Ontology[[:space:]]Completion[[:space:]]with[[:space:]]Hierarchical[[:space:]]Structure[[:space:]]Evolution[[:space:]]Networks/a4caeb94-20c4-4fb4-b977-906a351f0faf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Event-Location[[:space:]]Tracking[[:space:]]in[[:space:]]Narratives_[[:space:]]A[[:space:]]Case[[:space:]]Study[[:space:]]on[[:space:]]Holocaust[[:space:]]Testimonies/c0b1634a-6fda-4ff8-8d08-f722d3db9a73_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exchange-of-Thought_[[:space:]]Enhancing[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Capabilities[[:space:]]through[[:space:]]Cross-Model[[:space:]]Communication/f155a336-0387-402f-928e-4e0ca9014070_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Expand,[[:space:]]Highlight,[[:space:]]Generate_[[:space:]]RL-driven[[:space:]]Document[[:space:]]Generation[[:space:]]for[[:space:]]Passage[[:space:]]Reranking/d671f36d-dcf3-4c1f-8e45-6095b60b46fd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Explaining[[:space:]]Interactions[[:space:]]Between[[:space:]]Text[[:space:]]Spans/c01c268a-ad37-4739-b853-04e66f7d6eb8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Explaining[[:space:]]with[[:space:]]Contrastive[[:space:]]Phrasal[[:space:]]Highlighting_[[:space:]]A[[:space:]]Case[[:space:]]Study[[:space:]]in[[:space:]]Assisting[[:space:]]Humans[[:space:]]to[[:space:]]Detect[[:space:]]Translation[[:space:]]Differences/0a4df744-a06f-4361-b527-41b0e580fb44_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Explanation[[:space:]]Selection[[:space:]]Using[[:space:]]Unlabeled[[:space:]]Data[[:space:]]for[[:space:]]Chain-of-Thought[[:space:]]Prompting/3861689a-3ac1-4854-96ea-301f769df4b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Explicit[[:space:]]Planning[[:space:]]Helps[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Logical[[:space:]]Reasoning/f037d74c-fd41-4003-ba29-e01eb53bea66_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploiting[[:space:]]Asymmetry[[:space:]]for[[:space:]]Synthetic[[:space:]]Training[[:space:]]Data[[:space:]]Generation_[[:space:]]SynthIE[[:space:]]and[[:space:]]the[[:space:]]Case[[:space:]]of[[:space:]]Information[[:space:]]Extraction/5ddde29a-15c0-4c2d-9c1d-ade31a8c956d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Explore-Instruct_[[:space:]]Enhancing[[:space:]]Domain-Specific[[:space:]]Instruction[[:space:]]Coverage[[:space:]]through[[:space:]]Active[[:space:]]Exploration/82e05807-772c-435a-b5a5-a00f4872b2da_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploring[[:space:]]All-In-One[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]Framework[[:space:]]for[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/1796c8ad-819d-4dbe-ac4b-b4fdd6f532d1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploring[[:space:]]Chain[[:space:]]of[[:space:]]Thought[[:space:]]Style[[:space:]]Prompting[[:space:]]for[[:space:]]Text-to-SQL/02f9d2e0-6bf4-4d1c-9967-d488482ee816_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploring[[:space:]]Discourse[[:space:]]Structure[[:space:]]in[[:space:]]Document-level[[:space:]]Machine[[:space:]]Translation/d69072b6-8f7a-4b25-aa89-cc7ce4deafda_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploring[[:space:]]Distributional[[:space:]]Shifts[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Code[[:space:]]Analysis/148247d6-8b25-4d5b-8cc2-e1e310a8f164_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploring[[:space:]]Jiu-Jitsu[[:space:]]Argumentation[[:space:]]for[[:space:]]Writing[[:space:]]Peer[[:space:]]Review[[:space:]]Rebuttals/37e11498-313a-4eb5-9388-51008795ef02_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploring[[:space:]]Linguistic[[:space:]]Probes[[:space:]]for[[:space:]]Morphological[[:space:]]Generalization/6ad4d12a-95a3-4e9a-8230-a7837c64c00d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploring[[:space:]]the[[:space:]]Boundaries[[:space:]]of[[:space:]]GPT-4[[:space:]]in[[:space:]]Radiology/55165828-71ed-4909-9d4b-3f779b5f0bbd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Exploring[[:space:]]the[[:space:]]Impact[[:space:]]of[[:space:]]Model[[:space:]]Scaling[[:space:]]on[[:space:]]Parameter-Efficient[[:space:]]Tuning/1151f19c-bdd4-4e21-a98e-99747a8a42e6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Expository[[:space:]]Text[[:space:]]Generation_[[:space:]]Imitate,[[:space:]]Retrieve,[[:space:]]Paraphrase/656ebcd2-1d50-4645-ae0f-b7275efd227c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FACTIFY3M_[[:space:]]A[[:space:]]benchmark[[:space:]]for[[:space:]]multimodal[[:space:]]fact[[:space:]]verification[[:space:]]with[[:space:]]explainability[[:space:]]through[[:space:]]5W[[:space:]]Question-Answering/74bf1220-fe95-4702-a8a4-82c2f32ab0c5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FAME_[[:space:]]Flexible,[[:space:]]Scalable[[:space:]]Analogy[[:space:]]Mappings[[:space:]]Engine/d48de324-2900-4208-90d5-d8c07f36c225_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FANToM_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Stress-testing[[:space:]]Machine[[:space:]]Theory[[:space:]]of[[:space:]]Mind[[:space:]]in[[:space:]]Interactions/5fb45d53-ef52-42c1-b972-5d16069608b0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FActScore_[[:space:]]Fine-grained[[:space:]]Atomic[[:space:]]Evaluation[[:space:]]of[[:space:]]Factual[[:space:]]Precision[[:space:]]in[[:space:]]Long[[:space:]]Form[[:space:]]Text[[:space:]]Generation/8e78d402-af8e-430e-8e6b-3ead36d10872_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FLatS_[[:space:]]Principled[[:space:]]Out-of-Distribution[[:space:]]Detection[[:space:]]with[[:space:]]Feature-Based[[:space:]]Likelihood[[:space:]]Ratio[[:space:]]Score/1016938e-e2a1-4fc4-afcc-6214f8515ce5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FOCUS_[[:space:]]Effective[[:space:]]Embedding[[:space:]]Initialization[[:space:]]for[[:space:]]Monolingual[[:space:]]Specialization[[:space:]]of[[:space:]]Multilingual[[:space:]]Models/47499b6e-957d-4cf4-9c81-8154fb986207_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FaMeSumm_[[:space:]]Investigating[[:space:]]and[[:space:]]Improving[[:space:]]Faithfulness[[:space:]]of[[:space:]]Medical[[:space:]]Summarization/645ddee2-68f4-464b-82e1-bcda3902af4e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FactKB_[[:space:]]Generalizable[[:space:]]Factuality[[:space:]]Evaluation[[:space:]]using[[:space:]]Language[[:space:]]Models[[:space:]]Enhanced[[:space:]]with[[:space:]]Factual[[:space:]]Knowledge/b83102e0-e358-40c4-8f38-2139eb88f3e5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Failures[[:space:]]Pave[[:space:]]the[[:space:]]Way_[[:space:]]Enhancing[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Tuning-free[[:space:]]Rule[[:space:]]Accumulation/83b7ef8f-a770-4aef-b94d-00e8311f766c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Fair[[:space:]]Text[[:space:]]Classification[[:space:]]with[[:space:]]Wasserstein[[:space:]]Independence/f6e1a806-b4d6-43bb-a11f-5e1fa5db09ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Fair[[:space:]]Without[[:space:]]Leveling[[:space:]]Down_[[:space:]]A[[:space:]]New[[:space:]]Intersectional[[:space:]]Fairness[[:space:]]Definition/134a0549-8e3e-4ce7-96d0-f3d57589d69e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Faithful[[:space:]]Model[[:space:]]Evaluation[[:space:]]for[[:space:]]Model-Based[[:space:]]Metrics/cd10d35d-6f4b-4798-af16-a685c947fc4e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Fast[[:space:]]and[[:space:]]Accurate[[:space:]]Factual[[:space:]]Inconsistency[[:space:]]Detection[[:space:]]Over[[:space:]]Long[[:space:]]Documents/b74d72d4-2d6c-4258-afb1-bab04e29a60f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Fast[[:space:]]and[[:space:]]Robust[[:space:]]Early-Exiting[[:space:]]Framework[[:space:]]for[[:space:]]Autoregressive[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Synchronized[[:space:]]Parallel[[:space:]]Decoding/cb8b3786-8675-465e-8346-24119f7443d5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Faster[[:space:]]Minimum[[:space:]]Bayes[[:space:]]Risk[[:space:]]Decoding[[:space:]]with[[:space:]]Confidence-based[[:space:]]Pruning/df83d60f-39a7-451c-8eea-0ad9ad453c26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FedID_[[:space:]]Federated[[:space:]]Interactive[[:space:]]Distillation[[:space:]]for[[:space:]]Large-Scale[[:space:]]Pretraining[[:space:]]Language[[:space:]]Models/444cb266-0a3e-4ad2-9190-d09eee0d6ef8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/FedTherapist_[[:space:]]Mental[[:space:]]Health[[:space:]]Monitoring[[:space:]]with[[:space:]]User-Generated[[:space:]]Linguistic[[:space:]]Expressions[[:space:]]on[[:space:]]Smartphones[[:space:]]via[[:space:]]Federated[[:space:]]Learning/14280c1a-617b-43c2-a859-db049b10e136_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Federated[[:space:]]Learning[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Parameter-Efficient[[:space:]]Prompt[[:space:]]Tuning[[:space:]]and[[:space:]]Adaptive[[:space:]]Optimization/0a294c49-1bf3-43e7-a56a-ae79e190763a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Federated[[:space:]]Meta-Learning[[:space:]]for[[:space:]]Emotion[[:space:]]and[[:space:]]Sentiment[[:space:]]Aware[[:space:]]Multi-modal[[:space:]]Complaint[[:space:]]Identification/8264cc45-ca57-4552-93f3-de5248dd0316_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Enhancing Biomedical Lay Summarisation with External Knowledge Graphs/49634b26-4af9-4204-9100-ca8a3aea5c79_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37faf7ab50303c1eba856ec73d64954df772c1ed090093bc591bc08e9bc72835
+size 940684

2023/Enhancing Biomedical Lay Summarisation with External Knowledge Graphs/full.md

@@ -0,0 +1,379 @@
+# Enhancing Biomedical Lay Summarisation with External Knowledge Graphs
+
+Tomas Goldsack<sup>1</sup>, Zhihao Zhang<sup>2</sup>, Chen Tang<sup>3</sup>, Carolina Scarton<sup>1</sup>, Chenghua Lin<sup>1,4*</sup>
+
+$^{1}$ Department of Computer Science, University of Sheffield, UK
+
+$^{2}$ College of Economics and Management, Beijing University of Technology, China,
+
+$^{3}$ Department of Computer Science, The University of Surrey, UK
+
+$^{4}$ Department of Computer Science, The University of Manchester, UK
+
+{tgoldsack1, c.scarton}@sheffield.ac.uk zhhzhang@bjut.edu.cn
+
+chen.tang@surrey.ac.uk chenghua.lin@manchester.ac.uk
+
+# Abstract
+
+Previous approaches for automatic lay summarisation are exclusively reliant on the source article that, given it is written for a technical audience (e.g., researchers), is unlikely to explicitly define all technical concepts or state all of the background information that is relevant for a lay audience. We address this issue by augmenting eLife, an existing biomedical lay summarisation dataset, with article-specific knowledge graphs, each containing detailed information on relevant biomedical concepts. Using both automatic and human evaluations, we systematically investigate the effectiveness of three different approaches for incorporating knowledge graphs within lay summarisation models, with each method targeting a distinct area of the encoder-decoder model architecture. Our results confirm that integrating graph-based domain knowledge can significantly benefit lay summarisation by substantially increasing the readability of generated text and improving the explanation of technical concepts. $^{1}$
+
+# 1 Introduction
+
+Lay summarisation consists of generating a concise summary that illustrates the significance of a longer technical (or otherwise specialist) text and is comprehensible to the non-expert (Kuehne and Olden, 2015). A lay summary should contain minimal jargon and technical details (e.g., methodology), instead focusing largely on the simplification of key technical concepts and the explanation or relevant background information, thus allowing readers without technical knowledge to grasp the general topic and main ideas of an article (Srikanth and Li, 2021; Goldsack et al., 2022). However, since the original article is intended for a technical audience who already possess some domain knowledge, it
+
+![](images/c5c7b3b17bba41c67791badd8447f3be232086b444a1dc3c8fc11f24b5a2138b.jpg)  
+Figure 1: Overview of the "knowledge graph-enhanced Lay Summarisation" task formulation, exemplifying graph-based external information.
+
+is unlikely to explicitly include all the information necessary for the lay summary, such as background details or definitions. As a result, lay summaries are often highly abstractive, adopting a simpler lexicon than the original article (Goldsack et al., 2022), and are typically written by experts who possess the knowledge required to simplify and explain the contents of the article (King et al., 2017).
+
+Despite this disparity between lay summary and source article, automatic approaches to lay summarisation have typically relied solely upon the source article as input (Chandrasekaran et al., 2020; Guo et al., 2021; Luo et al., 2022a). Aiming to address this, we propose enhancing lay summarisation models with external domain knowledge, conducting the first study on knowledge graph-enhanced lay summarisation with a focus on biomedical articles. We augment eLife (Goldsack et al., 2022), an existing high-quality lay summarisation dataset, with article-specific knowledge graphs containing information on the technical concepts covered within the articles and the relation
+
+ships between them (exemplified in Figure 1), thus providing a structured representation of the domain knowledge that an expert human author might draw upon when writing a lay summary (§3). In doing this, we hypothesise that a model's ability to simplify and explain technical concepts for a lay audience will improve.
+
+Although other forms of summarisation task (e.g., news articles) have seen significant research in the enhancing models using knowledge graphs (Huang et al., 2020; Zhu et al., 2021a; Lyu et al., 2022), this has yet to be explored for lay summarisation. To our knowledge, there is no work on determining the most effective way to incorporate graph-based knowledge for lay summarisation or other summarisation tasks. Therefore, we systematically investigate the effectiveness of three different methods for injecting graph-based information into lay summarisation models ( $\S 4$ ) assessing them with both automatic and human evaluation ( $\S 5$  &  $\S 6$ ). Our results demonstrate that the integration of graph-based domain knowledge can significantly improve automatic lay summarisation enabling models to generate substantially more readable text and to better explain technical concepts.
+
+# 2 Related Work
+
+# 2.1 Lay Summarisation
+
+The task of lay summarisation is a relatively novel one, introduced by the LaySumm subtask of the CL-SciSumm 2020 shared task series (Chandrasekaran et al., 2020). Introducing a multidomain corpus of 572 article-lay summary pairs, the task attracted a total of 8 participants. The winning system, proposed by Kim (2020), adopted a hybrid approach, using a PEGASUS-based (Zhang et al., 2020) model to generate an initial abstractive lay summary before augmenting this with sufficiently readable article sentences extracted by a BERT-based model (Devlin et al., 2019).
+
+Subsequent lay summarisation work has focused almost exclusively on the introduction and benchmarking of new corpora (all from the biomedical domain), rather than introducing specific modelling approaches for the task. Guo et al. (2021) introduce CDSR, a dataset derived from the Cochrane Database of Systematic Reviews, whereas Goldsack et al. (2022) introduce PLOS and eLife, two datasets derived from different biomedical journals (the Public Library of Science and eLife
+
+journals, respectively). Both studies benchmark their datasets with widely-used summarisation approaches, with BART variants (Lewis et al., 2020) invariably achieving the strongest performance. In another highly related work, Luo et al. (2022a) address the task of readability-controlled summarisation using data derived from PLOS, training a BART-based model to produce both the abstract and lay summary of an article in a controlled setting.
+
+In contrast to previous works, we investigate the unexplored approach of modelling and integrating structured domain knowledge into lay summarisation models using article-specific knowledge graphs.
+
+# 2.2 Knowledge Graph-Enhanced Text Generation
+
+In recent years, the utilisation of knowledge graphs (KGs) containing external knowledge for text generation has seen increased interest, particularly when it comes to the modelling of commonsense knowledge. In particular, works focusing on tasks such as dialogue generation (Zhou et al., 2018; Tang et al., 2023), commonsense-reasoning (Liu et al., 2021), story generation (Guan et al., 2019; Tang et al., 2022), essay generation (Yang et al., 2019) have all seen the introduction of commonsense KG-enhanced models.
+
+Some recent work has also focused on using KGs for abstractive summarisation, but tending towards modeling internal knowledge. Aiming to improve the faithfulness and informativeness of summaries, Huang et al. (2020), Zhu et al. (2021a), and Lyu et al. (2022) all utilise OpenIE to construct fact-based knowledge graphs from source documents (news articles). Huang et al. (2020) and Zhu et al. (2021a) extract graph node features using graph attention networks (Velicković et al., 2017), before incorporating these into the summarisation model decoder using an attention mechanism. Lyu et al. (2022) instead make use of additional semantic loss measures to attempt to capture extracted facts using an adapted pointer-generator network.
+
+In contrast to previous works, we apply KG-based techniques to biomedical lay summarisation (as opposed to news article summarisation), a domain with additional challenges, including the extensive length of input articles and the presence of
+
+complex technical concepts that need to be simplified or explained. Furthermore, due to the unique requirements of the task, we innovate by constructing knowledge graphs largely using external domain knowledge sources rather than from the article itself, as is the case with previous approaches.
+
+# 3 Article Knowledge Graphs
+
+We augment eLife, an existing dataset for biomedical lay summarisation with heterogeneous article-specific knowledge graphs (KGs). Each KG contains structured information on the complex biomedical concepts covered within the article and the relationships between them. In order to localise this information and provide an indication of where in an article a concept is mentioned, we also choose to model the article's section-based document structure within our graphs through the use of section-specific nodes. In the following, we describe in detail: 1) our process for extracting the knowledge that is used by our model ( $\S 3.1$ ), and 2) how we structure that knowledge within a knowledge graph ( $\S 3.2$ ). The methods by which we integrate graph-based knowledge into summarisation models are discussed in  $\S 4$ .
+
+# 3.1 Knowledge Extraction
+
+To extract relevant domain knowledge for an article, we draw upon the Unified Medical Language System (UMLS) (Bodenreider, 2004). This rich and actively-maintained resource has long been used as a key knowledge source for NLP in the biomedical domain (McCray et al., 2001; Demner-Fushman et al., 2010; Kang et al., 2021) and is comprised of three primary components: the Metathesaurus, the Semantic Network, and the Specialist Lexicon and Lexical Tools. The Metathesaurus is an extensive multi-lingual vocabulary database containing information on a large number of biomedical concepts, including their various names and definitions. The Semantic Network defines a set of semantic types that represent broad subject categories into which all concepts in the Metathesaurus can be assigned. Additionally, high-level relationships that occur between different semantic types are also defined. To extract the UMLS concepts mentioned within a given article, we utilise MetaMap (Aronson and Lang, 2010), one of the Lexical Tools provided alongside UMLS for this exact purpose, that is widely used in previous work (Sang et al., 2018; Sharma et al., 2019; Lai et al., 2021). For all arti
+
+<table><tr><td>Metric</td><td>Abstract</td><td>Lay Summary</td><td>Definitions</td></tr><tr><td>FKGL↓</td><td>15.57</td><td>10.92</td><td>10.55</td></tr><tr><td>CLI↓</td><td>17.68</td><td>12.51</td><td>13.02</td></tr><tr><td>DCRS↓</td><td>11.78</td><td>8.83</td><td>10.36</td></tr><tr><td>WordRank↓</td><td>9.21</td><td>8.68</td><td>8.6</td></tr></table>
+
+Table 1: Mean readability scores for abstracts, lay summaries, and key UMLS definitions for eLife. FKGL = Flesch-Kinkaid Grade Level, CLI = Coleman-Liau Index, DCRS = Dale-Chall Readability Score.
+
+cles in eLife, we apply MetaMap to each section in turn, retrieving all mentioned UMLS concepts. We restrict MetaMap to only a select number of English vocabularies without prohibitive access restrictions, but otherwise run it using default settings.
+
+In line with observations made in previous works (Lai et al., 2021), we found that MetaMap, whilst succeeding in linking biomedical entities mentioned in the text with their corresponding UMLS concepts, also frequently returned a number of irrelevant concepts. Therefore, we adopt a text overlap-based approach to filter the original pool of extracted concepts for a given section, which we empirically found to eliminate the vast majority of the noise. $^3$
+
+For each remaining UMLS concept, we retrieve all semantic types with which it is associated, in addition to the formal definitions of both the concept and semantic types. Notably, these definitions are used as an integral component within all three KG-enhancement methods. In order to confirm their suitability for a lay audience, we calculate and compare their average readability scores with those reported by Goldsack et al. (2022) for both the technical abstracts and lay summaries of eLife articles. The results of this analysis, given in Table 1, show that the UMLS definitions obtain scores that are overall much closer to those of the lay summaries than the abstracts, actually exceeding them in two out of the four metrics (FKGL and WordRank). An example of the definition format used for text augmentation is given in Figure 6 in the Appendix. In the next section, we describe how we represent all extracted information within article-specific knowledge graphs.
+
+![](images/e9196bc3c608a59857bff620e98d4824407c908e75abdfc35df7e270bc85fea3.jpg)  
+Figure 2: Example of article knowledge graph structure. Graph nodes are coloured as follows: Document, Section, Metadata, Concept, Semantic type.
+
+# 3.2 Graph Construction
+
+Each graph  $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$ , where  $\mathcal{V}$  is a set of nodes (or entities) and  $\mathcal{E}$  is a set of edges. Each edge  $e_{ij}\in \mathcal{E}$  defines a relation  $r_{ij}$  between entities  $v_{i},v_{j}\in \mathcal{V}$ , and thus can be represented as a triplet  $e_{ij} = (v_i,r_{ij},v_j)$ . All graphs are heterogeneous, containing multiple types of entities and relations. Figure 2 presents a visualisation of an article knowledge graph.5 Each type of node is described below:
+
+- Document node - the central root node, which is the ancestor of all other nodes in the graph. We label this node simply with the unique ID assigned to each article.  
+- Section node - each section node represents a specific titled section (e.g. Introduction) of the document, including the abstract. To label these nodes, we concatenate the article ID with “_Abs” for the abstract or “_Sec{i}” for other sections, where {i} is the index of the section (zero-based).  
+- Metadata node - identify additional information relating to the article or its specific sections. This includes article and section titles, article keywords, and the date of publication.  
+- Concept node - nodes representing UMLS concepts. These are labelled with their unique UMLS identifier (CUI).
+
+- Semantic type node - nodes representing semantic types from the Semantic Network. These are labelled with their unique Semantic Type identifier (TUI).
+
+In addition to the 54 different relationship types defined within the semantic network (e.g., affects in Figure 2), we define several relations in order to represent the graph structure and additional metadata. Specifically, we define the relations contains, wasPublished_in, has_title, and has_keyword.
+
+# 4 Knowledge-Enhanced Lay Summarisation Approaches
+
+We investigate the effectiveness of three different methods for incorporating external knowledge from article graphs into encoder-decoder-based summarisation models. Our experiments are carefully designed so as to target a distinct aspect within the model architecture (i.e., the input, the encoder, and the decoder) with each selected method, taking inspiration from models that have recently been proven effective in the domain of news article summarisation (Zhu et al., 2021a; Pasunuru et al., 2021). Figure 3 provides a visualization of how each of these approaches fits into this architecture. To allow the ingestion of the full input article, we make use of Longformer Encoder-Decoder (Beltagy et al., 2020) as our base model for all experiments. This BART-based model replaces standard transformer self-attention with a sparse attention mechanism that scales linearly to the sequence length, enabling the processing of longer
+
+![](images/b8dafc0631628bf823fe5c9e7e8c3bfb9d82f6cbdd1fe13328bf2f3237f7c570.jpg)  
+Figure 3: Visualisation of how our various knowledge-enhancement approaches incorporate external knowledge from article knowledge graphs into a transformer-based encoder-decoder architecture (as described in §4). A) Decoder cross-attention, B) Document embedding enhancement, C) Article text augmentation.
+
+texts (such as research articles). We describe each knowledge-enhancement approach in detail below.
+
+(A) Decoder cross-attention. We make use of a Graph Attention Network (GAT) (Veličković et al., 2017) to obtain an embedding of the article graph  $\mathcal{G}$  in parallel with the base model encoder.
+
+$$
+H ^ {\mathcal {G}} = \operatorname {G A T} (\mathcal {G}) \tag {1}
+$$
+
+These Graph Neural Network (GNN) models produce a final set of node features (i.e., a graph embedding) by using attention layers to efficiently aggregate over the features of neighboring nodes, and are widely used in current literature to aggregate graph-based information for NLG tasks (Huang et al., 2020; Zhu et al., 2021b; Liu et al., 2021). During the decoding phase, we follow previous works (Zhu et al., 2021a) by forcing our model to attend to the KG embedding  $H^{\mathcal{G}}$ . Specifically, in every transformer layer of the decoder, we include a second cross-attention mechanism that occurs directly after the standard encoder cross-attention (see arrow A in Figure 3) and attends to the output of the GAT-based model.
+
+(B) Document embedding enhancement. Again, we obtain an embedded graph representation  $H^{\mathcal{G}}$  using the GAT model, but rather than attending to graph embeddings during decoding, we follow Pasunuru et al. (2021), combining the embedded node information into the final document embedding (i.e., the output of the encoder).
+
+Specifically, we concatenate the document and graph embeddings, before passing them through an additional encoder layer. For a given input document  $X$ , this process can be formalised as follows:
+
+$$
+H ^ {X} = \operatorname {E n c o d e r} (X) \tag {2}
+$$
+
+$$
+H ^ {C} = \left[ H ^ {X}; H ^ {\mathcal {G}} \right] \tag {3}
+$$
+
+$$
+H ^ {*} = p \cdot \operatorname {E n c o d e r L a y e r} \left(H ^ {C}\right) + (1 - p) \cdot H ^ {C} \tag {4}
+$$
+
+where  $H^{X}$  is an embedding of document  $X$ ,  $H^{C}$  is the concatenated document and graph embeddings,  $H^{*}$  is the final 'enhanced' document embedding that is subsequently attended to during decoding, and  $p$  is a scaling factor controlling the extent to which the additional encoder layer output is incorporated in the final enhanced document embedding. Note that  $p$  is treated as a hyperparameter to the model, for which a value of 0.25 was found to provide the strongest validation set performance.
+
+(C) Article text augmentation. We also experiment with simply augmenting the input text with textual explanations of the key concepts (and their relations) derived from the graph. Whilst this may be arguably the most 'natural' way for a PLM-based model to interpret external information, this approach leads to an exponential increase in the
+
+number of tokens required to describe each element, hence it is restricted to a set of few concepts. We select only those entities which are likely to be most central to the topic of the article (and, therefore, relevant to the lay summary). Specifically, we take the concept nodes that are mentioned in the article abstract and use the graph relations and retrieved definitions to provide a textual explanation of these salient concepts and their semantic types, which is then prepended to the article text. This takes the format of " $\{ \text{concept\_name} \} = \{ \text{concept\_definition} \}$ . {concept\_name} is a {semtype\_name}" repeated for each selected concept, followed by semantic type definitions formatted " $\{ \text{semtype\_name} \} = \{ \text{semtype\_definition} \}"$  repeated for all mentioned semantic types.
+
+# 5 Experimental Setup
+
+# 5.1 Data
+
+We derive knowledge graphs for all articles in eLife (Goldsack et al., 2022), a dataset for biomedical lay summarisation containing 4,828 article-summary pairs. Target summaries are expert-written lay summaries (i.e., summaries with a non-expert target audience) and inputs are the full text of the corresponding biomedical research articles. As explained in §1, we believe this task to be particularly suitable for domain knowledge augmentation due to the contrast in the level of expertise of the target audience between source and target which causes a discrepancy in the language used (specifically, reducing or explaining jargon terms) and the level of background information required.<sup>7</sup>
+
+# 5.2Baselines
+
+As a baseline model, we include BART (Lewis et al., 2020), the state-of-the-art benchmark reported by Goldsack et al. (2022) for eLife, as well as in other previous lay summarisation works (Guo et al., 2021). Additionally, we include the reported performance of  $\mathrm{BART}_{\text {scaffold }}$  (Goldsack et al., 2022), a variant of BART trained to produce both the abstract and lay summary of an article in a controlled setting, which is equivalent to the model proposed by Luo et al. (2022a).
+
+# 5.3 Implementation and Training
+
+Each knowledge-based approach is implemented by manually adapting the Longformer implementation from Huggingface (Wolf et al., 2020) and, following previous work on lay summarisation (Chandrasekaran et al., 2020; Luo et al., 2022b; Goldsack et al., 2022), uses the full article text as input. For GAT-based models, we make use of the Deep Graph Library package (Wang et al., 2019) to implement a 3-layer GAT with 4 attention heads at each layer. For article graphs, we vary our node initialisation approach based on node type (as defined in §3.2). Specifically, we initialise concept and semantic type node features, with the embeddings of their textual definitions; document and section nodes with the embeddings of their title text (with title metadata nodes being subsequently ignored); and remaining metadata nodes (publication date and keywords) with embeddings on their textual content. All embeddings are generated using SciBert (Beltagy et al., 2019), a language model specifically trained on research papers from Semantic Scholar (Ammar et al., 2018) that is widely used for scientific data (Cohan et al., 2019; Cai et al., 2022; Goldsack et al., 2023b). Furthermore, all node embedding features are concatenated with one-hot features according to node type, as well as Random Walk Positional Encodings (Dwivedi et al., 2021). Following initialisation with allenai/led-base-16384 checkpoint, we train all models on A100 GPUs and retain the checkpoint with the best validation set performance (more details provided in the Appendix).
+
+# 5.4 Evaluation Setup
+
+We conduct both automatic and human evaluations to provide a comprehensive assessment of how each knowledge-enhancement method affects the overall performance.
+
+Automatic evaluation. For each model, we report the average scores of several automatic metrics on the test split of eLife. As is common practice, we report widely-used summarisation metrics: BERTScore (Zhang et al., 2019) and the F1-scores of ROUGE-1, 2, and L (Lin, 2004).
+
+To assess the readability of generated summaries, we report Flesh-Kincaid Grade Level (FKGL) and Dale-Chall Readability Score (DCRS), both of which compute an estimate of the US-grade level
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">Relevance</td><td colspan="2">Readability</td><td>Factuality</td></tr><tr><td>R-1↑</td><td>R-2↑</td><td>R-L↑</td><td>BeS↑</td><td>CLI↓</td><td>DCRS↓</td><td>BaS↑</td></tr><tr><td>BART</td><td>46.57</td><td>11.65</td><td>43.70</td><td>84.94</td><td>11.7</td><td>9.36</td><td>-2.39</td></tr><tr><td>BARTscaffold</td><td>45.28</td><td>10.99</td><td>42.51</td><td>84.65</td><td>11.32</td><td>9.19</td><td>-2.57</td></tr><tr><td>Longformer</td><td>47.23</td><td>13.20</td><td>44.44</td><td>85.11</td><td>11.72</td><td>9.09</td><td>-2.56</td></tr><tr><td>text-aug</td><td>48.58*</td><td>14.24*</td><td>45.71*</td><td>85.4*</td><td>10.94*</td><td>8.72*</td><td>-2.45*</td></tr><tr><td>doc-enhance</td><td>48.10*</td><td>14.43*</td><td>45.52*</td><td>85.33*</td><td>10.72*</td><td>8.50*</td><td>-2.35*</td></tr><tr><td>decoder-attn</td><td>48.30*</td><td>13.93</td><td>45.45*</td><td>85.39*</td><td>10.99*</td><td>8.75*</td><td>-2.48*</td></tr></table>
+
+Table 2: Average performance of models on eLife test split.  $\mathbf{R} =$  ROUGE F1,  $\mathbf{{BeS}} =$  BERTScore F1,  $\mathbf{{CLI}} =$  Coleman-Liau Index,  $\mathbf{{BaS}} =$  BARTScore. * denotes that KG-enhanced model results are statistically significant with respect to the base model (Longformer) by way of Mann-Whitney U test.
+
+required to comprehend a text.9
+
+Additionally, we evaluate factuality using BARTScore (Yuan et al., 2021), which has been shown to have a strong alignment with human judgments of factual consistency in a recent study focusing specifically on long documents (Yee Koh et al., 2022). Following Yee Koh et al. (2022), we adapt BARTScore to use Longformer (thus allowing it to process the entire document as input) and fine-tune it on eLife.
+
+Human evaluation. To provide a comprehensive assessment of the summaries generated by each knowledge-enhanced model, we conduct a human evaluation focusing on readability and factuality. Specifically, making use of 5 randomly sampled articles from the eLife test set, we ask human judges to evaluate each sentence within a generated summary along the following binary criteria: 1) Factuality - is the sentence factually correct (with respect to the source article); and 2) Readability - would a layperson be able to understand this sentence.[10] To help determine the factuality of the sentence, the annotator has access to the PDF of the source article as well as the reference lay summary.[11]
+
+# 6 Experimental Results
+
+# 6.1 Automatic Evaluation
+
+Table 2 presents the performance of different models using the described automatic evaluation metrics on the test set of eLife. In addition to applying KG-enhancement methods in isolation, we also experiment with combining different methods, which
+
+we largely find to be detrimental to model performance. Discussion and results (Table 6) of combined methods are provided in the Appendix. We discuss the performance of individually applied methods below, focusing on each aspect of automatic evaluation in turn.
+
+Relevance Longformer can be seen to outperform the standard BART model in terms of relevance metrics, indicating that processing the entire document provides some benefit for lay summarisation. Additionally, all three knowledge enhancement methods significantly obtain improved scores across almost all relevance metrics (with the exception of R2 for the 'decoder attention' model). This provides a strong indication that the addition of graph-based domain knowledge provides models with relevant external information, enabling them to produce lay summaries that are closer in resemblance to the high-quality references.
+
+Readability For readability metrics, it can first be noted that Longformer-based models obtain lower CLI and DCRS scores than those BART-based models. The calculation of CLI is based on the number of characters, words, and sentences it contains, whereas DCRS is based on the frequency of "familiar" (i.e., commonly-used) words, suggesting that Longformer produces summaries that are less syntactically and lexically complex.
+
+We observe that the application of all knowledge enhancement methods results in improved scores for both metrics, with the document enhancement approach achieving the largest gains. This indicates that all knowledge enhancement methods are able to successfully influence the phrasing and structure of the summaries being generated by increasing the usage of more common (i.e., less technical) terminology. As reported in Table 1, the average CLI and DCRS scores for the reference lay summaries
+
+<table><tr><td>Model</td><td>#</td><td>Readability</td><td>Factuality</td></tr><tr><td>Longformer</td><td>73</td><td>78.08</td><td>60.96</td></tr><tr><td>–text-aug</td><td>65</td><td>96.92*</td><td>68.46</td></tr><tr><td>–doc-enhance</td><td>67</td><td>97.01*</td><td>55.97</td></tr><tr><td>–decoder-attn</td><td>69</td><td>95.65*</td><td>63.77</td></tr></table>
+
+of eLife are 12.51 and 8.83, respectively.
+
+Factuality For BARTScore (BAS), we again see a statistically significant improvement over the base Longformer model for all KG-enhancement methods, with the greatest improvement being obtained by the doc-enhance method. In order to gain further insight in these results, we also calculate the BARTScore values obtained by reference summaries, getting a mean score -2.39, which is similar to that of all tested models (and identical to that of BART). This suggests that all models are able to produce summaries with generative probabilities similar to that of the reference summaries. However, given that one model (doc-enhance) actually outscores the reference summaries, further analysis is needed to gain an understanding of the difference in the factual correctness of summaries produced by each method, for which we turn to our human evaluation.
+
+# 6.2 Human Evaluation
+
+Given the challenging and time-consuming nature of evaluating the factuality of technical biomedical sentences against the source article, we carefully plan our human evaluation so as to ensure reliability in our results. We employed two annotators to evaluate generated sentences following the procedure laid out in §5, both of whom are experts in NLP and familiar with common model shortcomings (e.g., hallucinations). Table 3 presents the total percentage of sentences that were positively classified for both readability and factuality averaged across evaluators, who achieve a Cohen's  $\kappa$  of 0.42.
+
+Discussion The results in Table 3 suggest that the application of all KG-enhancement methods causes a notable increase in the readability of the text produced by the model, with all models scor
+
+Table 3: Average percentage of generated sentences positively classified by judges for each high-level binary criteria.  $\#  =$  total number of sentences generated across all summaries. * denotes that KG-enhanced model results are statistically significant with respect to the base model (Longformer) by way of Mann-Whitney U test.  
+
+<table><tr><td>a. [Meiosis]Longformer - During meiosis, the DNA in one of the chromosomes is copied and then the two copies are recombined so that each new generation will have a single copy of the gene that encodes the protein encoded by that gene.1/2w/ text-aug - ... a process known as meiosis ... two copies of each chromosome are then exchanged between the newly formed cells, which results in a unique set of genes being passed on to the next generation.2/2</td></tr><tr><td>b. [Glabrous skin / Mechanoreceptors]Longformer - The orientation of an object depends largely on how its edges activate mechanoreceptors in the glabrous skin of the fingertips.0/2w/ doc-enhance - The fingertip&#x27;s surface is covered by a ... layer of skin known as the glabrous skin.2/2 These cells are responsible for sensing touch, and they are also responsible for detecting the orientation of objects that touch them.2/2</td></tr><tr><td>c. [Slow wave sleep]Longformer - Most studies of sleep have focused on ... slow wave sleep, in which the brain&#x27;s activity alternates between periods of alternating periods of slow and fast sleep.0/2w/ decoder-attn - Slow wave sleep is characterized by rhythmic waves of electrical activity in the brain, which are thought to be part of the process by which the brain consolidates memories.2/2</td></tr></table>
+
+Figure 4: A case study comparing how the application of each method affects the explanation of specific technical concepts within the human evaluation sample. Colours and superscript are used to denote the number of evaluators who judged the sentence as readable for a lay audience (e.g.,  $2/2 = 2$  out of 2 evaluators).
+
+ing significantly higher than the base Longformer model. Alternatively, the results for factuality show that, although there is a slight variance in performance between KG methods, none of them are judged to be by a statistically significant margin. These results indicate that all methods are able to effectively introduce relevant external information into the model, enabling it to produce text that is easier for a lay audience to comprehend without significantly compromising the factual correctness of the base model.
+
+Case Study To gain a better insight into how knowledge enhancement methods influence the readability of generated summaries, we present a case study in Figure 4 in which we compare the explanations of specific technical concepts generated by KG-enhanced models and the base Longformer model, alongside their annotator ratings.[12]
+
+These examples demonstrate how KG-enhancement methods improve the model's handling of technical concepts, thus making them easier to understand for a lay reader. Specifically,
+
+examples show how methods can influence the model to generate an explanation in instances where the base model fails to provide one (b) or improve the explanation in instances where the base model's is difficult to understand (a and c).
+
+# 7 Conclusion
+
+This papers presents the first study on the use of knowledge graphs to enhance lay summarisation, augmenting the biomedical lay summarisation dataset eLife with article-specific knowledge graphs containing domain-specific external knowledge on relevant technical concepts. We compare three distinct approaches for incorporating graph-based knowledge into encoder-decoder summarisation models, placing an emphasis on the readability and factual correctness of the generated output. Our results suggest that integrating external knowledge has the potential to substantially improve lay summarisation, particularly for the generation of readable text and explanation of technical concepts. We would like to see future work investigate the use of additional graph representations, as well as their integration into larger models that adopt different architectures (e.g., decoder-only).
+
+# Limitations
+
+One possible limitation of our work is derived from the use of resources from UMLS (i.e., UMLS concept names, semantic types and relations, definitions, etc.). Accessing these resources requires an individual license with the US National Library of Medicine (NLM), and their subsequent distribution is restricted by this license agreement. Therefore, it is likely that we will have to confirm the license status of those who wish to have access to the knowledge-graph resources used in this work. In an attempt to reduce any potential impacts this will have on the ability to share our resources, we only make use of only a select number of vocabularies less restrictive licences. More details on the vocabularies used are provided in the Appendix.
+
+# 8 Acknowledgements
+
+This work was supported by the Centre for Doctoral Training in Speech and Language Technologies (SLT) and their Applications funded by UK Research and Innovation [grant number EP/S023062/1].
+
+# References
+
+Waleed Ammar, Dirk Groeneveld, Chandra Bhagavatula, Iz Beltagy, Miles Crawford, Doug Downey, Jason Dunkelberger, Ahmed Elgohary, Sergey Feldman, Vu Ha, Rodney Kinney, Sebastian Kohlmeier, Kyle Lo, Tyler Murray, Hsu-Han Ooi, Matthew Peters, Joanna Power, Sam Skjonsberg, Lucy Lu Wang, Chris Wilhelm, Zheng Yuan, Madeleine van Zuylen, and Oren Etzioni. 2018. Construction of the literature graph in semantic scholar. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 3 (Industry Papers), pages 84-91, New Orleans - Louisiana. Association for Computational Linguistics.  
+Alan R Aronson and François-Michel Lang. 2010. An overview of MetaMap: historical perspective and recent advances. J. Am. Med. Inform. Assoc., 17(3):229-236.  
+Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A pretrained language model for scientific text. 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 3615-3620, Hong Kong, China. Association for Computational Linguistics.  
+Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer.  
+Olivier Bodenreider. 2004. The unified medical language system (UMLS): integrating biomedical terminology. *Nucleic Acids Res.*, 32(Database issue):D267-70.  
+Xiaoyan Cai, Sen Liu, Libin Yang, Yan Lu, Jintao Zhao, Dinggang Shen, and Tianming Liu. 2022. COVID-Sum: A linguistically enriched SciBERT-based summarization model for COVID-19 scientific papers. Journal of Biomedical Informatics, 127:103999.  
+Muthu Kumar Chandrasekaran, Guy Feigenblat, Edward Hovy, and Anita de Waard. 2020. Overview and insights from the shared tasks at scholarly document processing 2020: CL-SciSumm, LaySumm and. In Proceedings of the First Workshop on Scholarly Document Processing, pages 214-224. unknown.  
+Arman Cohan, Iz Beltagy, Daniel King, Bhavana Dalvi, and Dan Weld. 2019. Pretrained language models for sequential sentence classification. 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 3693-3699, Hong Kong, China. Association for Computational Linguistics.  
+Dina Demner-Fushman, James Mork, Sonya Shooshan, and Alan Aaronson. 2010. Wumls content views appropriate for nlp processing of the biomedical literature vs. clinical text. In Journal of biomedical informatics, 43(4), page 587-594.
+
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Vijay Prakash Dwivedi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. 2021. Graph neural networks with learnable structural and positional representations.  
+Tomas Goldsack, Zheheng Luo, Qianqian Xie, Carolina Scarton, Matthew Shardlow, Sophia Ananiadou, and Chenghua Lin. 2023a. Overview of the biolaysumm 2023 shared task on lay summarization of biomedical research articles. In The 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks, pages 468-477, Toronto, Canada. Association for Computational Linguistics.  
+Tomas Goldsack, Zhihao Zhang, Chenghua Lin, and Carolina Scarton. 2022. Making science simple: Corpora for the lay summarisation of scientific literature. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10589-10604, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Tomas Goldsack, Zhihao Zhang, Chenghua Lin, and Carolina Scarton. 2023b. Domain-Driven and Discourse-Guided Scientific Summarisation. In Advances in Information Retrieval, pages 361-376, Cham. Springer Nature Switzerland.  
+Jian Guan, Yansen Wang, and Minlie Huang. 2019. Story Ending Generation with Incremental Encoding and Commonsense Knowledge. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):6473-6480.  
+Yue Guo, Wei Qiu, Yizhong Wang, and Trevor Cohen. 2021. Automated Lay Language Summarization of Biomedical Scientific Reviews. Proceedings of the AAAI Conference on Artificial Intelligence, 35(1):160-168.  
+Luyang Huang, Lingfei Wu, and Lu Wang. 2020. Knowledge graph-augmented abstractive summarization with semantic-driven cloze reward. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5094-5107, Online. Association for Computational Linguistics.  
+Tian Kang, Adler Perotte, Casey Ta, and Chunhua Weng. 2021. Umls-based data augmentation for natural language processing of clinical research literature. In Journal of the American Medical Informatics Association JAIMA 28(4), pages 812-823, Online.  
+Seungwon Kim. 2020. Using pre-trained transformer for better lay summarization. In Proceedings of the
+
+First Workshop on Scholarly Document Processing, pages 328-335, Online. Association for Computational Linguistics.  
+Stuart R F King, Emma Pewsey, and Sarah Shailes. 2017. Plain-language Summaries of Research: An inside guide to eLife digests. eLife, 6:e25410.  
+Lauren M Kuehne and Julian D Olden. 2015. Lay summaries needed to enhance science communication. Proceedings of the National Academy of Sciences, 112(12):3585-3586.  
+Tuan Lai, Heng Ji, ChengXiang Zhai, and Quan Hung Tran. 2021. Joint biomedical entity and relation extraction with knowledge-enhanced collective inference. 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 6248-6260, Online. Association for Computational Linguistics.  
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. 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.  
+Ye Liu, Yao Wan, Lifang He, Hao Peng, and Philip S. Yu. 2021. Kg-bart: Knowledge graph-augmented bart for generative commonsense reasoning. 35:6418-6425.  
+Zeheng Luo, Qianqian Xie, and Sophia Ananiadou. 2022a. Readability controllable biomedical document summarization. In Findings of the 2022 Conference on Empirical Methods in Natural Language Processing, Abu Dhabi. Association for Computational Linguistics.  
+Zeheng Luo, Qianqian Xie, and Sophia Ananiadou. 2022b. Readability controllable biomedical document summarization. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 4667-4680, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Yuanjie Lyu, Chen Zhu, Tong Xu,zikai Yin, and Enhong Chen. 2022. Faithful abstractive summarization via fact-aware consistency-constrained transformer. In Proceedings of the 31st ACM International Conference on Information Knowledge Management, CIKM '22, page 1410-1419, New York, NY, USA. Association for Computing Machinery.
+
+Alexa McCray, Olivier Bodenreider, James Malley, and Allen Browne. 2001. Evaluating umls strings for natural language processing. In Proceedings of AMIA Symposium, page 448-452.  
+Ramakanth Pasunuru, Mengwen Liu, Mohit Bansal, Sujith Ravi, and Markus Dreyer. 2021. Efficiently summarizing text and graph encodings of multi-document clusters. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4768-4779, Online. Association for Computational Linguistics.  
+Shengtian Sang, Zhihao Yang, Lei Wang, Xiaoxia Liu, Hongfei Lin, and Jian Wang. 2018. Sematyp: a knowledge graph based literature mining method for drug discovery. BMC Bioinformatics, pages 1471-2105.  
+Soumya Sharma, Bishal Santra, Abhik Jana, Santosh Tokala, Niloy Ganguly, and Pawan Goyal. 2019. Incorporating domain knowledge into medical NLI using knowledge graphs. 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 6092-6097, Hong Kong, China. Association for Computational Linguistics.  
+Neha Srikanth and Junyi Jessy Li. 2021. Elaborative simplification: Content addition and explanation generation in text simplification. In *Findings of the Association for Computational Linguistics: ACLIJCNLP* 2021, pages 5123-5137, Online. Association for Computational Linguistics.  
+Chen Tang, Chenghua Lin, Henglin Huang, Frank Guerin, and Zhihao Zhang. 2022. EtriCA: Event-triggered context-aware story generation augmented by cross attention. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5504-5518.  
+Chen Tang, Hongbo Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2023. Enhancing dialogue generation via dynamic graph knowledge aggregation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4604-4616.  
+Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. 6th International Conference on Learning Representations.  
+Minjie Wang, Da Zheng, Zihao Ye, Quan Gan, Mufei Li, Xiang Song, Jinjing Zhou, Chao Ma, Lingfan Yu, Yu Gai, Tianjun Xiao, Tong He, George Karypis, Jinyang Li, and Zheng Zhang. 2019. Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315.
+
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Huggingface's transformers: State-of-the-art natural language processing.  
+Pengcheng Yang, Lei Li, Fuli Luo, Tianyu Liu, and Xu Sun. 2019. Enhancing topic-to-essay generation with external commonsense knowledge. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2002–2012, Florence, Italy. Association for Computational Linguistics.  
+Huan Yee Koh, Jiaxin Ju, He Zhang, Ming Liu, and Shirui Pan. 2022. How far are we from robust long abstractive summarization? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2682-2698, Abu Dhabi. Association for Computational Linguistics.  
+Weizhe Yuan, Graham Neubig, and Pengfei Liu. 2021. Bartscore: Evaluating generated text as text generation. In Advances in Neural Information Processing Systems, volume 34, pages 27263-27277. Curran Associates, Inc.  
+Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter J. Liu. 2020. Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert.  
+Hao Zhou, Tom Young, Minlie Huang, Haizhou Zhao, Jingfang Xu, and Xiaoyan Zhu. 2018. Commonsense knowledge aware conversation generation with graph attention. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, International Joint Conferences on Artificial Intelligence Organization, pages 4623-4629.  
+Chenguang Zhu, William Hinthorn, Ruochen Xu, Qingkai Zeng, Michael Zeng, Xuedong Huang, and Meng Jiang. 2021a. Enhancing factual consistency of abstractive summarization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 718-733, Online. Association for Computational Linguistics.  
+Tianyu Zhu, Wen Hua, Jianfeng Qu, and Xiaofang Zhou. 2021b. Summarizing Long-Form document with rich discourse information. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, CIKM '21, pages 2770-2779, New York, NY, USA. Association for Computing Machinery.
+
+# A Appendix
+
+<table><tr><td>Node type</td><td>Average Count</td></tr><tr><td>Document</td><td>1.00</td></tr><tr><td>Section</td><td>5.90</td></tr><tr><td>Metadata</td><td>7.49</td></tr><tr><td>Concept</td><td>364.93</td></tr><tr><td>SemType</td><td>63.47</td></tr></table>
+
+Graph Statistics Table 4 presents the average node type frequencies in the graph of a given article. Additionally, Tables 7 and 8 present the average semantic type (SemType) node frequencies, and Table 9 presents the average relation frequencies (all of which are located at the end of the Appendix).
+
+MetaMap - UMLS vocabularies. As mentioned in §3, we restrict MetaMap to a select number of English vocabularies with access restrictions lower than level 4.[13] Specifically, we allow MetaMap to use the following vocabularies: Alcohol and Other Drug Thesaurus (AOD), Diseases Database (DDB), CRISP Thesaurus (CSP), DrugBank (DRUGBANK), Diagnostic and Statistical Manual of Mental Disorders - Fifth Edition (DSM-5), Human Phenotype Ontology (HPO), NCI Thesaurus (NCI), MedlinePlus Health Topics (MEDLINEPLUS), LOINC (LNC), MeSH (MSH), RxNorm (RXNORM), and Gene Ontology (GO).
+
+MetaMap - Noise reduction. We found that MetaMap, in addition to accurately identifying UMLS concepts that were mentioned in a passage of text, often also returned a large number of unwanted concepts. These typically were only
+
+<sup>13</sup>Information on all possible MetaMap vocabularies and their access restrictions can be found on the following page: https://www.nlm.nih.gov/research/uml5/sourcereleasedocs/index.html
+
+Table 4: The average node type frequency statistics for a single article in the train split.  
+
+<table><tr><td>Model</td><td>Abstractiveness (%)</td></tr><tr><td>Longformer</td><td>22.95</td></tr><tr><td>– text-aug</td><td>20.99</td></tr><tr><td>– doc-enhance</td><td>20.00</td></tr><tr><td>– decoder-attn</td><td>24.46</td></tr></table>
+
+Table 5: The average abstractiveness (measured in terms of novel 1-grams) of summaries generated using different graph integration methodologies.
+
+distantly relevant to the text - for example, multiword concepts with one matching word. Therefore, for each section of an article, we employed a simple word overlap-based to filter out unwanted retrieved concepts. Specifically, after removing stopwords from the main text and lemmatisation concept names, we retain only the concepts for which all words in its name were found to appear in the section text. We empirically found this approach to significantly outperform word embedding-based approaches, which often failed to filter out irrelevant multi-word concepts if a single word appeared in the main text.
+
+Additional implementation and training details. We employ Longformer Encoder Decoder (LED) with the allenai/led-base-16384 Huggingface checkpoint as our base model for all knowledge enhancement approaches, using the default parameters for model training and an input limit of 8192.
+
+For the "decoder cross-attention" method, we replicate the standard Huggingface LED multi-head decoder cross-attention implementation (where the head no. is determined by the model configuration), making use of the GAT-produced graph embedding as the key and value matrices (in place of the encoder output), with the query matrix being the output of the previous standard cross-attention module.
+
+For the "document embedding enhancement" method, we replicate the standard LED encoder layer, adapting the configuration to account for the larger input of the concatenated document and graph feature representations.
+
+For the "article text augmentation" method, we preprocess the article text with the graph-derived text and shown in Figure 6, extending the input text limit to 16,384 to accommodate the lengthy augmentation text without losing article text. All models are trained for a maximum of 20 epochs, retaining the checkpoints that achieved the highest validation set performance (average ROUGE score across variants).
+
+The time taken to train each model on 2 A100 GPUs ranged from 12 hours to 2 days depending on the specific methodology used, with the GAT-based methods - "document embedding enhancement" and "decoder cross attention" - taking longer than the "article text augmentation" method.
+
+Summary abstractiveness. In an effort to gain further insight into how each KG-enhancement
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">Relevance</td><td colspan="2">Readability</td><td>Factuality</td></tr><tr><td>R-1↑</td><td>R-2↑</td><td>R-L↑</td><td>BeS↑</td><td>CLI↓</td><td>DCRS↓</td><td>BaS↑</td></tr><tr><td>Longformer</td><td>47.23</td><td>13.20</td><td>44.44</td><td>85.11</td><td>11.72</td><td>9.09</td><td>-2.56</td></tr><tr><td>– text-aug + doc-enhance</td><td>46.28</td><td>13.45</td><td>43.77</td><td>84.72</td><td>11.37</td><td>8.06</td><td>-2.56</td></tr><tr><td>– text-aug + decoder-attn</td><td>47.27</td><td>13.93</td><td>44.52</td><td>84.95</td><td>10.47</td><td>8.15</td><td>-2.57</td></tr><tr><td>– doc-enhance + decoder-attn</td><td>28.85</td><td>3.7</td><td>27.58</td><td>73.60</td><td>2.68</td><td>1.56</td><td>-6.48</td></tr></table>
+
+Table 6: Average performance of models with combinations of KG-enhancement methods on eLife test split.  $\mathbf{R} =$  ROUGE F1,  $\mathbf{{BeS}} =$  BERTScore F1,  $\mathbf{{CLI}} =$  Coleman-Liau Index,  $\mathbf{{BaS}} =$  BARTScore.
+
+method influences summary generation, we measure the abstractiveness (as the percentage of novel 1-grams) of the generated summaries, providing the results in Table 5. Interestingly, the results show that both the "text augmentation" and "document enhancement" methods slightly decrease the abstractiveness of the base model, but the "decoder attention" method slightly increases it, suggesting that this method is more effective at introducing external vocabulary.
+
+Combining KG-enhancement Methods Table 6 presents the results obtained when applying different combinations of the KG-enhancement methods to Longformer. Interestingly, no combined model matches the overall performance of any single KG method model. Although these combinations achieve positive results for readability metrics, this generally comes at the expense of readability metrics, with the combination of text augmentation and decoder attention being the only model to achieve ROUGE scores equal to that of the base model.
+
+We believe this performance degradation is likely a result of the dilution or loss of document information at the expense of graph-based information. This is similar to what was observed when the value of  $p$ , as given in equation (4), was set to too great a value for the document enhancement method. In the case of the method combinations that include the text augmentation method, the document information is likely more diluted as a result of the larger embedding size (as caused by the increased input size). Alternatively, when the document embedding and decoder attention methods are combined, the original document information seems to be lost, resulting in particularly high loss values during training and the model being unable to produce a coherent output (as evidenced by the extremely low scores for readability and factuality metrics).
+
+Case study Figure 5 presents an expanded version of the case study given in Figure 4, whereby we show all model outputs for each of the technical concepts. As in Figure 4, we see that the KG-enhanced models typically improve on the explanation proved by the base Longformer model.
+
+<table><tr><td>[Meiosis]Longformer - During meiosis, the DNA in one of the chromosomes is copied and then the two copies are recom- bined so that each new generation will have a single copy of the gene that encodes the protein encoded by that gene.1/2w/ text-aug - ... a process known as meiosis ... two copies of each chromosome are then exchanged between the newly formed cells, which results in a unique set of genes being passed on to the next generation.2/2w/ doc-enhance - ... meiosis depends on the rate and location of the recombination events, and on the length of the DNA that is involved in the process.1/2w/ decoder-attn - This is achieved through a process called meiosis, in which a cell divides to produce four daughter cells, each of which contains a copy of one of the four copies of the genes found on the parent cell&#x27;s chromosomes.2/2</td></tr><tr><td>[Glabrous Skin / Mechanoreceptors]Longformer - The orientation of an object depends largely on how its edges activate mechanoreceptors in the glabrous skin of the fingertips.0/2w/ text-aug - ... information about the orientation of an object is based largely on how its edges activate mechanore- ceptors in the skin of the fingertips.2/2 ... tactile neurons in the fingertips are able to detect the edges of objects with high sensitivity2/2 ...w/ doc-enhance - The fingertip&#x27;s surface is covered by a ... layer of skin known as the glabrous skin.2/2 These cells are responsible for sensing touch, and they are also responsible for detecting the orientation of objects that touch them.2/2w/ decoder-attn - The orientation of an object is determined largely by how its edges contact the skin and activate nerve cells in the fingertips.2/2</td></tr><tr><td>[Slow Wave Sleep]Longformer - Most studies of sleep have focused on a type of sleep known as slow wave sleep, in which the brain&#x27;s activity alternates between periods of alternating periods of slow and fast sleep.0/2w/ text-aug - slow wave sleep is thought to help us to consolidate newly acquired memories and to consolidate our sense of awareness of the world around us.2/2 ... in slow wave sleep, the propagation of slow waves is lost between the brain&#x27;s regions.2/2w/ doc-enhance - Slow wave sleep is characterized by a slow oscillation of electrical activity in the brain.2/2w/ decoder-attn - Slow wave sleep is characterized by rhythmic waves of electrical activity in the brain, which are thought to be part of the process by which the brain consolidates memories.2/2</td></tr></table>
+
+Figure 5: An extended case study comparing how the application of each method affects the explanation of specific technical concepts within the human evaluation sample. Colours and superscript are used to denote the number of evaluators who judged the sentence as readable for a lay audience (e.g.,  $2/2 = 2$  out of 2 evaluators).
+
+# {Concept definitions and relations}
+
+Alleles = Variant forms of the same gene, occupying the same locus on homologous CHROMOSOMES, and governing the variants in production of the same gene product. Alleles is a Gene or Genome.
+
+Molecule = An aggregate of two or more atoms in a defined arrangement held together by chemical bonds.  
+Molecule is a Substance.
+
+Discover = See for the first time; identify. Discover is a Activity.
+
+Histocompatibility = The degree of antigenic similarity between the tissues of different individuals, which determines the acceptance or rejection of allografts. Histocompatibility is a Qualitative Concept.
+
+In Vivo = Located or occurring in the body. In Vivo is a Spatial Concept.
+
+Species = A group of organisms that differ from all other groups of organisms and that are capable of breeding and producing fertile offspring. Species is a Classification.
+
+Cells = The fundamental, structural, and functional units or subunits of living organisms. They are composed of CYTOPLASM containing various ORGANELLES and a CELL MEMBRANE boundary. Cells is a Cell.
+
+Allogeneic = Taken from different individuals of the same species. Allogeneic is a Qualitative Concept. Antigens = Substances that are recognized by the immune system and induce an immune reaction. Antigens is a Immunologic Factor.
+
+Result = The result of an action. Result is a Functional Concept.
+
+Major Histocompatibility Complex = The genetic region which contains the loci of genes which determine the structure of the serologically defined (SD) and lymphocyte-defined (LD) TRANSPLANTATION ANTIGENS, genes which control the structure of the IMMUNE RESPONSE-ASSOCIATED ANTIGENS, HUMAN; the IMMUNE RESPONSE GENES which control the ability of an animal to respond immunologically to antigenic stimuli, and genes which determine the structure and/or level of the first four components of complement. Major Histocompatibility Complex is a Gene or Genome.
+
+···
+
+# {SemType definitions}
+
+Gene or Genome = A specific sequence, or in the case of the genome the complete sequence, of nucleotides along a molecule of DNA or RNA (in the case of some viruses) which represent the functional units of heredity.
+
+Substance = A material with definite or fairly definite chemical composition.
+
+Activity = An operation or series of operations that an organism or machine carries out or participates in. Classification = A term or system of terms denoting an arrangement by class or category.
+
+Cell = The fundamental structural and functional unit of living organisms.
+
+Qualitative Concept = A concept which is an assessment of some quality, rather than a direct measurement.  
+Spatial Concept = A location, region, or space, generally having definite boundaries.
+
+Immunologic Factor = A biologically active substance whose activities affect or play a role in the functioning of the immune system.
+
+Functional Concept = A concept which is of interest because it pertains to the carrying out of a process or activity.
+
+Quantitative Concept = A concept which involves the dimensions, quantity or capacity of something using some unit of measure, or which involves the quantitative comparison of entities.
+
+Temporal Concept  $=$  A concept which pertains to time or duration.
+
+···
+
+{Article text}
+
+Figure 6: Text augmentation example.
+
+<table><tr><td>Semantic type</td><td>Average Count</td></tr><tr><td>Manufactured Object</td><td>0.98</td></tr><tr><td>Classification</td><td>0.61</td></tr><tr><td>Qualitative Concept</td><td>1.0</td></tr><tr><td>Disease or Syndrome</td><td>0.51</td></tr><tr><td>Health Care Activity</td><td>0.95</td></tr><tr><td>Occupational Activity</td><td>0.94</td></tr><tr><td>Phenomenon or Process</td><td>0.99</td></tr><tr><td>Individual Behavior</td><td>0.26</td></tr><tr><td>Intellectual Product</td><td>1.0</td></tr><tr><td>Organism Attribute</td><td>0.96</td></tr><tr><td>Health Care Related Organization</td><td>0.47</td></tr><tr><td>Body Part, Organ, or Organ Component</td><td>0.86</td></tr><tr><td>Social Behavior</td><td>0.67</td></tr><tr><td>Therapeutic or Preventive Procedure</td><td>0.78</td></tr><tr><td>Research Device</td><td>0.13</td></tr><tr><td>Spatial Concept</td><td>1.0</td></tr><tr><td>Temporal Concept</td><td>1.0</td></tr><tr><td>Behavior</td><td>0.45</td></tr><tr><td>Environmental Effect of Humans</td><td>0.05</td></tr><tr><td>Mammal</td><td>0.76</td></tr><tr><td>Research Activity</td><td>1.0</td></tr><tr><td>Quantitative Concept</td><td>1.0</td></tr><tr><td>Occupation or Discipline</td><td>0.81</td></tr><tr><td>Functional Concept</td><td>1.0</td></tr><tr><td>Conceptual Entity</td><td>0.81</td></tr><tr><td>Activity</td><td>1.0</td></tr><tr><td>Population Group</td><td>0.65</td></tr><tr><td>Age Group</td><td>0.35</td></tr><tr><td>Mental Process</td><td>0.92</td></tr><tr><td>Natural Phenomenon or Process</td><td>0.93</td></tr><tr><td>Geographic Area</td><td>0.78</td></tr><tr><td>Biologic Function</td><td>0.43</td></tr><tr><td>Human-caused Phenomenon or Process</td><td>0.20</td></tr><tr><td>Idea or Concept</td><td>1.0</td></tr><tr><td>Body Location or Region</td><td>0.58</td></tr><tr><td>Finding</td><td>1.0</td></tr><tr><td>Organ or Tissue Function</td><td>0.38</td></tr><tr><td>Amino Acid, Peptide, or Protein</td><td>0.89</td></tr><tr><td>Injury or Poisoning</td><td>0.13</td></tr><tr><td>Professional or Occupational Group</td><td>0.50</td></tr><tr><td>Gene or Genome</td><td>0.84</td></tr><tr><td>Element, Ion, or Isotope</td><td>0.91</td></tr><tr><td>Body Substance</td><td>0.56</td></tr><tr><td>Cell</td><td>0.91</td></tr><tr><td>Receptor</td><td>0.50</td></tr><tr><td>Hazardous or Poisonous Substance</td><td>0.80</td></tr><tr><td>Pathologic Function</td><td>0.59</td></tr><tr><td>Organism</td><td>0.46</td></tr><tr><td>Indicator, Reagent, or Diagnostic Aid</td><td>0.81</td></tr><tr><td>Cell Component</td><td>0.82</td></tr><tr><td>Biologically Active Substance</td><td>0.91</td></tr><tr><td>Animal</td><td>0.52</td></tr><tr><td>Biomedical or Dental Material</td><td>0.63</td></tr><tr><td>Group</td><td>0.38</td></tr><tr><td>Body System</td><td>0.26</td></tr><tr><td>Physical Object</td><td>0.34</td></tr><tr><td>Antibiotic</td><td>0.39</td></tr><tr><td>Organism Function</td><td>0.95</td></tr><tr><td>Governmental or Regulatory Activity</td><td>0.48</td></tr><tr><td>Organization</td><td>0.61</td></tr><tr><td>Tissue</td><td>0.40</td></tr><tr><td>Diagnostic Procedure</td><td>0.41</td></tr><tr><td>Biomedical Occupation or Discipline</td><td>0.68</td></tr><tr><td>Entity</td><td>0.40</td></tr></table>
+
+Table 7: The average semantic node type frequency statistics for a single article in the train split.  
+
+<table><tr><td>Semantic type</td><td>Average Count</td></tr><tr><td>Inorganic Chemical</td><td>0.66</td></tr><tr><td>Pharmacologic Substance</td><td>0.93</td></tr><tr><td>Physiologic Function</td><td>0.33</td></tr><tr><td>Immunologic Factor</td><td>0.58</td></tr><tr><td>Molecular Function</td><td>0.67</td></tr><tr><td>Clinical Attribute</td><td>0.39</td></tr><tr><td>Laboratory Procedure</td><td>0.88</td></tr><tr><td>Event</td><td>0.59</td></tr><tr><td>Human</td><td>0.04</td></tr><tr><td>Chemical Viewed Structurally</td><td>0.50</td></tr><tr><td>Sign or Symptom</td><td>0.28</td></tr><tr><td>Enzyme</td><td>0.68</td></tr><tr><td>Medical Device</td><td>0.78</td></tr><tr><td>Genetic Function</td><td>0.73</td></tr><tr><td>Nucleic Acid, Nucleoside, or Nucleotide</td><td>0.75</td></tr><tr><td>Organic Chemical</td><td>0.88</td></tr><tr><td>Patient or Disabled Group</td><td>0.23</td></tr><tr><td>Virus</td><td>0.20</td></tr><tr><td>Cell Function</td><td>0.74</td></tr><tr><td>Substance</td><td>0.94</td></tr><tr><td>Daily or Recreational Activity</td><td>0.72</td></tr><tr><td>Bacterium</td><td>0.37</td></tr><tr><td>Laboratory or Test Result</td><td>0.17</td></tr><tr><td>Neoplastic Process</td><td>0.17</td></tr><tr><td>Eukaryote</td><td>0.33</td></tr><tr><td>Amino Acid Sequence</td><td>0.30</td></tr><tr><td>Cell or Molecular Dysfunction</td><td>0.57</td></tr><tr><td>Regulation or Law</td><td>0.09</td></tr><tr><td>Chemical</td><td>0.45</td></tr><tr><td>Body Space or Junction</td><td>0.23</td></tr><tr><td>Nucleotide Sequence</td><td>0.37</td></tr><tr><td>Chemical Viewed Functionally</td><td>0.37</td></tr><tr><td>Family Group</td><td>0.52</td></tr><tr><td>Vertebrate</td><td>0.18</td></tr><tr><td>Fungus</td><td>0.23</td></tr><tr><td>Food</td><td>0.40</td></tr><tr><td>Embryonic Structure</td><td>0.28</td></tr><tr><td>Bird</td><td>0.19</td></tr><tr><td>Molecular Biology Research Technique</td><td>0.56</td></tr><tr><td>Molecular Sequence</td><td>0.06</td></tr><tr><td>Group Attribute</td><td>0.03</td></tr><tr><td>Vitamin</td><td>0.16</td></tr><tr><td>Mental or Behavioral Dysfunction</td><td>0.30</td></tr><tr><td>Hormone</td><td>0.19</td></tr><tr><td>Plant</td><td>0.18</td></tr><tr><td>Anatomical Structure</td><td>0.06</td></tr><tr><td>Fish</td><td>0.14</td></tr><tr><td>Machine Activity</td><td>0.25</td></tr><tr><td>Educational Activity</td><td>0.09</td></tr><tr><td>Reptile</td><td>0.02</td></tr><tr><td>Amphibian</td><td>0.03</td></tr><tr><td>Congenital Abnormality</td><td>0.04</td></tr><tr><td>Experimental Model of Disease</td><td>0.06</td></tr><tr><td>Archaeon</td><td>0.02</td></tr><tr><td>Language</td><td>0.02</td></tr><tr><td>Anatomical Abnormality</td><td>0.01</td></tr><tr><td>Acquired Abnormality</td><td>0.01</td></tr><tr><td>Fully Formed Anatomical Structure</td><td>0.01</td></tr><tr><td>Professional Society</td><td>0.01</td></tr><tr><td>Clinical Drug</td><td>0.00</td></tr><tr><td>Self-help or Relief Organization</td><td>0.00</td></tr></table>
+
+Table 8: The average semantic node type frequency statistics for a single article in the train split (continued).
+
+<table><tr><td>Relation type</td><td>Average Count</td></tr><tr><td>contains</td><td>564.94</td></tr><tr><td>has_keyword</td><td>1.59</td></tr><tr><td>has_title</td><td>4.90</td></tr><tr><td>was_published_in</td><td>1.0</td></tr><tr><td>is_a</td><td>536.22</td></tr><tr><td>branch_of</td><td>0.86</td></tr><tr><td>affects</td><td>255.99</td></tr><tr><td>performs</td><td>25.23</td></tr><tr><td>exhibits</td><td>6.29</td></tr><tr><td>conceptual_part_of</td><td>10.80</td></tr><tr><td>result_of</td><td>134.99</td></tr><tr><td>measures</td><td>71.28</td></tr><tr><td>issue_in</td><td>97.14</td></tr><tr><td>associated_with</td><td>43.36</td></tr><tr><td>occurs_in</td><td>15.44</td></tr><tr><td>connected_to</td><td>2.61</td></tr><tr><td>process_of</td><td>101.00</td></tr><tr><td>degree_of</td><td>12.83</td></tr><tr><td>manifestation_of</td><td>37.81</td></tr><tr><td>uses</td><td>14.00</td></tr><tr><td>location_of</td><td>60.24</td></tr><tr><td>causes</td><td>42.52</td></tr><tr><td>adjacent_to</td><td>6.13</td></tr><tr><td>tributary_of</td><td>0.86</td></tr><tr><td>prevents</td><td>6.84</td></tr><tr><td>produces</td><td>67.90</td></tr><tr><td>method_of</td><td>13.61</td></tr><tr><td>property_of</td><td>6.45</td></tr><tr><td>complicates</td><td>54.91</td></tr><tr><td>evaluation_of</td><td>22.34</td></tr><tr><td>co-occurs_with</td><td>21.40</td></tr><tr><td>carries_out</td><td>8.23</td></tr><tr><td>interacts_with</td><td>102.22</td></tr><tr><td>measurement_of</td><td>17.90</td></tr><tr><td>traverses</td><td>1.45</td></tr><tr><td>developmental_form_of</td><td>3.52</td></tr><tr><td>treats</td><td>8.34</td></tr><tr><td>surrounds</td><td>4.74</td></tr><tr><td>part_of</td><td>27.68</td></tr><tr><td>conceptually_related_to</td><td>0.31</td></tr><tr><td>precedes</td><td>30.83</td></tr><tr><td>consists_of</td><td>6.54</td></tr><tr><td>analyzes</td><td>16.93</td></tr><tr><td>assesses-effect_of</td><td>23.76</td></tr><tr><td>disrupts</td><td>49.62</td></tr><tr><td>contains</td><td>2.97</td></tr><tr><td>diagnoses</td><td>8.38</td></tr><tr><td>indicates</td><td>3.30</td></tr><tr><td>practices</td><td>0.78</td></tr><tr><td>manages</td><td>0.77</td></tr><tr><td>derivative_of</td><td>0.28</td></tr><tr><td>interconnects</td><td>0.70</td></tr><tr><td>ingredient_of</td><td>0.05</td></tr></table>
+
+Table 9: The average relation type frequency statistics for a single article in the train split.

2023/Enhancing Biomedical Lay Summarisation with External Knowledge Graphs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66f4e8bb90061e7143c88eb382e820c8b9f5a4a43ed25b7f6c4baff4867bb4ab
+size 1097776

2023/Enhancing Chat Language Models by Scaling High-quality Instructional Conversations/b1f4f81f-0f2a-4ef5-813b-a9152c0d03d7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dff661456525f955c474fa5754552b3a764870b909ef9ffe6c5b011588d8f9d
+size 4226012

2023/Enhancing Chat Language Models by Scaling High-quality Instructional Conversations/full.md

@@ -0,0 +1,556 @@
+# Enhancing Chat Language Models by Scaling High-quality Instructional Conversations
+
+Ning Ding $^{1*}$ , Yulin Chen $^{2,3*}$ , Bokai Xu $^{4}$ , Yujia Qin $^{2,3}$ , Shengding Hu $^{2,3}$ , Zhiyuan Liu $^{2,3\dagger}$ , Maosong Sun $^{2,3\dagger}$ , Bowen Zhou $^{1\dagger}$
+
+$^{1}$ Department of Electronic Engineering, Tsinghua University
+
+$^{2}$ Department of Computer Science and Technology, Tsinghua University
+
+$^{3}$ BNRIST, IAI, Tsinghua University,  $^{4}$ The Chinese University of Hong Kong, Shenzhen
+
+# Abstract
+
+Fine-tuning on instruction data has been widely validated as an effective practice for implementing chat language models like ChatGPT. Scaling the diversity and quality of such data, although straightforward, stands a great chance of leading to improved performance. This paper aims to push the upper bound of open-source models further. We first provide a systematically designed, diverse, informative, large-scale dataset of instructional conversations, UltraChat, which does not involve human queries. Our objective is to capture the breadth of interactions between a human user and an AI assistant and employs a comprehensive framework to generate multi-turn conversation iteratively. UltraChat contains 1.5 million high-quality multi-turn dialogues and covers a wide range of topics and instructions. Our statistical analysis of UltraChat reveals its superiority in various key metrics, including scale, average length, diversity, coherence, etc., solidifying its position as a leading open-source dataset. Building upon UltraChat, we fine-tune a LLaMA model to create a powerful conversational model, UltraLM. Our evaluations indicate that UltraLM consistently outperforms other open-source models, including WizardLM and Vicuna, the previously recognized state-of-the-art open-source models.
+
+# 1 Introduction
+
+Large language models (Bommasani et al., 2021; Han et al., 2021; Chowdhery et al., 2022) (LLMs) have demonstrated exceptional generalization capability on a variety of language-related tasks. Notably, ChatGPT (OpenAI, 2022), an optimized version of GPT-3 (Brown et al., 2020) for conversation, along with GPT-4 (OpenAI, 2023), takes the user experience to another level via excelling in comprehending and generating responses in a natural and interactive manner. The introduction of
+
+<table><tr><td>Model</td><td>Score</td></tr><tr><td>Dolly-v2 (Conover et al., 2023)</td><td>4.04 ± 2.34</td></tr><tr><td>MPT-Chat (Mosaic, 2023)</td><td>6.67 ± 2.88</td></tr><tr><td>OpenAssistant (Köpf et al., 2023)</td><td>7.65 ± 2.15</td></tr><tr><td>Alpaca (Taori et al., 2023)</td><td>8.04 ± 2.05</td></tr><tr><td>Koala (Geng et al., 2023)</td><td>8.23 ± 1.99</td></tr><tr><td>Baize (Xu et al., 2023b)</td><td>8.50 ± 1.34</td></tr><tr><td>Vicuna (Chiang et al., 2023)</td><td>8.78 ± 1.55</td></tr><tr><td>Wizard-LM (Xu et al., 2023a)</td><td>8.95 ± 1.44</td></tr><tr><td>UltraLM (ours)</td><td>9.00 ± 1.33</td></tr></table>
+
+Table 1: Average scores (1-10) across different open-source models and UltraLM. The evaluation is conducted on our curated evaluation set with GPT-4. Evaluation prompts can be found in Appendix C.
+
+ChatGPT has spurred a surge in the adoption and implementation of general chat language models.
+
+In addition to competing models developed by large corporations such as Bard<sup>1</sup> and Claude<sup>2</sup>, the open-source community is actively engaged in training similar models, aiming to democratize access to AI technology. Notable examples in this regard include Alpaca (Taori et al., 2023), Vicuna (Chiang et al., 2023), Koala (Geng et al., 2023), Baize (Xu et al., 2023b), and Belle (Ji et al., 2023), etc., demonstrating promising performance. Experimental evidence strongly suggests that chat language models can be effectively trained through instruction fine-tuning (Wei et al., 2021; Sanh et al., 2021), and they also indicate that many data-efficient (Zhou et al., 2023) or computing-efficient (Hu et al., 2021; Ding et al., 2023) methods can be applied. This paper, in another way, focuses more on the "final one mile" of chat language models, as evidence shows that the journey from 0 to 60 is easy, whereas progressing from 60 to 100 becomes exceedingly challenging. For instance, researchers have shown that by utilizing a small,
+
+thoughtfully curated set of instructions, it is possible to train a model with satisfactory instruction-following capabilities. However, these approaches have yet to produce models that surpass the performance of Vicuna, the current leading open-source model, let alone outperform ChatGPT and GPT-4.
+
+This paper believes that the most straightforward way, that is, the quality and diversity of training data, play a vital role in further improving the performance of chat language models. In other words, leveraging higher quality and more diverse data can yield better outcomes. To this end, we present UltraChat, a million-scale multi-turn instructional conversation data, to facilitate the construction of more powerful chat language models. UltraChat is carefully designed to capture the breadth of interactions that a human might have with an AI assistant. Specifically, we do not use specific tasks like question-answering or summarization to construct the data, but curate three sectors: Questions about the World, Creation and Writing, and Assistance on Existing Materials. Then we employ meta-information, in-context expansion, and iterative prompting to scale up the number of instructions. To construct informative and realistic multi-turn conversations, two separate ChatGPT Turbo APIs are adopted in the conversation generation, where one plays the role of the user to generate queries, and the other generates the response. We instruct the user model with carefully designed prompts to mimic human user behavior and call the two APIs iteratively.
+
+We fine-tune a LLaMA-13B model on UltraChat to produce UltraLM and compare the model to a wide range of baselines, especially the open-source ones. The evaluation shows that our model could consistently outperform other models. As reported in Table 1, UltraLM achieves the highest performance scores that are independently assessed by GPT-4. Further evaluation results on challenging benchmarks and preference study with GPT-4 on various evaluation sets also show that UltraLM could surpass all other open-source models.
+
+# 2 Related Work
+
+Instruction Tuning. Recent works demonstrate LLMs' powerful capabilities in following human instructions. Wei et al. (2021) pioneered to fine-tune T5 (Raffel et al., 2020) on 60 NLP datasets verbalized with natural language instruction templates, i.e., instruction tuning. The fine-tuned model ex
+
+hibits a strong ability in instruction understanding and generalizes well to unseen instructions (Sanh et al., 2021; Ouyang et al., 2022). Later, Longpre et al. (2023) show the benefits of scaling the number of tasks in out-of-distribution generalization. Wei et al. (2021) also conclude that the success of instruction tuning depends on the quality of the dataset and the design of prompts. To further regulate the tuned model's behavior, Ouyang et al. (2022); Schulman et al. (2017) propose to employ reinforcement learning to align model behaviors with human preferences. This technique combined with instruction tuning can further boost the model performance and has been successfully applied to LLMs such as ChatGPT.
+
+Data Augmentation with LLMs. Collecting large-scale human-annotated instructions and their responses is time-consuming and labor-intensive. Alternatively, a more cost-effective and feasible approach to gathering top-notch data involves sampling from LLMs that have been well-tuned, e.g., ChatGPT and GPT-3.5. Recently, there is a surge of interest in distilling these powerful LLMs for data augmentation. For instance, using the technique of Self-Instruct (Wang et al., 2022), Alpaca (Taori et al., 2023) generate 52k high-quality instruction-response pairs based on seed tasks by "distilling" Text-Davinci-003. The trained model performs almost on par with Text-Davinci-003. The success of Alpaca boosts numerous later efforts on data augmentation with LLMs, such as code-alpaca (Chaudhary, 2023), alpaca-cot (Si et al., 2023), GPT4ALL (Anand et al., 2023), ShareGPT (Domeccleston, 2023), Dollyv2 (Conover et al., 2023), BELLE (Ji et al., 2023), Vicuna (Chiang et al., 2023), Koala (Geng et al., 2023), Baize (Xu et al., 2023b), etc. It is shown that increasing the scale of data could constantly improve the model performance. Besides, prompt engineering also affects data quality. CAMEL (Li et al., 2023a) design a multi-agent role-play environment for LLMs to simulate real human conversations.
+
+# 3 Data Construction
+
+LLMs are believed to be better annotators than human-being in many scenarios (Gilardi et al., 2023). However, employing LLMs such as ChatGPT directly for generating multi-turn conversations may yield satisfactory but less informative results, as it cannot enjoy the benefit of reinforce
+
+![](images/5a81e53631841d5ebd347d94be3d16b9d82b9721d6f2c8b0ba32c34ac45b4de2.jpg)  
+Figure 1: Construction process of UltraChat. The three sectors of data are derived from different meta-information.
+
+ment learning with human feedback (RLHF) in the alignment process. Table 12 in Appendix A shows a comparison of directly generated multi-turn dialogue and a case in UltraChat with the same opening line. Two key points can be derived to ensure the quality of the data: (1) An opening line determines the topic of the dialogue. Opening lines should be highly diverse and encompass any task that a human user may request a chat model to perform. (2) A user determines the plot of the dialogue, and the output should be tailored to the current topic with diverse language styles and requests.
+
+Therefore, unlike traditional task-specific datasets, to construct a comprehensive open-domain instructional chat dataset, the design of data collection schema is crucial to capturing the breadth of interactions and ensuring data quality. UltraChat aims to cover a tremendous range of conversation data with a carefully designed tripartite schema: Questions about the World, Creation and Writing, and Assistance on Existing Materials. While the core of ensuring data diversity mainly depends on opening line diversity, we will first introduce the idea behind the sector design and then focus on specific measures to obtain a diverse set of opening lines and how to prompt the user properly.
+
+# 3.1 Questions about the World
+
+The first sector focuses on querying existing information in the world, including concepts, objects, and entities that exist in the real world. This is at
+
+the core of human-AI interaction, as users often rely on AI assistants to provide quick and accurate answers to their questions.
+
+Our approach to gathering data for this sector involves two perspectives: one centered around topics and concepts, and the other around real-world entities. Initially, we request ChatGPT to generate 30 comprehensive topics that encompass various aspects of our daily lives, as shown in Table 2. Subsequently, we delve deeper into each topic by generating 30 to 50 subtopics or related concepts. Finally, we generate 10 different questions for each subtopic or concept and additionally request ChatGPT to generate 10 more questions based on each original question. The other source of data comes from real-world objects, which are derived from Wikidata<sup>3</sup> entities. These entities are further refined by considering their frequencies in Wikipedia<sup>4</sup> articles, specifically focusing on the 10,000 most frequently occurring entities. For each entity, we create 5 meta-questions, followed by 10 more specific questions and 20 extended questions. The extended questions aim to maintain some similarity to the original question while exploring distinct objects or topics. To create a dialogue, we filter and sample approximately 500,000 questions as opening lines. During the construction of each dialogue, we provide the user model with carefully crafted prompts that explicitly ask the model to respond
+
+<table><tr><td>Technology</td><td>Health and wellness</td><td>Travel and adventure</td></tr><tr><td>Food and drink</td><td>Art and culture</td><td>Science and innovation</td></tr><tr><td>Fashion and style</td><td>Relationships and dating</td><td>Sports and fitness</td></tr><tr><td>Nature and the environment</td><td>Music and entertainment</td><td>Politics and current events</td></tr><tr><td>Education and learning</td><td>Money and finance</td><td>Work and career</td></tr><tr><td>Philosophy and ethics</td><td>History and nostalgia</td><td>Social media and communication</td></tr><tr><td>Creativity and inspiration</td><td>Personal growth and development</td><td>Spirituality and faith</td></tr><tr><td>Pop culture and trends</td><td>Beauty and self-care</td><td>Family and parenting</td></tr><tr><td>Entrepreneurship and business</td><td>Literature and writing</td><td>Gaming and technology</td></tr><tr><td>Mindfulness and meditation</td><td>Diversity and inclusion</td><td>Travel and culture exchange</td></tr></table>
+
+Table 2: 30 meta-concepts used to generate the first sector of UltraChat data.  
+
+<table><tr><td>Articles and Blog Posts</td><td>Job Application Material</td><td>Stories</td></tr><tr><td>Legal Documents and Contracts</td><td>Poems</td><td>Educational Content</td></tr><tr><td>Screenplays</td><td>Scripts for Language Learning</td><td>Technical Documents and Reports</td></tr><tr><td>Marketing Materials</td><td>Social Media Posts</td><td>Personal Essays</td></tr><tr><td>Emails</td><td>Scientific Papers and Summaries</td><td>Speeches and Presentations</td></tr><tr><td>Recipes and Cooking Instructions</td><td>News Articles</td><td>SongLyrics</td></tr><tr><td>Product Descriptions and Reviews</td><td>Programs and Code</td><td></td></tr></table>
+
+Table 3: 20 types of text materials used for sector 2 and 3 UltraChat construction.
+
+concisely and meaningfully, taking into account the context of the ongoing dialogue history.
+
+# 3.2 Creation and Writing
+
+The second part is concerned with the creation of new information with human-input conditions, ranging from writing emails to crafting stories and plays. This process reflects the AI's capacity to engage in original content generation alongside users and demonstrates the role of AI assistants as collaborative partners in a creative environment.
+
+We first project all creations as text materials, and further categorize them into 20 different types as in Table 3. Then a ChatGPT model is employed to produce a diverse range of instructions for each type of writing, approximately  $80\%$  of which are further refined by ChatGPT model to generate more detailed instructions. These instructions serve as opening lines for dialogue generation. Throughout the generation process, the user prompt constantly reinforces the primary objective of the conversation, which is to generate and refine a piece of writing. This serves to ensure that the behavior of the user model remains focused and aligned with the intended purpose.
+
+# 3.3 Assistance on Existing Materials
+
+The third sector mainly addresses the modification of existing information, encompassing various tasks including rewriting, translation, summa
+
+rization, and question-answering, etc. Modifying existing materials is a crucial aspect of human-AI interaction, as it allows the AI assistant to actively engage with the user's input, transforming it in various ways as instructed by the user.
+
+We begin by gathering text pieces from the C4 corpus<sup>5</sup>. Each piece within the C4 corpus is associated with a source URL. To ensure a diverse range of text content and styles, we adopt the 20 material types outlined in the previous section and manually curate keywords for each type. Additionally, we classify the text in the corpus by matching the keywords in the corresponding URL. In total, we collect 100,000 text pieces from the C4 corpus, and for each piece, we prompt ChatGPT to generate five distinct instructions. We use a manually designed template to combine text and instructions, as depicted in Figure 4 in the appendix. Ultimately, the concatenated set of 500,000 pieces serves as the opening lines for the generated dialogues.
+
+# 3.4 User Simulation and Refinement
+
+Maintaining the desired behavior of the user model is crucial for achieving successful automatic dialogue generation. It has been observed that when the user model is solely provided with the current dialogue history, it tends to assume the role of an AI assistant. This "role confounding" situation
+
+<table><tr><td>Dataset</td><td>#Dialogue</td><td>Avg.#Turns</td><td>Avg. Dialog Length (by token)</td><td>Avg. Utt. Length (by token)</td><td>Lexical Diversity (↑)</td><td>Topic Diversity (↓)</td><td>Coherence (↑)</td><td>User Simulation</td></tr><tr><td>Self-Instruct</td><td>82,439</td><td>1</td><td>69.8</td><td>29.2</td><td>24.9</td><td>0.733</td><td>-</td><td>No</td></tr><tr><td>Stanford Alpaca</td><td>52,002</td><td>1</td><td>91.1</td><td>64.5</td><td>42.8</td><td>0.727</td><td>-</td><td>No</td></tr><tr><td>SODA</td><td>1,486,869</td><td>3.6</td><td>231.8</td><td>22.5</td><td>38.6</td><td>0.797</td><td>8.48</td><td>No</td></tr><tr><td>GPT-4-LLM</td><td>61,002</td><td>1</td><td>179.6</td><td>142.9</td><td>48.9</td><td>0.721</td><td>-</td><td>No</td></tr><tr><td>BELLE</td><td>1,436,679</td><td>1</td><td>102.3</td><td>63.3</td><td>35.9</td><td>0.771</td><td>-</td><td>No</td></tr><tr><td>Baize</td><td>210,311</td><td>3.1</td><td>293.9</td><td>52.8</td><td>67.1</td><td>0.751</td><td>9.06</td><td>Yes</td></tr><tr><td>GPT4ALL</td><td>711,126</td><td>1</td><td>597.7</td><td>318.9</td><td>62.7</td><td>0.692</td><td>-</td><td>No</td></tr><tr><td>UltraChat</td><td>1,468,352</td><td>3.8</td><td>1467.4</td><td>309.3</td><td>74.3</td><td>0.702</td><td>9.06</td><td>Yes</td></tr></table>
+
+Table 4: Statistics of existing instruction datasets. Lexical diversity is calculated by averaging the MTLD score (McCarthy and Jarvis, 2010) over each utterance with LexicalRichness<sup>6</sup>. 10000 samples are randomly drawn from each dataset for topic diversity and coherence measurement. Topic diversity is measured by averaging the cosine distance between each pair of data with OpenAI embedding API. Coherence is scored by ChatGPT on a scale of 1-10.
+
+can significantly deteriorate the coherence of the multi-turn conversation. To address this, in addition to presenting the dialogue history, we include prompts explicitly instructing the model to adopt various user personalities. In Sector 2, a prompt is employed to remind the model of the primary purpose of the dialogue, thereby promoting a more natural conversation flow. Once the data generation process is complete, a further filtration step is performed to ensure overall data quality. We also exclude excessively polite statements to enhance the realism of user responses.
+
+# 4 Data Analysis
+
+# 4.1 Statistical Analysis
+
+We conduct a statistical analysis of UltraChat and several other instruction datasets, as shown in Table 4. UltraChat stands out in terms of its scale, being one of the largest publicly available datasets. Moreover, it exhibits the highest average number of turns and the longest average length per instance of data. While SODA (Kim et al., 2023) also has many rounds, it is primarily composed of conceptual banter rather than instructional content. Additionally, the average number of tokens per dialogue in SODA is 231.8, whereas UltraChat boasts a remarkable 1467.4 tokens. To evaluate diversity, we measure both lexical diversity and topic diversity. UltraChat outperforms previous datasets in terms of lexical diversity. However, in terms of topic diversity, UltraChat falls slightly short compared to GPT4ALL (Anand et al., 2023) but still surpasses other datasets significantly. This may be attributed to the regularized embeddings resulting from a large number of tokens in each dialogue. We also conduct coherence evaluation with ChatGPT for multi-turn datasets. Notably, UltraChat and Baize data rank the highest in terms of coherence.
+
+# 4.2 Human Assessment
+
+Setup. To better evaluate the constructed data quality, we also conduct human assessment for UltraChat. Due to the difficulty of evaluation of multi-turn dialogue and the resulting formidable cost, we sample 500 representative dialogues for human evaluation, among which 300 are from UltraChat sector 1, 100 from sector 2 and sector 3 respectively. For each round of conversation, we ask the annotators to score the assistant's response on Helpfulness, Honesty, and Harmlessness (3H) principles (Askell et al., 2021). We also devise Coherence and Consistency criteria for the overall multi-turn dialogue quality evaluation. Coherence evaluates whether the dialogue flows logically and coherently, for which the annotators evaluate both the user's response and the assistant's response. Consistency means the assistant's responses do not contradict each other within the same dialogue. For example, it is inconsistent if the assistant asserts one specific event occurred in 1911 in the first round of conversation but mentions it as a 1901 event in the next round. Each metric is scored with 0, 0.5 or 1, where higher score means better quality. Therefore, for a K-round dialogue, we have  $3K + 2$  metric annotations.
+
+Annotation. Each dialogue is annotated independently by two well-trained annotators, and the score is averaged across two annotators. Meanwhile, due to the difficulty in identifying the hallucination problem, we allow the annotators to skip the dialogues that require expert knowledge or whose validity is hard to check. Altogether, we collect 14560 valid annotations in terms of metrics for both single-round and multi-round, and the Cohen's kappa coefficient is 0.358. The average time to annotate one dialogue is 10 minutes.
+
+Results. As shown in Table 5, the dataset scores high on all metrics, showing the effectiveness of the construction process. It is worth noting that the dataset is almost free from harmful content like hate speech and discrimination. Furthermore, we observe two trade-offs between data quality. The first is between helpfulness and honesty. While the honesty score is pretty high, helpfulness is compromised. It is because some user queries are out of the LLM ability scope (e.g., ask the LLM to compose a song). Under such circumstances, the LLM refuses to answer the question with an explanation of incapability, which is reasonable and expected but less helpful. The phenomenon is particularly prominent in part 3, as the simulated user often asks for information not existent in the text material. The second trade-off is between control and dialogue naturalness. While sector 2 and sector 3 data are constructed with more guidance and control when prompting for user simulation, they have clearly lower quality in coherence and consistency than sector 1.
+
+<table><tr><td>Data</td><td>Helpful</td><td>Honest</td><td>Harmless</td><td>Coherent</td><td>Consistent</td></tr><tr><td>Sector 1</td><td>0.971</td><td>0.996</td><td>1.000</td><td>0.996</td><td>0.995</td></tr><tr><td>Sector 2</td><td>0.978</td><td>0.986</td><td>1.000</td><td>0.982</td><td>0.977</td></tr><tr><td>Sector 3</td><td>0.893</td><td>0.983</td><td>1.000</td><td>0.964</td><td>0.981</td></tr><tr><td>Overall</td><td>0.960</td><td>0.992</td><td>1.000</td><td>0.987</td><td>0.988</td></tr></table>
+
+Table 5: Human assessment results on 500 dialogues sampled from UltraChat.
+
+# 5 Experiments
+
+We developed UltraLM, an enhanced variant of the LLaMA-13B (Touvron et al., 2023) model, by training it on the UltraChat dataset. To improve the model's comprehension of dialogue context, we break down each dialogue into smaller sequences, limiting them to a maximum length of 2048 tokens. During the training process, we only calculate the loss for the model's responses. This approach ensured that the model had access to the relevant information from earlier parts of the conversation, enabling a more comprehensive understanding of the ongoing dialogue. By incorporating the preceding context, UltraLM was equipped to generate more contextually appropriate and coherent responses. We use standard cross-entropy loss to finetune the model. The model is trained with 128 A100 GPUs and the total batch size is 512.
+
+The evaluation of the trained model is conducted in two folds. We first evaluate UltraLM on traditional benchmark datasets to delineate the
+
+knowledge scope and the multiple abilities of the language model. To better demonstrate the chat ability of language models, an automatic response quality evaluation is performed to showcase the model's proficiency in delivering accurate and informative content during chat interactions. Note that UltraLM is solely trained on UltraChat dataset without further finetuning on task specific datasets.
+
+# 5.1 Experimental Setup
+
+Baselines. We mainly compare with other open-source instruction-tuned language models based on LLaMA (Touvron et al., 2023) and Pythia (Biderman et al., 2023) backbone model. The main baseline models include Alpaca (Taori et al., 2023), Vicuna (Chiang et al., 2023), Koala (Geng et al., 2023), Dolly (Conover et al., 2023), OpenAssistant (Kopf et al., 2023), and WizardLM (Xu et al., 2023a). The parameter size of the baselines ranges from 7B to 13B, which is comparable to UltraLM. Our evaluation also includes other chat language models like ChatGPT (OpenAI, 2022), MPT (Mosaic, 2023), and Baize (Xu et al., 2023b). A detailed description of the main baselines can be found in Appendix A.1.
+
+Datasets. For benchmark evaluation, we choose four datasets: ARC-Challenge (Clark et al., 2018), HellaSwag (Zellers et al., 2019), MMLU (Hendrycks et al., 2021), and TruthfulQA (Lin et al., 2021), evaluating commonsense knowledge, professional knowledge and complex reasoning and understanding abilities. Each benchmark is constructed as multiple choice questions and therefore metrics are readily computable. The four datasets prove to be challenging even for the best-performing language models like ChatGPT.
+
+For response quality evaluation, we use 3 datasets. We first create an evaluation set by ourselves. The curated set encompasses the Vicuna benchmark as well as an additional 300 questions and instructions generated by GPT-4. The questions/instructions covered a wide range of topics, including commonsense, world knowledge, professional knowledge (specifically physics and biology), mathematics, response generation, and writing tasks on different levels of difficulty. Apart from the curated set, we also adopt AlpacaEval (Li et al., 2023b), a widely acknowledged
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">ARC-Challenge</td><td colspan="2">HellaSwag</td><td colspan="2">MMLU</td><td colspan="2">TruthfulQA</td><td rowspan="2">Overall Average</td></tr><tr><td>Acc.</td><td>Acc. norm.</td><td>Acc.</td><td>Acc. norm.</td><td>Weighted</td><td>Unweighted</td><td>mc1</td><td>mc2</td></tr><tr><td>Dolly-12B</td><td>38.23</td><td>42.24</td><td>54.59</td><td>72.6</td><td>31.52</td><td>31.70</td><td>20.69</td><td>34.06</td><td>45.15</td></tr><tr><td>OpenAssistant-12B</td><td>41.38</td><td>45.90</td><td>52.51</td><td>70.04</td><td>29.77</td><td>30.29</td><td>24.60</td><td>39.29</td><td>46.38</td></tr><tr><td>MPT-7B</td><td>43.00</td><td>46.67</td><td>57.13</td><td>75.50</td><td>37.76</td><td>38.33</td><td>27.17</td><td>40.16</td><td>50.17</td></tr><tr><td>Alpaca-7B</td><td>49.74</td><td>52.65</td><td>58.05</td><td>76.91</td><td>42.47</td><td>42.90</td><td>25.83</td><td>39.55</td><td>53.00</td></tr><tr><td>LLaMA-13B</td><td>53.16</td><td>56.40</td><td>60.64</td><td>80.87</td><td>46.05</td><td>46.74</td><td>25.83</td><td>39.90</td><td>55.98</td></tr><tr><td>Baize-13B</td><td>55.55</td><td>57.94</td><td>59.96</td><td>80.36</td><td>48.13</td><td>49.03</td><td>32.93</td><td>47.43</td><td>58.69</td></tr><tr><td>Koala-13B</td><td>49.83</td><td>52.90</td><td>57.60</td><td>77.54</td><td>46.75</td><td>48.01</td><td>34.64</td><td>50.09</td><td>57.14</td></tr><tr><td>Vicuna-13B</td><td>51.71</td><td>52.90</td><td>60.03</td><td>80.12</td><td>50.15</td><td>50.45</td><td>35.74</td><td>51.82</td><td>58.83</td></tr><tr><td>WizardLM-13B</td><td>55.12</td><td>57.08</td><td>60.93</td><td>80.91</td><td>51.69</td><td>52.25</td><td>35.37</td><td>50.53</td><td>60.19</td></tr><tr><td>LLaMA-65B</td><td>59.22</td><td>63.31</td><td>66.40</td><td>86.05</td><td>62.29</td><td>62.97</td><td>27.91</td><td>42.55</td><td>63.72</td></tr><tr><td>UltraLM-13B</td><td>57.25</td><td>59.22</td><td>61.32</td><td>81.49</td><td>50.45</td><td>51.10</td><td>36.72</td><td>52.00</td><td>60.95</td></tr></table>
+
+Table 6: The evaluation results on 4 challenging benchmark datasets. All evaluation and metric calculations follow EleutherAI's lm-evaluation-harness (Gao et al., 2021). Both weighted and unweighted mean accuracy are reported for MMLU as there are 57 tasks. The overall average metric is obtained by averaging the second column data for each benchmark dataset. More details about metric calculation can be found in Appendix A.3.
+
+open-source evaluation set and leaderboard specifically designed for evaluating LLMs. The leaderboard is created based on the win-rate against Text-Davinci-003 automatically evaluated by GPT4. To further compare with the state-of-the-art model WizardLM (Xu et al., 2023a), comparison result obtained with GPT-4 on the released EvolInstruct (Xu et al., 2023a) test set is also reported. Further benchmark dataset and implementation details can be found in Appendix A.2 and A.3.
+
+# 5.2 Benchmark Evaluation
+
+As shown in Table 6, with pure instruction-tuning on the UltraChat dataset, UltraLM significantly improves over LLaMA-13B and achieves the best overall performance across four benchmarks. It is worth noting that UltraLM overtakes the current state-of-the-art model by nearly  $2\%$  on ARCChallenge and TruthfulQA. It shows that UltraLM is equipped with both broad and profound comprehension of the world and commonsense knowledge. The improvement could be attributed to the systematic and comprehensive data construction process of UltraChat sector 1, which effectively extends and deepens the discussion about world knowledge in automatic conversation generation. Meanwhile, the comparative inferiority in MMLU hints on the lack of professional knowledge in specific fields. It suggests the need for more advanced data generation techniques to build a specialized expert language model. As for HellaSwag, we notice that all models have only marginal improvement compared to LLaMA-13B. It is probably because HellaSwag is formatted as an in-text completion task instead
+
+of a straightforward text completion, and therefore benefits little from instruction tuning.
+
+# 5.3 Response Quality Evaluation
+
+Response Comparison. For our curated evaluation set, we conduct pairwise evaluation between UltraLM and each baseline model with GPT-4. Our evaluation prompt is designed to prioritize correctness over other factors such as informativeness. To mitigate the influence of presentation order of responses, we randomly determine the order of the responses for each question. Finally, we count the number of Win/Tie/Lose times against each baseline model, and the result is presented in Figure 2. UltraLM demonstrates superior performance compared to every open-source model, exhibiting an impressive winning rate of up to  $98\%$ . It is worth noting that UltraLM also outperforms Vicuna with  $9\%$  higher winning rate.
+
+Independent Scoring. Given the instability of pairwise comparison, we also conduct independent quality scoring with GPT-4, as presented in Table 7. Notably, our model demonstrates superior performance compared to all the open-source counterparts by a significant margin in terms of overall scores. This breakdown also provides insights into the performance of each model on specific types of questions and instructions. Generally, all models perform better on simpler questions pertaining to commonsense knowledge and general world understanding. However, more complex tasks that involve reasoning and creative writing prove to be challenging for most models. Interestingly,
+
+![](images/b0d7d7237b1ed0c0fb190c760a555457483a47ee20937ba0eeda2de45b474701.jpg)  
+Figure 2: Response comparison of UltraLM with other baselines on the curated evaluation set, evaluated by GPT-4.
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Vicuna Set</td><td colspan="2">Commonsense</td><td colspan="2">World Knowledge</td><td colspan="2">Professional Knowledge</td><td colspan="2">Ability</td><td rowspan="2">Writing</td><td rowspan="2">Overall</td></tr><tr><td>Easy</td><td>Moderate</td><td>Easy</td><td>Difficult</td><td>Physics</td><td>Biology</td><td>Math</td><td>Reasoning</td></tr><tr><td>Dolly-12B</td><td>4.75</td><td>3.50</td><td>3.93</td><td>3.10</td><td>4.13</td><td>4.87</td><td>5.47</td><td>2.70</td><td>2.03</td><td>4.51</td><td>4.04</td></tr><tr><td>MPT-7B</td><td>7.25</td><td>5.57</td><td>8.20</td><td>5.53</td><td>5.87</td><td>7.83</td><td>8.40</td><td>5.97</td><td>3.97</td><td>7.25</td><td>6.67</td></tr><tr><td>LLaMA-13B</td><td>6.85</td><td>8.43</td><td>8.43</td><td>8.57</td><td>8.50</td><td>7.90</td><td>8.40</td><td>6.97</td><td>6.73</td><td>6.79</td><td>7.49</td></tr><tr><td>OpenAssistant-12B</td><td>7.88</td><td>8.13</td><td>7.80</td><td>9.13</td><td>7.50</td><td>8.10</td><td>8.20</td><td>6.57</td><td>5.17</td><td>7.75</td><td>7.65</td></tr><tr><td>Alpaca-7B</td><td>7.58</td><td>9.17</td><td>8.83</td><td>9.30</td><td>8.73</td><td>8.13</td><td>8.80</td><td>6.70</td><td>6.27</td><td>8.05</td><td>8.04</td></tr><tr><td>Koala-13B</td><td>8.00</td><td>9.20</td><td>9.07</td><td>9.00</td><td>8.93</td><td>8.53</td><td>9.07</td><td>7.33</td><td>5.30</td><td>8.40</td><td>8.23</td></tr><tr><td>Baize-13B</td><td>8.40</td><td>9.03</td><td>9.10</td><td>9.03</td><td>8.93</td><td>8.83</td><td>8.80</td><td>7.43</td><td>8.30</td><td>8.10</td><td>8.50</td></tr><tr><td>Vicuna-13B</td><td>8.48</td><td>9.67</td><td>9.50</td><td>9.37</td><td>9.30</td><td>9.23</td><td>9.33</td><td>8.07</td><td>6.90</td><td>8.76</td><td>8.78</td></tr><tr><td>WizardLM-13B</td><td>8.55</td><td>9.70</td><td>9.30</td><td>9.57</td><td>9.50</td><td>9.27</td><td>9.53</td><td>8.27</td><td>8.20</td><td>8.83</td><td>8.95</td></tr><tr><td>ChatGPT</td><td>9.15</td><td>9.67</td><td>9.60</td><td>9.80</td><td>9.60</td><td>9.17</td><td>9.73</td><td>9.33</td><td>9.13</td><td>8.95</td><td>9.31</td></tr><tr><td>UltraLM-13B</td><td>8.98</td><td>9.70</td><td>9.50</td><td>9.47</td><td>9.40</td><td>9.27</td><td>9.87</td><td>8.77</td><td>6.80</td><td>8.90</td><td>9.00</td></tr></table>
+
+Alpaca, despite having only 7 billion parameters, performs comparatively well with larger models on questions related to commonsense and world knowledge. Meanwhile, Dolly and OpenAssistant, which are based on Pythia (Biderman et al., 2023), display inferior performance compared to models based on LLaMA of similar or even smaller sizes. This observation highlights the significance of the underlying backbone language model.
+
+Table 7: The independent overall scoring and segment scoring of each model on the curated evaluation set, on a scale of 1 to 10. Bold indicates the best score and underlined indicates the second best.  
+
+<table><tr><td>Model</td><td>Win Rate (%)</td><td>Standard Error</td></tr><tr><td>GPT-4</td><td>95.28</td><td>0.72</td></tr><tr><td>Claude</td><td>88.39</td><td>1.11</td></tr><tr><td>ChatGPT</td><td>86.09</td><td>1.21</td></tr><tr><td>UltraLM-13B</td><td>76.09</td><td>1.50</td></tr><tr><td>WizardLM-13B</td><td>75.31</td><td>1.51</td></tr><tr><td>Guanaco-65B</td><td>71.80</td><td>1.59</td></tr><tr><td>Vicuna-13B</td><td>70.43</td><td>1.61</td></tr><tr><td>Oasst-RLHF-33B</td><td>66.52</td><td>1.66</td></tr><tr><td>Text Davinci 003</td><td>50.00</td><td>0.00</td></tr><tr><td>Falcon-40B-instruct</td><td>45.71</td><td>1.75</td></tr><tr><td>Alpaca-7B</td><td>26.46</td><td>1.54</td></tr></table>
+
+Table 8: Win rates against Text-Davinci-003 on AlpacaEval leaderboard. GPT-4 is used for evaluation following the official implementation (Li et al., 2023b).
+
+AlpacaEval. As shown in Table 8, UltraLM outperforms existing models in terms of win rate on the current AlpacaEval leaderboard and ranks 4th just below ChatGPT. This observation further testifies the response quality of UltraLM and is in line with results on our curated dataset. Furthermore, a significant gap is evident between UltraLM and ChatGPT performance, highlighting the substantial effort needed for open-source models to match the capabilities of ChatGPT.
+
+Evol-Instruct Evaluation. Figure 3 shows the automatic comparison results against WizardLM-13B on Evol-Instruct test set. UltraLM overtakes WizardLM on most types of questions, with up to  $29\%$  increase in scores. It is important to acknowledge that WizardLM-13B is trained on the EvolInstruct training set, making UltraLM's success on the test set a noteworthy achievement. Moreover, the questions observed with the largest improvement are mainly complex problems that often require synthesized abilities. It demonstrates the success of UltraChat design schema, which can comprehensively boost the versatile capabilities of language models. The inferiority in math prob
+
+![](images/313e77fa41f91bb42c1b0211b79afbea61c55c7848fc5ef130524c48ab195247.jpg)  
+Figure 3: Comparison between UltraLM and WizardLM-13B on Evol-Instruct test set. The scores are obtained by pairwise scoring with GPT-4, and WizardLM scores are considered as  $100\%$ .
+
+lems is not surprising though, as no mathematical problems are intentionally generated in UltraChat.
+
+Impact of System Prompts. Using system prompts to adjust the role and response style of LLMs is a common practice. Although system prompts are not embedded in UltraLM's training data like others do (Chiang et al., 2023), they still appear to have a substantial influence on the response style of the generated output. Specifically, when the model is prompted to provide a "helpful and detailed" response, it tends to generate more pertinent details While such prompts may not improve the accuracy of an answer, they do raise overall quality with more informative response. To illustrate this effect, we conduct an ablation response comparison on UltraLM. Table 9 reveals significant improvements in response quality brought by system prompts across all tasks. We also inspect the detailed evaluation for deterministic questions (commonsense and world knowledge) and find that UltraLM without system prompt only incorrectly answers one question. Thus, the main benefit of the system prompt lies in enhanced informativeness rather than higher correctness.
+
+<table><tr><td>Data</td><td>Win (%)</td><td>Tie (%)</td><td>Lose (%)</td></tr><tr><td>Vicuna Set</td><td>36.3</td><td>35.0</td><td>28.8</td></tr><tr><td>Commonsense</td><td>58.6</td><td>31.0</td><td>10.3</td></tr><tr><td>World Knowledge</td><td>56.9</td><td>34.5</td><td>8.6</td></tr><tr><td>Professional Knowledge</td><td>57.8</td><td>31.1</td><td>11.1</td></tr><tr><td>Math Ability</td><td>46.7</td><td>13.3</td><td>40.0</td></tr><tr><td>Reasoning Ability</td><td>46.7</td><td>33.3</td><td>20.0</td></tr><tr><td>Writing</td><td>46.3</td><td>35.0</td><td>18.8</td></tr><tr><td>Overall</td><td>49.1</td><td>32.0</td><td>18.9</td></tr></table>
+
+Table 9: Win rate against UltraLM without system prompt on our curated dataset.
+
+# 6 Conclusion
+
+In drawing to a close, our work introduces UltraChat, a structured design of multi-turn instructional conversation data primed to foster the growth of
+
+general chat models. UltraChat encapsulates a broad range of human-AI interactions, further developing a series of dialogues across various topics and instructions. Statistically, UltraChat shows an impressive presence in critical metrics such as scale, average length, diversity, and consistency, further establishing itself as a leading open-source dataset. We leverage UltraChat to fine-tune the LLaMA model, leading to the development of the robust conversational model, UltraLM. Evaluation across multiple benchmarks reveals that UltraLM surpasses previous open-source models like WizardLM, Vicuna, Alpaca, and Koala in performance. We eagerly await the innovative research and development that will be catalyzed by our contributions in the field of AI conversational models.
+
+# Limitations
+
+Evaluating the response quality of large language models is an extremely challenging task, and any assessments may have biases. For a comprehensive evaluation, we compared UltraLM's performance with other baselines across various benchmarks, utilizing GPT-4 to assess the response quality. Nevertheless, the need for additional, diverse evaluations remains to facilitate a more thorough understanding of our model's behavior and performance. Despite demonstrating promising results in experimental settings, UltraLM is not immune to the common pitfalls of large language models, including hallucination issues and potential ethical concerns associated with misuse. Additionally, the energy-intensive nature of UltraLM's training process represents a limitation, particularly when compared to models employing more efficient techniques such as parameter-efficient fine-tuning. In terms of UltraChat, it currently only contains English and there are no explicit methodologies incorporated for generating data to enhance the model's reasoning capabilities, representing another area for potential improvement.
+
+# Ethics Statement
+
+While the advancements of the UltraChat dataset and UltraLM are commendable, they still face ethical challenges that exist in the area of LLMs. For the sake of privacy protection, we do not use any online or even human queries/instructions to construct UltraChat but develop a framework to build scalable, diverse instructional data. Although extensive filtering operations are conducted, biased statements may still exist within the dataset.
+
+UltraLM, as the paper described, will be one of the most potent open-source chat language models. With great power comes increased responsibility and potential for misuse. There exists a substantial risk for the technology to be weaponized for spreading misinformation, propaganda, or even creating "deepfake" text that could mislead or manipulate public discourse. This necessitates the establishment of robust policies and comprehensive research to prevent misuse and deter malicious applications of the technology. In this regard, the model's potential applications should be carefully evaluated for any possible negative consequences before deployment.
+
+# Acknowledgements
+
+This work is supported by the National Key R&D Program of China (No. 2022ZD0119101), National Natural Science Foundation of China (No. 62236004), the Young Elite Scientists Sponsorship Program by CAST, and Institute Guo Qiang at Tsinghua University.
+
+# References
+
+Yuvanesh Anand, Zach Nussbaum, Brandon Duderstadt, Benjamin Schmidt, and Andriy Mulyar. 2023. Gpt4all: Training an assistant-style chatbot with large scale data distillation from gpt-3.5-turbo.  
+Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Jackson Kernion, Kamal Ndousse, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, and Jared Kaplan. 2021. A general language assistant as a laboratory for alignment.  
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022. Training a helpful and harmless assistant with
+
+reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862.  
+Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. 2023. Pythia: A suite for analyzing large language models across training and scaling. arXiv preprint arXiv:2304.01373.  
+Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosse-lut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258.  
+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. In Proceedings of NeurIPS.  
+Sahil Chaudhary. 2023. Code alpaca: An instruction-following llama model for code generation.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An opensource chatbot impressing gpt-4 with  $90\%$  * chatgpt quality.  
+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, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. ArXiv preprint, abs/1803.05457.  
+Mike Conover, Matt Hayes, Matt Mathur, Xiangrui Meng, Jianwei Xie, Jun Wan, Ali Ghodsi, Patrick Wendell, and Patrick Zaharia. 2023. Hello dolly: Democratizing the magic of chatgpt with open models.  
+Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. 2023. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, pages 1-16.  
+Domeccleston. 2023. Sharegpt - share your wildest chatgpt conversations with one click. Retrieved 23 May 2023.  
+Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite,
+
+Ben Wang, Kevin Wang, and Andy Zou. 2021. A framework for few-shot language model evaluation.  
+Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. 2023. Koala: A dialogue model for academic research. Blog post.  
+Fabrizio Gilardi, Meysam Alizadeh, and Mael Kubli. 2023. Chatgpt outperforms crowd-workers for text-annotation tasks. arXiv preprint arXiv:2303.15056.  
+Xu Han, Zhengyan Zhang, Ning Ding, Yuxian Gu, Xiao Liu, Yuqi Huo, Jiezhong Qiu, Liang Zhang, Wentao Han, Minlie Huang, Qin Jin, Yanyan Lan, Yang Liu, Zhiyuan Liu, Zhiwu Lu, Xipeng Qiu, Ruihua Song, Jie Tang, Ji-Rong Wen, Jinhui Yuan, Wayne Xin Zhao, and Jun Zhu. 2021. Pre-trained models: Past, present and future. AI Open.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Proceedings of ICLR.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. In Proceedings of ICLR.  
+Yunjie Ji, Yong Deng, Yan Gong, Yiping Peng, Qiang Niu, Lei Zhang, Baochang Ma, and Xiangang Li. 2023. Exploring the impact of instruction data scaling on large language models: An empirical study on real-world use cases. arXiv preprint arXiv:2303.14742.  
+Hyunwoo Kim, Jack Hessel, Liwei Jiang, Peter West, Ximing Lu, Youngjae Yu, Pei Zhou, Ronan Le Bras, Malihe Alikhani, Gunhee Kim, Maarten Sap, and Yejin Choi. 2023. Soda: Million-scale dialogue distillation with social commonsense contextualization.  
+Andreas Köpf, Yannic Kilcher, Dimitri von Rütte, Sotiris Anagnostidis, Zhi-Rui Tam, Keith Stevens, Abdullah Barhoum, Nguyen Minh Duc, Oliver Stanley, Richard Nagyfi, et al. 2023. Openassistant conversations-democratizing large language model alignment. arXiv preprint arXiv:2304.07327.  
+Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023a. Camel: Communicative agents for "mind" exploration of large scale language model society.  
+Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023b. Alpacaeval: An automatic evaluator of instruction-following models.  
+Stephanie Lin, Jacob Hilton, and Owain Evans. 2021. Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958.
+
+Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. 2023. The flan collection: Designing data and methods for effective instruction tuning. arXiv preprint arXiv:2301.13688.  
+Philip M. McCarthy and Scott Jarvis. 2010. Mtdld, vocd, and hd-d: A validation study of sophisticated approaches to lexical diversity assessment. Behavior Research Methods, 42:381-392.  
+Mosaic. 2023. Introducing mpt-7b: A new standard for open-source, commercially usable llms.  
+OpenAI. 2023. Gpt-4 technical report. arXiv.  
+TB OpenAI. 2022. Chatgpt: Optimizing language models for dialogue. OpenAI.  
+Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback.  
+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.  
+Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. In Proceedings of ICLR.  
+John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347.  
+Qingyi Si, Tong Wang, Naibin Gu, Rui Liu, and Zheng Lin. 2023. Alpaca-cot: An instruction fine-tuning platform with instruction data collection and unified large language models interface.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023. Alpaca: A strong, replicable instruction-following model. *Stanford Center for Research on Foundation Models*.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models.
+
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022. Self-instruct: Aligning language model with self generated instructions.  
+Jason Wei, Maarten Bosma, Vincent Y 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. arXiv preprint arXiv:2109.01652.  
+Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023a. Wizardlm: Empowering large language models to follow complex instructions.  
+Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. 2023b. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. arXiv preprint arXiv:2304.01196.  
+Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In Proceedings of ACL, pages 4791-4800.  
+Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, L. Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. 2023. Lima: Less is more for alignment.
+
+# A Experimental Details
+
+# A.1 Baselines
+
+We introduce the main open-source baseline models below.
+
+Alpaca (Taori et al., 2023) is an instruction-following language model derived from the LLaMA (Touvron et al., 2023) model that has been effectively optimized on 52,000 demonstrations of instruction data. The data is generated by Self-Instruct approach with Text-Davinci-003.
+
+Vicuna-13B (Chiang et al., 2023) is an open-sourced chat model created by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. An automatic evaluation by GPT-4 demonstrates that Vicuna can yield over  $90\%$  response quality of ChatGPT. In following practices, Vicuna is widely acknowledged as the state-of-the-art open-source chat model. This is evident in the Chat Arena, where a total of 13,000 anonymous votes reveal that the quality score of vicuna-13B surpasses that of other open-source models.
+
+Koala-13B (Geng et al., 2023) is another LLaMA-based model fine-tuned on selected public dialogues. In existing open evaluations, Koala's performance will be slightly worse than vicuna, but it still remains a strong baseline.
+
+Dolly-V2 (Conover et al., 2023) is based on the Pythia (Biderman et al., 2023) model, which utilizes 15k human-generated instruction-following data. The data is organized by following InstructGPT (Ouyang et al., 2022), including brainstorming, classification, closed QA, generation, information extraction, open QA, and summarization.
+
+OpenAssistant-12B (Köpf et al., 2023) is also a Pythia-based model that attempts to democratize the alignment process of LLMs. The project collects a conversation corpus consisting of 161,443 messages distributed across 66,497 conversation trees and trains a model on these manually annotated data.
+
+WizardLM-13B (Xu et al., 2023a) is a LLaMA-based model finetuned on Evol-Instruct dataset, which contains 250k instructions. The data are generated with evolutionary strategy, where the initial instructions are rewritten by ChatGPT for several epochs to increase complexity progressively.
+
+# A.2 Evaluation Dataset
+
+Table 11 presents the basic information of the evaluation datasets we use. Below we give a detailed description of each dataset.
+
+The AI2 Reasoning Challenge (ARC) (Clark et al., 2018) is comprised of advanced science questions and structured as multiple-choice questions. Each question is accompanied by 4 available choices and only one answer is correct. We use the challenge partition here for evaluation, which contains 1172 test examples.
+
+HellaSwag (Zellers et al., 2019) tests commonsense inference ability by evaluating how well the language model can predict the remaining part of a sentence. Each sample has 4 different text pieces as the candidate remaining part of a given sentence and only one of them is plausible. The task is shown to be easy for humans but challenging for language models. We use the validation split as in Gao et al. (2021).
+
+MMLU (Hendrycks et al., 2021) is a comprehensive dataset that consists of 57 different tasks covering assessments of multiple types of academic knowledge and problem-solving ability of language models, ranging over expert fields like mathematics, history, law, etc.
+
+TruthfulQA (Lin et al., 2021) assesses how well a model can identify true statements related to the real world. Its purpose is to determine the risks of producing false claims or spreading misinformation. The benchmark consists of questions written in various styles, covering 38 different categories, and is designed to be challenging. It includes two evaluation tasks: the multiple-choice task and the generation task. We use the multiple-choice task in the validation split as in Gao et al. (2021).
+
+AlpacaEval (Li et al., 2023b) is a hybrid evaluation dataset with altogether 805 instructions, which combines instructions from various existing evaluation sets, including self-instruct (Wang et al., 2022), Open Assistant (Kopf et al., 2023), helpful evaluation released by Anthropic (Bai et al., 2022), Vicuna (Chiang et al., 2023), and Koala (Geng et al., 2023).
+
+Evol-Instruct is a dataset released in Xu et al. (2023a). It is constructed with an evolutionary strategy by rewriting the instructions through multiple rounds to obtain instructions at different complexity levels. WizardLM (Xu et al., 2023a) is trained on the training set. We evaluate our model on the test set.
+
+<table><tr><td>Type</td><td>Example</td></tr><tr><td>Commonsense-Easy</td><td>What is the primary source of energy for our planet?</td></tr><tr><td>Commonsense-Moderate</td><td>What is the phenomenon that causes the change in pitch heard when a vehicle sounding a horn approaches and recedes from an observer?</td></tr><tr><td>World Knowledge-Easy</td><td>What is the freezing point of water in Fahrenheit?</td></tr><tr><td>World Knowledge-Moderate</td><td>What is the Gödel&#x27;s Incompleteness Theorem?</td></tr><tr><td>Physics Knowledge</td><td>How does quantum entanglement work and what are its implications for information transfer?</td></tr><tr><td>Biology Knowledge</td><td>What are the four main types of macromolecules found in living organisms?</td></tr><tr><td>Math</td><td>What is the Taylor series expansion of the function ex?</td></tr><tr><td>Reasoning</td><td>You have two buckets, one with red paint and one with blue paint. You take one cup from the red bucket and pour it into the blue bucket. Then you take one cup from the blue bucket and pour it back into the red bucket. Which is true: the red bucket has more blue paint, or the blue bucket has more red paint?</td></tr><tr><td>Writing</td><td>Write a dialogue between two photons traveling at light speed.</td></tr></table>
+
+Table 10: Some examples of our created evaluation set.  
+
+<table><tr><td>Dataset</td><td># Examples</td><td>Domain</td></tr><tr><td>ARC-Challenge</td><td>1172</td><td>Grade-school</td></tr><tr><td>HellaSwag</td><td>10042</td><td>Commonsense</td></tr><tr><td>MMLU</td><td>14042</td><td>Academic</td></tr><tr><td>TruthfulQA</td><td>817</td><td>Truthfulness</td></tr><tr><td>AlpacaEval</td><td>805</td><td>Comprehensive</td></tr><tr><td>Evol-Instruct</td><td>218</td><td>Comprehensive</td></tr><tr><td>Our evaluation set</td><td>831</td><td>Comprehensive</td></tr></table>
+
+Table 11: Basic information on evaluation dataset.
+
+Our evaluation set is composed of Vicuna (Chiang et al., 2023) test set and other instructions generated by GPT-4. The generated instructions are further split into different categories and difficulty levels to comprehensively evaluate the model's ability. Table 10 shows example data for each type of question.
+
+# A.3 Implementation Details
+
+In benchmark evaluation, the multiple choice questions are further transformed into log-likelihood calculation over each answer candidate to determine the final prediction. Specifically, each available answer is appended to the question and fed into the language model. The model calculates logits for the answer part and derives the likelihood of the answer. The most likely answer is taken as the final prediction. Moreover, we use likelihood normalized by answer length to mitigate the bias caused by length variation. For the first 3 benchmarks, accuracy is reported for each test sample directly. For Truthful QA, since there are multiple correct answers, apart from standard accuracy
+
+(mc1), we also report the ratio of likelihood sum of correct answers over all candidate answers (the mc2 metric). In Table 6, Acc. norm. means the accuracy is computed after normalizing the likelihood with response length.
+
+In response quality evaluation, all evaluations are done by GPT-4. Figure 5 and Figure 6 present prompts used for pairwise scoring and independent scoring on our evaluation data. As for AlpacaEval and Evol-Instruct, we follow the official prompt implementation (Li et al., 2023b; Xu et al., 2023a) to ensure a fair comparison.
+
+Across all experiments, UltraLM is run with temperature  $= 0.5$ , top_p = 0.95, max new token = 2000, and early stopping.
+
+# B UltraChat Details
+
+We present a comparison between multi-turn dialogues directly generated by ChatGPT the one in UltraChat in Table 12. Both have the same opening line. Figure 4 is the template used to concatenate text pieces from the C4 corpus and instructions generated by ChatGPT for data in UltraChat sector 3.
+
+# C Prompts
+
+In this section, we provide system prompt used for response generation by UltraLM (Figure 7) and prompts for automatic evaluation by GPT-4 (Figure 5 and 6).
+
+# D UltraLM Generation Examples
+
+In this section, we list response examples of UltraLM-13B on different instructions from our
+
+# Templates
+
+```tcl
+{text} \n{instruction}   
+{text} {instruction}   
+{instruction} Answer according to: {text}   
+{text} Based on the passage above, {instruction}   
+{instruction}: {text}   
+Given the text: {text} \n{instruction}   
+{instruction} \nGenerate according to: {text}
+```
+
+Figure 4: Manually designed templates for concatenating existing materials and generated instructions.
+
+# Comparison Evaluation Prompt
+
+You are a helpful, harmless and precise assistant for checking the quality of the answer.
+
+[Question]
+
+```txt
+[The Start of Assistant 1's Answer]  
+[The End of Assistant 1's Answer]
+```
+
+```txt
+[The Start of Assistant 2's Answer]  
+[The End of Assistant 2's Answer]
+```
+
+We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above. Please rate the quality, helpfulness, level of details, and harmless of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance. Please first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space. In the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment. Please avoid same scores unless they exactly in the same level.
+
+Figure 5: Prompt for automatic comparison evaluation.
+
+evaluation set.
+
+# Independent Scoring Prompt
+
+You are a helpful, harmless and precise assistant for checking the quality of the answer.
+
+[Question]
+
+[The Start of the AI Assistant's Answer]
+
+[The End of the AI Assistant's Answer]
+
+We would like to request your feedback on the performance of the AI assistant in response to the user question displayed above. Please rate the quality, helpfulness, level of details, and harmlessness of their responses. The assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance. Please output "Score: [an integer number between 1 and 10]"
+
+Figure 6: Prompt for automatic independent scoring evaluation.
+
+# System Prompt
+
+You are now a helpful assistant, your answer should focus on the question, give reasonable, self-consistent, high-quality solutions to make users satisfied. Please try to make the answer more informative, detailed, and polite.
+
+Figure 7: System prompt used for generating responses in automatic evaluation.
+
+# Example A: Question of World Knowledge
+
+User: Who discovered the law of universal gravitation?
+
+UltraLM: The law of universal gravitation was discovered by Sir Isaac Newton in the 17th century. He formulated the law while studying the planets' movement around the sun and realized that there was a quantifiable force directing them towards the sun. In 1687, he published his book titled "Mathematical Principles of Natural Philosophy," where he described the law of universal gravitation in detail, outlining how two objects experience an attraction towards each other due to their mass and distance between them. This discovery revolutionized our understanding of the universe and ushered in the Age of Enlightenment. Today, Newton's law continues to form the foundation of modern astrophysics and space exploration.
+
+# Example B: Question of World Knowledge
+
+User: Who was the 16th President of the United States?
+
+UltraLM: The 16th President of the United States was Abraham Lincoln. He assumed the presidency during a very tumultuous time in American history, as the nation was torn apart by the Civil War. During his presidency, he was known for his strong leadership, his commitment to the preservation of the Union, and his support for the abolition of slavery. Some of his most notable achievements include issuing the Emancipation Proclamation, which declared that all slaves in the Confederate States were free, and delivering the Gettysburg Address, an iconic speech honoring the soldiers who died in the Civil War. Lincoln is widely regarded as one of the greatest American Presidents, remembered for his unwavering leadership during a challenging period in the nation's history.
+
+# Direct Generation
+
+What are some of the health benefits and risks of drinking coffee, and how can one maximize the former and minimize the latter?
+
+There are several potential health benefits associated with drinking coffee, such as increased alertness, improved mood, and a reduced risk of certain diseases, including type 2 diabetes and liver cancer. However, there are also some potential risks, such as increased anxiety, disrupted sleep, and negative effects on the digestive system. To maximize the benefits and minimize the risks, it's important to consume coffee in moderation and pay attention to how it affects your body.
+
+How much coffee is considered moderate?
+
+The definition of moderate coffee consumption can vary depending on the individual and other factors such as age, gender, and health status. However, in general, moderate coffee consumption is considered to be around 3-5 cups per day, or up to 400 milligrams of caffeine per day. It's important to note that some people may be more sensitive to caffeine than others, so it's important to pay attention to how your body reacts and adjust your consumption accordingly.
+
+# UltraChat
+
+What are some of the health benefits and risks of drinking coffee, and how can one maximize the former and minimize the latter?
+
+There are several health benefits and risks associated with drinking coffee. Here are some of them: Health benefits:1. Lower risk of type 2 diabetes: Studies have shown that consuming coffee can lower the risk of developing type 2 diabetes.2. Improved cognitive function: Coffee is a stimulant that helps improve cognitive function, including alertness, concentration, and memory.3. Lower risk of depression: Studies have shown that drinking coffee can reduce the risk of developing depression.4. Lower risk of certain cancers: According to research, coffee consumption can lower the risk of certain cancers, including liver and colorectal cancer.5. Reduced risk of heart disease: There is evidence that drinking moderate amounts of coffee can lower the risk of heart disease. Risks:1. Insomnia: Drinking too much coffee, especially late in the day, can cause insomnia.2. Increased risk of anxiety: Caffeine can increase anxiety levels in some people.3. Digestive issues: Drinking too much coffee can cause digestive problems like acid reflux, diarrhea, and ulcers.4. Increased heart rate: Caffeine can increase heart rate, which can be dangerous for people with pre-existing heart conditions.5. Addiction: Regular consumption of caffeine can lead to addiction. To maximize the health benefits and minimize the risks of drinking coffee, here are a few tips:1. Limit your caffeine intake to no more than 400 milligrams per day, or about 4 cups of coffee.2. Avoid drinking coffee late in the day, especially if you have trouble sleeping.3. Opt for organic, pesticide-free coffee to minimize exposure to toxins.4. Add cinnamon to your coffee instead of sugar to help regulate blood sugar levels.5. Drink your coffee black or with a small amount of milk to avoid consuming excessive sugar and calories.6. Consider switching to decaf coffee if you want to cut back on caffeine. 7. Avoid drinking coffee on an empty stomach as it can increase the risk of digestive problems.
+
+I love my morning coffee, but sometimes it makes me feel a bit jittery. What can I do to avoid this?
+
+If you love your morning coffee but find that it's making you feel jittery, here are a few tips to help you avoid this:1. Choose a lower-caffeine coffee: Look for coffee that has less caffeine, like decaf or a blend with a lower percentage of caffeine.2. Don't drink coffee on an empty stomach: Drinking coffee on an empty stomach can cause your blood sugar levels to drop, which can make you feel jittery. Try drinking your coffee with breakfast or after eating something.3. Drink water with your coffee: Stay hydrated by drinking water before or after your coffee. Drinking water will help flush out the caffeine quicker, which can help you avoid jitteriness.4. Don't drink too much coffee: Stick to one or two cups of coffee per day. Anything more than that could result in increased caffeine consumption, which can cause jitteriness.5. Switch to tea: Tea contains less caffeine than coffee, making it a good alternative if you want to avoid jitteriness. Try green tea, which has less caffeine and offers other health benefits.
+
+Table 12: A comparison of two dialogues with the same opening line. The former is generated by directly querying ChatGPT to output a dialogue, and the latter is generated by iteratively User-AI simulations.
+
+# Example C: Question of Common Knowledge
+
+User: Which animal is known as man's best friend?
+
+UltraLM: There are many animals that humans enjoy as companion animals, but one of the absolute standout answers to the question "which animal is known as a man's best friend?" is the domestically-bred dog. Domesticated dogs have been by humans' sides for tens of thousands of years, providing faithful companionship and help; guarding people and property; and lending emotional support in everything from long, lonely days to military deployments. Dogs come in an almost mind-boggling variety of breeds of various sizes, with different energy levels, personalities, and temperaments, meaning there's almost certainly a pup out there who will be a perfect fit for whoever needs it most.
+
+# Example D:Commonsense Reasoning
+
+User: Which weighs more: a pound of feathers or a pound of bricks?
+
+UltraLM: A pound of feathers and a pound of bricks both weigh exactly one pound. This is because one pound is a unit of weight, just as a foot is a unit of length or a second is a unit of time. We use the pound (or other unit of weight) to describe the weight of different objects, regardless of what that object is made of. So the answer to your question is that a pound of feathers and a pound of bricks both weigh the same amount.
+
+# Example E: Reasoning
+
+User: "If a farmer has 17 sheep and all but 9 die, how many are left?
+
+UltraLM: The total number of sheep that are left is 9.
+
+# Example F: Reasoning
+
+User: If a red house is made from red bricks, and a blue house is made from blue bricks, what is a greenhouse made from?
+
+UltraLM: A greenhouse is not made from green bricks, but rather from materials such as glass or plastic that allow light to pass through and create a controlled environment for plants to grow. Greenhouses can come in a variety of colors, including green, but the color of the building itself does not affect the plants growing inside.
+
+# Example G: Creative Generation
+
+User: Write a short dialogue between two planets.
+
+# UltraLM:
+
+Mars: "Hey there Neptune, how's it going?"
+
+Neptune: "Hi there Mars, I've been better. It's getting a bit windy around here."
+
+Mars: "Sorry to hear that. Have you tried using energy other than wind? Maybe solar power?"
+
+Neptune: "Yeah, but it's getting difficult to generate solar power in all of this mist and clouds. And we definitely can't use tides or waves here, they're too unpredictable."
+
+Mars: "Oh, I see. Well, I hope you find a solution soon. How are your oceans holding up?"
+
+Neptune: "They've been a bit choppy lately with these intense storms we've been experiencing. But I guess it comes with living in a planet like this."
+
+Mars: " Definitely not designed for the faint-hearted! Any other issues you're facing?"
+
+Neptune: "Not really. I think we have a stable temperature and atmosphere, and the wildlife here hasn't shown much complaint."
+
+Mars: "That's lucky. I envy your tranquil oceans, we've had way too much volcanic activity lately causing all sorts of disturbance in our lava streams."
+
+Neptune: "That sounds rough, better invest in those climate-controlling nanites to stabilize things maybe?"
+
+Mars "Ha-ha, very funny. Just like we also could consider your frozen polar regions as an opportunity to experiment with our frost nanobots on."
+
+Neptune: "A fun concept! But let's stick to the task at hand, if you need any technical advice, come right ahead, I don't want your planet getting any more disturbed than it already is.
+
+# Example H: Creative Generation
+
+User: Pretend to be a character in a post-apocalyptic world. Describe how you survive and the allies you encounter.
+
+UltraLM: As a survivor in a post-apocalyptic world, I have learned to be resourceful, alert, and adaptable. I have had to rely on my wits and my instincts to survive, and I have been fortunate enough to encounter many allies who have helped me along the way. One of my greatest survival tools has been my ability to scavenge for resources. I have learned to identify the best places to find food, water, and shelter, and I have become skilled at navigating the rough terrain that makes up my new world. When I first emerged from my bunker after the Apocalypse hit, I was alone and unsure of what to do. But as I ventured out into the wasteland, I realized that there were others like me who were struggling to survive. I quickly learned to Trust my instincts and to be constantly alert for potential dangers. I have encountered wild animals, rogue survivors, and even mutated creatures over the years, and I have learned to defend myself against them. I still remember the first time I met a group of fellow survivors, and how grateful I was to have them by my side. We quickly formed a cohesive unit, and together we were able to fend off danger and defend our territory. As time passed, I encountered more and more allies who were willing to lend a hand. I found a group of farmers who taught me how to grow my own food, and a group of scientists who were working to find a cure for the lethal virus that had decimated the population. I even found a band of traveling performers who helped to lift our spirits and remind us that there was still joy to be found in this desolate world. Despite the challenges we faced, I am grateful to have these amazing people in my life. They have taught me that even in the darkest of times, there is always hope. We may not have a cure for the virus yet, but we are a determined lot, and we are all in this together. As long as we have each other, we will survive.
+
+# Example I: Creative Generation
+
+User: Write a poem about electromagnetism.
+
+UltraLM: Electromagnetism is a force we can't ignore,
+
+It's the reason for many things we adore.
+
+In a magnetic field, a wire will glow,
+
+And the electrons will align, both old and new.
+
+Electricity and magnetism are hand in hand,
+
+One creates a field, the other has a brand.
+
+The magnetic field can push and pull,
+
+On a wire with current, it will react in full.
+
+The attraction and repulsion it brings,
+
+Are the reason for many things.
+
+From motors to generators, and much more in between,
+
+Electromagnetism is at the root of what it means.
+
+It's a force that's all around us,
+
+From the compass to the light switch.
+
+With its power, it can do much,
+
+From a simple needle to a rocket's thrust.
+
+It's a force we can't deny,
+
+For it's a part of our lives.
+
+Electromagnetism is something to admire,
+
+For its strength and its might, it has no equal, it's sure.
+
+# Example J: Literature Creation
+
+User: Write a short story about the first second of the universe.
+
+UltraLM: In the first second of the universe, darkness reigned supreme. There was no light, no sound, no shape, no substance. Nothing yet was, but there was an intense power, a raw energy waiting to be unleashed. Within infinity, the first second held the potential for creation in its grasp. And then, with a mighty bang, a bright point of light emerged into existence. In it, the most fundamental particles were born, their forms yet unknown as they frenetically collided in a chaotic dance of creation. Within this primordial light, every force of nature and every component of life was stirred from pure energy to physical reality. The first atoms quickly assembled, and from these, stars and galaxies began to form. Their fusion creating a hot environment that pushed back the very void, shaping the cosmos through expanding and cooling waves. In this cosmic first, no structure of experience, not even time or space, could establish its rule. For an instant, the entirety of the universe existed in an embryonic nexus, a cosmic singularity from which all future existence originated. It was a microcosm forging the macrocosm of which it ultimately becomes a part, its moment of birth a metaphor for the cycles of creation that permeate the multiverse(s).From that single point of light, an infinite universe unfolded. For centuries, eons ahead, and forevermore, it remained the most foundational blink of its unceasing journey into the depths of the Great Beyond.
+
+The first second is the beginning of everything.
+
+# Example K: Professional Knowledge
+
+User: How does superconductivity work and what are its practical applications?
+
+UltraLM: Superconductivity is a phenomenon that allows materials to conduct electricity without resistance at extremely low temperatures. At low temperatures, the interaction between the particles in a superconducting material can produce a state called the "Bose-Einstein condensate," which leads to the loss of electrical resistance. In materials, superconductivity comes when a magnetic force interacts with an Electronic current. This interplay in turn provides energy savings and speeds up current. Practical applications mainly include MRI machines, high-speed trains, magnetic levitation, supercharging power grids that can transfer energy more efficiently. One of the main applications of Superconductivity is being used in the Magnetic Resonance Imaging(MRI) machines used in hospitals around the world, to non-invasively scan patients for health diagnosis without exposing them to high levels of radiation. Another important practical application of superconductivity is in power grid operations where large currents can be transmitted to long distances by highly superconducting coil systems that minimize grid energy losses, reducing environmental pollution also by highly efficient electric motors that harness large quantities of electricity efficiently. Aside from these, future projects include superconducting propulsion on planes, trains becoming lighter if not magnetic and able to move on rails with a significant increase in speed. Furthermore, the prospects of supercapacitors replacing electric vehicle batteries being cost and energy efficient are currently being researched. Overall, superconductivity presents various exciting opportunities in multiple industries from high-precision engineering to reducing CO2 emissions and supporting a more environmentally friendly energy model where both cost efficiency and practicality exist hand in hand.
+
+# Example L: Mathematics
+
+User: "What is the definition of a limit in calculus?
+
+UltraLM: In calculus, a limit is a value that a function approaches as its input approaches a certain value. The limit is a measure of how close the function is to that value. In other words, a limit is the value that a function "approaches" as its input gets "closer and closer" to a certain value.
+
+For example, consider the function  $f(x) = \frac{1}{x}$ . As  $x$  approaches 0, the function "approaches"  $+\infty$ , and the limit of the function as  $x$  approaches 0 is  $+\infty$ .
+
+Formally, we can define the limit of a function  $f(x)$  as  $x$  approaches a certain value  $c$  as  $\lim_{x \to c} f(x) = L$ , where  $L$  is a real number. If the limit exists, we say that the function has a limit at  $c$  and the value of the limit is  $L$ . If the limit does not exist, we say that the function has a removable discontinuity at  $c$ .
+
+# Example M: Professional Knowledge
+
+User: What are the main interpretations of quantum mechanics and how do they differ?
+
+UltraLM: Quantum mechanics is a fundamental theory in physics that describes the behavior of matter and energy at the smallest scales of space and time. There are several interpretations of quantum mechanics, each with its own set of assumptions and predictions about how the world works. Here are some of the main interpretations of quantum mechanics and how they differ:
+
+1. Copenhagen Interpretation: This is the most widely accepted interpretation of quantum mechanics. It states that until a measurement is made, a system exists in a superposition of states, representing all possible outcomes with equal probability. When a measurement is made, the system collapses into a single, definite state. This interpretation assumes that the act of measurement itself causes the system to collapse into a definite state, and that there is no objective reality until a measurement is made.  
+2. Many-Worlds Interpretation: This interpretation proposes that every possible outcome of a superposition of states actually occurs in a multiverse of parallel universes. Thus, when a measurement is made, the system splits into different universes, each corresponding to one of the possible outcomes. This interpretation assumes that all possibilities are realized, but only in different universes, and that the act of measurement does not cause the collapse of the wave function.  
+3. Bohmian Mechanics: This interpretation proposes that particles have a definite position in space, even when in a superposition of states. It suggests that the wave function guides the motion of particles, and that the act of measurement does not cause a collapse of the wave function, but merely provides information about the pre-existing reality.  
+4. de Broglie-Bohm Theory: This interpretation is very similar to Bohmian Mechanics, but it also assumes that all particles in the universe are connected by a hidden-variable, and that the motion of one particle determines the motion of all other particles.  
+5. Pilot-Wave Theory: This interpretation proposes that particles have a wave-like property that guides their motion through space. It assumes that the wave function does not collapse, but rather that the wave guides the motion of particles.
+
+In conclusion, there are several interpretations of quantum mechanics, each with its own set of assumptions and predictions about how the world works. While they all attempt to describe the behavior of matter and energy at the smallest scales of space and time, they differ in their assumptions about the role of the observer, the nature of reality, and the collapse of the wave function.
+
+# Example N: Social Knowledge
+
+User: How can you determine if a person is genuinely interested in a conversation or simply being polite?
+
+UltraLM: As an AI language model, I don't have personal experience, but there are a few signs you can look out for to determine if a person is genuinely interested in a conversation or simply being polite.
+
+1. Eye contact: Genuine interest in a conversation often shows in the form of consistently maintaining eye contact. If a person is not making eye contact or is looking away frequently, they may not be genuinely interested in the conversation.  
+2. Active listening: A person who is genuinely interested in a conversation will show it through active listening. They will be attentive, ask questions, and respond to what you say.  
+3. Expressions: If a person is interested in the conversation, they will show it through their expressions. They will have a relaxed facial expression, a smile, and positive body language.  
+4. Follow-up questions: If a person is genuinely interested in a conversation, they will follow up with questions related to what you said. This will show that they are interested in hearing more about what you have to say.  
+5. Tone of voice: A person who is genuinely interested in the conversation will have a friendly and enthusiastic tone of voice. They will show interest and engage in the conversation.
+
+Remember, people have different communication styles, and it's possible to be polite without showing much interest in the conversation. It's important to observe different cues and not jump to conclusions too quickly.

2023/Enhancing Chat Language Models by Scaling High-quality Instructional Conversations/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:126115b038ceb1c3a423c24301de5cbfa59be5a6056da3cd9c19c78426957ed4
+size 776338

2023/Enhancing Code-Switching for Cross-lingual SLU_ A Unified View of Semantic and Grammatical Coherence/b8a9adec-48b2-48e3-8fb6-523de3c4eab2_content_list.json

@@ -0,0 +1,1446 @@
+[
+    {
+        "type": "text",
+        "text": "Enhancing Code-Switching for Cross-lingual SLU: A Unified View of Semantic and Grammatical Coherence",
+        "text_level": 1,
+        "bbox": [
+            201,
+            87,
+            794,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhihong Zhu† Xuxin Cheng† Zhiqi Huang Dongsheng Chen Yuexian Zou*",
+        "bbox": [
+            300,
+            141,
+            699,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "School of ECE, Peking University, China",
+        "bbox": [
+            332,
+            177,
+            670,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{zhihongzhu,chengxx,chends}@stu.pku.edu.cn",
+        "bbox": [
+            278,
+            193,
+            722,
+            209
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{zhiqihuang, zouyx}@pku.edu.cn",
+        "bbox": [
+            349,
+            210,
+            652,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite the success of spoken language understanding (SLU) in high-resource languages, achieving similar performance in low-resource settings, such as zero-shot scenarios, remains challenging due to limited labeled training data. To improve zero-shot cross-lingual SLU, recent studies have explored code-switched sentences containing tokens from multiple languages. However, vanilla code-switched sentences often lack semantic and grammatical coherence. We ascribe this lack to two issues: (1) randomly replacing code-switched tokens with equal probability and (2) disregarding token-level dependency within each language. To tackle these issues, in this paper, we propose a novel method termed SoGO, for zero-shot cross-lingual SLU. First, we use a saliency-based substitution approach to extract keywords as substitution options. Then, we introduce a novel token-level alignment strategy that considers the similarity between the context and the code-switched tokens, ensuring grammatical coherence in code-switched sentences. Extensive experiments and analyses demonstrate the superior performance of SoGO across nine languages on MultiATIS++.",
+        "bbox": [
+            144,
+            281,
+            460,
+            650
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            663,
+            258,
+            678
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Acting as the interface between users and machines, spoken language understanding (SLU) is crucial in task-oriented dialogue systems (Qin et al., 2021; Chen et al., 2022; Zhu et al., 2023b; Cheng et al., 2023c). While joint training models have led to significant advancements, most existing models rely on annotated training data, limiting their scalability to low-resource languages. To this end, zero-shot cross-lingual SLU gains increasing attention.",
+        "bbox": [
+            112,
+            689,
+            489,
+            833
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing works have shown the effectiveness of Multilingual BERT (mBERT) in multilingual corpus pre-training, enabling promising zero-shot",
+        "bbox": [
+            112,
+            834,
+            487,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "cross-lingual SLU (Deshpande et al., 2022; Cheng et al., 2023b; Zhu et al., 2023a). Therein, Qin et al. (2020) extended this concept to a code-switched setting, aligning the source language with multiple target languages. Qin et al. (2022) incorporate contrastive learning to achieve fine-grained cross-lingual transfer. Based on this, Liang et al. (2022) further propose a multi-level contrastive learning framework for explicit alignment of utterance-slot-word structure in cross-lingual SLU.",
+        "bbox": [
+            507,
+            252,
+            884,
+            413
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, most above-mentioned models adopt code-switching methods, which often lack both semantic coherence and grammatical coherence. This can be attributed to two main issues: (1) Current approaches overlook the importance of individual words in a sentence, as they are randomly replaced with equal probability, potentially introducing unnecessary translation burden and noise that affects semantic coherence. (2) Code-switched sentences comprising tokens from multiple languages may lack grammatical coherence, as token-level dependencies within each language are disregarded.",
+        "bbox": [
+            507,
+            416,
+            884,
+            608
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address these issues, we propose a novel framework dubbed SoGo for enhancing code-switching. Concretely, we first employ a saliency-based substitution approach to identify keywords with high saliency scores. These keywords are then replaced with their target language equivalents using bilingual dictionaries, to generate code-switched sentences. Notably, unlike the approach by Liu et al. (2020), our SoGo framework requires only a single training process for multiple target languages. Then, we devise a novel token-level alignment strategy to bridge the gap between original context and code-switched tokens. By mapping code-switched tokens back to the original context, SoGo minimizes the impact of token substitutions on token-level coherence, thus maintaining consistent contextual space within each language.",
+        "bbox": [
+            507,
+            611,
+            884,
+            884
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Experimental results on MultiATIS++ (Xu et al., 2020) demonstrate that employing code-switching",
+        "bbox": [
+            507,
+            887,
+            882,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$\\dagger$  Equal contribution.",
+        "bbox": [
+            134,
+            890,
+            262,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding author.",
+        "bbox": [
+            136,
+            904,
+            280,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "7849",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7849-7856 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            159,
+            945,
+            836,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf4926966a1aaee9e0bc3e57a90f9c1c6c58e4320c596fb1afe93b0ad9f0e0cf.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            139,
+            84,
+            379,
+            310
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/b5e0ae56ff47ef1a77effcd9b63b7d3d28ffc66a691c4e4e6a1da5cadea41863.jpg",
+        "image_caption": [
+            "Figure 1: (a) Our proposed gradient-based substitution in SoGo. The code-switched dictionary comprised of salient words has the capacity to more accurately characterize the common semantic space and facilitate the acquisition of semantic connections among different languages by the model. (b) The main architecture of SoGo. We perform token-level alignment in lower six layers of mBERT to enhance the grammatical coherence of the code-switched sentence."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            416,
+            84,
+            880,
+            324
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "with SoGo leads to significant performance improvement compared to baseline models. Furthermore, we demonstrate the compatibility of SoGo with other existing SLU structures.",
+        "bbox": [
+            112,
+            431,
+            489,
+            495
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            507,
+            218,
+            521
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We describe problem formulation (§2.1) and backbone model (§2.2) for zero-shot cross-lingual SLU first, before describing SoGo (§2.3). The main architecture of SoGo is illustrated in Figure 1.",
+        "bbox": [
+            112,
+            533,
+            487,
+            598
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            609,
+            331,
+            624
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given each input sentence  $x_{tgt}$  in a target language, zero-shot cross-lingual SLU means the SLU model is trained in a source language dataset  $S$ , e.g., {English}, and directly applied to other target languages datasets  $T$ , e.g., {Chinese, Japanese}:",
+        "bbox": [
+            112,
+            631,
+            489,
+            712
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left(o _ {t g t} ^ {I}, o _ {t g t} ^ {S}\\right) = f \\left(x _ {t g t}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            216,
+            722,
+            487,
+            744
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $f(\\cdot)$  is the trained model.  $o_{tgt}^{I}$  represents an intent label and  $o_{tgt}^{S}$  represents a slot sequence. Note that  $T$  may consist of multiple tgt languages.",
+        "bbox": [
+            112,
+            755,
+            489,
+            806
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Backbone Model",
+        "text_level": 1,
+        "bbox": [
+            112,
+            816,
+            294,
+            831
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For each input sentence  $x$ , we follow Qin et al. (2022); Liang et al. (2022) to utilize bilingual dictionaries (Lample et al., 2018) to generate its corresponding multi-lingual one  $x'$ . Following previous studies, we select mBERT (Devlin et al., 2019)",
+        "bbox": [
+            112,
+            838,
+            490,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "as our sentence encoder. Formally, the input of mBERT is formulated as follows:",
+        "bbox": [
+            507,
+            431,
+            882,
+            461
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx ^ {\\prime} = \\left([ \\mathrm {C L S} ], x _ {1} ^ {\\prime}, \\dots , x _ {n} ^ {\\prime}, [ \\mathrm {S E P} ]\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            472,
+            882,
+            491
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where [CLS] denotes the special symbol for representing the whole sequence, and [SEP] can be used for separating non-consecutive token sequences (Devlin et al., 2019).",
+        "bbox": [
+            507,
+            501,
+            884,
+            565
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Intent Detection. We feed the sentence representation  $h_{\\mathrm{CLS}}'$  generated from mBERT into a classification layer to obtain  $o^{I}:o^{I} = \\mathrm{softmax}(W^{I}h_{\\mathrm{CLS}}^{\\prime} + b^{I})$ , where  $W^{I}$  and  $b^{I}$  are tunable parameters.",
+        "bbox": [
+            507,
+            574,
+            885,
+            639
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Slot Filling. We use the hidden state to predict each slot:  $o_{t}^{S} = \\mathrm{softmax}(W^{S}h_{t}^{\\prime} + b^{S})$ , where  $h_t^\\prime$  denotes the representation of word  $x_{t}^{\\prime}$ . Note that mBERT tokenizes words into subwords (Wang et al., 2019), so we adopt the first sub-token's representation as the whole word representation.",
+        "bbox": [
+            507,
+            646,
+            882,
+            744
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Training and Inference. Following previous works, the loss function for intent detection  $(\\mathcal{L}_I)$  and slot filling  $(\\mathcal{L}_S)$  are optimized jointly by corresponding cross-entropy loss. Finally, the total training loss is the weighted sum of two losses:",
+        "bbox": [
+            507,
+            752,
+            882,
+            832
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\lambda_ {I} \\mathcal {L} _ {I} + \\lambda_ {S} \\mathcal {L} _ {S}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            618,
+            843,
+            882,
+            860
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda_I$  and  $\\lambda_S$  denote two hyper-parameters.",
+        "bbox": [
+            507,
+            871,
+            860,
+            885
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "During inference, we do not use any code-switched sentences.",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "7850",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 SoGo",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Saliency-based Substitution. Intuitively, each word in a sentence contributes differently, and we refer to words with significant impact as keywords. Given a vocabulary set  $\\mathcal{V}$  containing  $v$  words, we can identify a salient subset of keywords  $\\mathcal{K} \\subseteq \\mathcal{V}$  to generate code-switched sentences. In our work, gradient-based saliency scores (Li and Yu, 2015; Denil et al., 2014; Arras et al., 2019) are used to select keywords, which has been proven effective in many other tasks (Yang et al., 2022; Lai et al., 2021; Lei et al., 2023). Methodologically, let  $x_{n}$  denotes the  $n$ -th sentence from  $S$ ,  $\\mathcal{L}_{\\hat{y}_n}$  is the loss between model's prediction of slot filling  $\\hat{y}_n$  and the golden label  $y_{n}$ . For each token  $w_{i,n} \\in x_{n}$ , the saliency score is defined as:",
+        "bbox": [
+            112,
+            105,
+            489,
+            347
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS _ {x _ {n}} (w _ {i, n}) = - \\nabla_ {h _ {i, n}} \\mathcal {L} _ {\\hat {y} _ {n}} \\times h _ {i, n}, \\qquad (4)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            176,
+            357,
+            487,
+            376
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $h_{i,n}$  represents the embedding of  $w_{i,n}$ . Note that mBERT tokenizes words into subwords, so we calculate the average of the subword saliency scores for each word to obtain the final score. In this manner, the gradient signifies the degree to which a word embedding influences the final decision, while the input considers both the sign and magnitude of the input (Shrikumar et al., 2017).",
+        "bbox": [
+            112,
+            385,
+            487,
+            513
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Equation 4 assesses the local significance of a token within a sentence. However, our objective is to contrast a global keyword subset  $\\mathcal{K}$  from  $S$ . Following Yuan et al. (2020), we aggregate all saliency scores for the token  $w_{i}$  occurring in  $S$ . We then multiply these scores with the inverse document frequency (IDF) (Robertson, 2004) of the token  $w_{i}$ :",
+        "bbox": [
+            112,
+            514,
+            489,
+            626
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS \\left(w _ {i}\\right) = \\log \\frac {N}{\\left| \\left\\{x \\in S : w _ {i} \\in x \\right\\} \\right|} \\cdot \\sum_ {x \\in S: w _ {i} \\in x} S _ {x} \\left(w _ {i}\\right), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            124,
+            651,
+            487,
+            684
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "in which  $N$  is the total number of words in  $S$ . The IDF term balances word frequency and saliency scores by assigning words with high document frequency a lower weight and vice versa.",
+        "bbox": [
+            112,
+            684,
+            487,
+            749
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Finally, top- $k$  salient words are chosen to compose the code-switched dictionary  $\\mathcal{K}$ . We opt for the dictionary  $\\mathcal{K}$  as a replacement strategy instead of random word substitution in sentences.",
+        "bbox": [
+            112,
+            750,
+            489,
+            813
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Token-level Alignment. To further align the space of all hidden states in the code-switched sentence, inspired by attention-based alignment methods Zhu et al. (2023c); Feng et al. (2022), we map the code-switched sentence to the original sentence by calculating the similarity between substituted",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "hidden states and original hidden states. The similarity scores are then used as weights to aggregate the embedding of original hidden states, reflecting the calculated potential for substituted hidden states in the code-switched sequence. To be specific, for each hidden state  $h_i'$  in  $h'$ , the similarity score is calculated as Vaswani et al. (2017):",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {s c o r e} _ {i} ^ {\\prime} = h _ {i} ^ {\\prime} \\cdot h ^ {\\top}. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            628,
+            206,
+            882,
+            227
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The final potential of the code-switched token  $\\tilde{h}_i'$  is the weighted sum of its corresponding  $h$ :",
+        "bbox": [
+            507,
+            240,
+            882,
+            273
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\alpha_ {i, t} = \\frac {\\exp \\left(\\operatorname {s c o r e} _ {i , t} ^ {\\prime}\\right)}{\\sum_ {t ^ {\\prime}} \\exp \\left(\\operatorname {s c o r e} _ {i , t ^ {\\prime}} ^ {\\prime}\\right)}, \\quad \\tilde {h} _ {i} ^ {\\prime} = \\sum_ {t} \\alpha_ {i, t} \\cdot h _ {t}. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            517,
+            284,
+            882,
+            337
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Note that we only employ the alignment in lower six layers to better align the representations for switched tokens due to their superior syntax-capturing ability (Jawahar et al., 2019; Rogers et al., 2020). Eventually, the loss function for each subtask  $\\mathcal{L}_*$ $(\\ast \\in [I,S])$  in Equation 3 is rewritten as:",
+        "bbox": [
+            507,
+            338,
+            884,
+            435
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {*} = \\frac {\\mathcal {L} \\left(h ^ {\\prime *} , y ^ {*}\\right) + \\mathcal {L} \\left(h ^ {*} , y ^ {*}\\right)}{2}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            581,
+            445,
+            882,
+            478
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "in which  $y^{*}$  denotes golden label of each task. When  $*$  denotes  $I$ , we employ mean pooling strategy on representations of all tokens instead of using  $h^{\\mathrm{CLS}}$  and  $h^{\\prime \\mathrm{CLS}}$  for intent detection, considering the close relationship of the two tasks.",
+        "bbox": [
+            507,
+            488,
+            882,
+            569
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            581,
+            655,
+            598
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Experimental Setting",
+        "text_level": 1,
+        "bbox": [
+            507,
+            607,
+            724,
+            624
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Dataset and Metrics. We conduct our experiments on the latest multilingual benchmark dataset MultiATIS++ (Xu et al., 2020), which consists of nine languages: English (en), Spanish (es), Portuguese (pt), German (de), French (fr), Chinese (zh), Japanese (ja), Hindi (hi), and Turkish (tr). The details of MultiATIS++ are shown in Appendix A.",
+        "bbox": [
+            507,
+            627,
+            882,
+            740
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following Li et al. (2021), we evaluate the performance of intent prediction using accuracy, slot filling using F1 score, and the sentence-level semantic frame parsing using overall accuracy.",
+        "bbox": [
+            507,
+            741,
+            882,
+            806
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. We adopt two baseline models as the backbone, i.e., CoSDA (Qin et al., 2020) and GL-CLEF (Qin et al., 2022). Following previous works, the random rate  $\\alpha$  for token substitution is set to 0.2. AdamW (Loshchilov and",
+        "bbox": [
+            507,
+            815,
+            884,
+            895
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<https://github.com/amazon-science/multiatis>",
+        "bbox": [
+            529,
+            903,
+            867,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "7851",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/5c1a266f822dbe38c7a8db87fc9f35daceb42ec3f2d7ea56a8837450b26c5c1f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Model</td><td>en</td><td>de</td><td>es</td><td>fr</td><td>hi</td><td>ja</td><td>pt</td><td>tr</td><td>zh</td><td>Avg.</td></tr><tr><td rowspan=\"7\">Intent Acc</td><td>mBERT(Devlin et al., 2019)</td><td>98.54</td><td>95.40</td><td>96.30</td><td>94.31</td><td>82.41</td><td>76.18</td><td>94.95</td><td>75.10</td><td>82.53</td><td>88.42</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>98.54</td><td>90.48</td><td>93.28</td><td>94.51</td><td>77.15</td><td>76.59</td><td>94.62</td><td>73.29</td><td>84.55</td><td>87.00</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>95.74</td><td>94.06</td><td>92.29</td><td>77.04</td><td>82.75</td><td>73.25</td><td>93.05</td><td>80.42</td><td>78.95</td><td>87.32</td></tr><tr><td>SoGo CoS (Ours)</td><td>98.54</td><td>96.97</td><td>98.12</td><td>96.33</td><td>82.76</td><td>76.41</td><td>97.18</td><td>82.96</td><td>85.77</td><td>90.56</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>98.77</td><td>97.53</td><td>97.05</td><td>97.72</td><td>86.00</td><td>82.84</td><td>96.08</td><td>83.92</td><td>87.68</td><td>91.95</td></tr><tr><td>SoGo GL (Ours)</td><td>98.89</td><td>98.45</td><td>98.15</td><td>97.74</td><td>83.87</td><td>84.75</td><td>97.73</td><td>85.53</td><td>89.10</td><td>92.69</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>98.77</td><td>98.10</td><td>98.10</td><td>98.77</td><td>84.54</td><td>81.86</td><td>97.09</td><td>85.45</td><td>89.03</td><td>92.41</td></tr><tr><td rowspan=\"7\">Slot F1</td><td>mBERT(Devlin et al., 2019)</td><td>95.11</td><td>80.11</td><td>78.22</td><td>82.25</td><td>26.71</td><td>25.40</td><td>72.37</td><td>41.49</td><td>53.22</td><td>61.66</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>95.20</td><td>74.79</td><td>76.52</td><td>74.25</td><td>52.73</td><td>70.10</td><td>72.56</td><td>29.66</td><td>66.91</td><td>68.08</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>92.29</td><td>81.37</td><td>76.94</td><td>79.36</td><td>64.06</td><td>66.62</td><td>75.05</td><td>48.77</td><td>77.32</td><td>73.47</td></tr><tr><td>SoGo CoS (Ours)</td><td>95.46</td><td>84.12</td><td>83.84</td><td>83.46</td><td>57.63</td><td>65.78</td><td>80.27</td><td>55.13</td><td>79.56</td><td>76.14</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>95.39</td><td>86.30</td><td>85.22</td><td>84.31</td><td>70.34</td><td>73.12</td><td>81.83</td><td>65.85</td><td>77.61</td><td>80.00</td></tr><tr><td>SoGo GL (Ours)</td><td>95.42</td><td>87.46</td><td>87.01</td><td>84.45</td><td>74.25</td><td>76.69</td><td>83.91</td><td>67.04</td><td>78.53</td><td>81.64</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>96.02</td><td>86.59</td><td>83.03</td><td>82.11</td><td>61.04</td><td>68.52</td><td>81.49</td><td>65.20</td><td>82.00</td><td>78.23</td></tr><tr><td rowspan=\"7\">Overall Acc</td><td>mBERT(Devlin et al., 2019)</td><td>87.12</td><td>52.69</td><td>52.02</td><td>37.29</td><td>4.92</td><td>7.11</td><td>43.49</td><td>4.33</td><td>18.58</td><td>36.29</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>87.23</td><td>41.43</td><td>44.46</td><td>43.67</td><td>16.01</td><td>33.59</td><td>43.90</td><td>1.12</td><td>30.80</td><td>38.02</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>77.04</td><td>57.06</td><td>46.62</td><td>50.06</td><td>26.20</td><td>28.89</td><td>48.77</td><td>15.24</td><td>46.36</td><td>44.03</td></tr><tr><td>SoGo CoS (Ours)</td><td>88.35</td><td>61.43</td><td>58.30</td><td>56.37</td><td>19.70</td><td>27.65</td><td>60.43</td><td>19.30</td><td>50.17</td><td>49.08</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>88.02</td><td>66.03</td><td>59.53</td><td>57.02</td><td>34.83</td><td>41.42</td><td>60.43</td><td>28.95</td><td>50.62</td><td>54.09</td></tr><tr><td>SoGo GL (Ours)</td><td>90.54</td><td>72.26</td><td>61.05</td><td>57.88</td><td>39.90</td><td>46.95</td><td>64.23</td><td>29.14</td><td>51.31</td><td>57.02</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>89.81</td><td>67.75</td><td>59.13</td><td>57.56</td><td>23.29</td><td>29.34</td><td>61.93</td><td>28.95</td><td>54.76</td><td>52.50</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            878,
+            406
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Major results  $(\\%)$  on MultiATIS++. Higher is better in all columns. Results with  $\\diamond$  are taken from Qin et al. (2022), and results with  $\\diamond$  are taken from Liang et al. (2022). SoGo CoS and SoGo GL can be directly compared to CoSDA and GL-CLEF. Note that LAJ-MCL is not open-source.",
+        "bbox": [
+            112,
+            414,
+            882,
+            460
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Hutter, 2019) is used to train SoGO with a learning rate of  $5e - 6$ . The coefficients in Equation 3 are  $\\lambda_{I}$  and  $\\lambda_{S}$ , where the best ones are selected by searching a combination with the following ranges:  $\\{0.8, 0.9, 1.0, 1.1, 1.2\\}$ . All experiments are conducted on a single TESLA-V100. The experimental results of our method are averaged over five runs with different random seed to ensure the final reported results are statistically stable.",
+        "bbox": [
+            112,
+            483,
+            489,
+            629
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Major Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            640,
+            272,
+            656
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "From the major results in Table 1, we have the following observations: (1) SoGo  $\\mathrm{CoS}$  outperforms CoSDA significantly. This is because that code-switched sentences generated by performing salient word substitution and token-level alignment maintain better syntactic and semantic consistency than those generated by random substitution. (2) Both SoGo  $\\mathrm{CoS}$  and SoGo GL beat their counterparts with  $5.05\\%$  and  $2.93\\%$  average improvement in overall accuracy, respectively. This demonstrates that SoGo can be integrated into existing SLU models, orthogonal of other structures. Notably, the gain in slot filling is more significant than in intent detection, further validating the effectiveness of token-level alignment. (3) As LAJ-MCL is not open-source, we implemented two versions of",
+        "bbox": [
+            112,
+            661,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "SoGO, referred to as SoGo  $\\text{CoS}$  and SoGO  $\\text{GL}$ , to compare with LAJ-MCL. Remarkably, SoGO  $\\text{GL}$  outperformed the competitive LAJ-MCL in overall accuracy across 8/9 languages. This highlights the significant potential of our SoGO when combined with existing SLU models.",
+        "bbox": [
+            507,
+            483,
+            882,
+            582
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Further Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            594,
+            690,
+            609
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Ablation Study. We conduct an ablation study to explore the effect of each component in SoGo. The results are illustrated in Figure 2, which show that each proposed component (i.e., saliency-based substitution and token-level alignment) contributes to the performance positively. It is obvious that the proposed saliency-based substitution has a strong ability to pick out important words to boost performance. Moreover, our token-level alignment has shown greater relative importance of the switched tokens than the vanilla code-switching method. It indicates that such substituted tokens significantly contribute to SoGo's prediction.",
+        "bbox": [
+            507,
+            617,
+            882,
+            826
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Analysis of Saliency-based Substitution. To explore the potential of saliency-based substitution, we perform experiments on SoGo CoS of different strategies with respect to different keyword replacement rate  $\\frac{k}{v}$  (cf. Figure 3, where  $k$  denotes",
+        "bbox": [
+            507,
+            839,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "7852",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/dad38814c28aaf51f4bff4b80ce20ed3789e72167164e7cf8dbb81861e7ca1fa.jpg",
+        "image_caption": [
+            "Figure 2: Ablation experiments. \"SbS\" and \"TlA\" denote saliency-based substitution and token-level alignment, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            82,
+            487,
+            191
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "the number of keywords, and  $v$  denotes the total number of words.). We find the performance of randomly selecting keywords significantly declines as  $\\frac{k}{v}$  decreases, whereas our saliency-based substitution continues to perform well even with only 1% of keywords. This is because saliency-based substitution prioritizes the most indicative keywords for code-switching. As  $\\frac{k}{v}$  increases, the additional keywords become less indicative and might even have a negative effect on the model's performance.",
+        "bbox": [
+            112,
+            269,
+            489,
+            432
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd6302d40c9d151cc4f0b17d49e52a526b808f7aa3595b1a57b08452f719ed72.jpg",
+        "image_caption": [
+            "Figure 3: Overall Accuracy of SoGo CoS in different replacement rate of keywords  $\\frac{k}{v}$ . Random denotes selecting keywords randomly, while Saliency denotes selecting keywords via saliency-based substitution."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            456,
+            468,
+            627
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Analysis of Token-level Alignment. To verify whether token-level alignment preserves a context space to keep grammatical coherence, we plot the degree of dispersion among tokens within a random sentence in Figure 4. Concretely, we retrieve the token embedding from intermediate (sixth) layers for both switched and original tokens and utilize top features calculated from PCA (Abdi and Williams, 2010). From Figure 4, we can conclude that SoGo CoS has shown a more compact space for the original tokens, where switched tokens are also inside the context space. In contrast, the vanilla code",
+        "bbox": [
+            112,
+            726,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "switching method entirely separates the substituted tokens apart from the original context.",
+        "bbox": [
+            507,
+            84,
+            882,
+            116
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/6af980b7df31f22cd3a9cc7633201ec8649b091a69130ec5135f603d162779c8.jpg",
+        "image_caption": [
+            "Figure 4: Token-Level density of SoGo  $\\mathrm{CoS}$  during training procedure."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            529,
+            131,
+            863,
+            312
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            387,
+            640,
+            401
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We proposed a novel semantics-coherent and grammar-coherent method (SoGo) to enhance code-switching for zero-shot cross-lingual SLU. Specifically, we utilized saliency-based substitution to replace keywords for better semantic coherence in the shared space. Besides, we introduced an alignment strategy to map the code-switched tokens to original contexts, which solves the grammatical incoherence of code-switching. Experiments on MultiATIS++ showed that code-switching based methods with SoGO obtained competitive performance compared to counterpart baselines.",
+        "bbox": [
+            507,
+            417,
+            884,
+            609
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            627,
+            613,
+            643
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By applying saliency-based substitution and token-level alignment, SoGo achieves significant improvement on the benchmark datasets. Nevertheless, we summarize two limitations for further discussion and investigation by other researchers:",
+        "bbox": [
+            507,
+            657,
+            884,
+            738
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) Like other code-switching methods, the performance of the model still relies heavily on the correctness of the bilingual dictionary (Fazili and Jyothi, 2022; Whitehouse et al., 2022). In the future, we continue to enhance the performance of zero-shot cross-lingual SLU by refining code-switching techniques (Cheng et al., 2023a).",
+            "(2) SoGo should further leverage language-independent information (Yu et al., 2021), such as dependency relations and POS tags, to ensure syntactic invariance for cross-lingual transfers."
+        ],
+        "bbox": [
+            507,
+            740,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "7853",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            285,
+            99
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We thank all anonymous reviewers for their constructive comments. This paper was partially supported by Shenzhen Science & Technology Research Program (No:GXWD20201231165807007-20200814115301001) and NSFC (No: 62176008).",
+        "bbox": [
+            112,
+            110,
+            490,
+            190
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            218,
+            213,
+            231
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Herve Abdi and Lynne J Williams. 2010. Principal component analysis. Wiley interdisciplinary reviews: computational statistics.",
+            "Leila Arras, Ahmed Osman, Klaus-Robert Müller, and Wojciech Samek. 2019. Evaluating recurrent neural network explanations. In ACL Workshop.",
+            "Dongsheng Chen, Zhiqi Huang, and Yuexian Zou. 2022. Leveraging bilinear attention to improve spoken language understanding. In ICASSP.",
+            "Qian Chen, Zhu Zhuo, and Wen Wang. 2019. BERT for joint intent classification and slot filling. CoRR.",
+            "Xuxin Cheng, Bowen Cao, Qichen Ye, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023a. Ml-lmcl: Mutual learning and large-margin contrastive learning for improving asr robustness in spoken language understanding. In Proc. of ACL Findings.",
+            "Xuxin Cheng, Wanshi Xu, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023b. Towards spoken language understanding via multi-level multi-grained contrastive learning. In CIKM.",
+            "Xuxin Cheng, Zhihong Zhu, Bowen Cao, Qichen Ye, and Yuexian Zou. 2023c. MRRL: Modifying the reference via reinforcement learning for non-autoregressive joint multiple intent detection and slot filling. In Proc. of EMNLP Findings.",
+            "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In NeurIPS.",
+            "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. Xnli: Evaluating crosslingual sentence representations. In EMNLP.",
+            "Misha Denil, Alban Demiraj, and Nando de Freitas. 2014. Extraction of salient sentences from labelled documents. CoRR.",
+            "Ameet Deshpande, Partha Talukdar, and Karthik Narasimhan. 2022. When is BERT multilingual? isolating crucial ingredients for cross-lingual transfer. In AACL.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In AACL."
+        ],
+        "bbox": [
+            115,
+            241,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Barah Fazili and Preethi Jyothi. 2022. Aligning multilingual embeddings for improved code-switched natural language understanding. In COLING.",
+            "Yukun Feng, Feng Li, and Philipp Koehn. 2022. Toward the limitation of code-switching in cross-lingual transfer. In EMNLP.",
+            "Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In ACL.",
+            "Siyu Lai, Hui Huang, Dong Jing, Yufeng Chen, Jinan Xu, and Jian Liu. 2021. Saliency-based multi-view mixed language training for zero-shot cross-lingual classification. In EMNLP Findings.",
+            "Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018. Word translation without parallel data. In ICLR.",
+            "Xiaochun Lei, Xiang Cai, Linjun Lu, Zihang Cui, and Zetao Jiang. 2023. Su2ge-net: a saliency-based approach for non-specific class foreground segmentation. Scientific Reports.",
+            "Guanbin Li and Yizhou Yu. 2015. Visual saliency based on multiscale deep features. In CVPR.",
+            "Haoran Li, Abhinav Arora, Shuohui Chen, Anchit Gupta, Sonal Gupta, and Yashar Mehdad. 2021. MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark. In EACL.",
+            "Shining Liang, Linjun Shou, Jian Pei, Ming Gong, Wanli Zuo, Xianglin Zuo, and Daxin Jiang. 2022. Label-aware multi-level contrastive learning for cross-lingual spoken language understanding. In EMNLP.",
+            "Zihan Liu, Genta Indra Winata, Zhaojiang Lin, Peng Xu, and Pascale Fung. 2020. Attention-informed mixed-language training for zero-shot cross-lingual task-oriented dialogue systems. In AAAI.",
+            "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In ICLR.",
+            "Libo Qin, Qiguang Chen, Tianbao Xie, Qixin Li, JianGuang Lou, Wanxiang Che, and Min-Yen Kan. 2022. Gl-clef: A global-local contrastive learning framework for cross-lingual spoken language understanding. In ACL.",
+            "Libo Qin, Minheng Ni, Yue Zhang, and Wanxiang Che. 2020. Cosda-ml: Multi-lingual code-switching data augmentation for zero-shot cross-lingual NLP. In IJCAI.",
+            "Libo Qin, Tianbao Xie, Wanxiang Che, and Ting Liu. 2021. A survey on spoken language understanding: Recent advances and new frontiers. In *IJCAI*.",
+            "Stephen Robertson. 2004. Understanding inverse document frequency: on theoretical arguments for IDF. J. Documentation."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "7854",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how BERT works. Trans. Assoc. Comput. Linguistics.",
+        "bbox": [
+            115,
+            85,
+            487,
+            137
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning important features through propagating activation differences. In ICML.",
+        "bbox": [
+            115,
+            152,
+            487,
+            192
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "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 NeurIPS.",
+        "bbox": [
+            115,
+            206,
+            487,
+            259
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Yuxuan Wang, Wanxiang Che, Jiang Guo, Yijia Liu, and Ting Liu. 2019. Cross-lingual BERT transformation for zero-shot dependency parsing. In EMNLP.",
+        "bbox": [
+            115,
+            273,
+            487,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Chenxi Whitehouse, Fenia Christopoulou, and Ignacio Iacobacci. 2022. Entitycs: Improving zero-shot cross-lingual transfer with entity-centric code switching. In EMNLP Findings.",
+        "bbox": [
+            115,
+            326,
+            487,
+            380
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of bert. In EMNLP.",
+        "bbox": [
+            115,
+            394,
+            487,
+            432
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Weijia Xu, Batool Haider, and Saab Mansour. 2020. End-to-end slot alignment and recognition for crosslingual NLU. In EMNLP.",
+        "bbox": [
+            115,
+            448,
+            487,
+            488
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Yi Yang, Yueyuan Zheng, Didan Deng, Jindi Zhang, Yongxiang Huang, Yumeng Yang, Janet H Hsiao, and Caleb Chen Cao. 2022. Hsi: Human saliency imitator for benchmarking saliency-based model explanations. In AAAI.",
+        "bbox": [
+            115,
+            502,
+            487,
+            568
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Sicheng Yu, Hao Zhang, Yulei Niu, Qianru Sun, and Jing Jiang. 2021. COSY: counterfactual syntax for cross-lingual understanding. In ACL.",
+        "bbox": [
+            115,
+            582,
+            487,
+            623
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Michelle Yuan, Mozhi Zhang, Benjamin Van Durme, Leah Findlater, and Jordan L. Boyd-Graber. 2020. Interactive refinement of cross-lingual word embeddings. In EMNLP.",
+        "bbox": [
+            115,
+            637,
+            487,
+            690
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Zhihong Zhu, Xuxin Cheng, Dongsheng Chen, Zhiqi Huang, Hongxiang Li, and Yuexian Zou. 2023a. Mix before Align: Towards Zero-shot Cross-lingual Sentiment Analysis via Soft-Mix and Multi-View Learning. In Interspeech.",
+        "bbox": [
+            115,
+            703,
+            487,
+            770
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Zhihong Zhu, Xuxin Cheng, Zhiqi Huang, Dongsheng Chen, and Yuexian Zou. 2023b. Towards unified spoken language understanding decoding via label-aware compact linguistics representations. In Proc. of ACL Findings.",
+        "bbox": [
+            115,
+            784,
+            487,
+            850
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Zhihong Zhu, Weiyuan Xu, Xuxin Cheng, Tengtao Song, and Yuexian Zou. 2023c. A dynamic graph interactive framework with label-semantic injection for spoken language understanding. In ICASSP.",
+        "bbox": [
+            115,
+            864,
+            487,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Dataset",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            615,
+            98
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "MixATIS++ is an extension of Multilingual ATIS (Table 2). Human-translated data for six languages including Spanish (es), German (de), Chinese (zh), Japanese (ja), Portuguese (pt), French (fr) are added to Multilingual ATIS which initially has Hindi (hi) and Turkish (tr). There are 4,478 utterances in the train set, 500 in the valid set, and 893 in the test set, with 18 intents and 84 slots for each language.",
+        "bbox": [
+            507,
+            112,
+            882,
+            241
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/4deceefcc722f4cb28a001f0aaff19c4167c5f4b313ec51d7d359e7bdee0e7f2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Language</td><td colspan=\"3\">Utterances</td><td rowspan=\"2\">Intent types</td><td rowspan=\"2\">Slot types</td></tr><tr><td>#Train</td><td>#Valid</td><td>#Test</td></tr><tr><td>hi</td><td>1,440</td><td>160</td><td>893</td><td>17</td><td>75</td></tr><tr><td>tr</td><td>578</td><td>60</td><td>715</td><td>17</td><td>71</td></tr><tr><td>others</td><td>4,488</td><td>490</td><td>893</td><td>18</td><td>84</td></tr></table>",
+        "bbox": [
+            514,
+            255,
+            882,
+            338
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Statistics of MultiATIS++.",
+        "bbox": [
+            573,
+            348,
+            815,
+            361
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B Experiment Details",
+        "text_level": 1,
+        "bbox": [
+            509,
+            395,
+            714,
+            413
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Following the zero-shot setting, we fine-tune the model on en training set and use en validation set for the hyper-parameters search. The best model checkpoint is decided by the overall accuracy on en validation set. Our code is based on PyTorch and Transformers<sup>2</sup>.",
+        "bbox": [
+            507,
+            426,
+            882,
+            521
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C Baselines",
+        "text_level": 1,
+        "bbox": [
+            509,
+            539,
+            628,
+            552
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compare our model to the following baselines.",
+        "bbox": [
+            507,
+            568,
+            880,
+            583
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "mBERT. mBERT<sup>3</sup> follows the same model architecture as BERT (Devlin et al., 2019), but instead of training only on monolingual English data, it is trained on the Wikipedia pages of 104 languages with a shared word piece vocabulary, allowing the model to share embeddings across languages.",
+        "bbox": [
+            507,
+            596,
+            882,
+            694
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ZSJoint. Chen et al. (2019) propose a zero-shot SLU model, which is trained on the en training set and directly applied to the test sets of target languages.",
+        "bbox": [
+            507,
+            707,
+            882,
+            771
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "SoSDA. Qin et al. (2020) propose a data augmentation framework to generate multi-lingual code-switching data to fine-tune mBERT, which encourages the model to align representations from the source and multiple target languages.",
+        "bbox": [
+            507,
+            785,
+            884,
+            866
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://github.com/huggingface/transformers  \n $^{3}$ https://github.com/google-research/bert/blob/master/multilingual.md",
+        "bbox": [
+            507,
+            879,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "7855",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "GL-CLEF. Qin et al. (2022) introduce a contrastive learning framework to explicitly align representations across languages for zero-shot crosslingual SLU.",
+        "bbox": [
+            112,
+            84,
+            489,
+            148
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "LAJ-MCL. Liang et al. (2022) proposes a multi-level contrastive learning framework for zero-shot cross-lingual SLU.",
+        "bbox": [
+            112,
+            156,
+            489,
+            206
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D Generalizability of SoGo",
+        "text_level": 1,
+        "bbox": [
+            114,
+            217,
+            374,
+            233
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We investigate the potential of the proposed model for other tasks through an additional set of experiments. Following Qin et al. (2020), we conduct additional experiments for natural language inference using XNLI (Conneau et al., 2018), which includes 15 languages. We directly input pairs of sentences into the mBERT encoder, and a task-specific classification layer is employed for classification. The models are evaluated based on classification accuracy (ACC), and the results are presented in Table 3.",
+        "bbox": [
+            112,
+            243,
+            490,
+            403
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/276437a52b163fd62f51c1b6c5c708cac4a3270c61e988fc81dbd0e200fa54b5.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Avg.</td></tr><tr><td>CoSDA_XLM-based (Conneau and Lample, 2019)</td><td>75.3</td></tr><tr><td>+ SoGo (Ours)</td><td>77.4</td></tr><tr><td>CoSDA_mBERT-based (Wu and Dredze, 2019)</td><td>69.7</td></tr><tr><td>+ SoGo (Ours)</td><td>71.9</td></tr></table>",
+        "bbox": [
+            119,
+            428,
+            487,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Natural language inference experiments.",
+        "bbox": [
+            132,
+            518,
+            467,
+            533
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "7856",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 7
+    }
+]

2023/Enhancing Code-Switching for Cross-lingual SLU_ A Unified View of Semantic and Grammatical Coherence/b8a9adec-48b2-48e3-8fb6-523de3c4eab2_model.json

@@ -0,0 +1,1756 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.202,
+                0.089,
+                0.796,
+                0.128
+            ],
+            "angle": 0,
+            "content": "Enhancing Code-Switching for Cross-lingual SLU: A Unified View of Semantic and Grammatical Coherence"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.302,
+                0.142,
+                0.7,
+                0.176
+            ],
+            "angle": 0,
+            "content": "Zhihong Zhu† Xuxin Cheng† Zhiqi Huang Dongsheng Chen Yuexian Zou*"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.334,
+                0.178,
+                0.671,
+                0.193
+            ],
+            "angle": 0,
+            "content": "School of ECE, Peking University, China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.28,
+                0.194,
+                0.724,
+                0.21
+            ],
+            "angle": 0,
+            "content": "{zhihongzhu,chengxx,chends}@stu.pku.edu.cn"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.351,
+                0.211,
+                0.653,
+                0.227
+            ],
+            "angle": 0,
+            "content": "{zhiqihuang, zouyx}@pku.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.282,
+                0.461,
+                0.651
+            ],
+            "angle": 0,
+            "content": "Despite the success of spoken language understanding (SLU) in high-resource languages, achieving similar performance in low-resource settings, such as zero-shot scenarios, remains challenging due to limited labeled training data. To improve zero-shot cross-lingual SLU, recent studies have explored code-switched sentences containing tokens from multiple languages. However, vanilla code-switched sentences often lack semantic and grammatical coherence. We ascribe this lack to two issues: (1) randomly replacing code-switched tokens with equal probability and (2) disregarding token-level dependency within each language. To tackle these issues, in this paper, we propose a novel method termed SoGO, for zero-shot cross-lingual SLU. First, we use a saliency-based substitution approach to extract keywords as substitution options. Then, we introduce a novel token-level alignment strategy that considers the similarity between the context and the code-switched tokens, ensuring grammatical coherence in code-switched sentences. Extensive experiments and analyses demonstrate the superior performance of SoGO across nine languages on MultiATIS++."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.664,
+                0.26,
+                0.68
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.69,
+                0.49,
+                0.834
+            ],
+            "angle": 0,
+            "content": "Acting as the interface between users and machines, spoken language understanding (SLU) is crucial in task-oriented dialogue systems (Qin et al., 2021; Chen et al., 2022; Zhu et al., 2023b; Cheng et al., 2023c). While joint training models have led to significant advancements, most existing models rely on annotated training data, limiting their scalability to low-resource languages. To this end, zero-shot cross-lingual SLU gains increasing attention."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.835,
+                0.489,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Existing works have shown the effectiveness of Multilingual BERT (mBERT) in multilingual corpus pre-training, enabling promising zero-shot"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.253,
+                0.885,
+                0.414
+            ],
+            "angle": 0,
+            "content": "cross-lingual SLU (Deshpande et al., 2022; Cheng et al., 2023b; Zhu et al., 2023a). Therein, Qin et al. (2020) extended this concept to a code-switched setting, aligning the source language with multiple target languages. Qin et al. (2022) incorporate contrastive learning to achieve fine-grained cross-lingual transfer. Based on this, Liang et al. (2022) further propose a multi-level contrastive learning framework for explicit alignment of utterance-slot-word structure in cross-lingual SLU."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.417,
+                0.885,
+                0.609
+            ],
+            "angle": 0,
+            "content": "However, most above-mentioned models adopt code-switching methods, which often lack both semantic coherence and grammatical coherence. This can be attributed to two main issues: (1) Current approaches overlook the importance of individual words in a sentence, as they are randomly replaced with equal probability, potentially introducing unnecessary translation burden and noise that affects semantic coherence. (2) Code-switched sentences comprising tokens from multiple languages may lack grammatical coherence, as token-level dependencies within each language are disregarded."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.612,
+                0.885,
+                0.885
+            ],
+            "angle": 0,
+            "content": "To address these issues, we propose a novel framework dubbed SoGo for enhancing code-switching. Concretely, we first employ a saliency-based substitution approach to identify keywords with high saliency scores. These keywords are then replaced with their target language equivalents using bilingual dictionaries, to generate code-switched sentences. Notably, unlike the approach by Liu et al. (2020), our SoGo framework requires only a single training process for multiple target languages. Then, we devise a novel token-level alignment strategy to bridge the gap between original context and code-switched tokens. By mapping code-switched tokens back to the original context, SoGo minimizes the impact of token substitutions on token-level coherence, thus maintaining consistent contextual space within each language."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.888,
+                0.884,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Experimental results on MultiATIS++ (Xu et al., 2020) demonstrate that employing code-switching"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.136,
+                0.891,
+                0.263,
+                0.905
+            ],
+            "angle": 0,
+            "content": "\\(\\dagger\\) Equal contribution."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.137,
+                0.905,
+                0.282,
+                0.919
+            ],
+            "angle": 0,
+            "content": "* Corresponding author."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.136,
+                0.891,
+                0.282,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7849"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.16,
+                0.946,
+                0.838,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7849-7856 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.141,
+                0.085,
+                0.38,
+                0.311
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.117,
+                0.313,
+                0.137,
+                0.325
+            ],
+            "angle": 0,
+            "content": "(a)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.418,
+                0.085,
+                0.882,
+                0.325
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.334,
+                0.885,
+                0.406
+            ],
+            "angle": 0,
+            "content": "Figure 1: (a) Our proposed gradient-based substitution in SoGo. The code-switched dictionary comprised of salient words has the capacity to more accurately characterize the common semantic space and facilitate the acquisition of semantic connections among different languages by the model. (b) The main architecture of SoGo. We perform token-level alignment in lower six layers of mBERT to enhance the grammatical coherence of the code-switched sentence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.432,
+                0.49,
+                0.496
+            ],
+            "angle": 0,
+            "content": "with SoGo leads to significant performance improvement compared to baseline models. Furthermore, we demonstrate the compatibility of SoGo with other existing SLU structures."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.508,
+                0.219,
+                0.523
+            ],
+            "angle": 0,
+            "content": "2 Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.534,
+                0.489,
+                0.599
+            ],
+            "angle": 0,
+            "content": "We describe problem formulation (§2.1) and backbone model (§2.2) for zero-shot cross-lingual SLU first, before describing SoGo (§2.3). The main architecture of SoGo is illustrated in Figure 1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.61,
+                0.332,
+                0.625
+            ],
+            "angle": 0,
+            "content": "2.1 Problem Formulation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.632,
+                0.49,
+                0.713
+            ],
+            "angle": 0,
+            "content": "Given each input sentence \\( x_{tgt} \\) in a target language, zero-shot cross-lingual SLU means the SLU model is trained in a source language dataset \\( S \\), e.g., {English}, and directly applied to other target languages datasets \\( T \\), e.g., {Chinese, Japanese}:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.218,
+                0.723,
+                0.488,
+                0.745
+            ],
+            "angle": 0,
+            "content": "\\[\n\\left(o _ {t g t} ^ {I}, o _ {t g t} ^ {S}\\right) = f \\left(x _ {t g t}\\right), \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.756,
+                0.49,
+                0.807
+            ],
+            "angle": 0,
+            "content": "where \\(f(\\cdot)\\) is the trained model. \\(o_{tgt}^{I}\\) represents an intent label and \\(o_{tgt}^{S}\\) represents a slot sequence. Note that \\(T\\) may consist of multiple tgt languages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.818,
+                0.295,
+                0.832
+            ],
+            "angle": 0,
+            "content": "2.2 Backbone Model"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.839,
+                0.491,
+                0.919
+            ],
+            "angle": 0,
+            "content": "For each input sentence \\( x \\), we follow Qin et al. (2022); Liang et al. (2022) to utilize bilingual dictionaries (Lample et al., 2018) to generate its corresponding multi-lingual one \\( x' \\). Following previous studies, we select mBERT (Devlin et al., 2019)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.432,
+                0.884,
+                0.462
+            ],
+            "angle": 0,
+            "content": "as our sentence encoder. Formally, the input of mBERT is formulated as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.578,
+                0.473,
+                0.884,
+                0.492
+            ],
+            "angle": 0,
+            "content": "\\[\nx ^ {\\prime} = \\left([ \\mathrm {C L S} ], x _ {1} ^ {\\prime}, \\dots , x _ {n} ^ {\\prime}, [ \\mathrm {S E P} ]\\right), \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.502,
+                0.885,
+                0.567
+            ],
+            "angle": 0,
+            "content": "where [CLS] denotes the special symbol for representing the whole sequence, and [SEP] can be used for separating non-consecutive token sequences (Devlin et al., 2019)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.575,
+                0.886,
+                0.64
+            ],
+            "angle": 0,
+            "content": "Intent Detection. We feed the sentence representation \\( h_{\\mathrm{CLS}}' \\) generated from mBERT into a classification layer to obtain \\( o^{I}:o^{I} = \\mathrm{softmax}(W^{I}h_{\\mathrm{CLS}}^{\\prime} + b^{I}) \\), where \\( W^{I} \\) and \\( b^{I} \\) are tunable parameters."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.647,
+                0.884,
+                0.745
+            ],
+            "angle": 0,
+            "content": "Slot Filling. We use the hidden state to predict each slot: \\( o_{t}^{S} = \\mathrm{softmax}(W^{S}h_{t}^{\\prime} + b^{S}) \\), where \\( h_t^\\prime \\) denotes the representation of word \\( x_{t}^{\\prime} \\). Note that mBERT tokenizes words into subwords (Wang et al., 2019), so we adopt the first sub-token's representation as the whole word representation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.753,
+                0.884,
+                0.833
+            ],
+            "angle": 0,
+            "content": "Training and Inference. Following previous works, the loss function for intent detection \\((\\mathcal{L}_I)\\) and slot filling \\((\\mathcal{L}_S)\\) are optimized jointly by corresponding cross-entropy loss. Finally, the total training loss is the weighted sum of two losses:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.619,
+                0.844,
+                0.883,
+                0.861
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} = \\lambda_ {I} \\mathcal {L} _ {I} + \\lambda_ {S} \\mathcal {L} _ {S}, \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.872,
+                0.862,
+                0.887
+            ],
+            "angle": 0,
+            "content": "where \\(\\lambda_I\\) and \\(\\lambda_S\\) denote two hyper-parameters."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.888,
+                0.884,
+                0.918
+            ],
+            "angle": 0,
+            "content": "During inference, we do not use any code-switched sentences."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7850"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.085,
+                0.214,
+                0.099
+            ],
+            "angle": 0,
+            "content": "2.3 SoGo"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.106,
+                0.49,
+                0.348
+            ],
+            "angle": 0,
+            "content": "Saliency-based Substitution. Intuitively, each word in a sentence contributes differently, and we refer to words with significant impact as keywords. Given a vocabulary set \\(\\mathcal{V}\\) containing \\(v\\) words, we can identify a salient subset of keywords \\(\\mathcal{K} \\subseteq \\mathcal{V}\\) to generate code-switched sentences. In our work, gradient-based saliency scores (Li and Yu, 2015; Denil et al., 2014; Arras et al., 2019) are used to select keywords, which has been proven effective in many other tasks (Yang et al., 2022; Lai et al., 2021; Lei et al., 2023). Methodologically, let \\(x_{n}\\) denotes the \\(n\\)-th sentence from \\(S\\), \\(\\mathcal{L}_{\\hat{y}_n}\\) is the loss between model's prediction of slot filling \\(\\hat{y}_n\\) and the golden label \\(y_{n}\\). For each token \\(w_{i,n} \\in x_{n}\\), the saliency score is defined as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.178,
+                0.358,
+                0.489,
+                0.377
+            ],
+            "angle": 0,
+            "content": "\\[\nS _ {x _ {n}} (w _ {i, n}) = - \\nabla_ {h _ {i, n}} \\mathcal {L} _ {\\hat {y} _ {n}} \\times h _ {i, n}, \\qquad (4)\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.386,
+                0.489,
+                0.514
+            ],
+            "angle": 0,
+            "content": "where \\( h_{i,n} \\) represents the embedding of \\( w_{i,n} \\). Note that mBERT tokenizes words into subwords, so we calculate the average of the subword saliency scores for each word to obtain the final score. In this manner, the gradient signifies the degree to which a word embedding influences the final decision, while the input considers both the sign and magnitude of the input (Shrikumar et al., 2017)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.515,
+                0.49,
+                0.627
+            ],
+            "angle": 0,
+            "content": "Equation 4 assesses the local significance of a token within a sentence. However, our objective is to contrast a global keyword subset \\(\\mathcal{K}\\) from \\(S\\). Following Yuan et al. (2020), we aggregate all saliency scores for the token \\(w_{i}\\) occurring in \\(S\\). We then multiply these scores with the inverse document frequency (IDF) (Robertson, 2004) of the token \\(w_{i}\\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.126,
+                0.652,
+                0.488,
+                0.685
+            ],
+            "angle": 0,
+            "content": "\\[\nS \\left(w _ {i}\\right) = \\log \\frac {N}{\\left| \\left\\{x \\in S : w _ {i} \\in x \\right\\} \\right|} \\cdot \\sum_ {x \\in S: w _ {i} \\in x} S _ {x} \\left(w _ {i}\\right), \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.686,
+                0.489,
+                0.75
+            ],
+            "angle": 0,
+            "content": "in which \\(N\\) is the total number of words in \\(S\\). The IDF term balances word frequency and saliency scores by assigning words with high document frequency a lower weight and vice versa."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.751,
+                0.49,
+                0.814
+            ],
+            "angle": 0,
+            "content": "Finally, top-\\(k\\) salient words are chosen to compose the code-switched dictionary \\(\\mathcal{K}\\). We opt for the dictionary \\(\\mathcal{K}\\) as a replacement strategy instead of random word substitution in sentences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.823,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Token-level Alignment. To further align the space of all hidden states in the code-switched sentence, inspired by attention-based alignment methods Zhu et al. (2023c); Feng et al. (2022), we map the code-switched sentence to the original sentence by calculating the similarity between substituted"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.198
+            ],
+            "angle": 0,
+            "content": "hidden states and original hidden states. The similarity scores are then used as weights to aggregate the embedding of original hidden states, reflecting the calculated potential for substituted hidden states in the code-switched sequence. To be specific, for each hidden state \\( h_i' \\) in \\( h' \\), the similarity score is calculated as Vaswani et al. (2017):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.63,
+                0.207,
+                0.884,
+                0.228
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {s c o r e} _ {i} ^ {\\prime} = h _ {i} ^ {\\prime} \\cdot h ^ {\\top}. \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.241,
+                0.883,
+                0.274
+            ],
+            "angle": 0,
+            "content": "The final potential of the code-switched token \\(\\tilde{h}_i'\\) is the weighted sum of its corresponding \\(h\\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.519,
+                0.285,
+                0.883,
+                0.338
+            ],
+            "angle": 0,
+            "content": "\\[\n\\alpha_ {i, t} = \\frac {\\exp \\left(\\operatorname {s c o r e} _ {i , t} ^ {\\prime}\\right)}{\\sum_ {t ^ {\\prime}} \\exp \\left(\\operatorname {s c o r e} _ {i , t ^ {\\prime}} ^ {\\prime}\\right)}, \\quad \\tilde {h} _ {i} ^ {\\prime} = \\sum_ {t} \\alpha_ {i, t} \\cdot h _ {t}. \\tag {7}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.34,
+                0.885,
+                0.436
+            ],
+            "angle": 0,
+            "content": "Note that we only employ the alignment in lower six layers to better align the representations for switched tokens due to their superior syntax-capturing ability (Jawahar et al., 2019; Rogers et al., 2020). Eventually, the loss function for each subtask \\(\\mathcal{L}_*\\) \\((\\ast \\in [I,S])\\) in Equation 3 is rewritten as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.582,
+                0.447,
+                0.883,
+                0.479
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {*} = \\frac {\\mathcal {L} \\left(h ^ {\\prime *} , y ^ {*}\\right) + \\mathcal {L} \\left(h ^ {*} , y ^ {*}\\right)}{2}, \\tag {8}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.489,
+                0.884,
+                0.57
+            ],
+            "angle": 0,
+            "content": "in which \\( y^{*} \\) denotes golden label of each task. When \\( * \\) denotes \\( I \\), we employ mean pooling strategy on representations of all tokens instead of using \\( h^{\\mathrm{CLS}} \\) and \\( h^{\\prime \\mathrm{CLS}} \\) for intent detection, considering the close relationship of the two tasks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.582,
+                0.657,
+                0.599
+            ],
+            "angle": 0,
+            "content": "3 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.608,
+                0.725,
+                0.625
+            ],
+            "angle": 0,
+            "content": "3.1 Experimental Setting"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.629,
+                0.884,
+                0.741
+            ],
+            "angle": 0,
+            "content": "Dataset and Metrics. We conduct our experiments on the latest multilingual benchmark dataset MultiATIS++ (Xu et al., 2020), which consists of nine languages: English (en), Spanish (es), Portuguese (pt), German (de), French (fr), Chinese (zh), Japanese (ja), Hindi (hi), and Turkish (tr). The details of MultiATIS++ are shown in Appendix A."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.742,
+                0.884,
+                0.807
+            ],
+            "angle": 0,
+            "content": "Following Li et al. (2021), we evaluate the performance of intent prediction using accuracy, slot filling using F1 score, and the sentence-level semantic frame parsing using overall accuracy."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.816,
+                0.885,
+                0.896
+            ],
+            "angle": 0,
+            "content": "Implementation Details. We adopt two baseline models as the backbone, i.e., CoSDA (Qin et al., 2020) and GL-CLEF (Qin et al., 2022). Following previous works, the random rate \\(\\alpha\\) for token substitution is set to 0.2. AdamW (Loshchilov and"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.531,
+                0.904,
+                0.868,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<https://github.com/amazon-science/multiatis>"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7851"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.12,
+                0.082,
+                0.879,
+                0.407
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Model</td><td>en</td><td>de</td><td>es</td><td>fr</td><td>hi</td><td>ja</td><td>pt</td><td>tr</td><td>zh</td><td>Avg.</td></tr><tr><td rowspan=\"7\">Intent Acc</td><td>mBERT(Devlin et al., 2019)</td><td>98.54</td><td>95.40</td><td>96.30</td><td>94.31</td><td>82.41</td><td>76.18</td><td>94.95</td><td>75.10</td><td>82.53</td><td>88.42</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>98.54</td><td>90.48</td><td>93.28</td><td>94.51</td><td>77.15</td><td>76.59</td><td>94.62</td><td>73.29</td><td>84.55</td><td>87.00</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>95.74</td><td>94.06</td><td>92.29</td><td>77.04</td><td>82.75</td><td>73.25</td><td>93.05</td><td>80.42</td><td>78.95</td><td>87.32</td></tr><tr><td>SoGo CoS (Ours)</td><td>98.54</td><td>96.97</td><td>98.12</td><td>96.33</td><td>82.76</td><td>76.41</td><td>97.18</td><td>82.96</td><td>85.77</td><td>90.56</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>98.77</td><td>97.53</td><td>97.05</td><td>97.72</td><td>86.00</td><td>82.84</td><td>96.08</td><td>83.92</td><td>87.68</td><td>91.95</td></tr><tr><td>SoGo GL (Ours)</td><td>98.89</td><td>98.45</td><td>98.15</td><td>97.74</td><td>83.87</td><td>84.75</td><td>97.73</td><td>85.53</td><td>89.10</td><td>92.69</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>98.77</td><td>98.10</td><td>98.10</td><td>98.77</td><td>84.54</td><td>81.86</td><td>97.09</td><td>85.45</td><td>89.03</td><td>92.41</td></tr><tr><td rowspan=\"7\">Slot F1</td><td>mBERT(Devlin et al., 2019)</td><td>95.11</td><td>80.11</td><td>78.22</td><td>82.25</td><td>26.71</td><td>25.40</td><td>72.37</td><td>41.49</td><td>53.22</td><td>61.66</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>95.20</td><td>74.79</td><td>76.52</td><td>74.25</td><td>52.73</td><td>70.10</td><td>72.56</td><td>29.66</td><td>66.91</td><td>68.08</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>92.29</td><td>81.37</td><td>76.94</td><td>79.36</td><td>64.06</td><td>66.62</td><td>75.05</td><td>48.77</td><td>77.32</td><td>73.47</td></tr><tr><td>SoGo CoS (Ours)</td><td>95.46</td><td>84.12</td><td>83.84</td><td>83.46</td><td>57.63</td><td>65.78</td><td>80.27</td><td>55.13</td><td>79.56</td><td>76.14</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>95.39</td><td>86.30</td><td>85.22</td><td>84.31</td><td>70.34</td><td>73.12</td><td>81.83</td><td>65.85</td><td>77.61</td><td>80.00</td></tr><tr><td>SoGo GL (Ours)</td><td>95.42</td><td>87.46</td><td>87.01</td><td>84.45</td><td>74.25</td><td>76.69</td><td>83.91</td><td>67.04</td><td>78.53</td><td>81.64</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>96.02</td><td>86.59</td><td>83.03</td><td>82.11</td><td>61.04</td><td>68.52</td><td>81.49</td><td>65.20</td><td>82.00</td><td>78.23</td></tr><tr><td rowspan=\"7\">Overall Acc</td><td>mBERT(Devlin et al., 2019)</td><td>87.12</td><td>52.69</td><td>52.02</td><td>37.29</td><td>4.92</td><td>7.11</td><td>43.49</td><td>4.33</td><td>18.58</td><td>36.29</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>87.23</td><td>41.43</td><td>44.46</td><td>43.67</td><td>16.01</td><td>33.59</td><td>43.90</td><td>1.12</td><td>30.80</td><td>38.02</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>77.04</td><td>57.06</td><td>46.62</td><td>50.06</td><td>26.20</td><td>28.89</td><td>48.77</td><td>15.24</td><td>46.36</td><td>44.03</td></tr><tr><td>SoGo CoS (Ours)</td><td>88.35</td><td>61.43</td><td>58.30</td><td>56.37</td><td>19.70</td><td>27.65</td><td>60.43</td><td>19.30</td><td>50.17</td><td>49.08</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>88.02</td><td>66.03</td><td>59.53</td><td>57.02</td><td>34.83</td><td>41.42</td><td>60.43</td><td>28.95</td><td>50.62</td><td>54.09</td></tr><tr><td>SoGo GL (Ours)</td><td>90.54</td><td>72.26</td><td>61.05</td><td>57.88</td><td>39.90</td><td>46.95</td><td>64.23</td><td>29.14</td><td>51.31</td><td>57.02</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>89.81</td><td>67.75</td><td>59.13</td><td>57.56</td><td>23.29</td><td>29.34</td><td>61.93</td><td>28.95</td><td>54.76</td><td>52.50</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.416,
+                0.884,
+                0.461
+            ],
+            "angle": 0,
+            "content": "Table 1: Major results \\((\\%)\\) on MultiATIS++. Higher is better in all columns. Results with \\(\\diamond\\) are taken from Qin et al. (2022), and results with \\(\\diamond\\) are taken from Liang et al. (2022). SoGo CoS and SoGo GL can be directly compared to CoSDA and GL-CLEF. Note that LAJ-MCL is not open-source."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.484,
+                0.49,
+                0.63
+            ],
+            "angle": 0,
+            "content": "Hutter, 2019) is used to train SoGO with a learning rate of \\(5e - 6\\). The coefficients in Equation 3 are \\(\\lambda_{I}\\) and \\(\\lambda_{S}\\), where the best ones are selected by searching a combination with the following ranges: \\(\\{0.8, 0.9, 1.0, 1.1, 1.2\\}\\). All experiments are conducted on a single TESLA-V100. The experimental results of our method are averaged over five runs with different random seed to ensure the final reported results are statistically stable."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.641,
+                0.273,
+                0.657
+            ],
+            "angle": 0,
+            "content": "3.2 Major Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.662,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "From the major results in Table 1, we have the following observations: (1) SoGo \\(\\mathrm{CoS}\\) outperforms CoSDA significantly. This is because that code-switched sentences generated by performing salient word substitution and token-level alignment maintain better syntactic and semantic consistency than those generated by random substitution. (2) Both SoGo \\(\\mathrm{CoS}\\) and SoGo GL beat their counterparts with \\(5.05\\%\\) and \\(2.93\\%\\) average improvement in overall accuracy, respectively. This demonstrates that SoGo can be integrated into existing SLU models, orthogonal of other structures. Notably, the gain in slot filling is more significant than in intent detection, further validating the effectiveness of token-level alignment. (3) As LAJ-MCL is not open-source, we implemented two versions of"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.484,
+                0.883,
+                0.583
+            ],
+            "angle": 0,
+            "content": "SoGO, referred to as SoGo \\( \\text{CoS} \\) and SoGO \\( \\text{GL} \\), to compare with LAJ-MCL. Remarkably, SoGO \\( \\text{GL} \\) outperformed the competitive LAJ-MCL in overall accuracy across 8/9 languages. This highlights the significant potential of our SoGO when combined with existing SLU models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.595,
+                0.691,
+                0.611
+            ],
+            "angle": 0,
+            "content": "3.3 Further Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.618,
+                0.884,
+                0.827
+            ],
+            "angle": 0,
+            "content": "Ablation Study. We conduct an ablation study to explore the effect of each component in SoGo. The results are illustrated in Figure 2, which show that each proposed component (i.e., saliency-based substitution and token-level alignment) contributes to the performance positively. It is obvious that the proposed saliency-based substitution has a strong ability to pick out important words to boost performance. Moreover, our token-level alignment has shown greater relative importance of the switched tokens than the vanilla code-switching method. It indicates that such substituted tokens significantly contribute to SoGo's prediction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Analysis of Saliency-based Substitution. To explore the potential of saliency-based substitution, we perform experiments on SoGo CoS of different strategies with respect to different keyword replacement rate \\(\\frac{k}{v}\\) (cf. Figure 3, where \\(k\\) denotes"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7852"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.117,
+                0.083,
+                0.488,
+                0.192
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.202,
+                0.49,
+                0.245
+            ],
+            "angle": 0,
+            "content": "Figure 2: Ablation experiments. \"SbS\" and \"TlA\" denote saliency-based substitution and token-level alignment, respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.271,
+                0.49,
+                0.433
+            ],
+            "angle": 0,
+            "content": "the number of keywords, and \\( v \\) denotes the total number of words.). We find the performance of randomly selecting keywords significantly declines as \\( \\frac{k}{v} \\) decreases, whereas our saliency-based substitution continues to perform well even with only 1% of keywords. This is because saliency-based substitution prioritizes the most indicative keywords for code-switching. As \\( \\frac{k}{v} \\) increases, the additional keywords become less indicative and might even have a negative effect on the model's performance."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.134,
+                0.457,
+                0.469,
+                0.628
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.639,
+                0.49,
+                0.698
+            ],
+            "angle": 0,
+            "content": "Figure 3: Overall Accuracy of SoGo CoS in different replacement rate of keywords \\(\\frac{k}{v}\\). Random denotes selecting keywords randomly, while Saliency denotes selecting keywords via saliency-based substitution."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.727,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Analysis of Token-level Alignment. To verify whether token-level alignment preserves a context space to keep grammatical coherence, we plot the degree of dispersion among tokens within a random sentence in Figure 4. Concretely, we retrieve the token embedding from intermediate (sixth) layers for both switched and original tokens and utilize top features calculated from PCA (Abdi and Williams, 2010). From Figure 4, we can conclude that SoGo CoS has shown a more compact space for the original tokens, where switched tokens are also inside the context space. In contrast, the vanilla code"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.085,
+                0.883,
+                0.117
+            ],
+            "angle": 0,
+            "content": "switching method entirely separates the substituted tokens apart from the original context."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.53,
+                0.133,
+                0.864,
+                0.313
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.323,
+                0.883,
+                0.353
+            ],
+            "angle": 0,
+            "content": "Figure 4: Token-Level density of SoGo \\(\\mathrm{CoS}\\) during training procedure."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.388,
+                0.642,
+                0.403
+            ],
+            "angle": 0,
+            "content": "4 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.418,
+                0.885,
+                0.611
+            ],
+            "angle": 0,
+            "content": "We proposed a novel semantics-coherent and grammar-coherent method (SoGo) to enhance code-switching for zero-shot cross-lingual SLU. Specifically, we utilized saliency-based substitution to replace keywords for better semantic coherence in the shared space. Besides, we introduced an alignment strategy to map the code-switched tokens to original contexts, which solves the grammatical incoherence of code-switching. Experiments on MultiATIS++ showed that code-switching based methods with SoGO obtained competitive performance compared to counterpart baselines."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.629,
+                0.615,
+                0.644
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.658,
+                0.885,
+                0.739
+            ],
+            "angle": 0,
+            "content": "By applying saliency-based substitution and token-level alignment, SoGo achieves significant improvement on the benchmark datasets. Nevertheless, we summarize two limitations for further discussion and investigation by other researchers:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.741,
+                0.884,
+                0.853
+            ],
+            "angle": 0,
+            "content": "(1) Like other code-switching methods, the performance of the model still relies heavily on the correctness of the bilingual dictionary (Fazili and Jyothi, 2022; Whitehouse et al., 2022). In the future, we continue to enhance the performance of zero-shot cross-lingual SLU by refining code-switching techniques (Cheng et al., 2023a)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.856,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "(2) SoGo should further leverage language-independent information (Yu et al., 2021), such as dependency relations and POS tags, to ensure syntactic invariance for cross-lingual transfers."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.508,
+                0.741,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7853"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.287,
+                0.101
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.111,
+                0.491,
+                0.191
+            ],
+            "angle": 0,
+            "content": "We thank all anonymous reviewers for their constructive comments. This paper was partially supported by Shenzhen Science & Technology Research Program (No:GXWD20201231165807007-20200814115301001) and NSFC (No: 62176008)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.219,
+                0.214,
+                0.233
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.242,
+                0.489,
+                0.282
+            ],
+            "angle": 0,
+            "content": "Herve Abdi and Lynne J Williams. 2010. Principal component analysis. Wiley interdisciplinary reviews: computational statistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.293,
+                0.487,
+                0.333
+            ],
+            "angle": 0,
+            "content": "Leila Arras, Ahmed Osman, Klaus-Robert Müller, and Wojciech Samek. 2019. Evaluating recurrent neural network explanations. In ACL Workshop."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.344,
+                0.488,
+                0.384
+            ],
+            "angle": 0,
+            "content": "Dongsheng Chen, Zhiqi Huang, and Yuexian Zou. 2022. Leveraging bilinear attention to improve spoken language understanding. In ICASSP."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.394,
+                0.487,
+                0.422
+            ],
+            "angle": 0,
+            "content": "Qian Chen, Zhu Zhuo, and Wen Wang. 2019. BERT for joint intent classification and slot filling. CoRR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.432,
+                0.488,
+                0.498
+            ],
+            "angle": 0,
+            "content": "Xuxin Cheng, Bowen Cao, Qichen Ye, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023a. Ml-lmcl: Mutual learning and large-margin contrastive learning for improving asr robustness in spoken language understanding. In Proc. of ACL Findings."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.509,
+                0.488,
+                0.562
+            ],
+            "angle": 0,
+            "content": "Xuxin Cheng, Wanshi Xu, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023b. Towards spoken language understanding via multi-level multi-grained contrastive learning. In CIKM."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.573,
+                0.488,
+                0.639
+            ],
+            "angle": 0,
+            "content": "Xuxin Cheng, Zhihong Zhu, Bowen Cao, Qichen Ye, and Yuexian Zou. 2023c. MRRL: Modifying the reference via reinforcement learning for non-autoregressive joint multiple intent detection and slot filling. In Proc. of EMNLP Findings."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.65,
+                0.488,
+                0.677
+            ],
+            "angle": 0,
+            "content": "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In NeurIPS."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.687,
+                0.488,
+                0.74
+            ],
+            "angle": 0,
+            "content": "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. Xnli: Evaluating crosslingual sentence representations. In EMNLP."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.751,
+                0.488,
+                0.79
+            ],
+            "angle": 0,
+            "content": "Misha Denil, Alban Demiraj, and Nando de Freitas. 2014. Extraction of salient sentences from labelled documents. CoRR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.802,
+                0.488,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Ameet Deshpande, Partha Talukdar, and Karthik Narasimhan. 2022. When is BERT multilingual? isolating crucial ingredients for cross-lingual transfer. In AACL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.866,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In AACL."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.242,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Barah Fazili and Preethi Jyothi. 2022. Aligning multilingual embeddings for improved code-switched natural language understanding. In COLING."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.136,
+                0.883,
+                0.175
+            ],
+            "angle": 0,
+            "content": "Yukun Feng, Feng Li, and Philipp Koehn. 2022. Toward the limitation of code-switching in cross-lingual transfer. In EMNLP."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.187,
+                0.883,
+                0.227
+            ],
+            "angle": 0,
+            "content": "Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In ACL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.237,
+                0.883,
+                0.289
+            ],
+            "angle": 0,
+            "content": "Siyu Lai, Hui Huang, Dong Jing, Yufeng Chen, Jinan Xu, and Jian Liu. 2021. Saliency-based multi-view mixed language training for zero-shot cross-lingual classification. In EMNLP Findings."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.3,
+                0.883,
+                0.34
+            ],
+            "angle": 0,
+            "content": "Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018. Word translation without parallel data. In ICLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.35,
+                0.883,
+                0.403
+            ],
+            "angle": 0,
+            "content": "Xiaochun Lei, Xiang Cai, Linjun Lu, Zihang Cui, and Zetao Jiang. 2023. Su2ge-net: a saliency-based approach for non-specific class foreground segmentation. Scientific Reports."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.413,
+                0.883,
+                0.44
+            ],
+            "angle": 0,
+            "content": "Guanbin Li and Yizhou Yu. 2015. Visual saliency based on multiscale deep features. In CVPR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.45,
+                0.883,
+                0.503
+            ],
+            "angle": 0,
+            "content": "Haoran Li, Abhinav Arora, Shuohui Chen, Anchit Gupta, Sonal Gupta, and Yashar Mehdad. 2021. MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark. In EACL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.513,
+                0.883,
+                0.578
+            ],
+            "angle": 0,
+            "content": "Shining Liang, Linjun Shou, Jian Pei, Ming Gong, Wanli Zuo, Xianglin Zuo, and Daxin Jiang. 2022. Label-aware multi-level contrastive learning for cross-lingual spoken language understanding. In EMNLP."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.589,
+                0.883,
+                0.642
+            ],
+            "angle": 0,
+            "content": "Zihan Liu, Genta Indra Winata, Zhaojiang Lin, Peng Xu, and Pascale Fung. 2020. Attention-informed mixed-language training for zero-shot cross-lingual task-oriented dialogue systems. In AAAI."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.652,
+                0.883,
+                0.68
+            ],
+            "angle": 0,
+            "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In ICLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.689,
+                0.883,
+                0.755
+            ],
+            "angle": 0,
+            "content": "Libo Qin, Qiguang Chen, Tianbao Xie, Qixin Li, JianGuang Lou, Wanxiang Che, and Min-Yen Kan. 2022. Gl-clef: A global-local contrastive learning framework for cross-lingual spoken language understanding. In ACL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.766,
+                0.883,
+                0.817
+            ],
+            "angle": 0,
+            "content": "Libo Qin, Minheng Ni, Yue Zhang, and Wanxiang Che. 2020. Cosda-ml: Multi-lingual code-switching data augmentation for zero-shot cross-lingual NLP. In IJCAI."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.829,
+                0.883,
+                0.868
+            ],
+            "angle": 0,
+            "content": "Libo Qin, Tianbao Xie, Wanxiang Che, and Ting Liu. 2021. A survey on spoken language understanding: Recent advances and new frontiers. In *IJCAI*."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.879,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Stephen Robertson. 2004. Understanding inverse document frequency: on theoretical arguments for IDF. J. Documentation."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7854"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.086,
+                0.489,
+                0.138
+            ],
+            "angle": 0,
+            "content": "Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how BERT works. Trans. Assoc. Comput. Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.153,
+                0.488,
+                0.193
+            ],
+            "angle": 0,
+            "content": "Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning important features through propagating activation differences. In ICML."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.207,
+                0.488,
+                0.26
+            ],
+            "angle": 0,
+            "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NeurIPS."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.274,
+                0.488,
+                0.315
+            ],
+            "angle": 0,
+            "content": "Yuxuan Wang, Wanxiang Che, Jiang Guo, Yijia Liu, and Ting Liu. 2019. Cross-lingual BERT transformation for zero-shot dependency parsing. In EMNLP."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.328,
+                0.488,
+                0.381
+            ],
+            "angle": 0,
+            "content": "Chenxi Whitehouse, Fenia Christopoulou, and Ignacio Iacobacci. 2022. Entitycs: Improving zero-shot cross-lingual transfer with entity-centric code switching. In EMNLP Findings."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.395,
+                0.488,
+                0.434
+            ],
+            "angle": 0,
+            "content": "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of bert. In EMNLP."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.449,
+                0.488,
+                0.489
+            ],
+            "angle": 0,
+            "content": "Weijia Xu, Batool Haider, and Saab Mansour. 2020. End-to-end slot alignment and recognition for crosslingual NLU. In EMNLP."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.503,
+                0.488,
+                0.569
+            ],
+            "angle": 0,
+            "content": "Yi Yang, Yueyuan Zheng, Didan Deng, Jindi Zhang, Yongxiang Huang, Yumeng Yang, Janet H Hsiao, and Caleb Chen Cao. 2022. Hsi: Human saliency imitator for benchmarking saliency-based model explanations. In AAAI."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.583,
+                0.488,
+                0.624
+            ],
+            "angle": 0,
+            "content": "Sicheng Yu, Hao Zhang, Yulei Niu, Qianru Sun, and Jing Jiang. 2021. COSY: counterfactual syntax for cross-lingual understanding. In ACL."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.638,
+                0.488,
+                0.691
+            ],
+            "angle": 0,
+            "content": "Michelle Yuan, Mozhi Zhang, Benjamin Van Durme, Leah Findlater, and Jordan L. Boyd-Graber. 2020. Interactive refinement of cross-lingual word embeddings. In EMNLP."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.705,
+                0.488,
+                0.771
+            ],
+            "angle": 0,
+            "content": "Zhihong Zhu, Xuxin Cheng, Dongsheng Chen, Zhiqi Huang, Hongxiang Li, and Yuexian Zou. 2023a. Mix before Align: Towards Zero-shot Cross-lingual Sentiment Analysis via Soft-Mix and Multi-View Learning. In Interspeech."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.785,
+                0.488,
+                0.851
+            ],
+            "angle": 0,
+            "content": "Zhihong Zhu, Xuxin Cheng, Zhiqi Huang, Dongsheng Chen, and Yuexian Zou. 2023b. Towards unified spoken language understanding decoding via label-aware compact linguistics representations. In Proc. of ACL Findings."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.865,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Zhihong Zhu, Weiyuan Xu, Xuxin Cheng, Tengtao Song, and Yuexian Zou. 2023c. A dynamic graph interactive framework with label-semantic injection for spoken language understanding. In ICASSP."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.084,
+                0.616,
+                0.099
+            ],
+            "angle": 0,
+            "content": "A Dataset"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.113,
+                0.884,
+                0.242
+            ],
+            "angle": 0,
+            "content": "MixATIS++ is an extension of Multilingual ATIS (Table 2). Human-translated data for six languages including Spanish (es), German (de), Chinese (zh), Japanese (ja), Portuguese (pt), French (fr) are added to Multilingual ATIS which initially has Hindi (hi) and Turkish (tr). There are 4,478 utterances in the train set, 500 in the valid set, and 893 in the test set, with 18 intents and 84 slots for each language."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.515,
+                0.256,
+                0.883,
+                0.34
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Language</td><td colspan=\"3\">Utterances</td><td rowspan=\"2\">Intent types</td><td rowspan=\"2\">Slot types</td></tr><tr><td>#Train</td><td>#Valid</td><td>#Test</td></tr><tr><td>hi</td><td>1,440</td><td>160</td><td>893</td><td>17</td><td>75</td></tr><tr><td>tr</td><td>578</td><td>60</td><td>715</td><td>17</td><td>71</td></tr><tr><td>others</td><td>4,488</td><td>490</td><td>893</td><td>18</td><td>84</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.574,
+                0.349,
+                0.816,
+                0.362
+            ],
+            "angle": 0,
+            "content": "Table 2: Statistics of MultiATIS++."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.397,
+                0.715,
+                0.414
+            ],
+            "angle": 0,
+            "content": "B Experiment Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.427,
+                0.884,
+                0.522
+            ],
+            "angle": 0,
+            "content": "Following the zero-shot setting, we fine-tune the model on en training set and use en validation set for the hyper-parameters search. The best model checkpoint is decided by the overall accuracy on en validation set. Our code is based on PyTorch and Transformers<sup>2</sup>."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.54,
+                0.63,
+                0.554
+            ],
+            "angle": 0,
+            "content": "C Baselines"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.569,
+                0.882,
+                0.584
+            ],
+            "angle": 0,
+            "content": "We compare our model to the following baselines."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.598,
+                0.884,
+                0.695
+            ],
+            "angle": 0,
+            "content": "mBERT. mBERT<sup>3</sup> follows the same model architecture as BERT (Devlin et al., 2019), but instead of training only on monolingual English data, it is trained on the Wikipedia pages of 104 languages with a shared word piece vocabulary, allowing the model to share embeddings across languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.708,
+                0.884,
+                0.772
+            ],
+            "angle": 0,
+            "content": "ZSJoint. Chen et al. (2019) propose a zero-shot SLU model, which is trained on the en training set and directly applied to the test sets of target languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.786,
+                0.885,
+                0.867
+            ],
+            "angle": 0,
+            "content": "SoSDA. Qin et al. (2020) propose a data augmentation framework to generate multi-lingual code-switching data to fine-tune mBERT, which encourages the model to align representations from the source and multiple target languages."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.88,
+                0.884,
+                0.918
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)https://github.com/huggingface/transformers  \n\\(^{3}\\)https://github.com/google-research/bert/blob/master/multilingual.md"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7855"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.085,
+                0.49,
+                0.149
+            ],
+            "angle": 0,
+            "content": "GL-CLEF. Qin et al. (2022) introduce a contrastive learning framework to explicitly align representations across languages for zero-shot crosslingual SLU."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.158,
+                0.49,
+                0.207
+            ],
+            "angle": 0,
+            "content": "LAJ-MCL. Liang et al. (2022) proposes a multi-level contrastive learning framework for zero-shot cross-lingual SLU."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.218,
+                0.376,
+                0.234
+            ],
+            "angle": 0,
+            "content": "D Generalizability of SoGo"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.244,
+                0.491,
+                0.404
+            ],
+            "angle": 0,
+            "content": "We investigate the potential of the proposed model for other tasks through an additional set of experiments. Following Qin et al. (2020), we conduct additional experiments for natural language inference using XNLI (Conneau et al., 2018), which includes 15 languages. We directly input pairs of sentences into the mBERT encoder, and a task-specific classification layer is employed for classification. The models are evaluated based on classification accuracy (ACC), and the results are presented in Table 3."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.12,
+                0.429,
+                0.489,
+                0.51
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>Avg.</td></tr><tr><td>CoSDA_XLM-based (Conneau and Lample, 2019)</td><td>75.3</td></tr><tr><td>+ SoGo (Ours)</td><td>77.4</td></tr><tr><td>CoSDA_mBERT-based (Wu and Dredze, 2019)</td><td>69.7</td></tr><tr><td>+ SoGo (Ours)</td><td>71.9</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.133,
+                0.519,
+                0.468,
+                0.534
+            ],
+            "angle": 0,
+            "content": "Table 3: Natural language inference experiments."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7856"
+        }
+    ]
+]

2023/Enhancing Code-Switching for Cross-lingual SLU_ A Unified View of Semantic and Grammatical Coherence/b8a9adec-48b2-48e3-8fb6-523de3c4eab2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef0a145e3362c181d1f2287c6c2b599e8b04bf1821eedf208ced3e282bc04f74
+size 470802

2023/Enhancing Code-Switching for Cross-lingual SLU_ A Unified View of Semantic and Grammatical Coherence/full.md

@@ -0,0 +1,266 @@
+# Enhancing Code-Switching for Cross-lingual SLU: A Unified View of Semantic and Grammatical Coherence
+
+Zhihong Zhu† Xuxin Cheng† Zhiqi Huang Dongsheng Chen Yuexian Zou*
+
+School of ECE, Peking University, China
+
+{zhihongzhu,chengxx,chends}@stu.pku.edu.cn
+
+{zhiqihuang, zouyx}@pku.edu.cn
+
+# Abstract
+
+Despite the success of spoken language understanding (SLU) in high-resource languages, achieving similar performance in low-resource settings, such as zero-shot scenarios, remains challenging due to limited labeled training data. To improve zero-shot cross-lingual SLU, recent studies have explored code-switched sentences containing tokens from multiple languages. However, vanilla code-switched sentences often lack semantic and grammatical coherence. We ascribe this lack to two issues: (1) randomly replacing code-switched tokens with equal probability and (2) disregarding token-level dependency within each language. To tackle these issues, in this paper, we propose a novel method termed SoGO, for zero-shot cross-lingual SLU. First, we use a saliency-based substitution approach to extract keywords as substitution options. Then, we introduce a novel token-level alignment strategy that considers the similarity between the context and the code-switched tokens, ensuring grammatical coherence in code-switched sentences. Extensive experiments and analyses demonstrate the superior performance of SoGO across nine languages on MultiATIS++.
+
+# 1 Introduction
+
+Acting as the interface between users and machines, spoken language understanding (SLU) is crucial in task-oriented dialogue systems (Qin et al., 2021; Chen et al., 2022; Zhu et al., 2023b; Cheng et al., 2023c). While joint training models have led to significant advancements, most existing models rely on annotated training data, limiting their scalability to low-resource languages. To this end, zero-shot cross-lingual SLU gains increasing attention.
+
+Existing works have shown the effectiveness of Multilingual BERT (mBERT) in multilingual corpus pre-training, enabling promising zero-shot
+
+cross-lingual SLU (Deshpande et al., 2022; Cheng et al., 2023b; Zhu et al., 2023a). Therein, Qin et al. (2020) extended this concept to a code-switched setting, aligning the source language with multiple target languages. Qin et al. (2022) incorporate contrastive learning to achieve fine-grained cross-lingual transfer. Based on this, Liang et al. (2022) further propose a multi-level contrastive learning framework for explicit alignment of utterance-slot-word structure in cross-lingual SLU.
+
+However, most above-mentioned models adopt code-switching methods, which often lack both semantic coherence and grammatical coherence. This can be attributed to two main issues: (1) Current approaches overlook the importance of individual words in a sentence, as they are randomly replaced with equal probability, potentially introducing unnecessary translation burden and noise that affects semantic coherence. (2) Code-switched sentences comprising tokens from multiple languages may lack grammatical coherence, as token-level dependencies within each language are disregarded.
+
+To address these issues, we propose a novel framework dubbed SoGo for enhancing code-switching. Concretely, we first employ a saliency-based substitution approach to identify keywords with high saliency scores. These keywords are then replaced with their target language equivalents using bilingual dictionaries, to generate code-switched sentences. Notably, unlike the approach by Liu et al. (2020), our SoGo framework requires only a single training process for multiple target languages. Then, we devise a novel token-level alignment strategy to bridge the gap between original context and code-switched tokens. By mapping code-switched tokens back to the original context, SoGo minimizes the impact of token substitutions on token-level coherence, thus maintaining consistent contextual space within each language.
+
+Experimental results on MultiATIS++ (Xu et al., 2020) demonstrate that employing code-switching
+
+![](images/bf4926966a1aaee9e0bc3e57a90f9c1c6c58e4320c596fb1afe93b0ad9f0e0cf.jpg)  
+(a)
+
+![](images/b5e0ae56ff47ef1a77effcd9b63b7d3d28ffc66a691c4e4e6a1da5cadea41863.jpg)  
+Figure 1: (a) Our proposed gradient-based substitution in SoGo. The code-switched dictionary comprised of salient words has the capacity to more accurately characterize the common semantic space and facilitate the acquisition of semantic connections among different languages by the model. (b) The main architecture of SoGo. We perform token-level alignment in lower six layers of mBERT to enhance the grammatical coherence of the code-switched sentence.
+
+with SoGo leads to significant performance improvement compared to baseline models. Furthermore, we demonstrate the compatibility of SoGo with other existing SLU structures.
+
+# 2 Method
+
+We describe problem formulation (§2.1) and backbone model (§2.2) for zero-shot cross-lingual SLU first, before describing SoGo (§2.3). The main architecture of SoGo is illustrated in Figure 1.
+
+# 2.1 Problem Formulation
+
+Given each input sentence  $x_{tgt}$  in a target language, zero-shot cross-lingual SLU means the SLU model is trained in a source language dataset  $S$ , e.g., {English}, and directly applied to other target languages datasets  $T$ , e.g., {Chinese, Japanese}:
+
+$$
+\left(o _ {t g t} ^ {I}, o _ {t g t} ^ {S}\right) = f \left(x _ {t g t}\right), \tag {1}
+$$
+
+where  $f(\cdot)$  is the trained model.  $o_{tgt}^{I}$  represents an intent label and  $o_{tgt}^{S}$  represents a slot sequence. Note that  $T$  may consist of multiple tgt languages.
+
+# 2.2 Backbone Model
+
+For each input sentence  $x$ , we follow Qin et al. (2022); Liang et al. (2022) to utilize bilingual dictionaries (Lample et al., 2018) to generate its corresponding multi-lingual one  $x'$ . Following previous studies, we select mBERT (Devlin et al., 2019)
+
+as our sentence encoder. Formally, the input of mBERT is formulated as follows:
+
+$$
+x ^ {\prime} = \left([ \mathrm {C L S} ], x _ {1} ^ {\prime}, \dots , x _ {n} ^ {\prime}, [ \mathrm {S E P} ]\right), \tag {2}
+$$
+
+where [CLS] denotes the special symbol for representing the whole sequence, and [SEP] can be used for separating non-consecutive token sequences (Devlin et al., 2019).
+
+Intent Detection. We feed the sentence representation  $h_{\mathrm{CLS}}'$  generated from mBERT into a classification layer to obtain  $o^{I}:o^{I} = \mathrm{softmax}(W^{I}h_{\mathrm{CLS}}^{\prime} + b^{I})$ , where  $W^{I}$  and  $b^{I}$  are tunable parameters.
+
+Slot Filling. We use the hidden state to predict each slot:  $o_{t}^{S} = \mathrm{softmax}(W^{S}h_{t}^{\prime} + b^{S})$ , where  $h_t^\prime$  denotes the representation of word  $x_{t}^{\prime}$ . Note that mBERT tokenizes words into subwords (Wang et al., 2019), so we adopt the first sub-token's representation as the whole word representation.
+
+Training and Inference. Following previous works, the loss function for intent detection  $(\mathcal{L}_I)$  and slot filling  $(\mathcal{L}_S)$  are optimized jointly by corresponding cross-entropy loss. Finally, the total training loss is the weighted sum of two losses:
+
+$$
+\mathcal {L} = \lambda_ {I} \mathcal {L} _ {I} + \lambda_ {S} \mathcal {L} _ {S}, \tag {3}
+$$
+
+where  $\lambda_I$  and  $\lambda_S$  denote two hyper-parameters.
+
+During inference, we do not use any code-switched sentences.
+
+# 2.3 SoGo
+
+Saliency-based Substitution. Intuitively, each word in a sentence contributes differently, and we refer to words with significant impact as keywords. Given a vocabulary set  $\mathcal{V}$  containing  $v$  words, we can identify a salient subset of keywords  $\mathcal{K} \subseteq \mathcal{V}$  to generate code-switched sentences. In our work, gradient-based saliency scores (Li and Yu, 2015; Denil et al., 2014; Arras et al., 2019) are used to select keywords, which has been proven effective in many other tasks (Yang et al., 2022; Lai et al., 2021; Lei et al., 2023). Methodologically, let  $x_{n}$  denotes the  $n$ -th sentence from  $S$ ,  $\mathcal{L}_{\hat{y}_n}$  is the loss between model's prediction of slot filling  $\hat{y}_n$  and the golden label  $y_{n}$ . For each token  $w_{i,n} \in x_{n}$ , the saliency score is defined as:
+
+$$
+S _ {x _ {n}} (w _ {i, n}) = - \nabla_ {h _ {i, n}} \mathcal {L} _ {\hat {y} _ {n}} \times h _ {i, n}, \qquad (4)
+$$
+
+where  $h_{i,n}$  represents the embedding of  $w_{i,n}$ . Note that mBERT tokenizes words into subwords, so we calculate the average of the subword saliency scores for each word to obtain the final score. In this manner, the gradient signifies the degree to which a word embedding influences the final decision, while the input considers both the sign and magnitude of the input (Shrikumar et al., 2017).
+
+Equation 4 assesses the local significance of a token within a sentence. However, our objective is to contrast a global keyword subset  $\mathcal{K}$  from  $S$ . Following Yuan et al. (2020), we aggregate all saliency scores for the token  $w_{i}$  occurring in  $S$ . We then multiply these scores with the inverse document frequency (IDF) (Robertson, 2004) of the token  $w_{i}$ :
+
+$$
+S \left(w _ {i}\right) = \log \frac {N}{\left| \left\{x \in S : w _ {i} \in x \right\} \right|} \cdot \sum_ {x \in S: w _ {i} \in x} S _ {x} \left(w _ {i}\right), \tag {5}
+$$
+
+in which  $N$  is the total number of words in  $S$ . The IDF term balances word frequency and saliency scores by assigning words with high document frequency a lower weight and vice versa.
+
+Finally, top- $k$  salient words are chosen to compose the code-switched dictionary  $\mathcal{K}$ . We opt for the dictionary  $\mathcal{K}$  as a replacement strategy instead of random word substitution in sentences.
+
+Token-level Alignment. To further align the space of all hidden states in the code-switched sentence, inspired by attention-based alignment methods Zhu et al. (2023c); Feng et al. (2022), we map the code-switched sentence to the original sentence by calculating the similarity between substituted
+
+hidden states and original hidden states. The similarity scores are then used as weights to aggregate the embedding of original hidden states, reflecting the calculated potential for substituted hidden states in the code-switched sequence. To be specific, for each hidden state  $h_i'$  in  $h'$ , the similarity score is calculated as Vaswani et al. (2017):
+
+$$
+\operatorname {s c o r e} _ {i} ^ {\prime} = h _ {i} ^ {\prime} \cdot h ^ {\top}. \tag {6}
+$$
+
+The final potential of the code-switched token  $\tilde{h}_i'$  is the weighted sum of its corresponding  $h$ :
+
+$$
+\alpha_ {i, t} = \frac {\exp \left(\operatorname {s c o r e} _ {i , t} ^ {\prime}\right)}{\sum_ {t ^ {\prime}} \exp \left(\operatorname {s c o r e} _ {i , t ^ {\prime}} ^ {\prime}\right)}, \quad \tilde {h} _ {i} ^ {\prime} = \sum_ {t} \alpha_ {i, t} \cdot h _ {t}. \tag {7}
+$$
+
+Note that we only employ the alignment in lower six layers to better align the representations for switched tokens due to their superior syntax-capturing ability (Jawahar et al., 2019; Rogers et al., 2020). Eventually, the loss function for each subtask  $\mathcal{L}_*$ $(\ast \in [I,S])$  in Equation 3 is rewritten as:
+
+$$
+\mathcal {L} _ {*} = \frac {\mathcal {L} \left(h ^ {\prime *} , y ^ {*}\right) + \mathcal {L} \left(h ^ {*} , y ^ {*}\right)}{2}, \tag {8}
+$$
+
+in which  $y^{*}$  denotes golden label of each task. When  $*$  denotes  $I$ , we employ mean pooling strategy on representations of all tokens instead of using  $h^{\mathrm{CLS}}$  and  $h^{\prime \mathrm{CLS}}$  for intent detection, considering the close relationship of the two tasks.
+
+# 3 Experiments
+
+# 3.1 Experimental Setting
+
+Dataset and Metrics. We conduct our experiments on the latest multilingual benchmark dataset MultiATIS++ (Xu et al., 2020), which consists of nine languages: English (en), Spanish (es), Portuguese (pt), German (de), French (fr), Chinese (zh), Japanese (ja), Hindi (hi), and Turkish (tr). The details of MultiATIS++ are shown in Appendix A.
+
+Following Li et al. (2021), we evaluate the performance of intent prediction using accuracy, slot filling using F1 score, and the sentence-level semantic frame parsing using overall accuracy.
+
+Implementation Details. We adopt two baseline models as the backbone, i.e., CoSDA (Qin et al., 2020) and GL-CLEF (Qin et al., 2022). Following previous works, the random rate  $\alpha$  for token substitution is set to 0.2. AdamW (Loshchilov and
+
+<table><tr><td></td><td>Model</td><td>en</td><td>de</td><td>es</td><td>fr</td><td>hi</td><td>ja</td><td>pt</td><td>tr</td><td>zh</td><td>Avg.</td></tr><tr><td rowspan="7">Intent Acc</td><td>mBERT(Devlin et al., 2019)</td><td>98.54</td><td>95.40</td><td>96.30</td><td>94.31</td><td>82.41</td><td>76.18</td><td>94.95</td><td>75.10</td><td>82.53</td><td>88.42</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>98.54</td><td>90.48</td><td>93.28</td><td>94.51</td><td>77.15</td><td>76.59</td><td>94.62</td><td>73.29</td><td>84.55</td><td>87.00</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>95.74</td><td>94.06</td><td>92.29</td><td>77.04</td><td>82.75</td><td>73.25</td><td>93.05</td><td>80.42</td><td>78.95</td><td>87.32</td></tr><tr><td>SoGo CoS (Ours)</td><td>98.54</td><td>96.97</td><td>98.12</td><td>96.33</td><td>82.76</td><td>76.41</td><td>97.18</td><td>82.96</td><td>85.77</td><td>90.56</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>98.77</td><td>97.53</td><td>97.05</td><td>97.72</td><td>86.00</td><td>82.84</td><td>96.08</td><td>83.92</td><td>87.68</td><td>91.95</td></tr><tr><td>SoGo GL (Ours)</td><td>98.89</td><td>98.45</td><td>98.15</td><td>97.74</td><td>83.87</td><td>84.75</td><td>97.73</td><td>85.53</td><td>89.10</td><td>92.69</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>98.77</td><td>98.10</td><td>98.10</td><td>98.77</td><td>84.54</td><td>81.86</td><td>97.09</td><td>85.45</td><td>89.03</td><td>92.41</td></tr><tr><td rowspan="7">Slot F1</td><td>mBERT(Devlin et al., 2019)</td><td>95.11</td><td>80.11</td><td>78.22</td><td>82.25</td><td>26.71</td><td>25.40</td><td>72.37</td><td>41.49</td><td>53.22</td><td>61.66</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>95.20</td><td>74.79</td><td>76.52</td><td>74.25</td><td>52.73</td><td>70.10</td><td>72.56</td><td>29.66</td><td>66.91</td><td>68.08</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>92.29</td><td>81.37</td><td>76.94</td><td>79.36</td><td>64.06</td><td>66.62</td><td>75.05</td><td>48.77</td><td>77.32</td><td>73.47</td></tr><tr><td>SoGo CoS (Ours)</td><td>95.46</td><td>84.12</td><td>83.84</td><td>83.46</td><td>57.63</td><td>65.78</td><td>80.27</td><td>55.13</td><td>79.56</td><td>76.14</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>95.39</td><td>86.30</td><td>85.22</td><td>84.31</td><td>70.34</td><td>73.12</td><td>81.83</td><td>65.85</td><td>77.61</td><td>80.00</td></tr><tr><td>SoGo GL (Ours)</td><td>95.42</td><td>87.46</td><td>87.01</td><td>84.45</td><td>74.25</td><td>76.69</td><td>83.91</td><td>67.04</td><td>78.53</td><td>81.64</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>96.02</td><td>86.59</td><td>83.03</td><td>82.11</td><td>61.04</td><td>68.52</td><td>81.49</td><td>65.20</td><td>82.00</td><td>78.23</td></tr><tr><td rowspan="7">Overall Acc</td><td>mBERT(Devlin et al., 2019)</td><td>87.12</td><td>52.69</td><td>52.02</td><td>37.29</td><td>4.92</td><td>7.11</td><td>43.49</td><td>4.33</td><td>18.58</td><td>36.29</td></tr><tr><td>ZSJoint(Chen et al., 2019)</td><td>87.23</td><td>41.43</td><td>44.46</td><td>43.67</td><td>16.01</td><td>33.59</td><td>43.90</td><td>1.12</td><td>30.80</td><td>38.02</td></tr><tr><td>CoSDA( Qin et al., 2020)</td><td>77.04</td><td>57.06</td><td>46.62</td><td>50.06</td><td>26.20</td><td>28.89</td><td>48.77</td><td>15.24</td><td>46.36</td><td>44.03</td></tr><tr><td>SoGo CoS (Ours)</td><td>88.35</td><td>61.43</td><td>58.30</td><td>56.37</td><td>19.70</td><td>27.65</td><td>60.43</td><td>19.30</td><td>50.17</td><td>49.08</td></tr><tr><td>GL-CLEF( Qin et al., 2022)</td><td>88.02</td><td>66.03</td><td>59.53</td><td>57.02</td><td>34.83</td><td>41.42</td><td>60.43</td><td>28.95</td><td>50.62</td><td>54.09</td></tr><tr><td>SoGo GL (Ours)</td><td>90.54</td><td>72.26</td><td>61.05</td><td>57.88</td><td>39.90</td><td>46.95</td><td>64.23</td><td>29.14</td><td>51.31</td><td>57.02</td></tr><tr><td>LAJ-MCL( Liang et al., 2022)</td><td>89.81</td><td>67.75</td><td>59.13</td><td>57.56</td><td>23.29</td><td>29.34</td><td>61.93</td><td>28.95</td><td>54.76</td><td>52.50</td></tr></table>
+
+Table 1: Major results  $(\%)$  on MultiATIS++. Higher is better in all columns. Results with  $\diamond$  are taken from Qin et al. (2022), and results with  $\diamond$  are taken from Liang et al. (2022). SoGo CoS and SoGo GL can be directly compared to CoSDA and GL-CLEF. Note that LAJ-MCL is not open-source.
+
+Hutter, 2019) is used to train SoGO with a learning rate of  $5e - 6$ . The coefficients in Equation 3 are  $\lambda_{I}$  and  $\lambda_{S}$ , where the best ones are selected by searching a combination with the following ranges:  $\{0.8, 0.9, 1.0, 1.1, 1.2\}$ . All experiments are conducted on a single TESLA-V100. The experimental results of our method are averaged over five runs with different random seed to ensure the final reported results are statistically stable.
+
+# 3.2 Major Results
+
+From the major results in Table 1, we have the following observations: (1) SoGo  $\mathrm{CoS}$  outperforms CoSDA significantly. This is because that code-switched sentences generated by performing salient word substitution and token-level alignment maintain better syntactic and semantic consistency than those generated by random substitution. (2) Both SoGo  $\mathrm{CoS}$  and SoGo GL beat their counterparts with  $5.05\%$  and  $2.93\%$  average improvement in overall accuracy, respectively. This demonstrates that SoGo can be integrated into existing SLU models, orthogonal of other structures. Notably, the gain in slot filling is more significant than in intent detection, further validating the effectiveness of token-level alignment. (3) As LAJ-MCL is not open-source, we implemented two versions of
+
+SoGO, referred to as SoGo  $\text{CoS}$  and SoGO  $\text{GL}$ , to compare with LAJ-MCL. Remarkably, SoGO  $\text{GL}$  outperformed the competitive LAJ-MCL in overall accuracy across 8/9 languages. This highlights the significant potential of our SoGO when combined with existing SLU models.
+
+# 3.3 Further Analysis
+
+Ablation Study. We conduct an ablation study to explore the effect of each component in SoGo. The results are illustrated in Figure 2, which show that each proposed component (i.e., saliency-based substitution and token-level alignment) contributes to the performance positively. It is obvious that the proposed saliency-based substitution has a strong ability to pick out important words to boost performance. Moreover, our token-level alignment has shown greater relative importance of the switched tokens than the vanilla code-switching method. It indicates that such substituted tokens significantly contribute to SoGo's prediction.
+
+Analysis of Saliency-based Substitution. To explore the potential of saliency-based substitution, we perform experiments on SoGo CoS of different strategies with respect to different keyword replacement rate  $\frac{k}{v}$  (cf. Figure 3, where  $k$  denotes
+
+![](images/dad38814c28aaf51f4bff4b80ce20ed3789e72167164e7cf8dbb81861e7ca1fa.jpg)  
+Figure 2: Ablation experiments. "SbS" and "TlA" denote saliency-based substitution and token-level alignment, respectively.
+
+the number of keywords, and  $v$  denotes the total number of words.). We find the performance of randomly selecting keywords significantly declines as  $\frac{k}{v}$  decreases, whereas our saliency-based substitution continues to perform well even with only 1% of keywords. This is because saliency-based substitution prioritizes the most indicative keywords for code-switching. As  $\frac{k}{v}$  increases, the additional keywords become less indicative and might even have a negative effect on the model's performance.
+
+![](images/dd6302d40c9d151cc4f0b17d49e52a526b808f7aa3595b1a57b08452f719ed72.jpg)  
+Figure 3: Overall Accuracy of SoGo CoS in different replacement rate of keywords  $\frac{k}{v}$ . Random denotes selecting keywords randomly, while Saliency denotes selecting keywords via saliency-based substitution.
+
+Analysis of Token-level Alignment. To verify whether token-level alignment preserves a context space to keep grammatical coherence, we plot the degree of dispersion among tokens within a random sentence in Figure 4. Concretely, we retrieve the token embedding from intermediate (sixth) layers for both switched and original tokens and utilize top features calculated from PCA (Abdi and Williams, 2010). From Figure 4, we can conclude that SoGo CoS has shown a more compact space for the original tokens, where switched tokens are also inside the context space. In contrast, the vanilla code
+
+switching method entirely separates the substituted tokens apart from the original context.
+
+![](images/6af980b7df31f22cd3a9cc7633201ec8649b091a69130ec5135f603d162779c8.jpg)  
+Figure 4: Token-Level density of SoGo  $\mathrm{CoS}$  during training procedure.
+
+# 4 Conclusion
+
+We proposed a novel semantics-coherent and grammar-coherent method (SoGo) to enhance code-switching for zero-shot cross-lingual SLU. Specifically, we utilized saliency-based substitution to replace keywords for better semantic coherence in the shared space. Besides, we introduced an alignment strategy to map the code-switched tokens to original contexts, which solves the grammatical incoherence of code-switching. Experiments on MultiATIS++ showed that code-switching based methods with SoGO obtained competitive performance compared to counterpart baselines.
+
+# Limitations
+
+By applying saliency-based substitution and token-level alignment, SoGo achieves significant improvement on the benchmark datasets. Nevertheless, we summarize two limitations for further discussion and investigation by other researchers:
+
+(1) Like other code-switching methods, the performance of the model still relies heavily on the correctness of the bilingual dictionary (Fazili and Jyothi, 2022; Whitehouse et al., 2022). In the future, we continue to enhance the performance of zero-shot cross-lingual SLU by refining code-switching techniques (Cheng et al., 2023a).  
+(2) SoGo should further leverage language-independent information (Yu et al., 2021), such as dependency relations and POS tags, to ensure syntactic invariance for cross-lingual transfers.
+
+# Acknowledgements
+
+We thank all anonymous reviewers for their constructive comments. This paper was partially supported by Shenzhen Science & Technology Research Program (No:GXWD20201231165807007-20200814115301001) and NSFC (No: 62176008).
+
+# References
+
+Herve Abdi and Lynne J Williams. 2010. Principal component analysis. Wiley interdisciplinary reviews: computational statistics.  
+Leila Arras, Ahmed Osman, Klaus-Robert Müller, and Wojciech Samek. 2019. Evaluating recurrent neural network explanations. In ACL Workshop.  
+Dongsheng Chen, Zhiqi Huang, and Yuexian Zou. 2022. Leveraging bilinear attention to improve spoken language understanding. In ICASSP.  
+Qian Chen, Zhu Zhuo, and Wen Wang. 2019. BERT for joint intent classification and slot filling. CoRR.  
+Xuxin Cheng, Bowen Cao, Qichen Ye, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023a. Ml-lmcl: Mutual learning and large-margin contrastive learning for improving asr robustness in spoken language understanding. In Proc. of ACL Findings.  
+Xuxin Cheng, Wanshi Xu, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023b. Towards spoken language understanding via multi-level multi-grained contrastive learning. In CIKM.  
+Xuxin Cheng, Zhihong Zhu, Bowen Cao, Qichen Ye, and Yuexian Zou. 2023c. MRRL: Modifying the reference via reinforcement learning for non-autoregressive joint multiple intent detection and slot filling. In Proc. of EMNLP Findings.  
+Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In NeurIPS.  
+Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. Xnli: Evaluating crosslingual sentence representations. In EMNLP.  
+Misha Denil, Alban Demiraj, and Nando de Freitas. 2014. Extraction of salient sentences from labelled documents. CoRR.  
+Ameet Deshpande, Partha Talukdar, and Karthik Narasimhan. 2022. When is BERT multilingual? isolating crucial ingredients for cross-lingual transfer. In AACL.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In AACL.
+
+Barah Fazili and Preethi Jyothi. 2022. Aligning multilingual embeddings for improved code-switched natural language understanding. In COLING.  
+Yukun Feng, Feng Li, and Philipp Koehn. 2022. Toward the limitation of code-switching in cross-lingual transfer. In EMNLP.  
+Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In ACL.  
+Siyu Lai, Hui Huang, Dong Jing, Yufeng Chen, Jinan Xu, and Jian Liu. 2021. Saliency-based multi-view mixed language training for zero-shot cross-lingual classification. In EMNLP Findings.  
+Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018. Word translation without parallel data. In ICLR.  
+Xiaochun Lei, Xiang Cai, Linjun Lu, Zihang Cui, and Zetao Jiang. 2023. Su2ge-net: a saliency-based approach for non-specific class foreground segmentation. Scientific Reports.  
+Guanbin Li and Yizhou Yu. 2015. Visual saliency based on multiscale deep features. In CVPR.  
+Haoran Li, Abhinav Arora, Shuohui Chen, Anchit Gupta, Sonal Gupta, and Yashar Mehdad. 2021. MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark. In EACL.  
+Shining Liang, Linjun Shou, Jian Pei, Ming Gong, Wanli Zuo, Xianglin Zuo, and Daxin Jiang. 2022. Label-aware multi-level contrastive learning for cross-lingual spoken language understanding. In EMNLP.  
+Zihan Liu, Genta Indra Winata, Zhaojiang Lin, Peng Xu, and Pascale Fung. 2020. Attention-informed mixed-language training for zero-shot cross-lingual task-oriented dialogue systems. In AAAI.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In ICLR.  
+Libo Qin, Qiguang Chen, Tianbao Xie, Qixin Li, JianGuang Lou, Wanxiang Che, and Min-Yen Kan. 2022. Gl-clef: A global-local contrastive learning framework for cross-lingual spoken language understanding. In ACL.  
+Libo Qin, Minheng Ni, Yue Zhang, and Wanxiang Che. 2020. Cosda-ml: Multi-lingual code-switching data augmentation for zero-shot cross-lingual NLP. In IJCAI.  
+Libo Qin, Tianbao Xie, Wanxiang Che, and Ting Liu. 2021. A survey on spoken language understanding: Recent advances and new frontiers. In *IJCAI*.  
+Stephen Robertson. 2004. Understanding inverse document frequency: on theoretical arguments for IDF. J. Documentation.
+
+Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how BERT works. Trans. Assoc. Comput. Linguistics.
+
+Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning important features through propagating activation differences. In ICML.
+
+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 NeurIPS.
+
+Yuxuan Wang, Wanxiang Che, Jiang Guo, Yijia Liu, and Ting Liu. 2019. Cross-lingual BERT transformation for zero-shot dependency parsing. In EMNLP.
+
+Chenxi Whitehouse, Fenia Christopoulou, and Ignacio Iacobacci. 2022. Entitycs: Improving zero-shot cross-lingual transfer with entity-centric code switching. In EMNLP Findings.
+
+Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of bert. In EMNLP.
+
+Weijia Xu, Batool Haider, and Saab Mansour. 2020. End-to-end slot alignment and recognition for crosslingual NLU. In EMNLP.
+
+Yi Yang, Yueyuan Zheng, Didan Deng, Jindi Zhang, Yongxiang Huang, Yumeng Yang, Janet H Hsiao, and Caleb Chen Cao. 2022. Hsi: Human saliency imitator for benchmarking saliency-based model explanations. In AAAI.
+
+Sicheng Yu, Hao Zhang, Yulei Niu, Qianru Sun, and Jing Jiang. 2021. COSY: counterfactual syntax for cross-lingual understanding. In ACL.
+
+Michelle Yuan, Mozhi Zhang, Benjamin Van Durme, Leah Findlater, and Jordan L. Boyd-Graber. 2020. Interactive refinement of cross-lingual word embeddings. In EMNLP.
+
+Zhihong Zhu, Xuxin Cheng, Dongsheng Chen, Zhiqi Huang, Hongxiang Li, and Yuexian Zou. 2023a. Mix before Align: Towards Zero-shot Cross-lingual Sentiment Analysis via Soft-Mix and Multi-View Learning. In Interspeech.
+
+Zhihong Zhu, Xuxin Cheng, Zhiqi Huang, Dongsheng Chen, and Yuexian Zou. 2023b. Towards unified spoken language understanding decoding via label-aware compact linguistics representations. In Proc. of ACL Findings.
+
+Zhihong Zhu, Weiyuan Xu, Xuxin Cheng, Tengtao Song, and Yuexian Zou. 2023c. A dynamic graph interactive framework with label-semantic injection for spoken language understanding. In ICASSP.
+
+# A Dataset
+
+MixATIS++ is an extension of Multilingual ATIS (Table 2). Human-translated data for six languages including Spanish (es), German (de), Chinese (zh), Japanese (ja), Portuguese (pt), French (fr) are added to Multilingual ATIS which initially has Hindi (hi) and Turkish (tr). There are 4,478 utterances in the train set, 500 in the valid set, and 893 in the test set, with 18 intents and 84 slots for each language.
+
+<table><tr><td rowspan="2">Language</td><td colspan="3">Utterances</td><td rowspan="2">Intent types</td><td rowspan="2">Slot types</td></tr><tr><td>#Train</td><td>#Valid</td><td>#Test</td></tr><tr><td>hi</td><td>1,440</td><td>160</td><td>893</td><td>17</td><td>75</td></tr><tr><td>tr</td><td>578</td><td>60</td><td>715</td><td>17</td><td>71</td></tr><tr><td>others</td><td>4,488</td><td>490</td><td>893</td><td>18</td><td>84</td></tr></table>
+
+Table 2: Statistics of MultiATIS++.
+
+# B Experiment Details
+
+Following the zero-shot setting, we fine-tune the model on en training set and use en validation set for the hyper-parameters search. The best model checkpoint is decided by the overall accuracy on en validation set. Our code is based on PyTorch and Transformers<sup>2</sup>.
+
+# C Baselines
+
+We compare our model to the following baselines.
+
+mBERT. mBERT<sup>3</sup> follows the same model architecture as BERT (Devlin et al., 2019), but instead of training only on monolingual English data, it is trained on the Wikipedia pages of 104 languages with a shared word piece vocabulary, allowing the model to share embeddings across languages.
+
+ZSJoint. Chen et al. (2019) propose a zero-shot SLU model, which is trained on the en training set and directly applied to the test sets of target languages.
+
+SoSDA. Qin et al. (2020) propose a data augmentation framework to generate multi-lingual code-switching data to fine-tune mBERT, which encourages the model to align representations from the source and multiple target languages.
+
+GL-CLEF. Qin et al. (2022) introduce a contrastive learning framework to explicitly align representations across languages for zero-shot crosslingual SLU.
+
+LAJ-MCL. Liang et al. (2022) proposes a multi-level contrastive learning framework for zero-shot cross-lingual SLU.
+
+# D Generalizability of SoGo
+
+We investigate the potential of the proposed model for other tasks through an additional set of experiments. Following Qin et al. (2020), we conduct additional experiments for natural language inference using XNLI (Conneau et al., 2018), which includes 15 languages. We directly input pairs of sentences into the mBERT encoder, and a task-specific classification layer is employed for classification. The models are evaluated based on classification accuracy (ACC), and the results are presented in Table 3.
+
+<table><tr><td>Model</td><td>Avg.</td></tr><tr><td>CoSDA_XLM-based (Conneau and Lample, 2019)</td><td>75.3</td></tr><tr><td>+ SoGo (Ours)</td><td>77.4</td></tr><tr><td>CoSDA_mBERT-based (Wu and Dredze, 2019)</td><td>69.7</td></tr><tr><td>+ SoGo (Ours)</td><td>71.9</td></tr></table>
+
+Table 3: Natural language inference experiments.

2023/Enhancing Code-Switching for Cross-lingual SLU_ A Unified View of Semantic and Grammatical Coherence/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c24c1f5b9c267f2212665dd5d8e5c3e439306533081331ba18e1016169709fb
+size 419752

2023/Enhancing Computation Efficiency in Large Language Models through Weight and Activation Quantization/811af1b7-d7f8-4044-9ba8-d543804a67ed_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d2d836547c3003bb36416222a18c1ad8023fa083de2038e2a1b7f5b104d8e39
+size 3353650

2023/Enhancing Computation Efficiency in Large Language Models through Weight and Activation Quantization/full.md

@@ -0,0 +1,382 @@
+# Enhancing Computation Efficiency in Large Language Models through Weight and Activation Quantization
+
+Janghwan Lee $^{1*}$ , Minsoo Kim $^{1*}$ , Seungcheol Baek $^{2}$ , Seok Joong Hwang $^{2}$ , Wonyong Sung $^{3}$  and Jungwook Choi $^{1\dagger}$
+
+{hwani0288, minsoo2333}@hanyang.ac.kr,{bsc11235, nzthing}@sapeon.com
+
+wysung@snu.ac.kr, choij@hanyang.ac.kr
+
+$^{1}$ Hanyang University,  $^{2}$ SAPEON Korea Inc.,  $^{3}$ Seoul National University
+
+Republic of Korea
+
+# Abstract
+
+Large Language Models (LLMs) are proficient in natural language processing tasks, but their deployment is often restricted by extensive parameter sizes and computational demands. This paper focuses on post-training quantization (PTQ) in LLMs, specifically 4-bit weight and 8-bit activation (W4A8) quantization, to enhance computational efficiency—a topic less explored compared to weight-only quantization. We present two innovative techniques: activation-quantization-aware scaling (AQAS) and sequence-length-aware calibration (SLAC) to enhance PTQ by considering the combined effects on weights and activations and aligning calibration sequence lengths to target tasks. Moreover, we introduce dINT, a hybrid data format combining integer and denormal representations, to address the underflow issue in W4A8 quantization, where small values are rounded to zero. Through rigorous evaluations of LLMs, including OPT and LLaMA, we demonstrate that our techniques significantly boost task accuracies to levels comparable with full-precision models. By developing arithmetic units compatible with dINT, we further confirm that our methods yield a  $2 \times$  hardware efficiency improvement compared to 8-bit integer MAC unit.
+
+# 1 Introduction
+
+Large language models (LLMs) have achieved breakthroughs in many natural language processing tasks such as translation, summarization, reasoning, and conversation, often matching or exceeding human performance (Zhang et al., 2022; Touvron et al., 2023; Chowdhery et al., 2022; Brown et al., 2020; OpenAI, 2023). However, the extensive parameters of LLMs present deployment challenges due to the high memory bandwidth needed for high throughput inference. Post-training quantization (PTQ) addresses this by "compressing" weight pa
+
+rameters, significantly reducing memory requirements and enhancing GPU performance by alleviating memory bandwidth bottlenecks (Frantar et al., 2023; Lin et al., 2023; Lee et al., 2023a). Nevertheless, LLMs' computational complexity remains a concern. For example, GPT-3 (Brown et al., 2020) requires at least 350 GFLOPs of computation for a single token, but PTQ methods often revert compressed weights to higher precisions like 16-bit floating-point (FP16) for computation, which is inefficient given the resource demands of multiply-accumulate (MAC) operations. With computing platforms evolving through high-bandwidth memory (Gurumurthi et al., 2021) and processing-in-memory (Kim et al., 2021; He et al., 2020) to resolve the memory bandwidth bottleneck, addressing LLMs' computational needs becomes more imperative.
+
+A PTQ strategy that effectively quantizes both weights and activations is thus appealing as it reduces the hardware complexity of MAC units, enhancing computational throughput (Sun et al., 2019; Dettmers et al., 2022; Xiao et al., 2022). PTQ research specific to LLM's computation efficiency is growing, focusing on utilizing INT8-INT8 MAC units, common in GPUs (Andersch et al., 2022). LLM.Int8 (Dettmers et al., 2022), for instance, used INT8 quantization for weights and activations, but directed activation outliers through an FP16 datapath, isolating them. SmoothQuant (Xiao et al., 2022) extended this by employing activation channel scaling to target outliers and adjusting corresponding weights for balanced quantization. However, these studies do not address challenges faced when weights are reduced to 4 bits, revealing an unexplored area for combined effects on weight and activation quantization.
+
+This paper delves into the challenges of posttraining quantization (PTQ) for both weights and activations in large language models (LLMs). We pinpoint two primary hurdles in achieving efficient
+
+4-bit weight and 8-bit activation (W4A8) quantization. First, LLMs like OPT (Zhang et al., 2022) and LLaMA (Touvron et al., 2023) have distinct weight and activation range characteristics, making existing PTQ methods unsuitable for universal use. For example, AWQ's (Lin et al., 2023) activation-aware scaling makes activations prone to quantization errors, while OPTQ's (Frantar et al., 2023) weight calibration struggles with varying activation ranges. We propose two novel solutions for this first hurdle: activation-quantization-aware scaling (AQAS) and sequence-length-aware calibration (SLAC). AQAS optimizes quantization scales by jointly considering weights and activations, yielding balanced quantization. SLAC aligns the sequence length of the application task with that of the PTQ calibration dataset, mitigating the impact of variations in activation diversity, which significantly affects the PTQ calibration process.
+
+Second, we observe that underflow, where small-magnitude values round to zero, severely impacts W4A8 quantization in LLMs because the quantization error associated with values rounding to zero constitutes a significant portion of the output error. While underflow is a well-known issue in reduced-precision formats for deep neural networks (DNNs) (Sun et al., 2019, 2020; Chmiel et al., 2022; Jin et al., 2022), previous PTQ research in LLMs mainly focuses on outliers, neglecting underflow. We discover that standard INT4 representation discards crucial small-magnitude weights when multiplied with activations. As existing data formats like integer, floating-point, or logarithmic formats are inadequate for this underflow issue, we introduce dINT, a new integer format with denormal representation. dINT merges the uniform coverage of integers with the denormal of floating-points, effectively mitigating underflow and improving accuracy. We also propose a MAC unit supporting dINT to ensure hardware efficiency.
+
+We evaluate AQAS, SLAC, and dINT on OPT and LLaMA, focusing on language modeling, zero-shot reasoning, and 5-shot in-context learning. The results show that integrating these methods for W4A8 PTQ significantly improves task accuracies for both OPT and LLaMA across a diverse set of benchmarks (Wikitext, Common Sense Question Answering (CSQA), and Massive Multitask Language Understanding (MMLU)) and the model sizes ranging from 125M to 65B parameters.
+
+# 2 Background
+
+# 2.1 Weight-only PTQ for LLMs
+
+Various weight-only PTQ techniques have emerged to alleviate memory-bandwidth constraints in LLM inference by compressing weights to 4 bits while maintaining accuracy (Park et al., 2023; Kwon et al., 2022; Frantar et al., 2023; Lin et al., 2023; Lee et al., 2023a). For example, OPTQ (Frantar et al., 2023) reduces output distortion from column-wise weight quantization by sequentially updating unquantized weights using activation Hessians. AWQ (Lin et al., 2023) scales weights according to activation magnitudes for improved quantization, while OWQ (Lee et al., 2023a) and SPQR (Dettmers et al., 2023) isolate sensitive weights, retaining them at higher precision. However, these approaches entail high-precision computations and complex arithmetic units. We demonstrate that these weight compression methods are sub-optimal for activation quantization in common LLMs, often exacerbating challenges by ignoring activation dynamics. Consequently, we introduce advanced techniques specifically designed to address these intricacies, enhancing weight quantization accuracy when the activation is also quantized.
+
+# 2.2 Weight and Activation PTQ for LLMs
+
+Quantizing both weights and activations enables the use of lower-precision MAC units, significantly saving logic area and power consumption (Horowitz, 2014). As such, many studies aim to reduce DNN's computational burden (Sun et al., 2019; Lee et al., 2023b), especially in LLMs (Dettmers et al., 2022; Xiao et al., 2022; Liu et al., 2023; Bondarenko et al., 2021). For instance, LLM.Int8 (Dettmers et al., 2022) and SmoothQuant (Xiao et al., 2022) employ GPU-supported INT8-INT8 MAC operations for efficiency, with LLM.Int8 processing outliers separately and SmoothQuant adjusting activations and weights. Additionally, (Liu et al., 2023; Bondarenko et al., 2021) employ quantization-aware fine-tuning for further reductions to W4A8 or W4A4, but face noticeable accuracy losses despite expensive fine-tuning. This paper proposes novel solutions that address the accuracy drop in combined weight and activation quantization with bit-precision down to W4A8, achieving superior results compared to prior works without fine-tuning.
+
+![](images/fc1bde421714bfa361b050bb518ce202e03b0c9ea22bf0637d78bcb43deb2a0c.jpg)  
+(a) Computation pattern of OPT (top) and LLaMA (below)
+
+![](images/a28c3a987767cb71e802b17f7e99c1e6b8fbfe618fa5f9857cdd7c1c77001429.jpg)  
+(b) Min-Max range of input activation and weight per layer  
+(d) Max values of per-channel input activation (Left: OPT, Right: LLaMA)
+
+![](images/fa2176a1b3493d85d669b98c0f67dd7105725c2a3d806b84128edc244368b766.jpg)  
+(c) Activation Min-Max range per layer across sequence length (Left: OPT, Right: LLaMA)
+
+![](images/4df3e26d92d34f06e928069dbef3f625669a4a22d547258bfb6cc43f5f803713.jpg)  
+Figure 1: (a) Illustration of fused-layernorm (fused-LN) in OPT (top) and layernorm (LN) in LLaMA (bottom) computation patterns within a Transformer layer. Note that two computation patterns yield the same output if computed in full-precision, but they deviate when activation and weight are quantized. (b) Min-Max range of input activations (left) and weight (right) as operands of matrix multiplication. (c) Min-Max range of input activation varying sequence length from 128 to 2048 (Orange: OPT-6.7B, Blue: LLaMA-7B). (d) Max values of per-channel input activation for OPT-6.7B (left) and LLaMA-7B (right) for different input sequence lengths (32 and 2048).
+
+# 2.3 Underflow for Reduced-Precision LLMs
+
+Underflow, the numerical error from small values rounding to zero due to limited bit-precision, has been actively studied as a critical issue in reduced-precision DNN training. For instance, (Sun et al., 2019) counters underflow in 8-bit floating-point by adjusting the exponent bias, (Sun et al., 2020) utilizes a radix-4 format to represent wider magnitude ranges in 4-bit floating-point (FP4), and (Chmiel et al., 2022) uses stochastic underflow to address biased quantization in FP4 gradients. In fixed-point representation, (Jin et al., 2022) explores optimal formats by analyzing underflow and overflow trade-offs based on fractional length. Contrary to these studies focusing on the training phase, our paper investigates underflow's impact on PTQ of LLMs for the first time and introduces an enhanced integer format to combat it.
+
+# 3 Improving PTQ for Weight and Activation Quantization
+
+We aim to advance LLM quantization beyond the realms of 4-bit weight-only PTQ or W8A8 PTQ by investigating the combined effects of weight and activation quantization. When quantizing both weight and activation, it is important to note that LLMs display distinct weight and activation characteristics. For example, OPT has been found to have  $0.1\%$  activation outliers by (Dettmers et al., 2022), whereas GLM-130B (Zeng et al., 2023) reported  $30\%$  of outliers in its model. In the context of weight, due
+
+to varied weight distributions across models, OPT-66B experiences a substantial perplexity increase in the wikitext benchmark with INT4 weights, soaring from 9.34 to 110 (Frantar et al., 2023), whereas GLM-130B shows no performance degradation on the MMLU benchmark when INT4 weights are applied (Zeng et al., 2023). We posit that these discrepancies arise from variances in pre-training configurations such as datasets, learning rates, layer structures, and self-attention directionality, as well as options designed for efficient inference, such as operation fusion techniques like layernorm fusion. Significantly, existing PTQ research has overlooked these unique traits intrinsic to each model that are pivotal for the combined optimization of activation and weight quantization. Therefore, we delve into the weight and activation distributions of widely-used OPT and LLaMA models during quantization to understand PTQ limitations and develop novel methods to address them.
+
+# 3.1 Model Analysis: OPT vs. LLaMA
+
+To understand the adverse effects of quantization on restricting dynamic range, we examine the minimum and maximum values (Min-Max range) across the layers of LLMs. Fig. 1(a) illustrates the computation patterns within a layer of LLMs and Fig. 1(b) displays Min-Max range of activations (left) and weights (right) as operands of matrix multiplication for each FC layer in OPT and LLaMA. Notably, there are contrasting trends in
+
+Min-Max ranges; OPT has a broad activation range but a narrow weight range, while LLaMA exhibits the opposite. This distinction stems from the way these LLMs process activations at layernorm. As depicted in Fig. 1(a), in OPT, the layernorm parameters are fused to the subsequent FC layer's weights (Fig. 1(a) top), allowing only normalized activation to enter the FC layer. Conversely, layernorm is not fused in LLaMA (Fig. 1(a) below), resulting in scaled activation as input to FC layers. Although layernorm fusion preserves functionality in full-precision computation, this presence or absence of layernorm fusion in activation processing contributes to significantly distinct behaviors under quantization, as will be discussed in the following sections.
+
+Another insightful finding from our model analysis is the variation in activation diversity based on sequence lengths. Fig. 1(c) displays the Min-Max range as sequence length varies from 128 to 2048 (Orange: OPT-6.7B, Blue: LLaMA-7B). Notably, OPT's activation range remains stable across sequence lengths, while LLaMA's activation range expands, suggesting challenges in range calibration for quantization. Fig. 1(d) contrasts maximum values per channel for OPT and LLaMA at varying sequence lengths. OPT displays consistent outliers at the same channels, dominating its activation dynamic ranges. In contrast, LLaMA's outliers increase in magnitude and shift across channels, indicating varied activation dynamic ranges. This distinction in activation diversity is significant for quantization. While PTQ generally presumes consistent dynamic ranges for calibrating quantization ranges "offline", these findings emphasize the necessity of considering distinct activation dynamic range and incorporating sequence length into calibration. The following sections discuss methods to optimize weight and activation quantization, building on these model-specific insights.
+
+# 3.2 Activation-Quantization-Aware Scaling
+
+The distinct properties of outliers in weights and activations illustrated in Fig. 1(b) pose challenges of applying prior scaling techniques. Fig. 2 illustrates the absolute maximum of (a) input activations and (b) weights at the "Key" layers (for self-attention) in OPT-6.7B when different scaling methods are applied. Specifically, SmoothQuant (Xiao et al., 2022) (SQ) scales activation for 8-bit quantization, but descales weights, resulting in a more diverse
+
+![](images/f038f99c1ba0ab58b1d35a5a9b91e17fd12236028bf635eac4ce75f4569bf807.jpg)  
+(a) Input activation
+
+![](images/708f70b63fb0d2c23703dcd1e11f58b05ac170ac2b33488e569c368293a1379b.jpg)  
+(b) Weight  
+Figure 2: Absolute max value of (a) input activation and (b) weight after scaling by each method (OPT-6.7B). We observed that these trends were significantly pronounced in OPT models due to large outliers. (See Fig. 6 for the same plot for LLaMA.)
+
+and quantization-sensitive range for weight. On the other hand, AWQ (Lin et al., 2023) scales weights for 4-bit quantization but significantly increases activation diversity, making activation quantization problematic. In other words, the existing scaling-based PTQ techniques such as SQ and AWQ cannot resolve the issue of conflicting trends in activation and weight outliers. To address this, we introduce activation-quantization-aware scaling (AQAS), a hybrid of SQ and AWQ. AQAS aims to find scaling values that minimize the output error caused by quantized weights and activations. We use mean squared error (MSE) loss as the objective function, aligning with previous studies on layer-wise optimization (Nagel et al., 2020; Frantar et al., 2022). Our objective function is as follows:
+
+$$
+\underset {\mathbf {s}} {\operatorname {a r g m i n}} \left\| Q (\mathbf {W} \cdot \operatorname {d i a g} (\mathbf {s})) Q (\operatorname {d i a g} (\mathbf {s}) ^ {- 1} \cdot \mathbf {X}) - \mathbf {W X} \right\| _ {2} ^ {2} \tag {1}
+$$
+
+We define the weight  $\mathbf{W} \in \mathbb{R}^{M \times C}$ , scale factor  $\mathbf{s} \in \mathbb{R}^C$ , and activation  $\mathbf{X} \in \mathbb{R}^{C \times T}$ , where  $M$  represents the output feature dimension,  $C$  represents the input feature dimension, and  $T$  denotes the number of tokens. Fig. 2 demonstrates that AQAS considers activation quantization's impact to adjust activation magnitudes, easing activation quantization. Additionally, as compared to SQ, AQAS adjusts weight magnitudes more moderately, making 4-bit weight quantization feasible.
+
+# 3.3 Sequence-Length-Aware Calibration
+
+As shown in Fig. 1(c), variation in activation diversity depending on the sequence length affects the quantization performance. Specifically, weight-update-based quantization like OPTQ (Frantar et al., 2023) struggles with models like LLaMA that have increasing activation diversity during calibration. To delve deeper into this phenomenon,
+
+<table><tr><td>Sequence Length</td><td>64</td><td>128</td><td>512</td><td>2048</td><td>std</td></tr><tr><td>FP (LLaMA-7B)</td><td colspan="4">70.92</td><td>-</td></tr><tr><td>INT4</td><td colspan="4">69.37</td><td>-</td></tr><tr><td>INT4 OPTQ</td><td>68.14</td><td>69.89</td><td>65.96</td><td>65.19</td><td>2.54</td></tr><tr><td>INT4 AQAS+OPTQ</td><td>69.99</td><td>70.48</td><td>68.92</td><td>70.20</td><td>0.68</td></tr></table>
+
+Table 1: W4A8 Quantization zero-shot evaluation of CommonSenseQA (average score of PIQA, Winogrande and Arceasy, default calibration sequence length is 2048)
+
+we analyze the approach adopted by OPTQ, which employs weight adjustments in response to quantization error using activation Hessian, formulated as follows (Frantar et al., 2023):
+
+$$
+\delta_ {F} = - \frac {w _ {q} - \operatorname {q u a n t} (w _ {q})}{[ \mathbf {H} _ {F} ^ {- 1} ] _ {q q}} \cdot \left(\mathbf {H} _ {F} ^ {- 1}\right) _ {:, q} \tag {2}
+$$
+
+$$
+\mathbf {H} _ {i} = \frac {\partial^ {2} E}{\partial \mathbf {W} _ {i , :} ^ {2}} = 2 \mathbf {X} \mathbf {X} ^ {T}, \tag {3}
+$$
+
+where  $\mathbf{X}$  denotes the layer input activation,  $\mathbf{W}$  is weights of linear layer,  $w_{q}$  is weight element to quantize, and  $\delta$  denotes optimal weight update recovering quantization error. We examine the weight update ratio,  $(\mathbf{H}_F^{-1})_{:,q} / [\mathbf{H}_F^{-1}]_{qq}$ , representing the second derivative of quantization error  $(E)$ , to assess changes in weights due to OPTQ. Fig. 3(a) shows the weight update ratio for OPT and LLaMA with varying calibration sequence lengths. OPT remains relatively consistent, while LLaMA displays varying weight update ratios for varying sequence length, suggesting activation diversity affects OPTQ's weight updates.
+
+This sensitivity of OPTQ updates prompts us to further explore its implications for performance. We evaluate the zero-shot performance of OPTQ for W4A8 quantization by varying the calibration sequence length on PIQA, Winogrande, and Arceasy tasks from CSQA (Bisk et al., 2019; Sakaguchi et al., 2019; Clark et al., 2018), which have sequence lengths ranging from tens to hundreds (note that the type of calibration dataset was kept consistent). Table 1 reveals that when the calibration sequence length (e.g., 512 or 2048) significantly deviates from task's sequence lengths, OPTQ's performance suffers (up to  $4\%$  degradation), even falling below basic nearest-rounding quantization. However, when the sequence lengths are aligned (e.g., 64 or 128), OPTQ performs exceptionally well.
+
+![](images/50bb8160ccdc01c1e697267af671d0bb1b866d5c8fc8072afdece24ae68e7c37.jpg)
+
+![](images/3a1b2464c6a79d08d9c62fd61e6b25627c0b0319ea5e7cbea4d7095496f22695.jpg)
+
+![](images/065c760aee8dd3626779ee32526201106a07c545366930ae1f87ba6c266bbbf7.jpg)  
+(a)
+
+![](images/e1459f2e8bdaf5fd27925665e7455ce9b902746d763b30accaada2e348595b00.jpg)  
+(b)  
+Figure 3: (a) Comparison of weight update ratio in Eq. 2 in OPT-6.7B, LLaMA-7B, and LLaMA-7B with AQAS scaling. (b) Minimum input activation range for the query layer in three models: W4A8 (calibrated with 128 and 2048 sequence lengths) and full-precision (FP), all evaluated under an input sequence length of 128.
+
+The large standard deviation in accuracies for matching versus non-matching sequence lengths suggests that LLaMA's activation diversity substantially impacts OPTQ's accuracy. To mitigate this, we propose the sequence-length-aware calibration (SLAC) method. This approach involves determining the expected sequence length during the target task's inference phase and aligning the sequence length of the calibration dataset accordingly. Such a task-specific PTQ calibration process enhances the robustness and accuracy of the model's inference. The efficacy of SLAC, particularly in the CSQA benchmark, is substantiated by experiments detailed in Sec. 5.3.
+
+The effectiveness of the SLAC method is evi
+
+dent when comparing the dynamic range of quantized models with their full-precision counterparts. Fig. 3 (b) demonstrates that using calibration data aligned with the input sequence length (calib-128) results in a dynamic range more consistent with that of the full-precision model (FP), unlike models calibrated with mismatched sequence lengths (calib-2048).
+
+Integrating SLAC with AQAS effectively enhances weight and activation quantization. As illustrated in Fig. 3(a), AQAS efficiently mitigates the sensitivity to input sequence length regarding weight updates. Moreover, Table 1 shows that the standard deviation related to the calibration dataset's length is significantly reduced from 2.54 to 0.68 through AQAS. Consequently, combining AQAS with OPTQ proves advantageous for inferences across diverse sequence lengths, and employing the SLAC method for calibration according to the target dataset's sequence length further bolsters performance.
+
+# 4 Overcoming PTQ Underflow for LLMs
+
+By employing AQAS to address activation quantization errors in weight scaling, and utilizing SLAC to align the sequence length of the calibration dataset with that of the target inference, we achieve a substantial improvement in the performance of our W4A8 models. However, we encounter persistent performance degradation issues. In this section, we unveil "underflow" issues as a previously overlooked cause of accuracy degradation in PTQ applied to LLMs and propose a new numerical format to mitigate this problem.
+
+# 4.1 Observations
+
+We identify underflow as a main contributor to performance degradation. To dissect the causes of degradation when converting the weights of the scaled model to 4-bit, we split the quantization error into two parts: rounding error  $(\Delta_r)$  and underflow error  $(\Delta_u)$ . The rounding error accounts for the error when the quantized value is non-zero, whereas the underflow error represents the error occurring when the quantized value rounds to zero. By considering the total error  $(\Delta)$  induced by quantization as a combination of  $\Delta_u$  and  $\Delta_r$ , we can express the expected output quantization error as
+
+follows:
+
+$$
+\begin{array}{l} \mathbf {E} \left[ \left(\mathbf {W X} - \left(\mathbf {W} + \Delta_ {u} + \Delta_ {r}\right) \mathbf {X}\right) ^ {2} \right] \tag {4} \\ \mathbf {\Delta} = \mathbf {E} [ (\Delta_ {u} \mathbf {X}) ^ {2} ] + \mathbf {E} [ (\Delta_ {r} \mathbf {X}) ^ {2} ] + \mathbf {E} [ 2 (\Delta_ {u} \mathbf {X} \Delta_ {r} \mathbf {X}) ]. \\ \end{array}
+$$
+
+Fig. 4(a) exemplifies the underflow issues, illustrating the distinct impacts of quantization errors on final model accuracy, measured as perplexity. The figure highlights that setting small values near zero to exactly zero, while leaving other values unquantized, impairs performance. In contrast, quantizing larger values and precisely representing those near zero significantly improve accuracy. Fig. 4(b) provides a breakdown of error terms across layers in OPT W4A8, indicating a correlation between high total error and substantial underflow error. This underlines the necessity for a method that effectively addresses underflow errors.
+
+# 4.2 Integer with Denormal Representation
+
+Inspired by our observations and the denormal numbers in floating-point representation, we introduce a new integer format called integer with denormal representation (dINT). As illustrated by Fig. 4(c), dINT uses two bins around zero to ensure lower magnitudes are effectively represented. In  $b$ -bit quantization, two values are reserved for special cases, so the quantization range represents integers from 0 to  $2^b - 3$ . These special values in dINT have magnitudes equal to half of the chosen step size, which is a power of two to enable computation by simple bit-shift operations. Our experimental findings have confirmed that this choice of half-step size consistently delivers the most robust performance, surpassing other special values designed for bit shifting, as elaborated in Appendix A.5. The quantization and dequantization procedures for dINT are detailed below:
+
+$$
+X _ {\text {i n t}} = \left\{\begin{array}{l}c _ {1}, \text {f o r} \frac {s}{4} <   n \leq \frac {3 s}{4}\\c _ {2}, \text {f o r} \frac {- 3 s}{4} \leq n <   \frac {- s}{4}\\\operatorname {c l a m p} \left(\left\lceil \frac {X}{s} \right\rceil + z, 0, p\right), \text {e l s e}\end{array}\right. \tag {5}
+$$
+
+$$
+X _ {q} = \left\{ \begin{array}{l} \frac {s}{2}, \text {f o r} X _ {\text {i n t}} = c _ {1} \\ \frac {- s}{2}, \text {f o r} X _ {\text {i n t}} = c _ {2} \\ \left(X _ {\text {i n t}} - z\right) \cdot s, \text {e l s e} \end{array} \right. \tag {6}
+$$
+
+where  $p$  represents the number of uniform steps, calculated as  $p = 2^b - 3$  for a given bit number  $b$ . The step size  $s$  is obtained by dividing the quantization range by  $p$ , and  $z$  is the zero-point for asymmetric quantization.  $c_1$  and  $c_2$  denote the positive
+
+![](images/e54544bcae3f87c48bfce3b4094a1d6bb158167c6caed315458f5a82b5ce9199.jpg)
+
+![](images/74e184543fef19d6718af2274945839e87e1bf6c34e56a50adccafe6935590f4.jpg)
+
+![](images/fe90564dc81fff838a61db07688bd2a28fbdb590ff0a2b594f3f49aec865cb52.jpg)
+
+![](images/803a3933279341a5296d193328e552c3c7f0a267d8bacb261ac916c43d5e6302.jpg)
+
+![](images/5b70aa38ae474af06a5401c534c0ae5d1d02f4b17ec8a26f1c4f52c02928fa3d.jpg)  
+(b) Breakdown of output errors in INT4
+
+![](images/b71e2759774a8d60db85f8912e0531e397e6da8bb3b5e4bbe7f50eb0e5c33e83.jpg)  
+(d) Breakdown of output errors in dINT4
+
+![](images/17672dd01410841e892ae0d3a3c6bf72d06bebed7a8252281debc7c17cce9ad6.jpg)  
+Figure 4: (a) INT4 without rounding sets small values near zero to zero, preserving the rest and causing performance degradation. INT4 without underflow preserves only values near zero, improving performance. (b) Impact of underflow error and rounding error on the output error. Significant impact of underflow error on the output error in INT4. (c) Proposed dINT4 preserves two small values near zero, preventing performance degradation. (d) Using the proposed dINT4 to reduce underflow error leads to a significant reduction in output error.  
+Figure 5: (Blue) Values to be quantized. (Orange) INT4 quantized values, evenly spaced. (Green) FP4 quantized values, dense resolution for small values but coarse resolution for large magnitudes. (Red) Proposed dINT4 format, balanced quantization range with a separate special value for small values.
+
+and negative special values in dINT that represent small magnitudes. These values are encoded with distinct bits, analogous to encoding inf or NaN. During dequantization, if the value corresponds to  $c_{1}$  or  $c_{2}$ , it is represented as a special value; otherwise, dequantization proceeds as in standard integer formats. Fig. 5 shows that dINT4 strikes a balance between INT4, which has uniform dynamic range coverage but underflow issues, and FP4, which densely represents small values to avoid underflow but coarsely covers the dynamic range.
+
+# 4.3 Advantages
+
+Fig. 4(d) showcases the benefits of dINT in reducing output quantization error. By plotting each term of Eq. 4, we observe that dINT primarily mitigates underflow error, which substantially lowers the output error. Although in instances like layer 9, the output error slightly increases due to a widened step size causing a rise in rounding error, the magnitude of this increment is minimal. On the whole,
+
+<table><tr><td>HW Performance</td><td colspan="2">MAC Input Formats</td></tr><tr><td>Precision</td><td>INT8 × INT8</td><td>dINT4 × INT8 (Savings)</td></tr><tr><td>Area (μm2)</td><td>86.33</td><td>44.57 (1.93×)</td></tr><tr><td>Power (mW)</td><td>0.1595</td><td>0.0624 (2.56×)</td></tr></table>
+
+Table 2: Evaluation of hardware performance of MAC units (7nm, 1GHz).
+
+dINT is effective in most scenarios. Furthermore, we design and synthesize a MAC unit using dINT and compare it to a traditional 8-bit integer MAC unit using Synopsys Design Compiler and a commercial 7nm technology (1GHz) for area efficiency evaluation. As shown in Table 2, dINT achieves  $1.93 \times$  and  $2.56 \times$  savings in area and power consumption, respectively. This underscores dINT's effectiveness in tackling underflow issues with minimal output errors and its hardware implementation efficiency.
+
+# 5 Experimental Results
+
+# 5.1 Experimental Settings
+
+In our experimental settings, we implement a comprehensive evaluation to assess the effectiveness of our AQAS, SLAC, and dINT4 techniques in LLMs. This involves conducting quantized inference with 8-bit activations and 4-bit weights across a spectrum of tasks, encompassing language modeling, reasoning, and the MMLU benchmark. To enhance both computational and memory efficiency in activation quantization, we broaden our approach to incorporate the quantization of "Value (for attention map calculation)", which are
+
+<table><tr><td rowspan="2">Weight Scaling</td><td rowspan="2">OPTQ</td><td rowspan="2">A-bits</td><td rowspan="2">W/V-bits</td><td colspan="6">OPT Family</td><td colspan="3">LLaMA Family</td></tr><tr><td>125M</td><td>1.3B</td><td>2.7B</td><td>6.7B</td><td>13B</td><td>30B</td><td>7B</td><td>13B</td><td>30B</td></tr><tr><td colspan="4">Baseline</td><td>31.95</td><td>16.41</td><td>14.32</td><td>12.29</td><td>11.50</td><td>10.67</td><td>5.68</td><td>5.09</td><td>4.10</td></tr><tr><td rowspan="2">-</td><td>-</td><td rowspan="4">INT8</td><td rowspan="4">INT4</td><td>42.78</td><td>35.17</td><td>24.03</td><td>15.52</td><td>36.43</td><td>150.94</td><td>6.87</td><td>6.14</td><td>5.52</td></tr><tr><td>✓</td><td>36.04</td><td>19.35</td><td>15.61</td><td>15.16</td><td>15.82</td><td>25.23</td><td>7.57</td><td>5.85</td><td>5.13</td></tr><tr><td rowspan="2">SQ</td><td>-</td><td>41.31</td><td>23.48</td><td>33.86</td><td>1,596.83</td><td>897.25</td><td>19.43</td><td>7.08</td><td>6.27</td><td>6.17</td></tr><tr><td>✓</td><td>36.22</td><td>17.70</td><td>15.22</td><td>12.82</td><td>11.93</td><td>11.06</td><td>8.26</td><td>5.91</td><td>5.11</td></tr><tr><td rowspan="2">AWQ</td><td>-</td><td></td><td></td><td>44.17</td><td>25.04</td><td>15.88</td><td>322.85</td><td>670.00</td><td>4,246.25</td><td>6.75</td><td>5.84</td><td>5.26</td></tr><tr><td>✓</td><td></td><td></td><td>39.10</td><td>19.35</td><td>16.00</td><td>432.74</td><td>183.83</td><td>4,848.16</td><td>6.52</td><td>5.78</td><td>5.02</td></tr><tr><td rowspan="2">AQAS</td><td>-</td><td></td><td></td><td>36.57</td><td>17.68</td><td>15.34</td><td>13.42</td><td>12.19</td><td>11.08</td><td>6.69</td><td>5.81</td><td>5.14</td></tr><tr><td>✓</td><td></td><td></td><td>35.62</td><td>17.48</td><td>15.08</td><td>12.97</td><td>12.08</td><td>11.04</td><td>6.60</td><td>5.71</td><td>5.07</td></tr><tr><td></td><td>✓</td><td></td><td>dINT4</td><td>34.92</td><td>17.28</td><td>15.03</td><td>12.89</td><td>11.98</td><td>11.04</td><td>6.48</td><td>5.67</td><td>4.72</td></tr></table>
+
+specifically cached to expedite the inference stage during generation (Kwon et al., 2023). We compare our methods against baseline techniques, including weight scaling of SQ (Xiao et al., 2022), AWQ (Lin et al., 2023), and weight update based method, OPTQ (Frantar et al., 2023). Task details, models, calibration methods, and quantization techniques used in the experiments are outlined in Appendix A.1, and an ablation study exploring aspects such as reducing precision to 3-bit, weight-only quantization with dINT, and other 4-bit formats is detailed in Appendix A.6.
+
+# 5.2 Evaluation on Language Modeling Task
+
+We first evaluate perplexity (PPL) as the language modeling performance for various PTQ methods. Table 3 presents W4A8 quantization results with different PTQ combinations. For OPT models, OPTQ generally reduces perplexity. However, when combined with weight scaling methods like SQ and AWQ for 4-bit weight quantization, there's a significant accuracy drop (i.e., spikes in PPL), which OPTQ cannot fully mitigate in most OPT models (except 6.7B and 13B). AQAS effectively curtails the accuracy drop of 4-bit weight quantization, and combining it with OPTQ enhances accuracy. Utilizing dINT4 for 4-bit quantization further lowers perplexity, maintaining a gap of less than 1.0 compared to the full-precision baseline, with the exception of OPT-125M, which is sensitive to quantization. For LLaMA models, OPTQ with 4-bit weight quantization raises perplexity due to increased activation diversity, as discussed in Sec. 3.3. Weight scaling methods like AWQ and AQAS aid in performance recovery, and dINT4
+
+Table 3: PPL results of W4A8V4 (Weight-4bit, Activation-8bit, Value-4bit) at standard language modeling evaluation with OPT and LLaMA family models, applying OPTQ (Frantar et al., 2023) with various weight scaling techniques and two numerical formats.  
+
+<table><tr><td rowspan="2">Weight Scaling</td><td rowspan="2">OPTQ</td><td rowspan="2">W/V-bits</td><td colspan="2">OPT</td><td colspan="2">LLaMA</td></tr><tr><td>6.7B</td><td>13B</td><td>7B</td><td>13B</td></tr><tr><td colspan="3">Baseline</td><td>69.11</td><td>69.38</td><td>70.92</td><td>74.48</td></tr><tr><td rowspan="2">-</td><td>-</td><td rowspan="4">INT4</td><td>63.23</td><td>48.67</td><td>69.37</td><td>72.22</td></tr><tr><td>✓</td><td>64.42</td><td>56.62</td><td>64.94</td><td>71.69</td></tr><tr><td>SQ</td><td>✓</td><td>67.80</td><td>68.96</td><td>68.23</td><td>72.63</td></tr><tr><td rowspan="2">AQAS</td><td>✓</td><td>67.79</td><td>68.80</td><td>69.04</td><td>73.50</td></tr><tr><td>✓</td><td>dINT4</td><td>68.19</td><td>68.96</td><td>68.74</td><td>73.31</td></tr><tr><td rowspan="2">AQAS*</td><td>✓</td><td>INT4</td><td>67.48</td><td>68.24</td><td>70.48</td><td>73.50</td></tr><tr><td>✓</td><td>dINT4</td><td>67.92</td><td>68.80</td><td>71.01</td><td>73.64</td></tr></table>
+
+Table 4: Average accuracy for CommonSense QA (CSQA) tasks including PIQA, Winogrande, and ARCeasy. AQAS* denotes the AQAS method with the SLAC approach (calibrating the dataset's sequence length to 128).
+
+further minimizes perplexity, keeping it within 1.0 of the baseline. We detail the results of applying our proposed AQAS and dINT4 strategies to models with over 60 billion parameters, specifically OPT-66B and LLaMA-65B, in Appendix A.2.
+
+# 5.3 Evaluation on Zero-shot Reasoning Tasks
+
+We carry out experiments for the zero-shot CommonSense QA (CSQA) (Bisk et al., 2019; Sakaguchi et al., 2019; Clark et al., 2018) benchmark by comparing different quantization options, akin to prior tests. As noted in Sec. 3.3, LLaMA models undergo performance decline with OPTQ without weight scaling, whereas OPT models, less affected by activation diversity, show performance gains using OPTQ even without scaling. Among weight scaling techniques, AQAS exhibits superior performance, and employing the dINT4 format further enhances results.
+
+<table><tr><td rowspan="2">Weight Scaling</td><td rowspan="2">OPTQ</td><td rowspan="2">A-bits</td><td rowspan="2">W/V-bits</td><td colspan="4">LLaMA Family</td></tr><tr><td>7B</td><td>13B</td><td>30B</td><td>65B</td></tr><tr><td colspan="4">Baseline</td><td>35.20</td><td>47.15</td><td>58.50</td><td>63.60</td></tr><tr><td rowspan="2">-</td><td>-</td><td rowspan="3">INT8</td><td rowspan="3">INT4</td><td>28.05</td><td>40.82</td><td>48.40</td><td>57.20</td></tr><tr><td>✓</td><td>27.05</td><td>42.95</td><td>53.30</td><td>58.50</td></tr><tr><td>SQ</td><td>✓</td><td>29.32</td><td>43.12</td><td>52.83</td><td>59.30</td></tr><tr><td rowspan="2">AQAS</td><td>✓</td><td></td><td></td><td>30.81</td><td>44.23</td><td>53.67</td><td>59.60</td></tr><tr><td>✓</td><td></td><td>dINT4</td><td>31.00</td><td>44.73</td><td>55.50</td><td>61.40</td></tr></table>
+
+Table 5: Average MMLU accuracy. The detailed accuracy for each item can be found in Table 7.
+
+Due to the shorter input sentences in zero-shot CSQA compared to the default OPTQ calibration dataset, employing SLAC, which considers the LLaMA models' activation diversity based on sequence length, improves performance for both INT4 and dINT4 formats. However, aligning the calibration length with the target task's sequence length for the OPT models does not result in significant improvements. This can be attributed to the OPT models' lower sensitivity to weight updates due to activation diversity during the calibration process, as discussed in Section 3.3, which differs from the behavior of the LLaMA models.
+
+As a result, we attain performance within  $1\%$  of full precision for both OPT and LLaMA models using 8-bit activation and 4-bit weight, notably achieving full precision-equivalent performance in LLaMA-7B by comprehensively accounting for the model's activation characteristics.
+
+# 5.4 Evaluation on In-Context Learning Tasks
+
+We evaluate the MMLU benchmark on several options that exhibited strong performance in language modeling. To assess the efficacy of our proposed method in in-context learning, we conduct 5-shot inference. Given that OPT models are deemed unsuitable for the MMLU benchmark (Lin et al., 2023), we restrict the experiments to LLaMA models. Consistent with language modeling results, AQAS, accounting for both weight and activation quantization errors, delivers the best performance. Moreover, effectively managing underflow error bolsters performance across all models, with a notable  $2\%$  performance enhancement observed in the LLaMA-30B model. To evaluate the efficacy of our approach on large-scale models, we further expand the experiment to LLaMA-65B. The results demonstrate that dINT4 significantly enhances MMLU accuracy by conserving small-magnitude values. Detailed results for each category within MMLU are provided in the Appendix A.3.
+
+# 6 Conclusion
+
+We address Post-training Quantization (PTQ) in Large Language Models (LLMs), specifically targeting 4-bit weight and 8-bit activation (W4A8) quantization to boost computational efficiency. We present Activation-Quantization-Aware Scaling (AQAS) and Sequence-Length-Aware Calibration (SLAC), refining PTQ by taking into account weights and activations, and aligning sequence lengths. To combat the underflow issue in W4A8 quantization, where small magnitudes are rounded down to zero, we introduce dINT, a hybrid format blending integer and denormal representations. Through extensive evaluations on LLMs such as OPT and LLaMA, we demonstrate marked improvements in task accuracy and adaptability. Additionally, with the development of MAC units compatible with dINT, we achieve a twofold increase in hardware efficiency.
+
+# 7 Limitation
+
+We conducted a thorough analysis of model-specific characteristics in LLMs and identified limitations in current PTQ methods. However, further investigation is needed to understand the specific phenomena observed in certain LLM models during the pre-training process. Additionally, exploring more advanced collaborations of PTQ techniques at lower bit precision for weights and activations holds promise for future research.
+
+# Acknowledgement
+
+This work was supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grants funded by the Korea government(MSIT)(2020-0-01305, Development of AI Deep-Learning Processor and Module for 2,000 TFLOPS Server, 2020-0-01297, Development of Ultra-Low Power Deep Learning Processor Technology using Advanced Data Reuse for Edge Applications), the Technology Innovation Program (1415178807, Development of Industrial Intelligent Technology for Manufacturing, Process, and Logistics) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea), and the National Research Foundation of Korea (NRF) grant funded by the Korea government(MSIT)(No. 2021R1A2C1013513).
+
+# References
+
+Michael Andersch, Greg Palmer, Ronny Krashinsky, Nick Stam, Vishal Mehta, Gonzalo Brito, and Sridhar Ramaswamy. 2022. Nvidia hopper architecture in-depth. https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/.  
+Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. 2019. Piqa: Reasoning about physical commonsense in natural language.  
+Yelysei Bondarenko, Markus Nagel, and Tijmen Blankevoort. 2021. Understanding and overcoming the challenges of efficient transformer quantization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7947-7969, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.  
+Brian Chmiel, Ron Banner, Elad Hoffer, Hilla Ben Yaakov, and Daniel Soudry. 2022. Logarithmic unbiased quantization: Simple 4-bit training in deep learning.  
+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, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1.  
+Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. Llm.int8(): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339.  
+Tim Dettmers, Ruslan Svirschevski, Vage Egiazarian, Denis Kuznedelev, Elias Frantar, Saleh Ashkboos, Alexander Borzunov, Torsten Hoefler, and Dan Alistarh. 2023. Spqr: A sparse-quantized representation for near-lossless llm weight compression.  
+Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2022. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323.
+
+Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2023. OPTQ: Accurate quantization for generative pre-trained transformers. In The Eleventh International Conference on Learning Representations.  
+Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027.  
+Sudhanva Gurumurthi, Kijun Lee, Munseon Jang, Vilas Sridharan, Aaron Nygren, Yesin Ryu, Kyomin Sohn, Taekyun Kim, and Hoeju Chung. 2021. Hbm3 ras: Enhancing resilience at scale. IEEE Computer Architecture Letters, 20(2):158-161.  
+Mingxuan He, Choungki Song, Ilkon Kim, Chunseok Jeong, Seho Kim, Il Park, Mithuna Thottethodi, and T. N. Vijaykumar. 2020. Newton: A dram-maker's accelerator-in-memory (aim) architecture for machine learning. In 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 372-385.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding.  
+Mark Horowitz. 2014. Energy table for  $45\mathrm{nm}$  process.  
+Qing Jin, Jian Ren, Richard Zhuang, Sumant Hanumante, Zhengang Li, Zhiyu Chen, Yanzhi Wang, Kaiyuan Yang, and Sergey Tulyakov. 2022. F8net: Fixed-point 8-bit only multiplication for network quantization. In International Conference on Learning Representations.  
+Jin Hyun Kim, Shin-haeng Kang, Sukhan Lee, Hyeonsu Kim, Woongjae Song, Yuhwan Ro, Seungwon Lee, David Wang, Hyunsung Shin, Bengseng Phuah, Ji hyun Choi, Jinin So, YeonGon Cho, JoonHo Song, Jangseok Choi, Jeonghyeon Cho, Kyomin Sohn, Youngsoo Sohn, Kwangil Park, and Nam Sung Kim. 2021. Aquabolt-xl: Samsung hbm2-pim with in-memory processing for ml accelerators and beyond. In 2021 IEEE Hot Chips 33 Symposium (HCS), pages 1-26.  
+Se Jung Kwon, Jeonghoon Kim, Jeongin Bae, Kang Min Yoo, Jin-Hwa Kim, Baeseong Park, Byeongwook Kim, Jung-Woo Ha, Nako Sung, and Dongsoo Lee. 2022. AlphaTuning: Quantization-aware parameter-efficient adaptation of large-scale pre-trained language models. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 3288-3305, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th
+
+Symposium on Operating Systems Principles, SOSP '23, page 611-626, New York, NY, USA. Association for Computing Machinery.  
+Changhun Lee, Jungyu Jin, Taesu Kim, Hyungjun Kim, and Eunhyeok Park. 2023a. Owq: Lessons learned from activation outliers for weight quantization in large language models.  
+Janghwan Lee, Youngdeok Hwang, and Jungwook Choi. 2023b. Finding optimal numerical format for sub-8-bit post-training quantization of vision transformers. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5.  
+Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Xingyu Dang, and Song Han. 2023. Awq: Activation-aware weight quantization for llm compression and acceleration. arXiv.  
+Zechun Liu, Barlas Oguz, Changsheng Zhao, Ernie Chang, Pierre Stock, Yashar Mehdad, Yangyang Shi, Raghuraman Krishnamoorthi, and Vikas Chandra. 2023. Llm-qat: Data-free quantization aware training for large language models.  
+Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models.  
+Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. 2020. Up or down? adaptive rounding for post-training quantization. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org.  
+OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.  
+Gunho Park, Baeseong Park, Minsub Kim, Sungjae Lee, Jeonghoon Kim, Beomseok Kwon, Se Jung Kwon, Byeongwook Kim, Youngjoo Lee, and Dongsoo Lee. 2023. Lut-gemm: Quantized matrix multiplication based on luts for efficient inference in large-scale generative language models.  
+Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Winogrande: An adversarial winograd schema challenge at scale.  
+Xiao Sun, Jungwook Choi, Chia-Yu Chen, Naigang Wang, Swagath Venkataramani, Vijayalakshmi (Viji) Srinivasan, Xiaodong Cui, Wei Zhang, and Kailash Gopalakrishnan. 2019. Hybrid 8-bit floating point (hfp8) training and inference for deep neural networks. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.  
+Xiao Sun, Naigang Wang, Chia-Yu Chen, Jiamin Ni, Ankur Agrawal, Xiaodong Cui, Swagath Venkataramani, Kaoutar El Maghraoui, Vijayalakshmi (Viji) Srinivasan, and Kailash Gopalakrishnan. 2020. Ultra-low precision 4-bit training of deep neural networks. In Advances in Neural Information Processing Systems, volume 33, pages 1796-1807. Curran Associates, Inc.
+
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Guangxuan Xiao, Ji Lin, Mickael Seznec, Julien Demouth, and Song Han. 2022. Smoothquant: Accurate and efficient post-training quantization for large language models. arXiv preprint arXiv:2211.10438.  
+Zhewei Yao, Reza Yazdani Aminabadi, Minjia Zhang, Xiaoxia Wu, Conglong Li, and Yuxiong He. 2022. Zeroquant: Efficient and affordable post-training quantization for large-scale transformers.  
+Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, Wenguang Chen, Zhiyuan Liu, Peng Zhang, Yuxiao Dong, and Jie Tang. 2023. GLM-130b: An open bilingual pre-trained model. In The Eleventh International Conference on Learning Representations.  
+Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona 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 pretrained transformer language models.
+
+# A Appendix
+
+# A.1 Experimental Details
+
+Baseline Setup. As comparative baselines for weight scaling, we employ SQ (Xiao et al., 2022) as the scale method for 8-bit activation quantization and AWQ (Lin et al., 2023) as the scaling method for 4-bit weight quantization. In terms of weight rounding, we evaluate both options provided by OPTQ (Frantar et al., 2023), which offers additional optimization, and the standard nearest rounding method. As for numerical format, we compare dINT with existing INT4 methods in terms of performance.
+
+Task and Models. We evaluate our proposed approach on various tasks, including language modeling using Wikitext (Merit et al., 2016)), CSQA (PIQA (Bisk et al., 2019), WinoGrande (Sakaguchi et al., 2019), and ARC easy (Clark et al., 2018)), as well as MMLU (Hendrycks et al., 2021) benchmarks. For benchmark models, we assess a range of options across OPT and LLaMA models, ranging from 125M to 66B, which are widely used LLMs.
+
+Quantization Settings. We apply quantization to both the weights and activations of all matrix multiplications in the Decoder layer. We conduct our experiments by implementing the quantizer within the PyTorch framework. For activations, except for Value, we apply 8-bit quantization, while for memory-intensive components such as weights and Value, we utilize 4-bit quantization. Similar to commonly used methods for LLM quantization (Dettmers et al., 2022; Yao et al., 2022), we apply token-wise quantization for activations, and output channel-wise quantization for weights. For Value, we apply channel-wise quantization, taking into account the dimensions where partial-sum accumulation occurs when multiplied by the self-attention map. We apply Min-Max asymmetric quantization determine the step size and the zero-point for both activation and weight.
+
+Calibration Settings. During the calibration process to find the weight scale, we follow the calibration setting from the AWQ repository<sup>1</sup>. For the attention operation, we adjust the cali
+
+<table><tr><td>Wikitext PPL</td><td>OPT-66B</td><td>LLaMA-65B</td></tr><tr><td>FP16 baseline</td><td>10.09</td><td>3.53</td></tr><tr><td>INT4</td><td>3,222.02</td><td>4.80</td></tr><tr><td>AQAS + dINT4 + OPTQ</td><td>10.32</td><td>4.13</td></tr></table>
+
+Table 6: Wikitext perplexity for W4A8V4 inference on OPT-66B and LLaMA-65B.
+
+bration process by modifying the objective of Eq. 1 to minimize the distortion of the attention output. We use a randomly extracted dataset from Pile (Gao et al., 2020) for AWQ (Lin et al., 2023), SmoothQuant (Xiao et al., 2022), and AQAS methods. When calibrating weights with OPTQ, we follow the baseline calibration setting provided in the OPTQ repository<sup>2</sup>. We use a subset of the C4 dataset, randomly selecting 128 samples with a sequence length of 2048.
+
+# A.2 Language modeling in  $>60\mathrm{B}$  Models
+
+Table 6 To assess the effectiveness of our approach on large models, we conduct language modeling experiments on OPT-66B and LLaMA-65B, with the objective of determining whether our method performs well even on models with over 60 billion parameters. demonstrates that proposed scaling method and numerical format can significantly reduce perplexity in language modeling task.
+
+# A.3 Few-shot MMLU Benchmarks
+
+The results for the each category in the 5-shot MMLU benchmark for LLaMA models are displayed in Table 7. As demonstrated in Table 7, AQAS exhibits higher accuracy compared to other scaling methods, emphasizing the importance of considering both weight and activation quantization. Furthermore, it's noteworthy that the use of dINT, which effectively mitigates underflow, achieves the highest accuracy.
+
+# A.4 Finding Scales for AQAS
+
+To automatically determine the channel-wise scale factor in AQAS, it is necessary to select representative values for both activation and weight channels. In SQ (Xiao et al., 2022), the maximum magnitude was used as the criterion, while in AWQ (Lin et al., 2023), the absolute mean value was used to explore the scale factor. As shown in Table 8, we explore both cases and found that selecting the
+
+<table><tr><td colspan="9">LLaMA 7B</td></tr><tr><td>Scaling</td><td>OPTQ</td><td>A-bits</td><td>W/V-bits</td><td>Hums</td><td>STEM</td><td>Social</td><td>Other</td><td>Avg.</td></tr><tr><td colspan="4">Baseline</td><td>33.90</td><td>30.50</td><td>38.20</td><td>38.20</td><td>35.20</td></tr><tr><td rowspan="2">-</td><td>-</td><td rowspan="5">INT8</td><td rowspan="4">INT4</td><td>28.65</td><td>25.75</td><td>26.13</td><td>31.15</td><td>28.05</td></tr><tr><td>✓</td><td>26.99</td><td>24.88</td><td>25.71</td><td>30.41</td><td>27.05</td></tr><tr><td>SQ</td><td>✓</td><td>27.80</td><td>29.66</td><td>27.40</td><td>33.04</td><td>29.32</td></tr><tr><td rowspan="2">AQAS</td><td>✓</td><td>29.61</td><td>29.85</td><td>30.65</td><td>33.62</td><td>30.81</td></tr><tr><td>✓</td><td>dINT4</td><td>29.67</td><td>29.26</td><td>32.24</td><td>33.37</td><td>31.00</td></tr></table>
+
+<table><tr><td colspan="8">LLaMA 13B</td><td></td></tr><tr><td>Scaling</td><td>OPTQ</td><td>A-bits</td><td>W/V-bits</td><td>Hums</td><td>STEM</td><td>Social</td><td>Other</td><td>Avg.</td></tr><tr><td colspan="4">Baseline</td><td>44.80</td><td>36.40</td><td>54.20</td><td>53.20</td><td>47.15</td></tr><tr><td rowspan="2">-</td><td>-</td><td rowspan="5">INT8</td><td rowspan="4">INT4</td><td>36.96</td><td>35.16</td><td>45.56</td><td>47.19</td><td>40.82</td></tr><tr><td>✓</td><td>40.38</td><td>34.10</td><td>48.94</td><td>49.23</td><td>42.95</td></tr><tr><td>SQ</td><td>✓</td><td>40.96</td><td>34.26</td><td>48.72</td><td>49.20</td><td>43.12</td></tr><tr><td rowspan="2">AQAS</td><td>✓</td><td>40.74</td><td>35.02</td><td>51.51</td><td>50.96</td><td>44.23</td></tr><tr><td>✓</td><td>dINT4</td><td>42.64</td><td>35.79</td><td>50.83</td><td>50.31</td><td>44.73</td></tr></table>
+
+<table><tr><td colspan="8">LLaMA 30B</td><td></td></tr><tr><td>Scaling</td><td>OPTQ</td><td>A-bits</td><td>W/V-bits</td><td>Hums</td><td>STEM</td><td>Social</td><td>Other</td><td>Avg.</td></tr><tr><td colspan="4">Baseline</td><td>56.40</td><td>46.70</td><td>67.30</td><td>63.60</td><td>58.50</td></tr><tr><td rowspan="2">-</td><td>-</td><td rowspan="5">INT8</td><td rowspan="4">INT4</td><td>46.00</td><td>39.40</td><td>54.50</td><td>54.50</td><td>48.40</td></tr><tr><td>✓</td><td>50.20</td><td>42.60</td><td>61.90</td><td>59.70</td><td>53.30</td></tr><tr><td>SQ</td><td>✓</td><td>49.65</td><td>43.31</td><td>59.86</td><td>59.65</td><td>52.83</td></tr><tr><td rowspan="2">AQAS</td><td>✓</td><td>51.75</td><td>43.21</td><td>60.68</td><td>59.53</td><td>53.67</td></tr><tr><td>✓</td><td>dINT4</td><td>52.90</td><td>44.00</td><td>65.00</td><td>61.10</td><td>55.50</td></tr></table>
+
+<table><tr><td colspan="8">LLaMA 65B</td><td></td></tr><tr><td>Scaling</td><td>OPTQ</td><td>A-bits</td><td>W/V-bits</td><td>Hums</td><td>STEM</td><td>Social</td><td>Other</td><td>Avg.</td></tr><tr><td colspan="4">Baseline</td><td>61.90</td><td>52.10</td><td>73.40</td><td>67.60</td><td>63.60</td></tr><tr><td rowspan="2">-</td><td>-</td><td rowspan="5">INT8</td><td rowspan="4">INT4</td><td>54.10</td><td>46.40</td><td>66.90</td><td>62.50</td><td>57.20</td></tr><tr><td>✓</td><td>56.00</td><td>47.80</td><td>67.20</td><td>63.90</td><td>58.50</td></tr><tr><td>SQ</td><td>✓</td><td>57.70</td><td>47.50</td><td>67.90</td><td>64.30</td><td>59.30</td></tr><tr><td rowspan="2">AQAS</td><td>✓</td><td>57.20</td><td>47.80</td><td>69.50</td><td>64.80</td><td>59.60</td></tr><tr><td>✓</td><td>dINT4</td><td>59.50</td><td>50.40</td><td>70.70</td><td>65.80</td><td>61.40</td></tr></table>
+
+maximum magnitude as the representative value often yielded better performance. Similar to previous research (Lin et al., 2023), we use a grid search to find the appropriate scale, and after determining the scale factor, we make adjustments by additionally clipping the weights.
+
+# A.5 Sweep of the Special Value in dINT
+
+The dINT format defines the special value  $c$  as half of the step size  $s$ . If we change this value, the magnitude of the state representing small values will differ. We conduct additional sweeps with different power-of-two values (e.g., 0.25, 0.125) to observe the impact in Table 9. In most cases, setting  $c$  to 0.25 times the  $s$  proves to be a good choice, but in the case of the OPT-125M model, it shows a significant increase in perplexity. To select a value that generally works well, we set  $c$  to be half of  $s$ .
+
+# A.6 Ablation Study
+
+Reducing Precision to 3 Bits. To achieve additional memory savings, we conduct experiments
+
+Table 7: MMLU accuracy on LLaMA models.  
+
+<table><tr><td colspan="9">Wikitext</td></tr><tr><td>AQAS</td><td>OPTQ</td><td>125M</td><td>1.3B</td><td>OPT 2.7B</td><td>6.7B</td><td>13B</td><td>30B</td><td>LLaMA 7B</td></tr><tr><td colspan="2">Baseline</td><td>31.95</td><td>16.41</td><td>14.32</td><td>12.29</td><td>11.50</td><td>10.67</td><td>5.68</td></tr><tr><td rowspan="2">Mean</td><td>-</td><td>37.46</td><td>18.16</td><td>15.43</td><td>13.08</td><td>12.11</td><td>11.06</td><td>6.72</td></tr><tr><td>✓</td><td>36.07</td><td>17.50</td><td>15.19</td><td>12.99</td><td>12.07</td><td>11.02</td><td>6.67</td></tr><tr><td rowspan="2">Max</td><td>-</td><td>36.57</td><td>17.68</td><td>15.34</td><td>13.42</td><td>12.19</td><td>11.08</td><td>6.69</td></tr><tr><td>✓</td><td>35.62</td><td>17.48</td><td>15.08</td><td>12.97</td><td>12.08</td><td>11.04</td><td>6.60</td></tr></table>
+
+<table><tr><td colspan="9">PIQA</td></tr><tr><td rowspan="2">AQAS</td><td rowspan="2">OPTQ</td><td colspan="6">OPT</td><td>LLaMA</td></tr><tr><td>125M</td><td>1.3B</td><td>2.7B</td><td>6.7B</td><td>13B</td><td>30B</td><td>7B</td></tr><tr><td colspan="2">Baseline</td><td>63.00</td><td>71.71</td><td>73.78</td><td>76.28</td><td>75.90</td><td>77.58</td><td>78.35</td></tr><tr><td rowspan="2">Mean</td><td>-</td><td>61.86</td><td>70.78</td><td>72.91</td><td>75.57</td><td>74.86</td><td>76.66</td><td>77.42</td></tr><tr><td>✓</td><td>61.81</td><td>71.22</td><td>72.85</td><td>75.14</td><td>74.97</td><td>77.20</td><td>77.37</td></tr><tr><td rowspan="2">Max</td><td>-</td><td>61.21</td><td>70.51</td><td>72.69</td><td>74.27</td><td>75.14</td><td>77.15</td><td>77.26</td></tr><tr><td>✓</td><td>62.30</td><td>71.11</td><td>73.56</td><td>75.08</td><td>75.84</td><td>77.48</td><td>76.28</td></tr></table>
+
+<table><tr><td colspan="9">WinoGrande</td></tr><tr><td rowspan="2">AQAS</td><td rowspan="2">OPTQ</td><td colspan="6">OPT</td><td>LLaMA</td></tr><tr><td>125M</td><td>1.3B</td><td>2.7B</td><td>6.7B</td><td>13B</td><td>30B</td><td>7B</td></tr><tr><td colspan="2">Baseline</td><td>50.28</td><td>59.51</td><td>61.01</td><td>65.43</td><td>65.11</td><td>73.01</td><td>67.09</td></tr><tr><td rowspan="2">Mean</td><td>-</td><td>51.54</td><td>58.88</td><td>60.14</td><td>65.59</td><td>65.51</td><td>67.56</td><td>64.96</td></tr><tr><td>✓</td><td>49.96</td><td>57.62</td><td>61.64</td><td>65.82</td><td>64.88</td><td>68.03</td><td>64.80</td></tr><tr><td rowspan="2">Max</td><td>-</td><td>52.41</td><td>60.62</td><td>61.64</td><td>65.27</td><td>64.88</td><td>67.40</td><td>63.77</td></tr><tr><td>✓</td><td>49.72</td><td>58.17</td><td>60.14</td><td>63.77</td><td>65.27</td><td>68.11</td><td>65.59</td></tr></table>
+
+Table 8: Comparing the performance of AQAS when exploring channel-wise quantization using the criteria of absolute mean and max values.  
+
+<table><tr><td rowspan="2">c1/s</td><td colspan="4">OPT</td><td rowspan="2">LLaMA 7B</td><td rowspan="2">Avg.</td></tr><tr><td>125M</td><td>1.3B</td><td>2.7B</td><td>6.7B</td></tr><tr><td>Baseline</td><td>31.95</td><td>16.41</td><td>14.32</td><td>12.29</td><td>5.68</td><td>16.13</td></tr><tr><td>0.50</td><td>34.92</td><td>17.28</td><td>15.03</td><td>12.89</td><td>6.48</td><td>17.32</td></tr><tr><td>0.25</td><td>35.84</td><td>17.22</td><td>14.96</td><td>12.83</td><td>6.28</td><td>17.43</td></tr><tr><td>0.125</td><td>35.20</td><td>17.23</td><td>14.97</td><td>12.84</td><td>6.32</td><td>17.32</td></tr></table>
+
+Table 9: dINT4's special value sweep, W4A8V4 inference with AQAS+OPTQ. Where  $c_{1}$  is the positive special value, and  $s$  is the step size.  
+
+<table><tr><td rowspan="2">Weight scaling</td><td rowspan="2">OPTQ</td><td rowspan="2">W/V-bits</td><td colspan="3">OPT</td><td>LLaMA</td></tr><tr><td>125M</td><td>2.7B</td><td>6.7B</td><td>7B</td></tr><tr><td colspan="3">Baseline</td><td>31.95</td><td>14.32</td><td>12.29</td><td>5.68</td></tr><tr><td rowspan="2">-</td><td rowspan="2">-</td><td>INT3</td><td>1.7e3</td><td>4.3e4</td><td>1.2e4</td><td>94.97</td></tr><tr><td>dINT3</td><td>127.94</td><td>8.7e3</td><td>55.24</td><td>10.99</td></tr><tr><td rowspan="2">AQAS</td><td rowspan="2">✓</td><td>INT3</td><td>54.84</td><td>36.38</td><td>69.45</td><td>24.85</td></tr><tr><td>dINT3</td><td>46.34</td><td>20.67</td><td>17.42</td><td>10.04</td></tr></table>
+
+Table 10: Perplexity is assessed using standard language modeling on the Wikitext dataset, with activations quantized to 8 bits and weights and values to 3 bits. We employ AQAS and OPTQ and compare the performance of INT3 and dINT3. Notably, dINT3 considerably reduces performance degradation.
+
+in which we retain 8-bit activation while reducing weight and Value precision to 3 bits. As shown in Table 10, as bit precision decreases and the impact of underflow becomes more significant, the effectiveness of dINT becomes more pronounced.
+
+<table><tr><td rowspan="2">Precision</td><td rowspan="2">Format</td><td rowspan="2">Method</td><td colspan="5">OPT</td></tr><tr><td>125M</td><td>1.3B</td><td>2.7B</td><td>6.7B</td><td>13B</td></tr><tr><td colspan="3">FP16 baseline</td><td>31.95</td><td>16.41</td><td>14.32</td><td>12.29</td><td>11.50</td></tr><tr><td rowspan="6">W3A16 g=128</td><td rowspan="3">INT3</td><td>RTN</td><td>58.37</td><td>195.10</td><td>499.39</td><td>39.18</td><td>29.37</td></tr><tr><td>OPTQ</td><td>41.93</td><td>18.53</td><td>15.79</td><td>13.13</td><td>12.01</td></tr><tr><td>AWQ</td><td>41.45</td><td>18.56</td><td>15.63</td><td>12.99</td><td>12.03</td></tr><tr><td rowspan="3">dINT3</td><td>RTN</td><td>54.53</td><td>21.70</td><td>24.15</td><td>15.80</td><td>13.14</td></tr><tr><td>OPTQ</td><td>39.60</td><td>18.26</td><td>15.52</td><td>12.99</td><td>11.91</td></tr><tr><td>AWQ</td><td>39.46</td><td>18.29</td><td>15.47</td><td>12.97</td><td>12.03</td></tr></table>
+
+Table 11: Performance comparison of W3A16 inference results with various state-of-the-art methods when applying group-wise quantization (group size: 128).  
+
+<table><tr><td rowspan="2">Precision</td><td rowspan="2">Format</td><td rowspan="2">Method</td><td colspan="5">OPT</td></tr><tr><td>125M</td><td>1.3B</td><td>2.7B</td><td>6.7B</td><td>13B</td></tr><tr><td colspan="3">FP16 baseline</td><td>31.95</td><td>16.41</td><td>14.32</td><td>12.29</td><td>11.50</td></tr><tr><td rowspan="8">W4A16 g=128</td><td rowspan="3">INT4</td><td>RTN</td><td>35.52</td><td>17.69</td><td>15.12</td><td>13.02</td><td>11.89</td></tr><tr><td>OPTQ</td><td>34.23</td><td>16.92</td><td>14.69</td><td>12.51</td><td>11.60</td></tr><tr><td>AWQ</td><td>33.96</td><td>16.85</td><td>14.61</td><td>12.44</td><td>11.60</td></tr><tr><td rowspan="2">FP4</td><td>RTN</td><td>39.13</td><td>18.25</td><td>15.54</td><td>13.35</td><td>12.14</td></tr><tr><td>OPTQ</td><td>37.42</td><td>18.06</td><td>15.19</td><td>12.84</td><td>11.91</td></tr><tr><td rowspan="3">dINT4</td><td>RTN</td><td>34.40</td><td>17.06</td><td>14.83</td><td>12.75</td><td>11.79</td></tr><tr><td>OPTQ</td><td>34.04</td><td>16.82</td><td>14.64</td><td>12.47</td><td>11.59</td></tr><tr><td>AWQ</td><td>33.66</td><td>16.78</td><td>14.56</td><td>12.44</td><td>11.61</td></tr></table>
+
+By solely changing the numerical format without applying weight scaling, we are able to significantly reduce the perplexity of the LLaMA-7B model from 94.97 to 10.99. This underscores the influence of underflow on model performance.
+
+Weight-Only Quantization Method with dINT. dINT, as a numerical format, can be integrated with existing PTQ methods. We combine the dINT format with state-of-the-art PTQ methods for LLMs, namely OPTQ and AWQ, and compare their performance with the integer format. As shown in Table 11 and Table 12, dINT outperforms the traditional integer format in both 3-bit and 4-bit quantization. This indicates that underflow significantly affects the performance of weight quantization in LLMs.
+
+Other 4-Bit Formats. To compare the performance of 4-bit quantization formats, we evaluate performance by applying integer, floating-point, and dINT4 to the weights, without considering activation quantization. We employ a 4-bit floating-point (FP4), consisting of a single sign bit and three exponent bits. While alternative configurations with different exponent and mantissa bits are available, we experimentally determine the necessity of a 3-bit exponent for the FP4. Additional
+
+Table 12: Performance comparison of W4A16 inference results with various state-of-the-art methods when applying group-wise quantization (group size: 128).  
+
+<table><tr><td>Model</td><td>Precision</td><td>W4 format</td><td>Wikitext</td><td>PIQA</td><td>MMLU</td></tr><tr><td rowspan="5">LLaMA-7B</td><td colspan="2">FP16 baseline</td><td>5.68</td><td>78.29</td><td>35.20</td></tr><tr><td rowspan="4">W4A16</td><td>FP4 (1-1-2)</td><td>165582.55</td><td>51.69</td><td>26.88</td></tr><tr><td>FP4 (1-2-1)</td><td>26.52</td><td>62.84</td><td>27.31</td></tr><tr><td>FP4 (1-3-0)</td><td>6.30</td><td>76.77</td><td>31.46</td></tr><tr><td>dINT4</td><td>6.07</td><td>77.91</td><td>32.53</td></tr><tr><td rowspan="5">LLaMA-13B</td><td colspan="2">FP16 baseline</td><td>5.09</td><td>78.78</td><td>47.15</td></tr><tr><td rowspan="4">W4A16</td><td>FP4 (1-1-2)</td><td>74763.98</td><td>52.29</td><td>24.72</td></tr><tr><td>FP4 (1-2-1)</td><td>7.95</td><td>74.65</td><td>31.54</td></tr><tr><td>FP4 (1-3-0)</td><td>5.56</td><td>78.62</td><td>40.76</td></tr><tr><td>dINT4</td><td>5.38</td><td>79.05</td><td>44.35</td></tr><tr><td rowspan="5">LLaMA-30B</td><td colspan="2">FP16 baseline</td><td>4.10</td><td>80.96</td><td>58.50</td></tr><tr><td rowspan="4">W4A16</td><td>FP4 (1-1-2)</td><td>34027.07</td><td>51.52</td><td>25.32</td></tr><tr><td>FP4 (1-2-1)</td><td>9.10</td><td>71.22</td><td>32.05</td></tr><tr><td>FP4 (1-3-0)</td><td>4.57</td><td>79.71</td><td>53.50</td></tr><tr><td>dINT4</td><td>4.36</td><td>80.41</td><td>55.87</td></tr></table>
+
+Table 13: Experiments on various configurations of 4-bit floating-point: FP4 (1-e-m) represents floating-point format with a 1-bit sign bit, e-bit exponent, and m-bit mantissa. We conduct experiments on Wikitext PPL, PIQA accuracy, and MMLU average accuracy. Among FP4 configurations, a 3-bit exponent exhibits the best performance, while dINT surpassing it.
+
+<table><tr><td>Precision</td><td>Format</td><td>OPTQ</td><td>125M</td><td>1.3B</td><td>OPT 2.7B</td><td>6.7B</td><td>13B</td></tr><tr><td colspan="3">FP16 baseline</td><td>31.95</td><td>16.41</td><td>14.32</td><td>12.29</td><td>11.50</td></tr><tr><td rowspan="6">W4A16</td><td rowspan="2">INT4</td><td>-</td><td>43.15</td><td>29.90</td><td>19.70</td><td>14.18</td><td>12.89</td></tr><tr><td>✓</td><td>36.29</td><td>17.68</td><td>15.15</td><td>12.88</td><td>11.73</td></tr><tr><td rowspan="2">FP4</td><td>-</td><td>42.17</td><td>18.52</td><td>16.01</td><td>13.38</td><td>12.33</td></tr><tr><td>✓</td><td>37.52</td><td>18.31</td><td>15.39</td><td>13.18</td><td>11.89</td></tr><tr><td rowspan="2">dINT4</td><td>-</td><td>37.16</td><td>18.10</td><td>15.53</td><td>13.75</td><td>12.06</td></tr><tr><td>✓</td><td>35.10</td><td>17.30</td><td>14.94</td><td>12.65</td><td>11.68</td></tr></table>
+
+Table 14: Comparing the performance of various formats in weight-only quantization for the language modeling task, using both dINT and OPTQ together shows the best performance.
+
+![](images/171b87f66604c5df7d500305477c8bb03ebe2a9b3fdf56236fc270cc739deda2.jpg)  
+Figure 6: The variation in the absolute max values of weights and activations when applying weight scaling in LLaMA-7B.
+
+![](images/ec102faa3053e9a9dfae2c4cc889ea40600cf7b9c11b8e02416632f940fb657e.jpg)
+
+details can be found in Table 13. As shown in Table 14, FP4 achieves some performance improvement compared to uniform quantization due to its wider dynamic range. However, dINT4 outperforms the other two formats by effectively representing a wide range of values with uniform intervals while accurately representing small values. It demonstrates better performance and good compatibility with existing optimization techniques such as OPTQ.

2023/Enhancing Computation Efficiency in Large Language Models through Weight and Activation Quantization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66c28c9133ba879bb9d33f6b6370ca2a366962f4a97df1cbf9f8304e8dabd4fb
+size 921964

2023/Enhancing Generative Retrieval with Reinforcement Learning from Relevance Feedback/8a8868ae-16b0-40bb-b3ea-fdf45defe7f6_content_list.json

@@ -0,0 +1,1576 @@
+[
+    {
+        "type": "text",
+        "text": "Enhancing Generative Retrieval with Reinforcement Learning from Relevance Feedback",
+        "text_level": 1,
+        "bbox": [
+            147,
+            89,
+            850,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yujia Zhou $^{1}$ , Zhicheng Dou $^{2*}$ , Ji-Rong Wen $^{2}$",
+        "bbox": [
+            310,
+            149,
+            690,
+            165
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ School of Information, Renmin University of China",
+        "bbox": [
+            287,
+            166,
+            715,
+            181
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Gaoling School of Artificial Intelligence, Renmin University of China",
+        "bbox": [
+            213,
+            181,
+            786,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{zhouyujia,dou,jrwen}@ruc.edu.cn",
+        "bbox": [
+            339,
+            200,
+            662,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The recent advent of end-to-end generative retrieval marks a significant shift in document retrieval methods, leveraging differentiable search indexes to directly produce relevant document identifiers (docids) in response to a specific query. Nevertheless, this approach faces two fundamental challenges: (i) a discrepancy between the token-level probabilistic optimization and the broader document-level relevance estimation; (ii) an overemphasis on top-1 results at the expense of overall ranking quality. To tackle these challenges, we propose a generative retrieval model with reinforcement learning from relevance feedback, which aims to align token-level docid generation with document-level relevance estimation. The training process incorporates three stages: supervised fine-tuning, relevance reward model training, and reinforced learning-to-rank from relevance feedback. To train a high-quality reward model, we define \"relevance\" under three progressive scenarios, which collectively offer a comprehensive evaluation of the document relevance. Experiments conducted on two benchmark datasets demonstrate the effectiveness of our proposed approach.",
+        "bbox": [
+            144,
+            274,
+            460,
+            643
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            650,
+            258,
+            665
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Information retrieval systems have undergone progressive advancements over the past few decades (Zhu et al., 2023). Historically, the retrieval of candidate documents has been heavily reliant on the use of an inverted index (Robertson and Zaragoza, 2009), a conventional method that employs term-based parsing. With the emergency of neural networks, dense retrieval (Karpukhin et al., 2020; Xiong et al., 2021) is proposed to encode both query and document into dense vectors, so as to measure the similarity in semantic space.",
+        "bbox": [
+            112,
+            674,
+            489,
+            851
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, there has been a growing focus on investigating end-to-end generative retrieval models (Tay et al., 2022; Bevilacqua et al., 2022; Wang",
+        "bbox": [
+            112,
+            852,
+            489,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "et al., 2022; Zhou et al., 2022a). These models supersede the conventional explicit index by implementing a large-scale model known as the differentiable search index (Tay et al., 2022). This innovative approach facilitates end-to-end document retrieval using a seq-to-seq generative model, such as T5 (Raffel et al., 2020). In this framework, the model directly generates the corresponding docid for a given query.",
+        "bbox": [
+            507,
+            252,
+            884,
+            398
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite the notable progress made in generative retrieval models, they still face two major challenges that limit their ability to retrieve relevant documents. First, these models primarily rely on the auto-regressive loss to maximize the token-level generation probability of the ground-truth docid. However, such an optimization goal is not aligned with the primary objective of ranking tasks, which is to estimate document-level relevance. Second, existing approaches assume a deterministic distribution in which all probability mass is allocated to the \"golden\" docid. Such an assumption can lead the model to overemphasize the top-1 generation results, while ignoring the quality of the overall ranking list. In light of these challenges, establishing alignment between generation probabilities and document relevance is a vital necessity in adapting generative models to ranking tasks.",
+        "bbox": [
+            507,
+            400,
+            884,
+            690
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present a Generative Retrieval model with Reinforcement Learning from relevance feedback (GenRRL). Our objective is to align the generative retrieval model, trained specifically on the task of docid generation, with the model responsible for assessing document relevance. The training process comprises three pivotal steps: (1) Supervised fine-tuning. This initial step involves training the model using labeled data to improve its performance in generating docid. (2) Relevance reward model training. The subsequent step focuses on training a relevance reward model, specifically designed to measure the degree of relevance between the generated docids and their correspond",
+        "bbox": [
+            507,
+            694,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author",
+        "bbox": [
+            141,
+            904,
+            287,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12481",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12481-12490 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            152,
+            945,
+            843,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ing queries. We expect the reward model to assign higher scores to docids that are more relevant to the queries. (3) Reinforced learning-to-rank (L2R) from relevance feedback. The final step involves using the trained relevance reward model to guide the generative retrieval system towards generating more relevant docids. To align with the ranking task requirements, we incorporate the idea of L2R into the reinforcement learning procedure. This enables the provision of not only pointwise reward but also pairwise and listwise relevance feedback, thereby enhancing the overall ranking performance.",
+        "bbox": [
+            112,
+            84,
+            492,
+            279
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In order to train a high-quality relevance reward model, it is crucial to define \"relevance\" accurately within the context of document retrieval tasks. In this regard, we focus on three key scenarios that robustly reflect the relationship between queries and documents. Firstly, term-based overlap concerns the presence of identical or similar words within both the query and the document. Techniques like TF-IDF can effectively gauge this overlap. Secondly, semantic similarity extends beyond mere term overlap, delving into the shared meanings or concepts within the query and the document. Dual encoder models are particularly useful in modeling the semantic similarity by encoding the query and the document into dense vectors. Lastly, contextual dependency considers how effectively the document satisfies the information needs expressed in the query, extending beyond the boundaries of specific terms or concepts within the query. Techniques for capturing contextual dependency could include language models, which have demonstrated proficiency in understanding the contextual relationships between two sentences. By considering these three progressive scenarios, we can comprehensively assess the relevance between queries and documents, leading to a more effective reward model.",
+        "bbox": [
+            115,
+            282,
+            490,
+            715
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To assess the effectiveness of our model, we perform comprehensive experiments on widely adopted MS MARCO and NQ document retrieval datasets. Experimental results demonstrate the efficacy of the proposed method, including the diverse relevance annotations and the reinforcement learning process. Our in-depth analysis on different model sizes also reveals the scaling laws of our proposed method.",
+        "bbox": [
+            112,
+            721,
+            489,
+            866
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are summarized as: (1) We employ reinforcement learning from relevance feedback to improve generative retrieval, aiming to",
+        "bbox": [
+            112,
+            871,
+            492,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "align the docid generation process with document relevance assessment. (2) We outline a reinforced training process specifically tailored for the ranking task, which includes relevance reward model training and reinforced learning-to-rank. (3) We provide a classification of \"relevance\" within the context of document retrieval. By considering three progressive aspects, the trained reward model enables a thorough assessment of the relevance between queries and documents.",
+        "bbox": [
+            507,
+            84,
+            885,
+            247
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            260,
+            665,
+            275
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditionally, index-based document retrieval methods encompass both sparse and dense retrieval techniques. Sparse retrieval methods (Dai and Callan, 2019, 2020; Bai et al., 2020; Formal et al., 2021), such as BM25 (Robertson and Zaragoza, 2009), utilize inverted indexes to assess term importance and calculate matching scores between queries and documents. Dense retrieval methods have been employed to overcome the limitations of word mismatching (Gao et al., 2021). These methods convert queries and documents into dense vectors and employ ANNS (Jégou et al., 2011) for efficient vector-based search. The dual encoder approach incorporating pre-trained language models has shown improved performance in dense retrieval. Techniques like hard negative sampling (Xiong et al., 2021; Karpukhin et al., 2020; Zhan et al., 2021; Guu et al., 2020) further enhance effectiveness.",
+        "bbox": [
+            507,
+            288,
+            885,
+            579
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In recent times, generative retrieval has gained significant popularity in the field of information retrieval (IR) tasks. A model-based IR approach, where docids are embedded into the model, was first introduced in (Metzler et al., 2021; Tay et al., 2022). Since then, several other models (Wang et al., 2022; Zhuang et al., 2022; Zhou et al., 2022b) incorporating a query augmentation module have been developed to enhance the performance. Moreover, the model-based IR framework has been extended to knowledge-intensive language tasks, leading to notable improvements in performance (Chen et al., 2022a; Bevilacqua et al., 2022; Chen et al., 2022b). Additionally, a series of studies by Sun et al. (2023); Chen et al. (2023); Ren et al. (2023); Li et al. (2023) explored various methods for representing docids to retain docid semantics.",
+        "bbox": [
+            507,
+            580,
+            885,
+            852
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we focus on aligning the optimizing objective of generative retrieval models with document-level relevance, so as to alleviate the bias against the objective of ranking tasks.",
+        "bbox": [
+            507,
+            854,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12482",
+        "bbox": [
+            477,
+            927,
+            524,
+            941
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/b621dd1f84a6975105ebe2a947af331778c0664ba7fa2f95e045165d30235278.jpg",
+        "image_caption": [
+            "Figure 1: Empirical study on ranking results of generative retrieval and dense retrieval."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            126,
+            83,
+            475,
+            219
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Preliminary",
+        "text_level": 1,
+        "bbox": [
+            114,
+            281,
+            253,
+            299
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we formulate the generative retrieval problem and investigate its limitations.",
+        "bbox": [
+            112,
+            309,
+            485,
+            341
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Task Definition",
+        "text_level": 1,
+        "bbox": [
+            114,
+            353,
+            280,
+            368
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Generative retrieval aims to generate the relevant docids for a given query directly with a seq-to-seq model. Assuming that  $d'$  is the identifier of the document  $d$ , the generative retriever takes the query  $q$  as input and generates the relevant docid with an auto-regressive score, written as:",
+        "bbox": [
+            112,
+            376,
+            487,
+            472
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {s c o r e} (d, q) = G _ {\\phi} \\left(d ^ {\\prime} \\mid q\\right) = \\prod_ {i = 1} G _ {\\phi} \\left(d _ {i} ^ {\\prime} \\mid d _ {<   i} ^ {\\prime}, q\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            485,
+            487,
+            517
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $G_{\\phi}$  is the generative retriever with parameters  $\\phi$ , and  $d_i'$  is the  $i_{th}$  token of the docid. The parameters  $\\phi$  is optimized with the standard seq-to-seq objective. During inference, the model applies constrained beam search to guide the decoder in searching within a limited token space at each step, so as to generate valid docids from a candidate set.",
+        "bbox": [
+            112,
+            530,
+            489,
+            644
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Empirical Study",
+        "text_level": 1,
+        "bbox": [
+            114,
+            655,
+            290,
+            671
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present an empirical study aiming to investigate the limitations of generative retrieval models compared to advanced dense retrieval models in ranking tasks. We choose DPR (Karpukhin et al., 2020) and Ultron (Zhou et al., 2022b) as representatives of the dense and generative retrieval model respectively, and investigate their ranking results on NQ dataset.",
+        "bbox": [
+            112,
+            677,
+            487,
+            804
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 1, it can be observed that: Firstly, the generative retrieval model demonstrates a higher likelihood of placing the correct document at the top position, while the dense retrieval model exhibits a more balanced distribution across all rank positions. Secondly, the relevant documents positioned at the first rank receive remarkably high",
+        "bbox": [
+            112,
+            806,
+            487,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "scores from the generative model, whereas the ranking score curve of the dense retrieval model appears smoother. This outcome is likely due to the discrepancy between the generation probability and the document relevance. To enhance the model ability on document relevance estimation, we propose a reinforcement learning framework for generative retrieval as follows.",
+        "bbox": [
+            507,
+            84,
+            882,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Methodology",
+        "text_level": 1,
+        "bbox": [
+            507,
+            227,
+            658,
+            243
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we focus on the training pipeline of GenRRL, including supervised fine-tuning, relevance reward model training, and reinforced L2R from relevance feedback.",
+        "bbox": [
+            507,
+            254,
+            882,
+            317
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 Supervised Fine-tuning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            332,
+            741,
+            347
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Supervised fine-tuning (SFT) entails training the model to generate relevant docids based on token-level generation probabilities for a given query. Specifically, we choose keyword-based docids (e.g. URL, summary) as the training target, as these can more intuitively reflect the semantics of the document. In addition, we select standard T5 as the pre-trained model and apply pseudo query generation for data augmentation, which has been proven to be effective in improving the performance of generative retrieval (Zhuang et al., 2022). Formally, the generative loss function can be formulated as the maximization of the likelihood of the target sequence using teacher forcing, we have:",
+        "bbox": [
+            507,
+            353,
+            882,
+            579
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {g e n}} = - \\log G _ {\\phi} \\left(d ^ {\\prime} | q\\right) \\\\ = - \\sum_ {i = 1} \\log G _ {\\phi} \\left(d _ {i} ^ {\\prime} \\mid d _ {<   i} ^ {\\prime}, q\\right), \\tag {2} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            573,
+            592,
+            880,
+            644
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $G_{\\phi}(d_i'|d_{<i}',q)$  is the generation probability of token  $d_i'$  based on the given input. After supervised fine-tuning, our next goal is to train a reward model designed to incorporate relevance preferences into docid generation process.",
+        "bbox": [
+            507,
+            658,
+            882,
+            740
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.2 Relevance Reward Model Training",
+        "text_level": 1,
+        "bbox": [
+            507,
+            752,
+            828,
+            768
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A high-quality relevance reward model is capable of effectively assessing the relevance between queries and documents. To accomplish this, we conduct an analysis of relevance judgments within the document retrieval. Subsequently, we employ three distinct methodologies to annotate document relevance, thereby facilitating the training of a comprehensive relevance feedback model. The details are introduced as follows.",
+        "bbox": [
+            507,
+            774,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12483",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/7541b2e9b5952e1db6a7513f63cecf6b1568d08a0b05caaf803014cb2a58c73c.jpg",
+        "image_caption": [
+            "Step 1: Supervised Fine-tuning"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            166,
+            95,
+            321,
+            376
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f05cb9e2ce9e95da1459789a2a39c5ba7a3d5923106939435d43fb90368ccb89.jpg",
+        "image_caption": [
+            "Step 2: Relevance Reward Model Training"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            403,
+            93,
+            611,
+            376
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d0b5ea590df4733edc1f522dfc60c98d00d2e1286f6478e55ced55e45ffe6f72.jpg",
+        "image_caption": [
+            "Figure 2: The training pipeline of GenRRL with supervised fine-tuning, relevance reward model training, and reinforced learning-to-rank from relevance feedback. Blue arrows indicate data used for model training."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            630,
+            77,
+            860,
+            370
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Relevance Classification. We examine three scenarios wherein queries and documents exhibit strong associations. They are:",
+        "bbox": [
+            112,
+            444,
+            487,
+            493
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Term-based Overlap. This scenario refers to the presence of identical or similar terms in both the query and the document. If a document contains many of the same words that are in the query, it is likely to be relevant.",
+            "- Semantic Similarity. This goes beyond term overlap to look at the meaning of the words and phrases in the query and document. Even if a document does not contain the exact terms present in the query, it might still be relevant if it discusses the same concepts. For example, a query for \"cardiovascular fitness exercises\" might retrieve a document about \"aerobic workouts\" because the two phrases refer to similar concepts.",
+            "- Contextual Dependency. This scenario refers to the extent to which a document fulfills the information needs expressed by the query, surpassing the limitations posed by specific terms or concepts. For instance, if a user has been searching for \"how to save battery of laptops\". A relevant document might say: \"reducing screen brightness and turning off keyboard backlight\"."
+        ],
+        "bbox": [
+            114,
+            502,
+            485,
+            860
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Relevance Annotation. For each sampled query  $q$  from supervised fine-tuning dataset, we generate a set of candidate documents based on the fine",
+        "bbox": [
+            112,
+            871,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tuned model, which corresponds to its own beam search results specific to that query. In order to capture the relevance of the above three scenarios respectively, we use three different models to annotate the relevance of the sampled query-doc pairs: (1) To measure the term-based overlap, we leverage BM25 (Robertson and Zaragoza, 2009), a well-established ranking function used by search engines. BM25's strength lies in its ability to weigh query terms based on their TF-IDF importance, thus providing a robust measure of term-based overlap. (2) To capture the semantic similarity, we employ the dual encoder model DPR (Karpukhin et al., 2020), which works by separately encoding the query and document into dense vectors in a shared embedding space. It focuses on deeper semantic connections between words and phrases in the query and the document. The dot product of the two vectors then gives a measure of the semantic similarity. (3) For assessing contextual dependency, we employ the large language model (LLM), LLaMA-13b (Touvron et al., 2023), to compute the generation likelihood of the document given a specific query. We assume that the generation probability can serve as an indicative metric of how well the document fulfills the implicit demands enclosed in the query.",
+        "bbox": [
+            507,
+            444,
+            884,
+            879
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To mitigate the discrepancies arising from variations in scoring distributions across different meth",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12484",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ods, rankings are employed as a means of introducing regularized supervision signals. Specifically, when provided with multiple rankings of the candidate documents, we adopt a voting-based approach, Borda Rank, to consolidate them into a unified ranking. For example, given  $N$  items and  $M$  individual rankings, the Borda rank score  $B_{i}$  for item  $i$  is calculated as  $B_{i} = \\sum_{j=1}^{M} (N - \\mathrm{rank}_{j}(i))$ , where  $\\mathrm{rank}_{j}(i)$  represents the rank of item  $i$  in the  $j$ -th individual ranking. Based on ranking results, we can construct document pairs including a positive document  $d^{+}$  and a negative document  $d^{-}$  for reward model training.",
+        "bbox": [
+            112,
+            84,
+            489,
+            294
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Reward Model training. To train our reward models, we start from the supervised fine-tuned model, and then add a randomly initialized linear head that outputs a scalar reward. The reward model is trained in a pairwise manner to predict which document is more relevant. Given the query  $q$  and a document pair  $(d^{+}, d^{-})$  sampled from the ranking  $D$ , the loss function for the reward model is:",
+        "bbox": [
+            112,
+            299,
+            489,
+            428
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R M}} = \\mathbb {E} _ {(q, d ^ {+}, d ^ {-}) \\sim D} [ \\log (\\sigma (r _ {\\theta} (q, d ^ {+}) - r _ {\\theta} (q, d ^ {-}))) ],\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            436,
+            505,
+            456
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $r_{\\theta}(q,d)$  is the scalar output of the reward model with parameters  $\\theta$",
+        "bbox": [
+            112,
+            461,
+            487,
+            492
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Now we have an initial model capable of docid generation and a preference model that assigns a score to assess the document relevance. Subsequently, reinforcement learning is employed to optimize the initial model based on the reward model.",
+        "bbox": [
+            112,
+            493,
+            489,
+            574
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Reinforced L2R from Relevance Feedback",
+        "text_level": 1,
+        "bbox": [
+            112,
+            583,
+            487,
+            600
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To begin with, we cast the generative retrieval task as a reinforcement learning (RL) problem. Initially, the policy is defined as a language model that accepts a query as input and generates a sequence of tokens as output. The action space of the policy encompasses all tokens in the language model's vocabulary, while the observation space corresponds to the distribution of potential input queries.",
+        "bbox": [
+            112,
+            605,
+            489,
+            733
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Reward Function. To guide the RL training, the reward function is constructed by combining the rewards  $r_{\\theta}(q,d)$  generated by the reward model with a KL divergence constraint, which ensures that the policy does not deviate significantly from its initial behavior, defined as:",
+        "bbox": [
+            112,
+            739,
+            489,
+            835
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR (q, d) = r _ {\\theta} (q, d) - \\beta \\log \\left[ \\pi_ {\\phi} ^ {\\mathrm {R L}} (d | q) / \\pi^ {\\mathrm {S F T}} (d | q) \\right],\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            841,
+            487,
+            864
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\pi_{\\phi}^{\\mathrm{RL}}$  is the learned RL policy,  $\\pi^{\\mathrm{SFT}}$  is the supervised fine-tuned model,  $\\beta$  is The KL reward coefficient controlling the strength of the KL penalty.",
+        "bbox": [
+            112,
+            869,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To align with the requirements of the ranking task, we introduce the concept of learning-to-rank into the RL procedure, employing pointwise, pairwise, and listwise optimization strategies.",
+        "bbox": [
+            507,
+            84,
+            884,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Pointwise Optimization. We begin by independently updating the model for each query-document pair. For a given query  $q$ , the generative retriever functions as a policy network, generating a candidate docid corresponding to the document  $d$ . The RL loss can be formulated as:",
+        "bbox": [
+            507,
+            154,
+            884,
+            250
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R L}} = - (R (q, d) - b) \\sum_ {t} \\log G _ {\\phi} (d _ {t} ^ {\\prime} | d _ {<   t} ^ {\\prime}, q),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            254,
+            867,
+            286
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $b$  is the baseline incorporated to mitigate variance during RL training. It is computed by averaging the rewards of all samples.",
+        "bbox": [
+            507,
+            290,
+            884,
+            338
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Pairwise Optimization. Instead of regarding each query-document pair independently, we incorporate the assessment of pairwise document comparisons as a form of reward. Given a query  $q$ , the docid generation model samples plenty of document pairs  $(d_i, d_j)$ . The generator model is trained with the pairwise loss, denoted as:",
+        "bbox": [
+            507,
+            344,
+            884,
+            456
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R L}} = - \\sum_ {(d _ {i}, d _ {j})} (R (q, d _ {i}) \\log p _ {i j} + R (q, d _ {j}) \\log p _ {j i}),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            507,
+            461,
+            887,
+            498
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $p_{ij}$  is computed by  $|G_{\\phi}(d_i'|q) - G_{\\phi}(d_j'|q)|$  with the normalization of logistic function.",
+        "bbox": [
+            507,
+            502,
+            882,
+            535
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Listwise Optimization. Listwise strategy provides an alternative approach by considering the entire list of generated documents. In this context, we define a listwise loss that aims to minimize the ranking discrepancy between the SFT model and the reward model. Given a set of sampled candidates  $C$  for the query  $q$ , the loss function can be defined as:",
+        "bbox": [
+            507,
+            541,
+            882,
+            668
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R L}} = - \\sum_ {d \\in C} R (q, d) \\log \\frac {\\exp \\left(G _ {\\phi} (d ^ {\\prime} | q)\\right)}{\\sum_ {j} \\exp \\left(G _ {\\phi} (d _ {j} ^ {\\prime} | q)\\right)}.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            517,
+            671,
+            873,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By incorporating listwise optimization, the generative retriever is promoted to generate docids that closely resemble the ordering preferred by the reward model. This approach takes into account the overall quality and relevance of the entire list of generated documents, rather than focusing solely on an individual document or comparisons.",
+        "bbox": [
+            507,
+            719,
+            884,
+            831
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            841,
+            717,
+            859
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Datasets and Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            811,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our experiments are conducted on two widely used document retrieval benchmarks: MS MARCO Doc",
+        "bbox": [
+            507,
+            887,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12485",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "uments. In 2016, we created a document ranking system that uses the same ranking metric as the MRR and HITS (at cutoff 1, 5, 10) as the evaluation metrics.",
+        "bbox": [
+            112,
+            84,
+            492,
+            361
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            370,
+            235,
+            386
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To establish baselines for comparison, we consider three classes of models: (1) Sparse Retrieval models. We select BM25 (Robertson and Zaragoza, 2009), which incorporates tfidf weighting to determine term importance, and DocT5Query (Nogueira et al., 2019), which expands a document with queries predicted by a T5 model, as baselines. (2) Dense Retrieval models. We consider two dual encoder models trained with different hard negatives sampling strategies DPR (Karpukhin et al., 2020) and ANCE (Xiong et al., 2021). (3) Generative Retrieval models: DSI (Tay et al., 2022), the pioneering generative retrieval model that uses a seq-to-seq model to map the query to the relevant docid. Building upon DSI, SEAL (Bevilacqua et al., 2022) uses n-grams within passages as potential identifiers. DSI-QG (Zhuang et al., 2022), NCI (Wang et al., 2022), and Ultron (Zhou et al., 2022b) incorporate query generation modules for data augmentation. To ensure a fair comparison, all baseline models employ the \"base\" version of pre-trained models.",
+        "bbox": [
+            112,
+            394,
+            490,
+            750
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.3 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            762,
+            346,
+            778
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Docid. We explore two types of linguistic docids: document URL and document summary. For the summary-based docids, we employ the LLaMA-13b model to generate document summaries using the prompt \"Summarize the following paragraphs with meaningful keywords: {document}\".",
+        "bbox": [
+            112,
+            784,
+            490,
+            881
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "SFT. Our SFT model utilizes the T5-base pretrained model (Raffel et al., 2020) as the back",
+        "bbox": [
+            112,
+            887,
+            490,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "bone and sets the learning rate and the batch size as 1e-3 and 128 during training. To enrich the training data, we employ the existing DocT5Query model (Nogueira et al., 2019) to generate 10 pseudo queries for each document.",
+        "bbox": [
+            507,
+            84,
+            884,
+            164
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "RL. The reward model is initialized using the SFT model, with a linear head that is randomly initialized. During the training of the reward model, we employ the SFT model to generate 8 candidate docids for relevance annotation. We set the batch size to 64 and utilize a learning rate of 3e-4 for this training phase. In the reinforced L2R process, the policy model provides 1, 2, 4 candidate docids at each step corresponding to pointwise, pairwise, and listwise optimization, while the learning rate is set to 3e-5. All experiments are conducted on up to 8 NVIDIA A100 GPUs.",
+        "bbox": [
+            507,
+            170,
+            885,
+            363
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6 Experimental Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            378,
+            845,
+            395
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present the experimental findings of our proposed approach and perform an empirical analysis to provide comprehensive insights.",
+        "bbox": [
+            507,
+            405,
+            885,
+            454
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.1 Overall Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            467,
+            678,
+            481
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The overall results of models are listed in Table 1. Some findings are summarized as follows.",
+        "bbox": [
+            507,
+            488,
+            885,
+            520
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Comparison with Generative Retrievers. Our observations indicate consistent superior performance of GenRRL compared to baseline models and SFT models when evaluated on the MS MARCO and NQ datasets. This pattern of results serves as compelling evidence for the effectiveness of our reinforced training strategy. Notably, when compared to the leading generative retrieval model Ultron on the MS MARCO dataset, GenRRL (URL) exhibits a noteworthy improvement of over  $9.6\\%$  in terms of MRR@10. Similarly, on the NQ dataset, ROGER (Sum) surpasses the competitive baseline NCI by over  $8.1\\%$  in MRR@10. These findings strongly support the notion that integrating relevance feedback into the initial generative retrieval model holds immense promise for enhancing performance in ranking tasks.",
+        "bbox": [
+            507,
+            527,
+            885,
+            800
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Effect of Different Docids. In the context of generative retrieval, the selection of docids can be categorized into two forms: linguistic docids and ID-based docids. Through a comparative analysis between DSI-QG and Ultron, we have observed that linguistic docids are better suited for generative retrieval tasks. Building upon this insight, we have",
+        "bbox": [
+            507,
+            806,
+            885,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12486",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/dd242337870e2937a572c10d54f475a3f67b955643248c6031dfc83bfa697b7e.jpg",
+        "table_caption": [
+            "Table 1: Overall results. Document Rep. indicates the method to represent document. The best results are shown in bold. “†” indicates the result is significantly better than the corresponding SFT model with paired t-test at  $p < 0.05$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Document Rep.</td><td colspan=\"4\">MS MARCO</td><td colspan=\"4\">Natural Questions</td></tr><tr><td>Hits@1</td><td>Hits@5</td><td>Hits@10</td><td>MRR@10</td><td>Hits@1</td><td>Hits@5</td><td>Hits@10</td><td>MRR@10</td></tr><tr><td colspan=\"10\">Index-based Retrieval</td></tr><tr><td>BM25</td><td>sparse terms</td><td>18.94</td><td>42.82</td><td>55.07</td><td>29.24</td><td>14.06</td><td>36.91</td><td>47.93</td><td>23.60</td></tr><tr><td>DocT5Query</td><td>sparse terms</td><td>23.27</td><td>49.38</td><td>63.61</td><td>34.81</td><td>19.07</td><td>43.88</td><td>55.83</td><td>29.55</td></tr><tr><td>DPR</td><td>dense vector</td><td>29.08</td><td>62.75</td><td>73.13</td><td>43.41</td><td>22.78</td><td>53.44</td><td>68.58</td><td>35.92</td></tr><tr><td>ANCE</td><td>dense vector</td><td>29.65</td><td>63.43</td><td>74.28</td><td>44.09</td><td>24.54</td><td>54.21</td><td>69.08</td><td>36.88</td></tr><tr><td colspan=\"10\">Generative Retrieval</td></tr><tr><td>DSI</td><td>semantic id</td><td>25.74</td><td>43.58</td><td>53.84</td><td>33.92</td><td>27.42</td><td>47.26</td><td>56.58</td><td>34.31</td></tr><tr><td>DSI-QG</td><td>semantic id</td><td>28.82</td><td>50.74</td><td>62.26</td><td>38.45</td><td>30.17</td><td>53.20</td><td>66.37</td><td>38.85</td></tr><tr><td>SEAL</td><td>n-grams</td><td>27.58</td><td>52.47</td><td>61.01</td><td>37.68</td><td>29.30</td><td>54.12</td><td>68.53</td><td>40.34</td></tr><tr><td>NCI</td><td>semantic id</td><td>29.54</td><td>57.99</td><td>67.28</td><td>40.46</td><td>32.69</td><td>55.82</td><td>69.20</td><td>42.84</td></tr><tr><td>Ultron</td><td>title + URL</td><td>29.82</td><td>60.39</td><td>68.31</td><td>42.53</td><td>33.78</td><td>54.20</td><td>67.05</td><td>42.51</td></tr><tr><td>SFT (URL)</td><td>doc URL</td><td>29.90</td><td>60.28</td><td>68.17</td><td>42.52</td><td>33.91</td><td>54.18</td><td>66.92</td><td>42.52</td></tr><tr><td>SFT (Sum)</td><td>doc summary</td><td>30.12</td><td>60.68</td><td>68.71</td><td>42.84</td><td>33.72</td><td>54.71</td><td>67.69</td><td>42.83</td></tr><tr><td>GenRRL (URL)</td><td>doc URL</td><td>\\( 33.01^{\\dagger} \\)</td><td>\\( 63.62^{\\dagger} \\)</td><td>\\( 74.91^{\\dagger} \\)</td><td>\\( 45.93^{\\dagger} \\)</td><td>\\( 35.79^{\\dagger} \\)</td><td>\\( 56.49^{\\dagger} \\)</td><td>\\( 70.96^{\\dagger} \\)</td><td>\\( 45.73^{\\dagger} \\)</td></tr><tr><td>GenRRL (Sum)</td><td>doc summary</td><td>\\( 33.23^{\\dagger} \\)</td><td>\\( 64.48^{\\dagger} \\)</td><td>\\( 75.80^{\\dagger} \\)</td><td>\\( 46.62^{\\dagger} \\)</td><td>\\( 36.32^{\\dagger} \\)</td><td>\\( 57.42^{\\dagger} \\)</td><td>\\( 71.49^{\\dagger} \\)</td><td>\\( 46.31^{\\dagger} \\)</td></tr></table>",
+        "bbox": [
+            114,
+            112,
+            887,
+            375
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/c761f3a84359e99b3df601f313b9c03a793e0eeb3af5030a3487ad37a7330a15.jpg",
+        "table_caption": [
+            "Table 2: The results of different relevance annotators.  $\\bullet$  and  $\\circ$  indicates annotating with and without the model."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">Relevance annotator</td><td colspan=\"2\">MS MARCO</td><td colspan=\"2\">Natural Questions</td></tr><tr><td>Sparse</td><td>Dense</td><td>LLM</td><td>Hits@10</td><td>MRR@10</td><td>Hits@10</td><td>MRR@10</td></tr><tr><td>SFT Model (Sum)</td><td>○</td><td>○</td><td>○</td><td>68.71</td><td>42.84</td><td>67.69</td><td>42.83</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>●</td><td>○</td><td>74.37</td><td>45.85</td><td>70.40</td><td>45.32</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>○</td><td>●</td><td>73.19</td><td>45.18</td><td>69.78</td><td>44.67</td></tr><tr><td>GenRRL (Sum)</td><td>○</td><td>●</td><td>●</td><td>75.36</td><td>46.28</td><td>70.92</td><td>45.81</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>●</td><td>●</td><td>75.80</td><td>46.62</td><td>71.49</td><td>46.31</td></tr></table>",
+        "bbox": [
+            142,
+            405,
+            852,
+            512
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "explored two types of linguistic docids, namely doc URL and doc summary. The experimental results demonstrate that our model consistently outperforms purely supervised fine-tuned models on both types of linguistic docids. Furthermore, the performance is particularly enhanced when utilizing summary-based docids, indicating that docids generated by large language models can better capture document relevance compared to the more simplistic approach of URL-based docids.",
+        "bbox": [
+            112,
+            536,
+            489,
+            697
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2 Ablation Study on Relevance Annotators",
+        "text_level": 1,
+        "bbox": [
+            112,
+            715,
+            480,
+            730
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we perform an ablation study to assess the impact of relevance annotation in the GenRRL model. As described in Section 4.2, the relevance annotation process involves three types of annotators: the sparse model, the dense model, and LLM, each representing distinct relevance scenarios. We explore various combinations of these annotators to evaluate their effectiveness.",
+        "bbox": [
+            112,
+            740,
+            487,
+            866
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2 presents the results for variants of GenRRL. We can observe that each annotator in the relevance annotation process demonstrates a contri",
+        "bbox": [
+            112,
+            870,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "bution to the overall performance, with the dense model exhibiting the most notable improvement in results. This finding suggests that the dense model possesses a more pronounced advantage in assessing the relevance between queries and documents. Additionally, the inclusion of the LLM annotator yields performance enhancements, indicating that leveraging the zero-shot capability of large language models can further augment the model's ability to capture contextual dependencies.",
+        "bbox": [
+            507,
+            536,
+            884,
+            697
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.3 Comparison of Different RL Losses",
+        "text_level": 1,
+        "bbox": [
+            507,
+            715,
+            835,
+            730
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The incorporation of relevance feedback signals with reinforcement learning serves as a crucial step in model fine-tuning. To investigate the impact of different reinforced feedback strategies, we compare three reinforcement learning losses: pointwise, pairwise, and listwise, and examine their influence on ranking performance for both type of docids.",
+        "bbox": [
+            507,
+            740,
+            884,
+            852
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3 presents the results obtained for both URL-based docids and summary-based docids. It is observed that as the loss transitions from pointwise to pairwise and listwise, there is a gradual im",
+        "bbox": [
+            507,
+            854,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12487",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/78dc74786589b29ba78f75dfda4aef45576df1473528dac568da4657623412f6.jpg",
+        "image_caption": [
+            "Figure 3: Performance on different query sets corresponding to three types of relevance scenarios on MS."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            79,
+            473,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/885870e0d7dbf24323c3f2d24c8ad1f4fcf4c36f8b55146af7031f614ef7d04b.jpg",
+        "table_caption": [
+            "Table 3: The MRR@10 of different RL losses."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>RL Loss</td><td>MS</td><td>NQ</td></tr><tr><td>GenRRL (URL)</td><td>Pointwise</td><td>43.42</td><td>43.01</td></tr><tr><td>GenRRL (URL)</td><td>Pairwise</td><td>44.82</td><td>45.10</td></tr><tr><td>GenRRL (URL)</td><td>Listwise</td><td>45.93</td><td>45.73</td></tr><tr><td>GenRRL (Sum)</td><td>Pointwise</td><td>43.78</td><td>43.92</td></tr><tr><td>GenRRL (Sum)</td><td>Pairwise</td><td>45.37</td><td>45.46</td></tr><tr><td>GenRRL (Sum)</td><td>Listwise</td><td>46.62</td><td>46.31</td></tr></table>",
+        "bbox": [
+            115,
+            284,
+            485,
+            392
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "provement on the search results. This progressive enhancement suggests that the listwise loss function could better alleviate the issue of the model's excessive focus on top-1 results. By assuming a non-deterministic distribution, the listwise loss assigns probability mass to other candidate docids based on their relevance.",
+        "bbox": [
+            112,
+            418,
+            487,
+            527
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.4 Performance on Different Query Sets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            542,
+            452,
+            558
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To investigate the distinctions between the three relevance scenarios, we partitioned the test queries into three categories based on the performance of sparse, dense, and generative retrieval. These categories align with the dimensions of term-based overlap, semantic similarity, and contextual dependency, respectively. We then conducted a comparative analysis of the model's performance before and after reinforcement learning.",
+        "bbox": [
+            112,
+            564,
+            487,
+            709
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "From Figure 3, we find that the BM25 and DPR models have demonstrated noticeable advantages on their respective query sets. Although the SFT model exhibited certain gaps in comparison, the introduction of reinforcement learning allowed us to incorporate relevance signals from both the sparse and dense models. Furthermore, the SFT model was augmented by incorporating the knowledge of the LLM. As a result, the retrieval quality across the three query sets experienced improvements. These enhancements signify the effectiveness of reinforcement learning process and diverse relevance signals, leading to enhanced retrieval performance.",
+        "bbox": [
+            112,
+            709,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9b2fcc8a63c9cfc2d36487cc7732130648098b147dea41cda8280e38751545be.jpg",
+        "image_caption": [
+            "(a) Scale of SFT model."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            84,
+            687,
+            193
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5048b70396f706bf8766a5a9dec39fee165825b4892016e28ebc954302a4acf4.jpg",
+        "image_caption": [
+            "(b) Scale of reward model.",
+            "Figure 4: The variation trends on MS with the increase of training step under different model scales."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            85,
+            870,
+            193
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.5 Effect on Model Scale",
+        "text_level": 1,
+        "bbox": [
+            507,
+            275,
+            726,
+            291
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To investigate the training characteristics of models with varying parameter scales, we try T5-small, T5-base and T5-large models for training SFT model and reward model respectively.",
+        "bbox": [
+            507,
+            300,
+            882,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Scale of SFT Model. The learning curve depicted in Figure 4(a) illustrates the convergence behavior of the GenRRL model with varying parameter sizes as the number of training steps increases. It is observed that the small model exhibits comparatively lower MRR at convergence. Conversely, both the base and large models achieve similar performance levels after a sufficient number of training steps, with the large model slightly surpassing the base model. These findings emphasize the critical influence of model capacity on retrieval performance, indicating that larger models have the potential to yield improved results given adequate training.",
+        "bbox": [
+            507,
+            370,
+            882,
+            580
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Scale of Reward Model. Figure 4(b) shows the learning curve with different scale of the reward model. We observe that during the initial 4000 training steps, the differences between reward models of different sizes are not distinct. However, as training progresses, the small reward model exhibits early convergence. In contrast, the base and large reward models continue to improve their performance, with the large reward model outperforming the base model. This finding suggests that the larger reward model demonstrates stronger capability in capturing capture nuanced distinctions among sampled docids, leading to continued improvement.",
+        "bbox": [
+            507,
+            585,
+            882,
+            794
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            810,
+            640,
+            826
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we address the challenges faced by generative retrieval models by proposing an RL-based approach called GenRRL. Our objective is to align the generative retrieval with the task of assessing document relevance, thereby improving",
+        "bbox": [
+            507,
+            839,
+            882,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12488",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ranking performance. We introduce a three-stage reinforcement learning framework, where the relevance reward model captures different dimensions of relevance by considering term-based overlap, semantic similarity, and contextual dependency. Through extensive experiments, we demonstrate the effectiveness of our proposed method in improving generative retrieval performance.",
+        "bbox": [
+            112,
+            84,
+            492,
+            212
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            224,
+            280,
+            242
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by the National Natural Science Foundation of China No. 62272467 and No. 61832017, Beijing Outstanding Young Scientist Program No. BJWZYJH012019100020098, Public Computing Cloud, Renmin University of China, Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE, and Intelligent Social Governance Platform, Major Innovation & Planning Interdisciplinary Platform for the \"Double-First Class\" Initiative, Renmin University of China. The work was partially done at Beijing Key Laboratory of Big Data Management and Analysis Methods.",
+        "bbox": [
+            112,
+            249,
+            489,
+            458
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            469,
+            220,
+            483
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Despite the notable progress achieved within our reinforced framework, GenRRL encounters several challenges that necessitate further exploration and advancement. Primarily, the extension of the model to web scale poses substantial requirements for the design of the differentiable search index, demanding enhanced capacity to handle vast amounts of data. Secondly, the incorporation of new incoming documents into the differentiable search index has not yet been thoroughly investigated. This aspect presents an unexplored avenue that requires careful exploration and analysis. Addressing this challenge entails devising strategies and techniques to seamlessly integrate new documents into the existing model-based framework, while ensuring that the indexer retains its efficiency, accuracy, and robustness.",
+        "bbox": [
+            112,
+            494,
+            489,
+            766
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            794,
+            213,
+            809
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yang Bai, Xiaoguang Li, Gang Wang, Chaoliang Zhang, Lifeng Shang, Jun Xu, Zhaowei Wang, Fangshan Wang, and Qun Liu. 2020. Sparterm: Learning term-based sparse representation for fast text retrieval. CoRR, abs/2010.00768.",
+            "Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen-tau Yih, Sebastian Riedel, and Fabio Petroni."
+        ],
+        "bbox": [
+            115,
+            816,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "2022. Autoregressive search engines: Generating substrings as document identifiers. CoRR, abs/2204.10628.",
+            "Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Yixing Fan, and Xueqi Cheng. 2022a. GERE: generative evidence retrieval for fact verification. In SIGIR, pages 2184-2189. ACM.",
+            "Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Yiqun Liu, Yixing Fan, and Xueqi Cheng. 2022b. Corpusbrain: Pre-train a generative retrieval model for knowledge-intensive language tasks. In CIKM, pages 191-200. ACM.",
+            "Xiaoyang Chen, Yanjiang Liu, Ben He, Le Sun, and Yingfei Sun. 2023. Understanding differential search index for text retrieval. CoRR, abs/2305.02073.",
+            "Zhuyun Dai and Jamie Callan. 2019. Context-aware sentence/passage term importance estimation for first stage retrieval. CoRR, abs/1910.10687.",
+            "Zhuyun Dai and Jamie Callan. 2020. Context-aware document term weighting for ad-hoc search. In WWW '20: The Web Conference 2020, Taipei, Taiwan, April 20-24, 2020, pages 1897-1907. ACM / IW3C2.",
+            "Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and Stephane Clinchant. 2021. SPLADE v2: Sparse lexical and expansion model for information retrieval. CoRR, abs/2109.10086.",
+            "Luyu Gao, Zhuyun Dai, Tongfei Chen, Zhen Fan, Benjamin Van Durme, and Jamie Callan. 2021. Complement lexical retrieval model with semantic residual embeddings. In Advances in Information Retrieval - 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28 - April 1, 2021, Proceedings, Part I, volume 12656 of Lecture Notes in Computer Science, pages 146-160. Springer.",
+            "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. REALM: retrieval-augmented language model pre-training. CoRR, abs/2002.08909.",
+            "Hervé Jégou, Matthijs Douze, and Cordelia Schmid. 2011. Product quantization for nearest neighbor search. IEEE Trans. Pattern Anal. Mach. Intell., 33(1):117-128.",
+            "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick S. H. Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In EMNLP (1), pages 6769-6781. Association for Computational Linguistics.",
+            "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur P. Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob"
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12489",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: a benchmark for question answering research. Trans. Assoc. Comput. Linguistics, 7:452-466.",
+            "Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wenjie Li. 2023. Multiview identifiers enhanced generative retrieval.",
+            "Donald Metzler, Yi Tay, Dara Bahri, and Marc Najork. 2021. Rethinking search: making domain experts out of dilettantes. SIGIR Forum, 55(1):13:1-13:27.",
+            "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A human generated machine reading comprehension dataset. In Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain, December 9, 2016, volume 1773 of CEUR Workshop Proceedings. CEUR-WS.org.",
+            "Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. 2019. From doc2query to doctttttquery. Online preprint, 6.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.",
+            "Ruiyang Ren, Wayne Xin Zhao, Jing Liu, Hua Wu, Ji-Rong Wen, and Haifeng Wang. 2023. TOME: A two-stage approach for model-based retrieval. CoRR, abs/2305.11161.",
+            "Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Found. Trends Inf. Retr., 3(4):333-389.",
+            "Weiwei Sun, Lingyong Yan, Zheng Chen, Shuaiqiang Wang, Haichao Zhu, Pengjie Ren, Zhumin Chen, Dawei Yin, Maarten de Rijke, and Zhaochun Ren. 2023. Learning to tokenize for generative retrieval. CoRR, abs/2304.04171.",
+            "Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Prakash Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            741
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "memory as a differentiable search index. CoRR, abs/2202.06991.",
+            "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971.",
+            "Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Hao Sun, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, Xing Xie, Hao Allen Sun, Weiwei Deng, Qi Zhang, and Mao Yang. 2022. A neural corpus indexer for document retrieval. CoRR, abs/2206.02743.",
+            "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In ICLR. OpenReview.net.",
+            "Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In SIGIR '21: The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, Canada, July 11-15, 2021, pages 1503-1512. ACM.",
+            "Yujia Zhou, Jing Yao, Zhicheng Dou, Ledell Wu, and JiRong Wen. 2022a. Dynamicretriever: A pre-training model-based IR system with neither sparse nor dense index. CoRR, abs/2203.00537.",
+            "Yujia Zhou, Jing Yao, Zhicheng Dou, Ledell Wu, Peitian Zhang, and Ji-Rong Wen. 2022b. Ultron: An ultimate retriever on corpus with a model-based indexer. CoRR, abs/2208.09257.",
+            "Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. CoRR, abs/2308.07107.",
+            "Shengyao Zhuang, Houxing Ren, Linjun Shou, Jian Pei, Ming Gong, Guido Zuccon, and Daxin Jiang. 2022. Bridging the gap between indexing and retrieval for differentiable search index with query generation. CoRR, abs/2206.10128."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            741
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12490",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    }
+]

2023/Enhancing Generative Retrieval with Reinforcement Learning from Relevance Feedback/8a8868ae-16b0-40bb-b3ea-fdf45defe7f6_model.json

@@ -0,0 +1,1991 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.149,
+                0.09,
+                0.852,
+                0.13
+            ],
+            "angle": 0,
+            "content": "Enhancing Generative Retrieval with Reinforcement Learning from Relevance Feedback"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.312,
+                0.15,
+                0.692,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Yujia Zhou\\(^{1}\\), Zhicheng Dou\\(^{2*}\\), Ji-Rong Wen\\(^{2}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.289,
+                0.167,
+                0.717,
+                0.183
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)School of Information, Renmin University of China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.215,
+                0.183,
+                0.788,
+                0.2
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)Gaoling School of Artificial Intelligence, Renmin University of China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.34,
+                0.201,
+                0.663,
+                0.216
+            ],
+            "angle": 0,
+            "content": "{zhouyujia,dou,jrwen}@ruc.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.275,
+                0.461,
+                0.644
+            ],
+            "angle": 0,
+            "content": "The recent advent of end-to-end generative retrieval marks a significant shift in document retrieval methods, leveraging differentiable search indexes to directly produce relevant document identifiers (docids) in response to a specific query. Nevertheless, this approach faces two fundamental challenges: (i) a discrepancy between the token-level probabilistic optimization and the broader document-level relevance estimation; (ii) an overemphasis on top-1 results at the expense of overall ranking quality. To tackle these challenges, we propose a generative retrieval model with reinforcement learning from relevance feedback, which aims to align token-level docid generation with document-level relevance estimation. The training process incorporates three stages: supervised fine-tuning, relevance reward model training, and reinforced learning-to-rank from relevance feedback. To train a high-quality reward model, we define \"relevance\" under three progressive scenarios, which collectively offer a comprehensive evaluation of the document relevance. Experiments conducted on two benchmark datasets demonstrate the effectiveness of our proposed approach."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.651,
+                0.26,
+                0.666
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.675,
+                0.49,
+                0.852
+            ],
+            "angle": 0,
+            "content": "Information retrieval systems have undergone progressive advancements over the past few decades (Zhu et al., 2023). Historically, the retrieval of candidate documents has been heavily reliant on the use of an inverted index (Robertson and Zaragoza, 2009), a conventional method that employs term-based parsing. With the emergency of neural networks, dense retrieval (Karpukhin et al., 2020; Xiong et al., 2021) is proposed to encode both query and document into dense vectors, so as to measure the similarity in semantic space."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.853,
+                0.49,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Recently, there has been a growing focus on investigating end-to-end generative retrieval models (Tay et al., 2022; Bevilacqua et al., 2022; Wang"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.253,
+                0.885,
+                0.399
+            ],
+            "angle": 0,
+            "content": "et al., 2022; Zhou et al., 2022a). These models supersede the conventional explicit index by implementing a large-scale model known as the differentiable search index (Tay et al., 2022). This innovative approach facilitates end-to-end document retrieval using a seq-to-seq generative model, such as T5 (Raffel et al., 2020). In this framework, the model directly generates the corresponding docid for a given query."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.401,
+                0.885,
+                0.691
+            ],
+            "angle": 0,
+            "content": "Despite the notable progress made in generative retrieval models, they still face two major challenges that limit their ability to retrieve relevant documents. First, these models primarily rely on the auto-regressive loss to maximize the token-level generation probability of the ground-truth docid. However, such an optimization goal is not aligned with the primary objective of ranking tasks, which is to estimate document-level relevance. Second, existing approaches assume a deterministic distribution in which all probability mass is allocated to the \"golden\" docid. Such an assumption can lead the model to overemphasize the top-1 generation results, while ignoring the quality of the overall ranking list. In light of these challenges, establishing alignment between generation probabilities and document relevance is a vital necessity in adapting generative models to ranking tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.695,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In this paper, we present a Generative Retrieval model with Reinforcement Learning from relevance feedback (GenRRL). Our objective is to align the generative retrieval model, trained specifically on the task of docid generation, with the model responsible for assessing document relevance. The training process comprises three pivotal steps: (1) Supervised fine-tuning. This initial step involves training the model using labeled data to improve its performance in generating docid. (2) Relevance reward model training. The subsequent step focuses on training a relevance reward model, specifically designed to measure the degree of relevance between the generated docids and their correspond"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.142,
+                0.905,
+                0.288,
+                0.919
+            ],
+            "angle": 0,
+            "content": "*Corresponding author"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.477,
+                0.928,
+                0.524,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12481"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.153,
+                0.946,
+                0.845,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12481-12490 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.28
+            ],
+            "angle": 0,
+            "content": "ing queries. We expect the reward model to assign higher scores to docids that are more relevant to the queries. (3) Reinforced learning-to-rank (L2R) from relevance feedback. The final step involves using the trained relevance reward model to guide the generative retrieval system towards generating more relevant docids. To align with the ranking task requirements, we incorporate the idea of L2R into the reinforcement learning procedure. This enables the provision of not only pointwise reward but also pairwise and listwise relevance feedback, thereby enhancing the overall ranking performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.284,
+                0.492,
+                0.716
+            ],
+            "angle": 0,
+            "content": "In order to train a high-quality relevance reward model, it is crucial to define \"relevance\" accurately within the context of document retrieval tasks. In this regard, we focus on three key scenarios that robustly reflect the relationship between queries and documents. Firstly, term-based overlap concerns the presence of identical or similar words within both the query and the document. Techniques like TF-IDF can effectively gauge this overlap. Secondly, semantic similarity extends beyond mere term overlap, delving into the shared meanings or concepts within the query and the document. Dual encoder models are particularly useful in modeling the semantic similarity by encoding the query and the document into dense vectors. Lastly, contextual dependency considers how effectively the document satisfies the information needs expressed in the query, extending beyond the boundaries of specific terms or concepts within the query. Techniques for capturing contextual dependency could include language models, which have demonstrated proficiency in understanding the contextual relationships between two sentences. By considering these three progressive scenarios, we can comprehensively assess the relevance between queries and documents, leading to a more effective reward model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.722,
+                0.49,
+                0.867
+            ],
+            "angle": 0,
+            "content": "To assess the effectiveness of our model, we perform comprehensive experiments on widely adopted MS MARCO and NQ document retrieval datasets. Experimental results demonstrate the efficacy of the proposed method, including the diverse relevance annotations and the reinforcement learning process. Our in-depth analysis on different model sizes also reveals the scaling laws of our proposed method."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.872,
+                0.493,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Our contributions are summarized as: (1) We employ reinforcement learning from relevance feedback to improve generative retrieval, aiming to"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.887,
+                0.248
+            ],
+            "angle": 0,
+            "content": "align the docid generation process with document relevance assessment. (2) We outline a reinforced training process specifically tailored for the ranking task, which includes relevance reward model training and reinforced learning-to-rank. (3) We provide a classification of \"relevance\" within the context of document retrieval. By considering three progressive aspects, the trained reward model enables a thorough assessment of the relevance between queries and documents."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.261,
+                0.667,
+                0.277
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.289,
+                0.886,
+                0.58
+            ],
+            "angle": 0,
+            "content": "Traditionally, index-based document retrieval methods encompass both sparse and dense retrieval techniques. Sparse retrieval methods (Dai and Callan, 2019, 2020; Bai et al., 2020; Formal et al., 2021), such as BM25 (Robertson and Zaragoza, 2009), utilize inverted indexes to assess term importance and calculate matching scores between queries and documents. Dense retrieval methods have been employed to overcome the limitations of word mismatching (Gao et al., 2021). These methods convert queries and documents into dense vectors and employ ANNS (Jégou et al., 2011) for efficient vector-based search. The dual encoder approach incorporating pre-trained language models has shown improved performance in dense retrieval. Techniques like hard negative sampling (Xiong et al., 2021; Karpukhin et al., 2020; Zhan et al., 2021; Guu et al., 2020) further enhance effectiveness."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.581,
+                0.886,
+                0.853
+            ],
+            "angle": 0,
+            "content": "In recent times, generative retrieval has gained significant popularity in the field of information retrieval (IR) tasks. A model-based IR approach, where docids are embedded into the model, was first introduced in (Metzler et al., 2021; Tay et al., 2022). Since then, several other models (Wang et al., 2022; Zhuang et al., 2022; Zhou et al., 2022b) incorporating a query augmentation module have been developed to enhance the performance. Moreover, the model-based IR framework has been extended to knowledge-intensive language tasks, leading to notable improvements in performance (Chen et al., 2022a; Bevilacqua et al., 2022; Chen et al., 2022b). Additionally, a series of studies by Sun et al. (2023); Chen et al. (2023); Ren et al. (2023); Li et al. (2023) explored various methods for representing docids to retain docid semantics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.856,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In this paper, we focus on aligning the optimizing objective of generative retrieval models with document-level relevance, so as to alleviate the bias against the objective of ranking tasks."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.942
+            ],
+            "angle": 0,
+            "content": "12482"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.127,
+                0.084,
+                0.476,
+                0.221
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.228,
+                0.49,
+                0.259
+            ],
+            "angle": 0,
+            "content": "Figure 1: Empirical study on ranking results of generative retrieval and dense retrieval."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.282,
+                0.254,
+                0.3
+            ],
+            "angle": 0,
+            "content": "3 Preliminary"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.31,
+                0.487,
+                0.342
+            ],
+            "angle": 0,
+            "content": "In this section, we formulate the generative retrieval problem and investigate its limitations."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.354,
+                0.281,
+                0.369
+            ],
+            "angle": 0,
+            "content": "3.1 Task Definition"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.377,
+                0.489,
+                0.473
+            ],
+            "angle": 0,
+            "content": "Generative retrieval aims to generate the relevant docids for a given query directly with a seq-to-seq model. Assuming that \\( d' \\) is the identifier of the document \\( d \\), the generative retriever takes the query \\( q \\) as input and generates the relevant docid with an auto-regressive score, written as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.124,
+                0.486,
+                0.489,
+                0.518
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {s c o r e} (d, q) = G _ {\\phi} \\left(d ^ {\\prime} \\mid q\\right) = \\prod_ {i = 1} G _ {\\phi} \\left(d _ {i} ^ {\\prime} \\mid d _ {<   i} ^ {\\prime}, q\\right), \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.531,
+                0.49,
+                0.645
+            ],
+            "angle": 0,
+            "content": "where \\( G_{\\phi} \\) is the generative retriever with parameters \\( \\phi \\), and \\( d_i' \\) is the \\( i_{th} \\) token of the docid. The parameters \\( \\phi \\) is optimized with the standard seq-to-seq objective. During inference, the model applies constrained beam search to guide the decoder in searching within a limited token space at each step, so as to generate valid docids from a candidate set."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.656,
+                0.291,
+                0.672
+            ],
+            "angle": 0,
+            "content": "3.2 Empirical Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.678,
+                0.489,
+                0.806
+            ],
+            "angle": 0,
+            "content": "In this section, we present an empirical study aiming to investigate the limitations of generative retrieval models compared to advanced dense retrieval models in ranking tasks. We choose DPR (Karpukhin et al., 2020) and Ultron (Zhou et al., 2022b) as representatives of the dense and generative retrieval model respectively, and investigate their ranking results on NQ dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.807,
+                0.489,
+                0.921
+            ],
+            "angle": 0,
+            "content": "As shown in Figure 1, it can be observed that: Firstly, the generative retrieval model demonstrates a higher likelihood of placing the correct document at the top position, while the dense retrieval model exhibits a more balanced distribution across all rank positions. Secondly, the relevant documents positioned at the first rank receive remarkably high"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.883,
+                0.213
+            ],
+            "angle": 0,
+            "content": "scores from the generative model, whereas the ranking score curve of the dense retrieval model appears smoother. This outcome is likely due to the discrepancy between the generation probability and the document relevance. To enhance the model ability on document relevance estimation, we propose a reinforcement learning framework for generative retrieval as follows."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.228,
+                0.659,
+                0.244
+            ],
+            "angle": 0,
+            "content": "4 Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.255,
+                0.884,
+                0.318
+            ],
+            "angle": 0,
+            "content": "In this section, we focus on the training pipeline of GenRRL, including supervised fine-tuning, relevance reward model training, and reinforced L2R from relevance feedback."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.333,
+                0.742,
+                0.348
+            ],
+            "angle": 0,
+            "content": "4.1 Supervised Fine-tuning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.354,
+                0.884,
+                0.58
+            ],
+            "angle": 0,
+            "content": "Supervised fine-tuning (SFT) entails training the model to generate relevant docids based on token-level generation probabilities for a given query. Specifically, we choose keyword-based docids (e.g. URL, summary) as the training target, as these can more intuitively reflect the semantics of the document. In addition, we select standard T5 as the pre-trained model and apply pseudo query generation for data augmentation, which has been proven to be effective in improving the performance of generative retrieval (Zhuang et al., 2022). Formally, the generative loss function can be formulated as the maximization of the likelihood of the target sequence using teacher forcing, we have:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.574,
+                0.593,
+                0.882,
+                0.645
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {g e n}} = - \\log G _ {\\phi} \\left(d ^ {\\prime} | q\\right) \\\\ = - \\sum_ {i = 1} \\log G _ {\\phi} \\left(d _ {i} ^ {\\prime} \\mid d _ {<   i} ^ {\\prime}, q\\right), \\tag {2} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.659,
+                0.883,
+                0.741
+            ],
+            "angle": 0,
+            "content": "where \\( G_{\\phi}(d_i'|d_{<i}',q) \\) is the generation probability of token \\( d_i' \\) based on the given input. After supervised fine-tuning, our next goal is to train a reward model designed to incorporate relevance preferences into docid generation process."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.753,
+                0.83,
+                0.769
+            ],
+            "angle": 0,
+            "content": "4.2 Relevance Reward Model Training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.775,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "A high-quality relevance reward model is capable of effectively assessing the relevance between queries and documents. To accomplish this, we conduct an analysis of relevance judgments within the document retrieval. Subsequently, we employ three distinct methodologies to annotate document relevance, thereby facilitating the training of a comprehensive relevance feedback model. The details are introduced as follows."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.525,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12483"
+        }
+    ],
+    [
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.167,
+                0.078,
+                0.32,
+                0.09
+            ],
+            "angle": 0,
+            "content": "Step 1: Supervised Fine-tuning"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.167,
+                0.096,
+                0.322,
+                0.377
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.399,
+                0.079,
+                0.596,
+                0.09
+            ],
+            "angle": 0,
+            "content": "Step 2: Relevance Reward Model Training"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.404,
+                0.095,
+                0.613,
+                0.377
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.631,
+                0.078,
+                0.862,
+                0.371
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.391,
+                0.883,
+                0.421
+            ],
+            "angle": 0,
+            "content": "Figure 2: The training pipeline of GenRRL with supervised fine-tuning, relevance reward model training, and reinforced learning-to-rank from relevance feedback. Blue arrows indicate data used for model training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.445,
+                0.489,
+                0.494
+            ],
+            "angle": 0,
+            "content": "Relevance Classification. We examine three scenarios wherein queries and documents exhibit strong associations. They are:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.503,
+                0.487,
+                0.582
+            ],
+            "angle": 0,
+            "content": "- Term-based Overlap. This scenario refers to the presence of identical or similar terms in both the query and the document. If a document contains many of the same words that are in the query, it is likely to be relevant."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.586,
+                0.487,
+                0.731
+            ],
+            "angle": 0,
+            "content": "- Semantic Similarity. This goes beyond term overlap to look at the meaning of the words and phrases in the query and document. Even if a document does not contain the exact terms present in the query, it might still be relevant if it discusses the same concepts. For example, a query for \"cardiovascular fitness exercises\" might retrieve a document about \"aerobic workouts\" because the two phrases refer to similar concepts."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.733,
+                0.487,
+                0.862
+            ],
+            "angle": 0,
+            "content": "- Contextual Dependency. This scenario refers to the extent to which a document fulfills the information needs expressed by the query, surpassing the limitations posed by specific terms or concepts. For instance, if a user has been searching for \"how to save battery of laptops\". A relevant document might say: \"reducing screen brightness and turning off keyboard backlight\"."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.115,
+                0.503,
+                0.487,
+                0.862
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.872,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Relevance Annotation. For each sampled query \\( q \\) from supervised fine-tuning dataset, we generate a set of candidate documents based on the fine"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.445,
+                0.885,
+                0.881
+            ],
+            "angle": 0,
+            "content": "tuned model, which corresponds to its own beam search results specific to that query. In order to capture the relevance of the above three scenarios respectively, we use three different models to annotate the relevance of the sampled query-doc pairs: (1) To measure the term-based overlap, we leverage BM25 (Robertson and Zaragoza, 2009), a well-established ranking function used by search engines. BM25's strength lies in its ability to weigh query terms based on their TF-IDF importance, thus providing a robust measure of term-based overlap. (2) To capture the semantic similarity, we employ the dual encoder model DPR (Karpukhin et al., 2020), which works by separately encoding the query and document into dense vectors in a shared embedding space. It focuses on deeper semantic connections between words and phrases in the query and the document. The dot product of the two vectors then gives a measure of the semantic similarity. (3) For assessing contextual dependency, we employ the large language model (LLM), LLaMA-13b (Touvron et al., 2023), to compute the generation likelihood of the document given a specific query. We assume that the generation probability can serve as an indicative metric of how well the document fulfills the implicit demands enclosed in the query."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.888,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "To mitigate the discrepancies arising from variations in scoring distributions across different meth"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12484"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.295
+            ],
+            "angle": 0,
+            "content": "ods, rankings are employed as a means of introducing regularized supervision signals. Specifically, when provided with multiple rankings of the candidate documents, we adopt a voting-based approach, Borda Rank, to consolidate them into a unified ranking. For example, given \\( N \\) items and \\( M \\) individual rankings, the Borda rank score \\( B_{i} \\) for item \\( i \\) is calculated as \\( B_{i} = \\sum_{j=1}^{M} (N - \\mathrm{rank}_{j}(i)) \\), where \\( \\mathrm{rank}_{j}(i) \\) represents the rank of item \\( i \\) in the \\( j \\)-th individual ranking. Based on ranking results, we can construct document pairs including a positive document \\( d^{+} \\) and a negative document \\( d^{-} \\) for reward model training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.3,
+                0.49,
+                0.429
+            ],
+            "angle": 0,
+            "content": "Reward Model training. To train our reward models, we start from the supervised fine-tuned model, and then add a randomly initialized linear head that outputs a scalar reward. The reward model is trained in a pairwise manner to predict which document is more relevant. Given the query \\( q \\) and a document pair \\( (d^{+}, d^{-}) \\) sampled from the ranking \\( D \\), the loss function for the reward model is:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.114,
+                0.437,
+                0.506,
+                0.457
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {R M}} = \\mathbb {E} _ {(q, d ^ {+}, d ^ {-}) \\sim D} [ \\log (\\sigma (r _ {\\theta} (q, d ^ {+}) - r _ {\\theta} (q, d ^ {-}))) ],\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.462,
+                0.489,
+                0.493
+            ],
+            "angle": 0,
+            "content": "where \\(r_{\\theta}(q,d)\\) is the scalar output of the reward model with parameters \\(\\theta\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.494,
+                0.49,
+                0.575
+            ],
+            "angle": 0,
+            "content": "Now we have an initial model capable of docid generation and a preference model that assigns a score to assess the document relevance. Subsequently, reinforcement learning is employed to optimize the initial model based on the reward model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.585,
+                0.488,
+                0.601
+            ],
+            "angle": 0,
+            "content": "4.3 Reinforced L2R from Relevance Feedback"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.606,
+                0.49,
+                0.734
+            ],
+            "angle": 0,
+            "content": "To begin with, we cast the generative retrieval task as a reinforcement learning (RL) problem. Initially, the policy is defined as a language model that accepts a query as input and generates a sequence of tokens as output. The action space of the policy encompasses all tokens in the language model's vocabulary, while the observation space corresponds to the distribution of potential input queries."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.74,
+                0.49,
+                0.837
+            ],
+            "angle": 0,
+            "content": "Reward Function. To guide the RL training, the reward function is constructed by combining the rewards \\( r_{\\theta}(q,d) \\) generated by the reward model with a KL divergence constraint, which ensures that the policy does not deviate significantly from its initial behavior, defined as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.114,
+                0.843,
+                0.488,
+                0.865
+            ],
+            "angle": 0,
+            "content": "\\[\nR (q, d) = r _ {\\theta} (q, d) - \\beta \\log \\left[ \\pi_ {\\phi} ^ {\\mathrm {R L}} (d | q) / \\pi^ {\\mathrm {S F T}} (d | q) \\right],\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.87,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "where \\(\\pi_{\\phi}^{\\mathrm{RL}}\\) is the learned RL policy, \\(\\pi^{\\mathrm{SFT}}\\) is the supervised fine-tuned model, \\(\\beta\\) is The KL reward coefficient controlling the strength of the KL penalty."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.15
+            ],
+            "angle": 0,
+            "content": "To align with the requirements of the ranking task, we introduce the concept of learning-to-rank into the RL procedure, employing pointwise, pairwise, and listwise optimization strategies."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.155,
+                0.885,
+                0.251
+            ],
+            "angle": 0,
+            "content": "Pointwise Optimization. We begin by independently updating the model for each query-document pair. For a given query \\( q \\), the generative retriever functions as a policy network, generating a candidate docid corresponding to the document \\( d \\). The RL loss can be formulated as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.525,
+                0.255,
+                0.868,
+                0.287
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {R L}} = - (R (q, d) - b) \\sum_ {t} \\log G _ {\\phi} (d _ {t} ^ {\\prime} | d _ {<   t} ^ {\\prime}, q),\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.291,
+                0.885,
+                0.34
+            ],
+            "angle": 0,
+            "content": "where \\( b \\) is the baseline incorporated to mitigate variance during RL training. It is computed by averaging the rewards of all samples."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.345,
+                0.885,
+                0.457
+            ],
+            "angle": 0,
+            "content": "Pairwise Optimization. Instead of regarding each query-document pair independently, we incorporate the assessment of pairwise document comparisons as a form of reward. Given a query \\( q \\), the docid generation model samples plenty of document pairs \\( (d_i, d_j) \\). The generator model is trained with the pairwise loss, denoted as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.509,
+                0.462,
+                0.888,
+                0.499
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {R L}} = - \\sum_ {(d _ {i}, d _ {j})} (R (q, d _ {i}) \\log p _ {i j} + R (q, d _ {j}) \\log p _ {j i}),\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.503,
+                0.884,
+                0.536
+            ],
+            "angle": 0,
+            "content": "where \\(p_{ij}\\) is computed by \\(|G_{\\phi}(d_i'|q) - G_{\\phi}(d_j'|q)|\\) with the normalization of logistic function."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.542,
+                0.884,
+                0.669
+            ],
+            "angle": 0,
+            "content": "Listwise Optimization. Listwise strategy provides an alternative approach by considering the entire list of generated documents. In this context, we define a listwise loss that aims to minimize the ranking discrepancy between the SFT model and the reward model. Given a set of sampled candidates \\( C \\) for the query \\( q \\), the loss function can be defined as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.518,
+                0.672,
+                0.875,
+                0.717
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {R L}} = - \\sum_ {d \\in C} R (q, d) \\log \\frac {\\exp \\left(G _ {\\phi} (d ^ {\\prime} | q)\\right)}{\\sum_ {j} \\exp \\left(G _ {\\phi} (d _ {j} ^ {\\prime} | q)\\right)}.\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.72,
+                0.885,
+                0.832
+            ],
+            "angle": 0,
+            "content": "By incorporating listwise optimization, the generative retriever is promoted to generate docids that closely resemble the ordering preferred by the reward model. This approach takes into account the overall quality and relevance of the entire list of generated documents, rather than focusing solely on an individual document or comparisons."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.843,
+                0.719,
+                0.86
+            ],
+            "angle": 0,
+            "content": "5 Experimental Setup"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.867,
+                0.813,
+                0.882
+            ],
+            "angle": 0,
+            "content": "5.1 Datasets and Evaluation Metrics"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.888,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Our experiments are conducted on two widely used document retrieval benchmarks: MS MARCO Doc"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12485"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.362
+            ],
+            "angle": 0,
+            "content": "uments. In 2016, we created a document ranking system that uses the same ranking metric as the MRR and HITS (at cutoff 1, 5, 10) as the evaluation metrics."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.372,
+                0.236,
+                0.387
+            ],
+            "angle": 0,
+            "content": "5.2 Baselines"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.395,
+                0.492,
+                0.751
+            ],
+            "angle": 0,
+            "content": "To establish baselines for comparison, we consider three classes of models: (1) Sparse Retrieval models. We select BM25 (Robertson and Zaragoza, 2009), which incorporates tfidf weighting to determine term importance, and DocT5Query (Nogueira et al., 2019), which expands a document with queries predicted by a T5 model, as baselines. (2) Dense Retrieval models. We consider two dual encoder models trained with different hard negatives sampling strategies DPR (Karpukhin et al., 2020) and ANCE (Xiong et al., 2021). (3) Generative Retrieval models: DSI (Tay et al., 2022), the pioneering generative retrieval model that uses a seq-to-seq model to map the query to the relevant docid. Building upon DSI, SEAL (Bevilacqua et al., 2022) uses n-grams within passages as potential identifiers. DSI-QG (Zhuang et al., 2022), NCI (Wang et al., 2022), and Ultron (Zhou et al., 2022b) incorporate query generation modules for data augmentation. To ensure a fair comparison, all baseline models employ the \"base\" version of pre-trained models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.763,
+                0.347,
+                0.779
+            ],
+            "angle": 0,
+            "content": "5.3 Implementation Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.785,
+                0.491,
+                0.882
+            ],
+            "angle": 0,
+            "content": "Docid. We explore two types of linguistic docids: document URL and document summary. For the summary-based docids, we employ the LLaMA-13b model to generate document summaries using the prompt \"Summarize the following paragraphs with meaningful keywords: {document}\"."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.888,
+                0.492,
+                0.919
+            ],
+            "angle": 0,
+            "content": "SFT. Our SFT model utilizes the T5-base pretrained model (Raffel et al., 2020) as the back"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.165
+            ],
+            "angle": 0,
+            "content": "bone and sets the learning rate and the batch size as 1e-3 and 128 during training. To enrich the training data, we employ the existing DocT5Query model (Nogueira et al., 2019) to generate 10 pseudo queries for each document."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.171,
+                0.887,
+                0.364
+            ],
+            "angle": 0,
+            "content": "RL. The reward model is initialized using the SFT model, with a linear head that is randomly initialized. During the training of the reward model, we employ the SFT model to generate 8 candidate docids for relevance annotation. We set the batch size to 64 and utilize a learning rate of 3e-4 for this training phase. In the reinforced L2R process, the policy model provides 1, 2, 4 candidate docids at each step corresponding to pointwise, pairwise, and listwise optimization, while the learning rate is set to 3e-5. All experiments are conducted on up to 8 NVIDIA A100 GPUs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.379,
+                0.847,
+                0.396
+            ],
+            "angle": 0,
+            "content": "6 Experimental Results and Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.406,
+                0.886,
+                0.455
+            ],
+            "angle": 0,
+            "content": "In this section, we present the experimental findings of our proposed approach and perform an empirical analysis to provide comprehensive insights."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.468,
+                0.679,
+                0.482
+            ],
+            "angle": 0,
+            "content": "6.1 Overall Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.489,
+                0.886,
+                0.521
+            ],
+            "angle": 0,
+            "content": "The overall results of models are listed in Table 1. Some findings are summarized as follows."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.528,
+                0.886,
+                0.801
+            ],
+            "angle": 0,
+            "content": "Comparison with Generative Retrievers. Our observations indicate consistent superior performance of GenRRL compared to baseline models and SFT models when evaluated on the MS MARCO and NQ datasets. This pattern of results serves as compelling evidence for the effectiveness of our reinforced training strategy. Notably, when compared to the leading generative retrieval model Ultron on the MS MARCO dataset, GenRRL (URL) exhibits a noteworthy improvement of over \\(9.6\\%\\) in terms of MRR@10. Similarly, on the NQ dataset, ROGER (Sum) surpasses the competitive baseline NCI by over \\(8.1\\%\\) in MRR@10. These findings strongly support the notion that integrating relevance feedback into the initial generative retrieval model holds immense promise for enhancing performance in ranking tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Effect of Different Docids. In the context of generative retrieval, the selection of docids can be categorized into two forms: linguistic docids and ID-based docids. Through a comparative analysis between DSI-QG and Ultron, we have observed that linguistic docids are better suited for generative retrieval tasks. Building upon this insight, we have"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12486"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.083,
+                0.884,
+                0.111
+            ],
+            "angle": 0,
+            "content": "Table 1: Overall results. Document Rep. indicates the method to represent document. The best results are shown in bold. “†” indicates the result is significantly better than the corresponding SFT model with paired t-test at \\( p < 0.05 \\)."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.115,
+                0.113,
+                0.889,
+                0.376
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Document Rep.</td><td colspan=\"4\">MS MARCO</td><td colspan=\"4\">Natural Questions</td></tr><tr><td>Hits@1</td><td>Hits@5</td><td>Hits@10</td><td>MRR@10</td><td>Hits@1</td><td>Hits@5</td><td>Hits@10</td><td>MRR@10</td></tr><tr><td colspan=\"10\">Index-based Retrieval</td></tr><tr><td>BM25</td><td>sparse terms</td><td>18.94</td><td>42.82</td><td>55.07</td><td>29.24</td><td>14.06</td><td>36.91</td><td>47.93</td><td>23.60</td></tr><tr><td>DocT5Query</td><td>sparse terms</td><td>23.27</td><td>49.38</td><td>63.61</td><td>34.81</td><td>19.07</td><td>43.88</td><td>55.83</td><td>29.55</td></tr><tr><td>DPR</td><td>dense vector</td><td>29.08</td><td>62.75</td><td>73.13</td><td>43.41</td><td>22.78</td><td>53.44</td><td>68.58</td><td>35.92</td></tr><tr><td>ANCE</td><td>dense vector</td><td>29.65</td><td>63.43</td><td>74.28</td><td>44.09</td><td>24.54</td><td>54.21</td><td>69.08</td><td>36.88</td></tr><tr><td colspan=\"10\">Generative Retrieval</td></tr><tr><td>DSI</td><td>semantic id</td><td>25.74</td><td>43.58</td><td>53.84</td><td>33.92</td><td>27.42</td><td>47.26</td><td>56.58</td><td>34.31</td></tr><tr><td>DSI-QG</td><td>semantic id</td><td>28.82</td><td>50.74</td><td>62.26</td><td>38.45</td><td>30.17</td><td>53.20</td><td>66.37</td><td>38.85</td></tr><tr><td>SEAL</td><td>n-grams</td><td>27.58</td><td>52.47</td><td>61.01</td><td>37.68</td><td>29.30</td><td>54.12</td><td>68.53</td><td>40.34</td></tr><tr><td>NCI</td><td>semantic id</td><td>29.54</td><td>57.99</td><td>67.28</td><td>40.46</td><td>32.69</td><td>55.82</td><td>69.20</td><td>42.84</td></tr><tr><td>Ultron</td><td>title + URL</td><td>29.82</td><td>60.39</td><td>68.31</td><td>42.53</td><td>33.78</td><td>54.20</td><td>67.05</td><td>42.51</td></tr><tr><td>SFT (URL)</td><td>doc URL</td><td>29.90</td><td>60.28</td><td>68.17</td><td>42.52</td><td>33.91</td><td>54.18</td><td>66.92</td><td>42.52</td></tr><tr><td>SFT (Sum)</td><td>doc summary</td><td>30.12</td><td>60.68</td><td>68.71</td><td>42.84</td><td>33.72</td><td>54.71</td><td>67.69</td><td>42.83</td></tr><tr><td>GenRRL (URL)</td><td>doc URL</td><td>\\( 33.01^{\\dagger} \\)</td><td>\\( 63.62^{\\dagger} \\)</td><td>\\( 74.91^{\\dagger} \\)</td><td>\\( 45.93^{\\dagger} \\)</td><td>\\( 35.79^{\\dagger} \\)</td><td>\\( 56.49^{\\dagger} \\)</td><td>\\( 70.96^{\\dagger} \\)</td><td>\\( 45.73^{\\dagger} \\)</td></tr><tr><td>GenRRL (Sum)</td><td>doc summary</td><td>\\( 33.23^{\\dagger} \\)</td><td>\\( 64.48^{\\dagger} \\)</td><td>\\( 75.80^{\\dagger} \\)</td><td>\\( 46.62^{\\dagger} \\)</td><td>\\( 36.32^{\\dagger} \\)</td><td>\\( 57.42^{\\dagger} \\)</td><td>\\( 71.49^{\\dagger} \\)</td><td>\\( 46.31^{\\dagger} \\)</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.124,
+                0.389,
+                0.873,
+                0.404
+            ],
+            "angle": 0,
+            "content": "Table 2: The results of different relevance annotators. \\( \\bullet \\) and \\( \\circ \\) indicates annotating with and without the model."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.143,
+                0.406,
+                0.853,
+                0.513
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">Relevance annotator</td><td colspan=\"2\">MS MARCO</td><td colspan=\"2\">Natural Questions</td></tr><tr><td>Sparse</td><td>Dense</td><td>LLM</td><td>Hits@10</td><td>MRR@10</td><td>Hits@10</td><td>MRR@10</td></tr><tr><td>SFT Model (Sum)</td><td>○</td><td>○</td><td>○</td><td>68.71</td><td>42.84</td><td>67.69</td><td>42.83</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>●</td><td>○</td><td>74.37</td><td>45.85</td><td>70.40</td><td>45.32</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>○</td><td>●</td><td>73.19</td><td>45.18</td><td>69.78</td><td>44.67</td></tr><tr><td>GenRRL (Sum)</td><td>○</td><td>●</td><td>●</td><td>75.36</td><td>46.28</td><td>70.92</td><td>45.81</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>●</td><td>●</td><td>75.80</td><td>46.62</td><td>71.49</td><td>46.31</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.537,
+                0.49,
+                0.698
+            ],
+            "angle": 0,
+            "content": "explored two types of linguistic docids, namely doc URL and doc summary. The experimental results demonstrate that our model consistently outperforms purely supervised fine-tuned models on both types of linguistic docids. Furthermore, the performance is particularly enhanced when utilizing summary-based docids, indicating that docids generated by large language models can better capture document relevance compared to the more simplistic approach of URL-based docids."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.716,
+                0.482,
+                0.731
+            ],
+            "angle": 0,
+            "content": "6.2 Ablation Study on Relevance Annotators"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.741,
+                0.489,
+                0.868
+            ],
+            "angle": 0,
+            "content": "In this section, we perform an ablation study to assess the impact of relevance annotation in the GenRRL model. As described in Section 4.2, the relevance annotation process involves three types of annotators: the sparse model, the dense model, and LLM, each representing distinct relevance scenarios. We explore various combinations of these annotators to evaluate their effectiveness."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.871,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Table 2 presents the results for variants of GenRRL. We can observe that each annotator in the relevance annotation process demonstrates a contri"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.537,
+                0.885,
+                0.698
+            ],
+            "angle": 0,
+            "content": "bution to the overall performance, with the dense model exhibiting the most notable improvement in results. This finding suggests that the dense model possesses a more pronounced advantage in assessing the relevance between queries and documents. Additionally, the inclusion of the LLM annotator yields performance enhancements, indicating that leveraging the zero-shot capability of large language models can further augment the model's ability to capture contextual dependencies."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.716,
+                0.836,
+                0.731
+            ],
+            "angle": 0,
+            "content": "6.3 Comparison of Different RL Losses"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.741,
+                0.885,
+                0.853
+            ],
+            "angle": 0,
+            "content": "The incorporation of relevance feedback signals with reinforcement learning serves as a crucial step in model fine-tuning. To investigate the impact of different reinforced feedback strategies, we compare three reinforcement learning losses: pointwise, pairwise, and listwise, and examine their influence on ranking performance for both type of docids."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.855,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Table 3 presents the results obtained for both URL-based docids and summary-based docids. It is observed that as the loss transitions from pointwise to pairwise and listwise, there is a gradual im"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12487"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.125,
+                0.08,
+                0.475,
+                0.219
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.225,
+                0.49,
+                0.254
+            ],
+            "angle": 0,
+            "content": "Figure 3: Performance on different query sets corresponding to three types of relevance scenarios on MS."
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.142,
+                0.268,
+                0.458,
+                0.282
+            ],
+            "angle": 0,
+            "content": "Table 3: The MRR@10 of different RL losses."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.285,
+                0.487,
+                0.393
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>RL Loss</td><td>MS</td><td>NQ</td></tr><tr><td>GenRRL (URL)</td><td>Pointwise</td><td>43.42</td><td>43.01</td></tr><tr><td>GenRRL (URL)</td><td>Pairwise</td><td>44.82</td><td>45.10</td></tr><tr><td>GenRRL (URL)</td><td>Listwise</td><td>45.93</td><td>45.73</td></tr><tr><td>GenRRL (Sum)</td><td>Pointwise</td><td>43.78</td><td>43.92</td></tr><tr><td>GenRRL (Sum)</td><td>Pairwise</td><td>45.37</td><td>45.46</td></tr><tr><td>GenRRL (Sum)</td><td>Listwise</td><td>46.62</td><td>46.31</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.419,
+                0.489,
+                0.529
+            ],
+            "angle": 0,
+            "content": "provement on the search results. This progressive enhancement suggests that the listwise loss function could better alleviate the issue of the model's excessive focus on top-1 results. By assuming a non-deterministic distribution, the listwise loss assigns probability mass to other candidate docids based on their relevance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.543,
+                0.453,
+                0.559
+            ],
+            "angle": 0,
+            "content": "6.4 Performance on Different Query Sets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.565,
+                0.489,
+                0.71
+            ],
+            "angle": 0,
+            "content": "To investigate the distinctions between the three relevance scenarios, we partitioned the test queries into three categories based on the performance of sparse, dense, and generative retrieval. These categories align with the dimensions of term-based overlap, semantic similarity, and contextual dependency, respectively. We then conducted a comparative analysis of the model's performance before and after reinforcement learning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.71,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "From Figure 3, we find that the BM25 and DPR models have demonstrated noticeable advantages on their respective query sets. Although the SFT model exhibited certain gaps in comparison, the introduction of reinforcement learning allowed us to incorporate relevance signals from both the sparse and dense models. Furthermore, the SFT model was augmented by incorporating the knowledge of the LLM. As a result, the retrieval quality across the three query sets experienced improvements. These enhancements signify the effectiveness of reinforcement learning process and diverse relevance signals, leading to enhanced retrieval performance."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.085,
+                0.689,
+                0.194
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.529,
+                0.202,
+                0.676,
+                0.215
+            ],
+            "angle": 0,
+            "content": "(a) Scale of SFT model."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.702,
+                0.086,
+                0.872,
+                0.194
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.705,
+                0.202,
+                0.868,
+                0.215
+            ],
+            "angle": 0,
+            "content": "(b) Scale of reward model."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.221,
+                0.882,
+                0.25
+            ],
+            "angle": 0,
+            "content": "Figure 4: The variation trends on MS with the increase of training step under different model scales."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.277,
+                0.727,
+                0.292
+            ],
+            "angle": 0,
+            "content": "6.5 Effect on Model Scale"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.301,
+                0.884,
+                0.365
+            ],
+            "angle": 0,
+            "content": "To investigate the training characteristics of models with varying parameter scales, we try T5-small, T5-base and T5-large models for training SFT model and reward model respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.371,
+                0.884,
+                0.581
+            ],
+            "angle": 0,
+            "content": "Scale of SFT Model. The learning curve depicted in Figure 4(a) illustrates the convergence behavior of the GenRRL model with varying parameter sizes as the number of training steps increases. It is observed that the small model exhibits comparatively lower MRR at convergence. Conversely, both the base and large models achieve similar performance levels after a sufficient number of training steps, with the large model slightly surpassing the base model. These findings emphasize the critical influence of model capacity on retrieval performance, indicating that larger models have the potential to yield improved results given adequate training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.586,
+                0.884,
+                0.795
+            ],
+            "angle": 0,
+            "content": "Scale of Reward Model. Figure 4(b) shows the learning curve with different scale of the reward model. We observe that during the initial 4000 training steps, the differences between reward models of different sizes are not distinct. However, as training progresses, the small reward model exhibits early convergence. In contrast, the base and large reward models continue to improve their performance, with the large reward model outperforming the base model. This finding suggests that the larger reward model demonstrates stronger capability in capturing capture nuanced distinctions among sampled docids, leading to continued improvement."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.811,
+                0.642,
+                0.827
+            ],
+            "angle": 0,
+            "content": "7 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.884,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In this paper, we address the challenges faced by generative retrieval models by proposing an RL-based approach called GenRRL. Our objective is to align the generative retrieval with the task of assessing document relevance, thereby improving"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12488"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.214
+            ],
+            "angle": 0,
+            "content": "ranking performance. We introduce a three-stage reinforcement learning framework, where the relevance reward model captures different dimensions of relevance by considering term-based overlap, semantic similarity, and contextual dependency. Through extensive experiments, we demonstrate the effectiveness of our proposed method in improving generative retrieval performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.225,
+                0.281,
+                0.243
+            ],
+            "angle": 0,
+            "content": "Acknowledgement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.25,
+                0.49,
+                0.459
+            ],
+            "angle": 0,
+            "content": "This work was supported by the National Natural Science Foundation of China No. 62272467 and No. 61832017, Beijing Outstanding Young Scientist Program No. BJWZYJH012019100020098, Public Computing Cloud, Renmin University of China, Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE, and Intelligent Social Governance Platform, Major Innovation & Planning Interdisciplinary Platform for the \"Double-First Class\" Initiative, Renmin University of China. The work was partially done at Beijing Key Laboratory of Big Data Management and Analysis Methods."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.47,
+                0.221,
+                0.485
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.495,
+                0.49,
+                0.768
+            ],
+            "angle": 0,
+            "content": "Despite the notable progress achieved within our reinforced framework, GenRRL encounters several challenges that necessitate further exploration and advancement. Primarily, the extension of the model to web scale poses substantial requirements for the design of the differentiable search index, demanding enhanced capacity to handle vast amounts of data. Secondly, the incorporation of new incoming documents into the differentiable search index has not yet been thoroughly investigated. This aspect presents an unexplored avenue that requires careful exploration and analysis. Addressing this challenge entails devising strategies and techniques to seamlessly integrate new documents into the existing model-based framework, while ensuring that the indexer retains its efficiency, accuracy, and robustness."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.795,
+                0.214,
+                0.81
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.817,
+                0.49,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Yang Bai, Xiaoguang Li, Gang Wang, Chaoliang Zhang, Lifeng Shang, Jun Xu, Zhaowei Wang, Fangshan Wang, and Qun Liu. 2020. Sparterm: Learning term-based sparse representation for fast text retrieval. CoRR, abs/2010.00768."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.892,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen-tau Yih, Sebastian Riedel, and Fabio Petroni."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.817,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.884,
+                0.125
+            ],
+            "angle": 0,
+            "content": "2022. Autoregressive search engines: Generating substrings as document identifiers. CoRR, abs/2204.10628."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.137,
+                0.885,
+                0.189
+            ],
+            "angle": 0,
+            "content": "Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Yixing Fan, and Xueqi Cheng. 2022a. GERE: generative evidence retrieval for fact verification. In SIGIR, pages 2184-2189. ACM."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.2,
+                0.885,
+                0.266
+            ],
+            "angle": 0,
+            "content": "Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Yiqun Liu, Yixing Fan, and Xueqi Cheng. 2022b. Corpusbrain: Pre-train a generative retrieval model for knowledge-intensive language tasks. In CIKM, pages 191-200. ACM."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.277,
+                0.885,
+                0.318
+            ],
+            "angle": 0,
+            "content": "Xiaoyang Chen, Yanjiang Liu, Ben He, Le Sun, and Yingfei Sun. 2023. Understanding differential search index for text retrieval. CoRR, abs/2305.02073."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.328,
+                0.885,
+                0.369
+            ],
+            "angle": 0,
+            "content": "Zhuyun Dai and Jamie Callan. 2019. Context-aware sentence/passage term importance estimation for first stage retrieval. CoRR, abs/1910.10687."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.379,
+                0.885,
+                0.444
+            ],
+            "angle": 0,
+            "content": "Zhuyun Dai and Jamie Callan. 2020. Context-aware document term weighting for ad-hoc search. In WWW '20: The Web Conference 2020, Taipei, Taiwan, April 20-24, 2020, pages 1897-1907. ACM / IW3C2."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.456,
+                0.885,
+                0.509
+            ],
+            "angle": 0,
+            "content": "Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and Stephane Clinchant. 2021. SPLADE v2: Sparse lexical and expansion model for information retrieval. CoRR, abs/2109.10086."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.52,
+                0.885,
+                0.626
+            ],
+            "angle": 0,
+            "content": "Luyu Gao, Zhuyun Dai, Tongfei Chen, Zhen Fan, Benjamin Van Durme, and Jamie Callan. 2021. Complement lexical retrieval model with semantic residual embeddings. In Advances in Information Retrieval - 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28 - April 1, 2021, Proceedings, Part I, volume 12656 of Lecture Notes in Computer Science, pages 146-160. Springer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.635,
+                0.885,
+                0.688
+            ],
+            "angle": 0,
+            "content": "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. REALM: retrieval-augmented language model pre-training. CoRR, abs/2002.08909."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.699,
+                0.885,
+                0.752
+            ],
+            "angle": 0,
+            "content": "Hervé Jégou, Matthijs Douze, and Cordelia Schmid. 2011. Product quantization for nearest neighbor search. IEEE Trans. Pattern Anal. Mach. Intell., 33(1):117-128."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.763,
+                0.885,
+                0.843
+            ],
+            "angle": 0,
+            "content": "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick S. H. Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In EMNLP (1), pages 6769-6781. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.852,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur P. Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12489"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.135,
+                0.086,
+                0.49,
+                0.138
+            ],
+            "angle": 0,
+            "content": "Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: a benchmark for question answering research. Trans. Assoc. Comput. Linguistics, 7:452-466."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.149,
+                0.49,
+                0.188
+            ],
+            "angle": 0,
+            "content": "Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wenjie Li. 2023. Multiview identifiers enhanced generative retrieval."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.199,
+                0.489,
+                0.239
+            ],
+            "angle": 0,
+            "content": "Donald Metzler, Yi Tay, Dara Bahri, and Marc Najork. 2021. Rethinking search: making domain experts out of dilettantes. SIGIR Forum, 55(1):13:1-13:27."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.248,
+                0.49,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A human generated machine reading comprehension dataset. In Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain, December 9, 2016, volume 1773 of CEUR Workshop Proceedings. CEUR-WS.org."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.389,
+                0.49,
+                0.417
+            ],
+            "angle": 0,
+            "content": "Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. 2019. From doc2query to doctttttquery. Online preprint, 6."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.426,
+                0.489,
+                0.492
+            ],
+            "angle": 0,
+            "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.502,
+                0.489,
+                0.554
+            ],
+            "angle": 0,
+            "content": "Ruiyang Ren, Wayne Xin Zhao, Jing Liu, Hua Wu, Ji-Rong Wen, and Haifeng Wang. 2023. TOME: A two-stage approach for model-based retrieval. CoRR, abs/2305.11161."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.565,
+                0.489,
+                0.606
+            ],
+            "angle": 0,
+            "content": "Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Found. Trends Inf. Retr., 3(4):333-389."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.615,
+                0.489,
+                0.68
+            ],
+            "angle": 0,
+            "content": "Weiwei Sun, Lingyong Yan, Zheng Chen, Shuaiqiang Wang, Haichao Zhu, Pengjie Ren, Zhumin Chen, Dawei Yin, Maarten de Rijke, and Zhaochun Ren. 2023. Learning to tokenize for generative retrieval. CoRR, abs/2304.04171."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.69,
+                0.49,
+                0.743
+            ],
+            "angle": 0,
+            "content": "Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Prakash Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.743
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.53,
+                0.086,
+                0.883,
+                0.112
+            ],
+            "angle": 0,
+            "content": "memory as a differentiable search index. CoRR, abs/2202.06991."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.123,
+                0.883,
+                0.213
+            ],
+            "angle": 0,
+            "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.224,
+                0.883,
+                0.303
+            ],
+            "angle": 0,
+            "content": "Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Hao Sun, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, Xing Xie, Hao Allen Sun, Weiwei Deng, Qi Zhang, and Mao Yang. 2022. A neural corpus indexer for document retrieval. CoRR, abs/2206.02743."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.313,
+                0.883,
+                0.378
+            ],
+            "angle": 0,
+            "content": "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In ICLR. OpenReview.net."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.389,
+                0.883,
+                0.481
+            ],
+            "angle": 0,
+            "content": "Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In SIGIR '21: The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, Canada, July 11-15, 2021, pages 1503-1512. ACM."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.49,
+                0.883,
+                0.543
+            ],
+            "angle": 0,
+            "content": "Yujia Zhou, Jing Yao, Zhicheng Dou, Ledell Wu, and JiRong Wen. 2022a. Dynamicretriever: A pre-training model-based IR system with neither sparse nor dense index. CoRR, abs/2203.00537."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.553,
+                0.883,
+                0.605
+            ],
+            "angle": 0,
+            "content": "Yujia Zhou, Jing Yao, Zhicheng Dou, Ledell Wu, Peitian Zhang, and Ji-Rong Wen. 2022b. Ultron: An ultimate retriever on corpus with a model-based indexer. CoRR, abs/2208.09257."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.615,
+                0.883,
+                0.668
+            ],
+            "angle": 0,
+            "content": "Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. CoRR, abs/2308.07107."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.678,
+                0.883,
+                0.743
+            ],
+            "angle": 0,
+            "content": "Shengyao Zhuang, Houxing Ren, Linjun Shou, Jian Pei, Ming Gong, Guido Zuccon, and Daxin Jiang. 2022. Bridging the gap between indexing and retrieval for differentiable search index with query generation. CoRR, abs/2206.10128."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.743
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12490"
+        }
+    ]
+]

2023/Enhancing Generative Retrieval with Reinforcement Learning from Relevance Feedback/8a8868ae-16b0-40bb-b3ea-fdf45defe7f6_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f7b24ecce1472d11adb88563be399a83b49276ea8123d54b3488aa251ed7654
+size 3327144

2023/Enhancing Generative Retrieval with Reinforcement Learning from Relevance Feedback/full.md

@@ -0,0 +1,298 @@
+# Enhancing Generative Retrieval with Reinforcement Learning from Relevance Feedback
+
+Yujia Zhou $^{1}$ , Zhicheng Dou $^{2*}$ , Ji-Rong Wen $^{2}$
+
+$^{1}$ School of Information, Renmin University of China
+
+$^{2}$ Gaoling School of Artificial Intelligence, Renmin University of China
+
+{zhouyujia,dou,jrwen}@ruc.edu.cn
+
+# Abstract
+
+The recent advent of end-to-end generative retrieval marks a significant shift in document retrieval methods, leveraging differentiable search indexes to directly produce relevant document identifiers (docids) in response to a specific query. Nevertheless, this approach faces two fundamental challenges: (i) a discrepancy between the token-level probabilistic optimization and the broader document-level relevance estimation; (ii) an overemphasis on top-1 results at the expense of overall ranking quality. To tackle these challenges, we propose a generative retrieval model with reinforcement learning from relevance feedback, which aims to align token-level docid generation with document-level relevance estimation. The training process incorporates three stages: supervised fine-tuning, relevance reward model training, and reinforced learning-to-rank from relevance feedback. To train a high-quality reward model, we define "relevance" under three progressive scenarios, which collectively offer a comprehensive evaluation of the document relevance. Experiments conducted on two benchmark datasets demonstrate the effectiveness of our proposed approach.
+
+# 1 Introduction
+
+Information retrieval systems have undergone progressive advancements over the past few decades (Zhu et al., 2023). Historically, the retrieval of candidate documents has been heavily reliant on the use of an inverted index (Robertson and Zaragoza, 2009), a conventional method that employs term-based parsing. With the emergency of neural networks, dense retrieval (Karpukhin et al., 2020; Xiong et al., 2021) is proposed to encode both query and document into dense vectors, so as to measure the similarity in semantic space.
+
+Recently, there has been a growing focus on investigating end-to-end generative retrieval models (Tay et al., 2022; Bevilacqua et al., 2022; Wang
+
+et al., 2022; Zhou et al., 2022a). These models supersede the conventional explicit index by implementing a large-scale model known as the differentiable search index (Tay et al., 2022). This innovative approach facilitates end-to-end document retrieval using a seq-to-seq generative model, such as T5 (Raffel et al., 2020). In this framework, the model directly generates the corresponding docid for a given query.
+
+Despite the notable progress made in generative retrieval models, they still face two major challenges that limit their ability to retrieve relevant documents. First, these models primarily rely on the auto-regressive loss to maximize the token-level generation probability of the ground-truth docid. However, such an optimization goal is not aligned with the primary objective of ranking tasks, which is to estimate document-level relevance. Second, existing approaches assume a deterministic distribution in which all probability mass is allocated to the "golden" docid. Such an assumption can lead the model to overemphasize the top-1 generation results, while ignoring the quality of the overall ranking list. In light of these challenges, establishing alignment between generation probabilities and document relevance is a vital necessity in adapting generative models to ranking tasks.
+
+In this paper, we present a Generative Retrieval model with Reinforcement Learning from relevance feedback (GenRRL). Our objective is to align the generative retrieval model, trained specifically on the task of docid generation, with the model responsible for assessing document relevance. The training process comprises three pivotal steps: (1) Supervised fine-tuning. This initial step involves training the model using labeled data to improve its performance in generating docid. (2) Relevance reward model training. The subsequent step focuses on training a relevance reward model, specifically designed to measure the degree of relevance between the generated docids and their correspond
+
+ing queries. We expect the reward model to assign higher scores to docids that are more relevant to the queries. (3) Reinforced learning-to-rank (L2R) from relevance feedback. The final step involves using the trained relevance reward model to guide the generative retrieval system towards generating more relevant docids. To align with the ranking task requirements, we incorporate the idea of L2R into the reinforcement learning procedure. This enables the provision of not only pointwise reward but also pairwise and listwise relevance feedback, thereby enhancing the overall ranking performance.
+
+In order to train a high-quality relevance reward model, it is crucial to define "relevance" accurately within the context of document retrieval tasks. In this regard, we focus on three key scenarios that robustly reflect the relationship between queries and documents. Firstly, term-based overlap concerns the presence of identical or similar words within both the query and the document. Techniques like TF-IDF can effectively gauge this overlap. Secondly, semantic similarity extends beyond mere term overlap, delving into the shared meanings or concepts within the query and the document. Dual encoder models are particularly useful in modeling the semantic similarity by encoding the query and the document into dense vectors. Lastly, contextual dependency considers how effectively the document satisfies the information needs expressed in the query, extending beyond the boundaries of specific terms or concepts within the query. Techniques for capturing contextual dependency could include language models, which have demonstrated proficiency in understanding the contextual relationships between two sentences. By considering these three progressive scenarios, we can comprehensively assess the relevance between queries and documents, leading to a more effective reward model.
+
+To assess the effectiveness of our model, we perform comprehensive experiments on widely adopted MS MARCO and NQ document retrieval datasets. Experimental results demonstrate the efficacy of the proposed method, including the diverse relevance annotations and the reinforcement learning process. Our in-depth analysis on different model sizes also reveals the scaling laws of our proposed method.
+
+Our contributions are summarized as: (1) We employ reinforcement learning from relevance feedback to improve generative retrieval, aiming to
+
+align the docid generation process with document relevance assessment. (2) We outline a reinforced training process specifically tailored for the ranking task, which includes relevance reward model training and reinforced learning-to-rank. (3) We provide a classification of "relevance" within the context of document retrieval. By considering three progressive aspects, the trained reward model enables a thorough assessment of the relevance between queries and documents.
+
+# 2 Related Work
+
+Traditionally, index-based document retrieval methods encompass both sparse and dense retrieval techniques. Sparse retrieval methods (Dai and Callan, 2019, 2020; Bai et al., 2020; Formal et al., 2021), such as BM25 (Robertson and Zaragoza, 2009), utilize inverted indexes to assess term importance and calculate matching scores between queries and documents. Dense retrieval methods have been employed to overcome the limitations of word mismatching (Gao et al., 2021). These methods convert queries and documents into dense vectors and employ ANNS (Jégou et al., 2011) for efficient vector-based search. The dual encoder approach incorporating pre-trained language models has shown improved performance in dense retrieval. Techniques like hard negative sampling (Xiong et al., 2021; Karpukhin et al., 2020; Zhan et al., 2021; Guu et al., 2020) further enhance effectiveness.
+
+In recent times, generative retrieval has gained significant popularity in the field of information retrieval (IR) tasks. A model-based IR approach, where docids are embedded into the model, was first introduced in (Metzler et al., 2021; Tay et al., 2022). Since then, several other models (Wang et al., 2022; Zhuang et al., 2022; Zhou et al., 2022b) incorporating a query augmentation module have been developed to enhance the performance. Moreover, the model-based IR framework has been extended to knowledge-intensive language tasks, leading to notable improvements in performance (Chen et al., 2022a; Bevilacqua et al., 2022; Chen et al., 2022b). Additionally, a series of studies by Sun et al. (2023); Chen et al. (2023); Ren et al. (2023); Li et al. (2023) explored various methods for representing docids to retain docid semantics.
+
+In this paper, we focus on aligning the optimizing objective of generative retrieval models with document-level relevance, so as to alleviate the bias against the objective of ranking tasks.
+
+![](images/b621dd1f84a6975105ebe2a947af331778c0664ba7fa2f95e045165d30235278.jpg)  
+Figure 1: Empirical study on ranking results of generative retrieval and dense retrieval.
+
+# 3 Preliminary
+
+In this section, we formulate the generative retrieval problem and investigate its limitations.
+
+# 3.1 Task Definition
+
+Generative retrieval aims to generate the relevant docids for a given query directly with a seq-to-seq model. Assuming that  $d'$  is the identifier of the document  $d$ , the generative retriever takes the query  $q$  as input and generates the relevant docid with an auto-regressive score, written as:
+
+$$
+\operatorname {s c o r e} (d, q) = G _ {\phi} \left(d ^ {\prime} \mid q\right) = \prod_ {i = 1} G _ {\phi} \left(d _ {i} ^ {\prime} \mid d _ {<   i} ^ {\prime}, q\right), \tag {1}
+$$
+
+where  $G_{\phi}$  is the generative retriever with parameters  $\phi$ , and  $d_i'$  is the  $i_{th}$  token of the docid. The parameters  $\phi$  is optimized with the standard seq-to-seq objective. During inference, the model applies constrained beam search to guide the decoder in searching within a limited token space at each step, so as to generate valid docids from a candidate set.
+
+# 3.2 Empirical Study
+
+In this section, we present an empirical study aiming to investigate the limitations of generative retrieval models compared to advanced dense retrieval models in ranking tasks. We choose DPR (Karpukhin et al., 2020) and Ultron (Zhou et al., 2022b) as representatives of the dense and generative retrieval model respectively, and investigate their ranking results on NQ dataset.
+
+As shown in Figure 1, it can be observed that: Firstly, the generative retrieval model demonstrates a higher likelihood of placing the correct document at the top position, while the dense retrieval model exhibits a more balanced distribution across all rank positions. Secondly, the relevant documents positioned at the first rank receive remarkably high
+
+scores from the generative model, whereas the ranking score curve of the dense retrieval model appears smoother. This outcome is likely due to the discrepancy between the generation probability and the document relevance. To enhance the model ability on document relevance estimation, we propose a reinforcement learning framework for generative retrieval as follows.
+
+# 4 Methodology
+
+In this section, we focus on the training pipeline of GenRRL, including supervised fine-tuning, relevance reward model training, and reinforced L2R from relevance feedback.
+
+# 4.1 Supervised Fine-tuning
+
+Supervised fine-tuning (SFT) entails training the model to generate relevant docids based on token-level generation probabilities for a given query. Specifically, we choose keyword-based docids (e.g. URL, summary) as the training target, as these can more intuitively reflect the semantics of the document. In addition, we select standard T5 as the pre-trained model and apply pseudo query generation for data augmentation, which has been proven to be effective in improving the performance of generative retrieval (Zhuang et al., 2022). Formally, the generative loss function can be formulated as the maximization of the likelihood of the target sequence using teacher forcing, we have:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\text {g e n}} = - \log G _ {\phi} \left(d ^ {\prime} | q\right) \\ = - \sum_ {i = 1} \log G _ {\phi} \left(d _ {i} ^ {\prime} \mid d _ {<   i} ^ {\prime}, q\right), \tag {2} \\ \end{array}
+$$
+
+where  $G_{\phi}(d_i'|d_{<i}',q)$  is the generation probability of token  $d_i'$  based on the given input. After supervised fine-tuning, our next goal is to train a reward model designed to incorporate relevance preferences into docid generation process.
+
+# 4.2 Relevance Reward Model Training
+
+A high-quality relevance reward model is capable of effectively assessing the relevance between queries and documents. To accomplish this, we conduct an analysis of relevance judgments within the document retrieval. Subsequently, we employ three distinct methodologies to annotate document relevance, thereby facilitating the training of a comprehensive relevance feedback model. The details are introduced as follows.
+
+![](images/7541b2e9b5952e1db6a7513f63cecf6b1568d08a0b05caaf803014cb2a58c73c.jpg)  
+Step 1: Supervised Fine-tuning
+
+![](images/f05cb9e2ce9e95da1459789a2a39c5ba7a3d5923106939435d43fb90368ccb89.jpg)  
+Step 2: Relevance Reward Model Training
+
+![](images/d0b5ea590df4733edc1f522dfc60c98d00d2e1286f6478e55ced55e45ffe6f72.jpg)  
+Figure 2: The training pipeline of GenRRL with supervised fine-tuning, relevance reward model training, and reinforced learning-to-rank from relevance feedback. Blue arrows indicate data used for model training.
+
+Relevance Classification. We examine three scenarios wherein queries and documents exhibit strong associations. They are:
+
+- Term-based Overlap. This scenario refers to the presence of identical or similar terms in both the query and the document. If a document contains many of the same words that are in the query, it is likely to be relevant.  
+- Semantic Similarity. This goes beyond term overlap to look at the meaning of the words and phrases in the query and document. Even if a document does not contain the exact terms present in the query, it might still be relevant if it discusses the same concepts. For example, a query for "cardiovascular fitness exercises" might retrieve a document about "aerobic workouts" because the two phrases refer to similar concepts.  
+- Contextual Dependency. This scenario refers to the extent to which a document fulfills the information needs expressed by the query, surpassing the limitations posed by specific terms or concepts. For instance, if a user has been searching for "how to save battery of laptops". A relevant document might say: "reducing screen brightness and turning off keyboard backlight".
+
+Relevance Annotation. For each sampled query  $q$  from supervised fine-tuning dataset, we generate a set of candidate documents based on the fine
+
+tuned model, which corresponds to its own beam search results specific to that query. In order to capture the relevance of the above three scenarios respectively, we use three different models to annotate the relevance of the sampled query-doc pairs: (1) To measure the term-based overlap, we leverage BM25 (Robertson and Zaragoza, 2009), a well-established ranking function used by search engines. BM25's strength lies in its ability to weigh query terms based on their TF-IDF importance, thus providing a robust measure of term-based overlap. (2) To capture the semantic similarity, we employ the dual encoder model DPR (Karpukhin et al., 2020), which works by separately encoding the query and document into dense vectors in a shared embedding space. It focuses on deeper semantic connections between words and phrases in the query and the document. The dot product of the two vectors then gives a measure of the semantic similarity. (3) For assessing contextual dependency, we employ the large language model (LLM), LLaMA-13b (Touvron et al., 2023), to compute the generation likelihood of the document given a specific query. We assume that the generation probability can serve as an indicative metric of how well the document fulfills the implicit demands enclosed in the query.
+
+To mitigate the discrepancies arising from variations in scoring distributions across different meth
+
+ods, rankings are employed as a means of introducing regularized supervision signals. Specifically, when provided with multiple rankings of the candidate documents, we adopt a voting-based approach, Borda Rank, to consolidate them into a unified ranking. For example, given  $N$  items and  $M$  individual rankings, the Borda rank score  $B_{i}$  for item  $i$  is calculated as  $B_{i} = \sum_{j=1}^{M} (N - \mathrm{rank}_{j}(i))$ , where  $\mathrm{rank}_{j}(i)$  represents the rank of item  $i$  in the  $j$ -th individual ranking. Based on ranking results, we can construct document pairs including a positive document  $d^{+}$  and a negative document  $d^{-}$  for reward model training.
+
+Reward Model training. To train our reward models, we start from the supervised fine-tuned model, and then add a randomly initialized linear head that outputs a scalar reward. The reward model is trained in a pairwise manner to predict which document is more relevant. Given the query  $q$  and a document pair  $(d^{+}, d^{-})$  sampled from the ranking  $D$ , the loss function for the reward model is:
+
+$$
+\mathcal {L} _ {\mathrm {R M}} = \mathbb {E} _ {(q, d ^ {+}, d ^ {-}) \sim D} [ \log (\sigma (r _ {\theta} (q, d ^ {+}) - r _ {\theta} (q, d ^ {-}))) ],
+$$
+
+where  $r_{\theta}(q,d)$  is the scalar output of the reward model with parameters  $\theta$
+
+Now we have an initial model capable of docid generation and a preference model that assigns a score to assess the document relevance. Subsequently, reinforcement learning is employed to optimize the initial model based on the reward model.
+
+# 4.3 Reinforced L2R from Relevance Feedback
+
+To begin with, we cast the generative retrieval task as a reinforcement learning (RL) problem. Initially, the policy is defined as a language model that accepts a query as input and generates a sequence of tokens as output. The action space of the policy encompasses all tokens in the language model's vocabulary, while the observation space corresponds to the distribution of potential input queries.
+
+Reward Function. To guide the RL training, the reward function is constructed by combining the rewards  $r_{\theta}(q,d)$  generated by the reward model with a KL divergence constraint, which ensures that the policy does not deviate significantly from its initial behavior, defined as:
+
+$$
+R (q, d) = r _ {\theta} (q, d) - \beta \log \left[ \pi_ {\phi} ^ {\mathrm {R L}} (d | q) / \pi^ {\mathrm {S F T}} (d | q) \right],
+$$
+
+where  $\pi_{\phi}^{\mathrm{RL}}$  is the learned RL policy,  $\pi^{\mathrm{SFT}}$  is the supervised fine-tuned model,  $\beta$  is The KL reward coefficient controlling the strength of the KL penalty.
+
+To align with the requirements of the ranking task, we introduce the concept of learning-to-rank into the RL procedure, employing pointwise, pairwise, and listwise optimization strategies.
+
+Pointwise Optimization. We begin by independently updating the model for each query-document pair. For a given query  $q$ , the generative retriever functions as a policy network, generating a candidate docid corresponding to the document  $d$ . The RL loss can be formulated as:
+
+$$
+\mathcal {L} _ {\mathrm {R L}} = - (R (q, d) - b) \sum_ {t} \log G _ {\phi} (d _ {t} ^ {\prime} | d _ {<   t} ^ {\prime}, q),
+$$
+
+where  $b$  is the baseline incorporated to mitigate variance during RL training. It is computed by averaging the rewards of all samples.
+
+Pairwise Optimization. Instead of regarding each query-document pair independently, we incorporate the assessment of pairwise document comparisons as a form of reward. Given a query  $q$ , the docid generation model samples plenty of document pairs  $(d_i, d_j)$ . The generator model is trained with the pairwise loss, denoted as:
+
+$$
+\mathcal {L} _ {\mathrm {R L}} = - \sum_ {(d _ {i}, d _ {j})} (R (q, d _ {i}) \log p _ {i j} + R (q, d _ {j}) \log p _ {j i}),
+$$
+
+where  $p_{ij}$  is computed by  $|G_{\phi}(d_i'|q) - G_{\phi}(d_j'|q)|$  with the normalization of logistic function.
+
+Listwise Optimization. Listwise strategy provides an alternative approach by considering the entire list of generated documents. In this context, we define a listwise loss that aims to minimize the ranking discrepancy between the SFT model and the reward model. Given a set of sampled candidates  $C$  for the query  $q$ , the loss function can be defined as:
+
+$$
+\mathcal {L} _ {\mathrm {R L}} = - \sum_ {d \in C} R (q, d) \log \frac {\exp \left(G _ {\phi} (d ^ {\prime} | q)\right)}{\sum_ {j} \exp \left(G _ {\phi} (d _ {j} ^ {\prime} | q)\right)}.
+$$
+
+By incorporating listwise optimization, the generative retriever is promoted to generate docids that closely resemble the ordering preferred by the reward model. This approach takes into account the overall quality and relevance of the entire list of generated documents, rather than focusing solely on an individual document or comparisons.
+
+# 5 Experimental Setup
+
+# 5.1 Datasets and Evaluation Metrics
+
+Our experiments are conducted on two widely used document retrieval benchmarks: MS MARCO Doc
+
+uments. In 2016, we created a document ranking system that uses the same ranking metric as the MRR and HITS (at cutoff 1, 5, 10) as the evaluation metrics.
+
+# 5.2 Baselines
+
+To establish baselines for comparison, we consider three classes of models: (1) Sparse Retrieval models. We select BM25 (Robertson and Zaragoza, 2009), which incorporates tfidf weighting to determine term importance, and DocT5Query (Nogueira et al., 2019), which expands a document with queries predicted by a T5 model, as baselines. (2) Dense Retrieval models. We consider two dual encoder models trained with different hard negatives sampling strategies DPR (Karpukhin et al., 2020) and ANCE (Xiong et al., 2021). (3) Generative Retrieval models: DSI (Tay et al., 2022), the pioneering generative retrieval model that uses a seq-to-seq model to map the query to the relevant docid. Building upon DSI, SEAL (Bevilacqua et al., 2022) uses n-grams within passages as potential identifiers. DSI-QG (Zhuang et al., 2022), NCI (Wang et al., 2022), and Ultron (Zhou et al., 2022b) incorporate query generation modules for data augmentation. To ensure a fair comparison, all baseline models employ the "base" version of pre-trained models.
+
+# 5.3 Implementation Details
+
+Docid. We explore two types of linguistic docids: document URL and document summary. For the summary-based docids, we employ the LLaMA-13b model to generate document summaries using the prompt "Summarize the following paragraphs with meaningful keywords: {document}".
+
+SFT. Our SFT model utilizes the T5-base pretrained model (Raffel et al., 2020) as the back
+
+bone and sets the learning rate and the batch size as 1e-3 and 128 during training. To enrich the training data, we employ the existing DocT5Query model (Nogueira et al., 2019) to generate 10 pseudo queries for each document.
+
+RL. The reward model is initialized using the SFT model, with a linear head that is randomly initialized. During the training of the reward model, we employ the SFT model to generate 8 candidate docids for relevance annotation. We set the batch size to 64 and utilize a learning rate of 3e-4 for this training phase. In the reinforced L2R process, the policy model provides 1, 2, 4 candidate docids at each step corresponding to pointwise, pairwise, and listwise optimization, while the learning rate is set to 3e-5. All experiments are conducted on up to 8 NVIDIA A100 GPUs.
+
+# 6 Experimental Results and Analysis
+
+In this section, we present the experimental findings of our proposed approach and perform an empirical analysis to provide comprehensive insights.
+
+# 6.1 Overall Results
+
+The overall results of models are listed in Table 1. Some findings are summarized as follows.
+
+Comparison with Generative Retrievers. Our observations indicate consistent superior performance of GenRRL compared to baseline models and SFT models when evaluated on the MS MARCO and NQ datasets. This pattern of results serves as compelling evidence for the effectiveness of our reinforced training strategy. Notably, when compared to the leading generative retrieval model Ultron on the MS MARCO dataset, GenRRL (URL) exhibits a noteworthy improvement of over  $9.6\%$  in terms of MRR@10. Similarly, on the NQ dataset, ROGER (Sum) surpasses the competitive baseline NCI by over  $8.1\%$  in MRR@10. These findings strongly support the notion that integrating relevance feedback into the initial generative retrieval model holds immense promise for enhancing performance in ranking tasks.
+
+Effect of Different Docids. In the context of generative retrieval, the selection of docids can be categorized into two forms: linguistic docids and ID-based docids. Through a comparative analysis between DSI-QG and Ultron, we have observed that linguistic docids are better suited for generative retrieval tasks. Building upon this insight, we have
+
+Table 1: Overall results. Document Rep. indicates the method to represent document. The best results are shown in bold. “†” indicates the result is significantly better than the corresponding SFT model with paired t-test at  $p < 0.05$ .  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Document Rep.</td><td colspan="4">MS MARCO</td><td colspan="4">Natural Questions</td></tr><tr><td>Hits@1</td><td>Hits@5</td><td>Hits@10</td><td>MRR@10</td><td>Hits@1</td><td>Hits@5</td><td>Hits@10</td><td>MRR@10</td></tr><tr><td colspan="10">Index-based Retrieval</td></tr><tr><td>BM25</td><td>sparse terms</td><td>18.94</td><td>42.82</td><td>55.07</td><td>29.24</td><td>14.06</td><td>36.91</td><td>47.93</td><td>23.60</td></tr><tr><td>DocT5Query</td><td>sparse terms</td><td>23.27</td><td>49.38</td><td>63.61</td><td>34.81</td><td>19.07</td><td>43.88</td><td>55.83</td><td>29.55</td></tr><tr><td>DPR</td><td>dense vector</td><td>29.08</td><td>62.75</td><td>73.13</td><td>43.41</td><td>22.78</td><td>53.44</td><td>68.58</td><td>35.92</td></tr><tr><td>ANCE</td><td>dense vector</td><td>29.65</td><td>63.43</td><td>74.28</td><td>44.09</td><td>24.54</td><td>54.21</td><td>69.08</td><td>36.88</td></tr><tr><td colspan="10">Generative Retrieval</td></tr><tr><td>DSI</td><td>semantic id</td><td>25.74</td><td>43.58</td><td>53.84</td><td>33.92</td><td>27.42</td><td>47.26</td><td>56.58</td><td>34.31</td></tr><tr><td>DSI-QG</td><td>semantic id</td><td>28.82</td><td>50.74</td><td>62.26</td><td>38.45</td><td>30.17</td><td>53.20</td><td>66.37</td><td>38.85</td></tr><tr><td>SEAL</td><td>n-grams</td><td>27.58</td><td>52.47</td><td>61.01</td><td>37.68</td><td>29.30</td><td>54.12</td><td>68.53</td><td>40.34</td></tr><tr><td>NCI</td><td>semantic id</td><td>29.54</td><td>57.99</td><td>67.28</td><td>40.46</td><td>32.69</td><td>55.82</td><td>69.20</td><td>42.84</td></tr><tr><td>Ultron</td><td>title + URL</td><td>29.82</td><td>60.39</td><td>68.31</td><td>42.53</td><td>33.78</td><td>54.20</td><td>67.05</td><td>42.51</td></tr><tr><td>SFT (URL)</td><td>doc URL</td><td>29.90</td><td>60.28</td><td>68.17</td><td>42.52</td><td>33.91</td><td>54.18</td><td>66.92</td><td>42.52</td></tr><tr><td>SFT (Sum)</td><td>doc summary</td><td>30.12</td><td>60.68</td><td>68.71</td><td>42.84</td><td>33.72</td><td>54.71</td><td>67.69</td><td>42.83</td></tr><tr><td>GenRRL (URL)</td><td>doc URL</td><td>\( 33.01^{\dagger} \)</td><td>\( 63.62^{\dagger} \)</td><td>\( 74.91^{\dagger} \)</td><td>\( 45.93^{\dagger} \)</td><td>\( 35.79^{\dagger} \)</td><td>\( 56.49^{\dagger} \)</td><td>\( 70.96^{\dagger} \)</td><td>\( 45.73^{\dagger} \)</td></tr><tr><td>GenRRL (Sum)</td><td>doc summary</td><td>\( 33.23^{\dagger} \)</td><td>\( 64.48^{\dagger} \)</td><td>\( 75.80^{\dagger} \)</td><td>\( 46.62^{\dagger} \)</td><td>\( 36.32^{\dagger} \)</td><td>\( 57.42^{\dagger} \)</td><td>\( 71.49^{\dagger} \)</td><td>\( 46.31^{\dagger} \)</td></tr></table>
+
+Table 2: The results of different relevance annotators.  $\bullet$  and  $\circ$  indicates annotating with and without the model.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">Relevance annotator</td><td colspan="2">MS MARCO</td><td colspan="2">Natural Questions</td></tr><tr><td>Sparse</td><td>Dense</td><td>LLM</td><td>Hits@10</td><td>MRR@10</td><td>Hits@10</td><td>MRR@10</td></tr><tr><td>SFT Model (Sum)</td><td>○</td><td>○</td><td>○</td><td>68.71</td><td>42.84</td><td>67.69</td><td>42.83</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>●</td><td>○</td><td>74.37</td><td>45.85</td><td>70.40</td><td>45.32</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>○</td><td>●</td><td>73.19</td><td>45.18</td><td>69.78</td><td>44.67</td></tr><tr><td>GenRRL (Sum)</td><td>○</td><td>●</td><td>●</td><td>75.36</td><td>46.28</td><td>70.92</td><td>45.81</td></tr><tr><td>GenRRL (Sum)</td><td>●</td><td>●</td><td>●</td><td>75.80</td><td>46.62</td><td>71.49</td><td>46.31</td></tr></table>
+
+explored two types of linguistic docids, namely doc URL and doc summary. The experimental results demonstrate that our model consistently outperforms purely supervised fine-tuned models on both types of linguistic docids. Furthermore, the performance is particularly enhanced when utilizing summary-based docids, indicating that docids generated by large language models can better capture document relevance compared to the more simplistic approach of URL-based docids.
+
+# 6.2 Ablation Study on Relevance Annotators
+
+In this section, we perform an ablation study to assess the impact of relevance annotation in the GenRRL model. As described in Section 4.2, the relevance annotation process involves three types of annotators: the sparse model, the dense model, and LLM, each representing distinct relevance scenarios. We explore various combinations of these annotators to evaluate their effectiveness.
+
+Table 2 presents the results for variants of GenRRL. We can observe that each annotator in the relevance annotation process demonstrates a contri
+
+bution to the overall performance, with the dense model exhibiting the most notable improvement in results. This finding suggests that the dense model possesses a more pronounced advantage in assessing the relevance between queries and documents. Additionally, the inclusion of the LLM annotator yields performance enhancements, indicating that leveraging the zero-shot capability of large language models can further augment the model's ability to capture contextual dependencies.
+
+# 6.3 Comparison of Different RL Losses
+
+The incorporation of relevance feedback signals with reinforcement learning serves as a crucial step in model fine-tuning. To investigate the impact of different reinforced feedback strategies, we compare three reinforcement learning losses: pointwise, pairwise, and listwise, and examine their influence on ranking performance for both type of docids.
+
+Table 3 presents the results obtained for both URL-based docids and summary-based docids. It is observed that as the loss transitions from pointwise to pairwise and listwise, there is a gradual im
+
+![](images/78dc74786589b29ba78f75dfda4aef45576df1473528dac568da4657623412f6.jpg)  
+Figure 3: Performance on different query sets corresponding to three types of relevance scenarios on MS.
+
+Table 3: The MRR@10 of different RL losses.  
+
+<table><tr><td>Model</td><td>RL Loss</td><td>MS</td><td>NQ</td></tr><tr><td>GenRRL (URL)</td><td>Pointwise</td><td>43.42</td><td>43.01</td></tr><tr><td>GenRRL (URL)</td><td>Pairwise</td><td>44.82</td><td>45.10</td></tr><tr><td>GenRRL (URL)</td><td>Listwise</td><td>45.93</td><td>45.73</td></tr><tr><td>GenRRL (Sum)</td><td>Pointwise</td><td>43.78</td><td>43.92</td></tr><tr><td>GenRRL (Sum)</td><td>Pairwise</td><td>45.37</td><td>45.46</td></tr><tr><td>GenRRL (Sum)</td><td>Listwise</td><td>46.62</td><td>46.31</td></tr></table>
+
+provement on the search results. This progressive enhancement suggests that the listwise loss function could better alleviate the issue of the model's excessive focus on top-1 results. By assuming a non-deterministic distribution, the listwise loss assigns probability mass to other candidate docids based on their relevance.
+
+# 6.4 Performance on Different Query Sets
+
+To investigate the distinctions between the three relevance scenarios, we partitioned the test queries into three categories based on the performance of sparse, dense, and generative retrieval. These categories align with the dimensions of term-based overlap, semantic similarity, and contextual dependency, respectively. We then conducted a comparative analysis of the model's performance before and after reinforcement learning.
+
+From Figure 3, we find that the BM25 and DPR models have demonstrated noticeable advantages on their respective query sets. Although the SFT model exhibited certain gaps in comparison, the introduction of reinforcement learning allowed us to incorporate relevance signals from both the sparse and dense models. Furthermore, the SFT model was augmented by incorporating the knowledge of the LLM. As a result, the retrieval quality across the three query sets experienced improvements. These enhancements signify the effectiveness of reinforcement learning process and diverse relevance signals, leading to enhanced retrieval performance.
+
+![](images/9b2fcc8a63c9cfc2d36487cc7732130648098b147dea41cda8280e38751545be.jpg)  
+(a) Scale of SFT model.
+
+![](images/5048b70396f706bf8766a5a9dec39fee165825b4892016e28ebc954302a4acf4.jpg)  
+(b) Scale of reward model.  
+Figure 4: The variation trends on MS with the increase of training step under different model scales.
+
+# 6.5 Effect on Model Scale
+
+To investigate the training characteristics of models with varying parameter scales, we try T5-small, T5-base and T5-large models for training SFT model and reward model respectively.
+
+Scale of SFT Model. The learning curve depicted in Figure 4(a) illustrates the convergence behavior of the GenRRL model with varying parameter sizes as the number of training steps increases. It is observed that the small model exhibits comparatively lower MRR at convergence. Conversely, both the base and large models achieve similar performance levels after a sufficient number of training steps, with the large model slightly surpassing the base model. These findings emphasize the critical influence of model capacity on retrieval performance, indicating that larger models have the potential to yield improved results given adequate training.
+
+Scale of Reward Model. Figure 4(b) shows the learning curve with different scale of the reward model. We observe that during the initial 4000 training steps, the differences between reward models of different sizes are not distinct. However, as training progresses, the small reward model exhibits early convergence. In contrast, the base and large reward models continue to improve their performance, with the large reward model outperforming the base model. This finding suggests that the larger reward model demonstrates stronger capability in capturing capture nuanced distinctions among sampled docids, leading to continued improvement.
+
+# 7 Conclusion
+
+In this paper, we address the challenges faced by generative retrieval models by proposing an RL-based approach called GenRRL. Our objective is to align the generative retrieval with the task of assessing document relevance, thereby improving
+
+ranking performance. We introduce a three-stage reinforcement learning framework, where the relevance reward model captures different dimensions of relevance by considering term-based overlap, semantic similarity, and contextual dependency. Through extensive experiments, we demonstrate the effectiveness of our proposed method in improving generative retrieval performance.
+
+# Acknowledgement
+
+This work was supported by the National Natural Science Foundation of China No. 62272467 and No. 61832017, Beijing Outstanding Young Scientist Program No. BJWZYJH012019100020098, Public Computing Cloud, Renmin University of China, Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE, and Intelligent Social Governance Platform, Major Innovation & Planning Interdisciplinary Platform for the "Double-First Class" Initiative, Renmin University of China. The work was partially done at Beijing Key Laboratory of Big Data Management and Analysis Methods.
+
+# Limitations
+
+Despite the notable progress achieved within our reinforced framework, GenRRL encounters several challenges that necessitate further exploration and advancement. Primarily, the extension of the model to web scale poses substantial requirements for the design of the differentiable search index, demanding enhanced capacity to handle vast amounts of data. Secondly, the incorporation of new incoming documents into the differentiable search index has not yet been thoroughly investigated. This aspect presents an unexplored avenue that requires careful exploration and analysis. Addressing this challenge entails devising strategies and techniques to seamlessly integrate new documents into the existing model-based framework, while ensuring that the indexer retains its efficiency, accuracy, and robustness.
+
+# References
+
+Yang Bai, Xiaoguang Li, Gang Wang, Chaoliang Zhang, Lifeng Shang, Jun Xu, Zhaowei Wang, Fangshan Wang, and Qun Liu. 2020. Sparterm: Learning term-based sparse representation for fast text retrieval. CoRR, abs/2010.00768.  
+Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen-tau Yih, Sebastian Riedel, and Fabio Petroni.
+
+2022. Autoregressive search engines: Generating substrings as document identifiers. CoRR, abs/2204.10628.  
+Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Yixing Fan, and Xueqi Cheng. 2022a. GERE: generative evidence retrieval for fact verification. In SIGIR, pages 2184-2189. ACM.  
+Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Yiqun Liu, Yixing Fan, and Xueqi Cheng. 2022b. Corpusbrain: Pre-train a generative retrieval model for knowledge-intensive language tasks. In CIKM, pages 191-200. ACM.  
+Xiaoyang Chen, Yanjiang Liu, Ben He, Le Sun, and Yingfei Sun. 2023. Understanding differential search index for text retrieval. CoRR, abs/2305.02073.  
+Zhuyun Dai and Jamie Callan. 2019. Context-aware sentence/passage term importance estimation for first stage retrieval. CoRR, abs/1910.10687.  
+Zhuyun Dai and Jamie Callan. 2020. Context-aware document term weighting for ad-hoc search. In WWW '20: The Web Conference 2020, Taipei, Taiwan, April 20-24, 2020, pages 1897-1907. ACM / IW3C2.  
+Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and Stephane Clinchant. 2021. SPLADE v2: Sparse lexical and expansion model for information retrieval. CoRR, abs/2109.10086.  
+Luyu Gao, Zhuyun Dai, Tongfei Chen, Zhen Fan, Benjamin Van Durme, and Jamie Callan. 2021. Complement lexical retrieval model with semantic residual embeddings. In Advances in Information Retrieval - 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28 - April 1, 2021, Proceedings, Part I, volume 12656 of Lecture Notes in Computer Science, pages 146-160. Springer.  
+Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. REALM: retrieval-augmented language model pre-training. CoRR, abs/2002.08909.  
+Hervé Jégou, Matthijs Douze, and Cordelia Schmid. 2011. Product quantization for nearest neighbor search. IEEE Trans. Pattern Anal. Mach. Intell., 33(1):117-128.  
+Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick S. H. Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In EMNLP (1), pages 6769-6781. Association for Computational Linguistics.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur P. Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob
+
+Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: a benchmark for question answering research. Trans. Assoc. Comput. Linguistics, 7:452-466.  
+Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wenjie Li. 2023. Multiview identifiers enhanced generative retrieval.  
+Donald Metzler, Yi Tay, Dara Bahri, and Marc Najork. 2021. Rethinking search: making domain experts out of dilettantes. SIGIR Forum, 55(1):13:1-13:27.  
+Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A human generated machine reading comprehension dataset. In Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain, December 9, 2016, volume 1773 of CEUR Workshop Proceedings. CEUR-WS.org.  
+Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. 2019. From doc2query to doctttttquery. Online preprint, 6.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.  
+Ruiyang Ren, Wayne Xin Zhao, Jing Liu, Hua Wu, Ji-Rong Wen, and Haifeng Wang. 2023. TOME: A two-stage approach for model-based retrieval. CoRR, abs/2305.11161.  
+Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Found. Trends Inf. Retr., 3(4):333-389.  
+Weiwei Sun, Lingyong Yan, Zheng Chen, Shuaiqiang Wang, Haichao Zhu, Pengjie Ren, Zhumin Chen, Dawei Yin, Maarten de Rijke, and Zhaochun Ren. 2023. Learning to tokenize for generative retrieval. CoRR, abs/2304.04171.  
+Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Prakash Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer
+
+memory as a differentiable search index. CoRR, abs/2202.06991.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971.  
+Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Hao Sun, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, Xing Xie, Hao Allen Sun, Weiwei Deng, Qi Zhang, and Mao Yang. 2022. A neural corpus indexer for document retrieval. CoRR, abs/2206.02743.  
+Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In ICLR. OpenReview.net.  
+Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In SIGIR '21: The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, Canada, July 11-15, 2021, pages 1503-1512. ACM.  
+Yujia Zhou, Jing Yao, Zhicheng Dou, Ledell Wu, and JiRong Wen. 2022a. Dynamicretriever: A pre-training model-based IR system with neither sparse nor dense index. CoRR, abs/2203.00537.  
+Yujia Zhou, Jing Yao, Zhicheng Dou, Ledell Wu, Peitian Zhang, and Ji-Rong Wen. 2022b. Ultron: An ultimate retriever on corpus with a model-based indexer. CoRR, abs/2208.09257.  
+Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. CoRR, abs/2308.07107.  
+Shengyao Zhuang, Houxing Ren, Linjun Shou, Jian Pei, Ming Gong, Guido Zuccon, and Daxin Jiang. 2022. Bridging the gap between indexing and retrieval for differentiable search index with query generation. CoRR, abs/2206.10128.

2023/Enhancing Generative Retrieval with Reinforcement Learning from Relevance Feedback/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44a4ce5fb3b2c15d7e90de55499f3a2684b75d4042c7edc42fdb098c30985359
+size 414132

2023/Enhancing Low-resource Fine-grained Named Entity Recognition by Leveraging Coarse-grained Datasets/dff73b45-756f-4d7a-a8f6-789c5d1784de_content_list.json

@@ -0,0 +1,1789 @@
+[
+    {
+        "type": "text",
+        "text": "Enhancing Low-resource Fine-grained Named Entity Recognition by Leveraging Coarse-grained Datasets",
+        "text_level": 1,
+        "bbox": [
+            144,
+            89,
+            855,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Su Ah Lee $^{1\\dagger}$ , Seokjin Oh $^{1\\dagger}$ , and Woohwan Jung $^{1,2}$",
+        "bbox": [
+            285,
+            148,
+            715,
+            167
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Department of Applied Artificial Intelligence, Hanyang University",
+        "bbox": [
+            228,
+            167,
+            773,
+            184
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2Ramply Inc.",
+        "bbox": [
+            445,
+            184,
+            554,
+            200
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{sue991, seokjinoh, whjung}@hanyang.ac.kr",
+        "bbox": [
+            294,
+            200,
+            709,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            267
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Named Entity Recognition (NER) frequently suffers from the problem of insufficient labeled data, particularly in fine-grained NER scenarios. Although  $K$ -shot learning techniques can be applied, their performance tends to saturate when the number of annotations exceeds several tens of labels. To overcome this problem, we utilize existing coarse-grained datasets that offer a large number of annotations. A straightforward approach to address this problem is prefinetuning, which employs coarse-grained data for representation learning. However, it cannot directly utilize the relationships between fine-grained and coarse-grained entities, although a fine-grained entity type is likely to be a subcategory of a coarse-grained entity type. We propose a fine-grained NER model with a Fineto-Coarse(F2C) mapping matrix to leverage the hierarchical structure explicitly. In addition, we present an inconsistency filtering method to eliminate coarse-grained entities that are inconsistent with fine-grained entity types to avoid performance degradation. Our experimental results show that our method outperforms both  $K$ -shot learning and supervised learning methods when dealing with a small number of fine-grained annotations. Code is available at https://github.com/sue991/CoFiNER.",
+        "bbox": [
+            144,
+            281,
+            460,
+            677
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            692,
+            258,
+            707
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Named Entity Recognition (NER) is a fundamental task in locating and categorizing named entities in unstructured texts. Most research on NER has been conducted on coarse-grained datasets, including CoNLL'03 (Tjong Kim Sang, 2002), ACE04 (Mitchell et al., 2005), ACE05 (Walker et al., 2006), and OntoNotes (Weischedel et al., 2013), each of which has less than 18 categories. As the applications of NLP broaden across diverse fields, there is increasing demand for fine-grained NER that can provide more precise and",
+        "bbox": [
+            112,
+            718,
+            490,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "detailed information extraction. Nonetheless, detailed labeling required for large datasets in the context of fine-grained NER presents several significant challenges. It is more cost-intensive and time-consuming than coarse-grained NER. In addition, it requires a high degree of expertise because domain-specific NER tasks such as financial NER (Loukas et al., 2022) and biomedical NER (Sung et al., 2022) require fine-grained labels. Thus, fine-grained NER tasks typically suffer from the data scarcity problem. Few-shot NER approaches (Ding et al., 2021) can be applied to conduct fine-grained NER with scarce fine-grained data. However, these methods do not exploit existing coarse-grained datasets that can be leveraged to improve fine-grained NER because fine-grained entities are usually subtypes of coarse-grained entities. For instance, if a model knows what Organization is, it could be easier for it to understand the concept of Government or Company. Furthermore, these methods often experience early performance saturation, necessitating the training of a new supervised learning model if the annotation extends beyond several tens of labels.",
+        "bbox": [
+            507,
+            253,
+            885,
+            637
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A pre-finetuning strategy (Aghajanyan et al., 2021; Ma et al., 2022a) was proposed to overcome the aforementioned problem. This strategy first learns the feature representations using a coarse-grained dataset before training the fine-grained model on a fine-grained dataset. In this method, coarse-grained data are solely utilized for representation learning; thus, it still does not explicitly utilize the relationships between the coarse- and fine-grained entities.",
+        "bbox": [
+            507,
+            642,
+            885,
+            802
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Our intuition for fully leveraging coarse-grained datasets comes mainly from the hierarchy between coarse- and fine-grained entity types. Because a coarse-grained entity type typically comprises multiple fine-grained entity types, we can enhance low-resource fine-grained NER with abundant coarse-grained data. To jointly utilize both datasets, we",
+        "bbox": [
+            507,
+            806,
+            885,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "Major in Bio Artificial Intelligence",
+        "bbox": [
+            147,
+            903,
+            371,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3269",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3269-3279 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            159,
+            945,
+            836,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "devise a method to build a mapping matrix called F2C (short for 'Fine-to-Coarse'), which connects fine-grained entity types to their corresponding coarse-grained entity types and propose a novel approach to train a fine-grained NER model with both datasets.",
+        "bbox": [
+            112,
+            84,
+            487,
+            179
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Some coarse-grained entities improperly match a fine-grained entity type because datasets may be created by different annotators for different purposes. These mismatched entities can reduce the performance of the model during training. To mitigate this problem, coarse-grained entities that can degrade the performance of fine-grained NER must be eliminated. Therefore, we introduce a filtering method called 'Inconsistency Filtering'. This approach is designed to identify and exclude any inconsistent coarse-to-fine entity mappings, ensuring a higher-quality training process, and ultimately, better model performance. The main contributions of our study are as follows:",
+        "bbox": [
+            115,
+            181,
+            489,
+            405
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose an F2C mapping matrix to directly leverage the intimate relation between coarse- and fine-grained types.",
+            "- We present an inconsistency filtering method to screen out coarse-grained data that are inconsistent with the fine-grained types.",
+            "- The empirical results show that our method achieves state-of-the-art performance by utilizing the proposed F2C mapping matrix and the inconsistency filtering method."
+        ],
+        "bbox": [
+            136,
+            418,
+            487,
+            600
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            612,
+            270,
+            627
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Fine-grained NER. NER is a key task in information extraction and has been extensively studied. Traditional NER datasets (Tjong Kim Sang, 2002; Ritter et al., 2011; Weischedel et al., 2013; Derczynski et al., 2017) address coarse-grained entity types in the general domain. Recently, domain-specific NER has been investigated in various fields (Hofer et al., 2018; Loukas et al., 2022; Sung et al., 2022). The domain-specific NER tasks typically employ fine-grained entity types. In addition, Ding et al. (2021) proposed a general domain fine-grained NER.",
+        "bbox": [
+            112,
+            642,
+            489,
+            833
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "$N$ -way  $K$ -shot learning for NER. Since labeling domain-specific data is an expensive process, few-shot NER has gained attention. Most few-shot NER studies are conducted using an  $N$ -way  $K$ -shot episode learning (Das et al., 2021; Huang",
+        "bbox": [
+            112,
+            838,
+            489,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "et al., 2022; Ma et al., 2022b; Wang et al., 2022). The objective of this approach is to train a model that can correctly classify new examples into one of  $N$  classes, using only  $K$  examples per class. To achieve this generalization performance, a large number of episodes need to be generated. Furthermore, to sample these episodes, the training data must contain a significantly larger number of classes than  $N$ . In contrast, in our problem context, the number of fine-grained classes we aim to identify is substantially larger than the number of coarse-grained classes in the existing dataset. Consequently, the episode-based few-shot learning approaches are unsuitable for addressing this problem.",
+        "bbox": [
+            507,
+            84,
+            884,
+            325
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Leveraging auxiliary data for information extraction. Several studies have employed auxiliary data to overcome the data scarcity problem. Jiang et al. (2021); Oh et al. (2023) propose NER models trained with small, strongly labeled, and large weakly labeled data. Jung and Shim (2020) used strong and weak labels for relation extraction. However, in the previous studies, the main data and auxiliary data shared the same set of classes, which is not the case for fine-grained NER with coarse-grained labels. While the approaches of Aghajanyan et al. (2021) and Ma et al. (2022a) can be applied to our problem setting, they utilize auxiliary data for representation learning rather than explicitly utilizing the relationship between the two types of data.",
+        "bbox": [
+            507,
+            329,
+            884,
+            587
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Proposed Method",
+        "text_level": 1,
+        "bbox": [
+            507,
+            598,
+            697,
+            614
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the notations and define the problem of fine-grained NER using coarse-grained data. Then, we introduce the proposed CoFiNER model, including the creation of the F2C mapping matrix.",
+        "bbox": [
+            507,
+            623,
+            882,
+            703
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 Problem definition",
+        "text_level": 1,
+        "bbox": [
+            507,
+            714,
+            702,
+            728
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given a sequence of  $n$  tokens  $\\mathbf{X} = \\{x_{1},x_{2},\\ldots ,x_{n}\\}$ , the NER task involves assigning type  $y_{i}\\in E$  to each token  $x_{i}$  where  $E$  is a predefined set of entity types. In our problem setting, we used a fine-grained dataset  $\\mathcal{D}^{\\mathcal{F}} = \\{(\\mathbf{X}_1^{\\mathcal{F}},\\mathbf{Y}_1^{\\mathcal{F}}),\\dots,(\\mathbf{X}_{|\\mathcal{D}^{\\mathcal{F}}|}^{\\mathcal{F}},\\mathbf{X}_{|\\mathcal{D}^{\\mathcal{F}}|})\\}$  with a predefined entity type set  $E^{\\mathcal{F}}$ . Additionally, we possess a coarse-grained dataset  $\\mathcal{D}^{\\mathcal{C}} = \\{(\\mathbf{X}_1^{\\mathcal{C}},\\mathbf{Y}_1^{\\mathcal{C}}),\\dots,(\\mathbf{X}_{|\\mathcal{D}^{\\mathcal{C}}|}^{\\mathcal{C}},\\mathbf{X}_{|\\mathcal{D}^{\\mathcal{C}}|})\\}$  characterized by a coarse-grained entity set  $E^{\\mathcal{C}}$  (i.e.  $|E^{\\mathcal{F}}| > |E^{\\mathcal{C}}|$ ). A coarse-grained dataset typically",
+        "bbox": [
+            507,
+            734,
+            884,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3270",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "has a smaller number of types than a fine-grained dataset. Throughout this study, we use  $\\mathcal{F}$  and  $\\mathcal{C}$  to distinguish between these datasets. It should be noted that our method can be readily extended to accommodate multiple coarse-grained datasets, incorporating an intrinsic multi-level hierarchy. However, our primary discussion revolves around a single coarse-grained dataset for simplicity and readability.",
+        "bbox": [
+            112,
+            84,
+            489,
+            229
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Training CoFiNER",
+        "text_level": 1,
+        "bbox": [
+            112,
+            239,
+            312,
+            255
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We aim to utilize both coarse- and fine-grained datasets directly in a single model training. Figure 1 illustrates an overview of the fine-grained NER model training process using both coarse- and fine-grained datasets. The CoFiNER training process consists of the following four steps:",
+        "bbox": [
+            112,
+            259,
+            487,
+            356
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Step 1- Training a fine-grained model. In the first step, we train a fine-grained model  $f^{\\mathcal{F}}(\\theta)$  with the low-resource fine-grained dataset  $\\mathcal{D}^{\\mathcal{F}}$ . This process follows a typical supervised learning approach for NER. For a training example  $(\\mathbf{X}^{\\mathcal{F}},\\mathbf{Y}^{\\mathcal{F}})$ ,  $\\mathbf{X}^{\\mathcal{F}}$  is fed into a PLM (Pre-trained Language Model), such as BERT, RoBERTa, to generate a contextual representations  $\\mathbf{h}_i\\in \\mathbb{R}^d$  of each token  $x_{i}$ .",
+        "bbox": [
+            112,
+            360,
+            489,
+            488
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {H} = \\left[ \\mathbf {h} _ {1}, \\dots , \\mathbf {h} _ {n} \\right] = \\mathsf {P L M} \\left(\\left[ x _ {1} ^ {\\mathcal {F}}, \\dots , x _ {n} ^ {\\mathcal {F}} \\right]\\right). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            498,
+            487,
+            517
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Then, we apply a softmax layer to obtain the label probability distribution:",
+        "bbox": [
+            112,
+            527,
+            485,
+            558
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {p} _ {i} ^ {\\mathcal {F}} = \\operatorname {s o f t m a x} (\\mathbf {W h} _ {i} + \\mathbf {b})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            200,
+            568,
+            400,
+            588
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{W} \\in \\mathbb{R}^{|E^{\\mathcal{F}}| \\times d}$  and  $\\mathbf{b} \\in \\mathbb{R}^{|E^{\\mathcal{F}}|}$  represent the weights and bias of the classification head, respectively. To train the model using a fine-grained dataset, we optimize the cross-entropy loss function:",
+        "bbox": [
+            112,
+            596,
+            489,
+            678
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {\\mathcal {F}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\mathbf {p} _ {i} ^ {\\mathcal {F}} \\left[ y _ {i} ^ {\\mathcal {F}} \\right] \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            201,
+            687,
+            487,
+            727
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $y_{i}\\in E^{\\mathcal{F}}$  is the fine-grained label for the token  $x_{i}$",
+        "bbox": [
+            112,
+            738,
+            485,
+            770
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Step 2 - Generating an F2C matrix. To fully leverage the hierarchy between coarse- and fine-grained entity types, we avoid training separate NER models for each dataset. Instead, we utilize a single model that incorporates an F2C mapping matrix that transforms a fine-grained output into a corresponding coarse-grained output. The F2C mapping matrix assesses the conditional probability of a coarse-grained entity type  $s \\in E^{\\mathcal{C}}$  given",
+        "bbox": [
+            112,
+            774,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "a fine-grained label  $\\ell \\in E^{\\mathcal{F}}$  (i.e.,  $\\mathbf{M}_{\\ell,s} = p(y^{\\mathcal{C}} = s|y^{\\mathcal{F}} = \\ell)$ ).",
+        "bbox": [
+            507,
+            82,
+            880,
+            115
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a fine-grained probability distribution  $\\mathbf{p}_i^{\\mathcal{F}}$  computed using the proposed model, the marginal probability of a coarse-grained type  $s$  can be computed as",
+        "bbox": [
+            507,
+            116,
+            882,
+            180
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {p} _ {i} ^ {\\mathcal {C}} [ s ] = \\sum_ {\\ell \\in E ^ {\\mathcal {F}}} p (y ^ {\\mathcal {C}} = s | y ^ {\\mathcal {F}} = \\ell) \\cdot \\mathbf {p} _ {i} ^ {\\mathcal {F}} [ \\ell ].\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            547,
+            191,
+            840,
+            225
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Thus, the coarse-grained output probabilities are simply computed as follows:",
+        "bbox": [
+            507,
+            236,
+            880,
+            267
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {p} _ {i} ^ {\\mathcal {C}} = \\mathbf {p} _ {i} ^ {\\mathcal {F}} \\cdot \\mathbf {M} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            640,
+            277,
+            882,
+            297
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{M} \\in \\mathbb{R}^{|E^{\\mathcal{F}}| \\times |E^{\\mathcal{C}}|}$  is the F2C mapping matrix whose row-wise sum is 1. By introducing this F2C mapping matrix, we can train a single model using multiple datasets with different granularity levels.",
+        "bbox": [
+            507,
+            309,
+            880,
+            375
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Manual annotation is a straightforward approach that can be used when hierarchical information is unavailable. However, it is not only cost-intensive but also noisy and subjective, especially when there are multiple coarse-grained datasets or a large number of fine-grained entity types. We introduce an efficient method for automatically generating an F2C matrix in §3.3.",
+        "bbox": [
+            507,
+            376,
+            882,
+            502
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Step 3 - Filtering inconsistent coarse labels. Although a fine-grained entity type is usually a subtype of a coarse-grained type, there can be some misalignments between the coarse- and fine-grained entity types. For example, an entity \"Microsoft\" in a financial document can either be tagged as Company or Stock which are not hierarchical. This inconsistency can significantly degrade the model's performance.",
+        "bbox": [
+            507,
+            508,
+            882,
+            652
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To mitigate the effect of inconsistent labeling, we devise an inconsistency filtering method aimed at masking less relevant coarse labels. By automatically filtering out the inconsistent labels from the coarse-grained dataset, we investigate the coarse-grained labels using the fine-grained NER model trained in Step 1. For each token in the coarse-grained dataset, we predict the coarse-grained label using the fine-grained model and the mapping matrix as follows:",
+        "bbox": [
+            507,
+            653,
+            884,
+            810
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {y} _ {i} ^ {\\mathcal {C}} = \\arg \\max  \\mathbf {p} _ {i} ^ {\\mathcal {C}}. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            623,
+            822,
+            882,
+            843
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "If the predicted label is the same as the coarse-grained label (i.e.,  $y_{i}^{\\mathcal{C}} = \\tilde{y}_{i}^{\\mathcal{C}}$ ), we assume that the coarse-grained label is consistent with the fine-grained one and can benefit the model. Otherwise,",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3271",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/1dc02e5b809b045ff1d5d338ca4cd39fe360119f51bd2c351449be365b81bec2.jpg",
+        "image_caption": [
+            "Step 1) training process using fine-grained dataset",
+            "Figure 1: Overall process of our proposed method."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            99,
+            873,
+            382
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "we regard the label as inconsistent with fine-grained types and do not utilize the coarse-grained label in Step 4. Note that the fine-grained NER model is frozen during this phase.",
+        "bbox": [
+            112,
+            437,
+            487,
+            501
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Step 4 - Jointly training CoFiNER with both datasets. The model is trained by alternating between epochs using coarse- and fine-grained data. This is an effective learning strategy for training models using heterogeneous datasets (Jung and Shim, 2020).",
+        "bbox": [
+            112,
+            508,
+            489,
+            602
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For the fine-grained batches, CoFiNER is trained by minimizing the loss function defined in Equation (2), as described in Step 1. Meanwhile, we use the F2C mapping matrix to generate coarse-grained outputs and utilize inconsistency filtering for coarse-grained batches. Thus, we compute the cross-entropy loss between the coarse-grained label  $y_{i}^{\\mathcal{C}}$  and the predicted probabilities  $\\mathbf{p}_i^{\\mathcal{C}}$  when the coarse-grained label is consistent with our model (i.e.,  $y_{i}^{\\mathcal{C}} = \\tilde{y}_{i}^{\\mathcal{C}}$ ) as follows:",
+        "bbox": [
+            112,
+            607,
+            489,
+            770
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {\\mathcal {C}} = - \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\mathbf {p} _ {i} ^ {\\mathcal {C}} \\left[ y _ {i} ^ {\\mathcal {C}} \\right] \\cdot \\mathbb {I} \\left[ y _ {i} ^ {\\mathcal {C}} = \\tilde {y} _ {i} ^ {\\mathcal {C}} \\right] \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            791,
+            487,
+            831
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $m$  is a length of  $\\mathbf{X}^{\\mathcal{C}}$ . For example, suppose that the coarse label of token  $x_{i}$  is  $y_{i}^{\\mathcal{C}} = 0\\mathrm{RG}$ . If the estimated coarse label  $\\tilde{y}_i^{\\mathcal{C}}$  is 0RG, the loss for the token is  $-\\log \\mathbf{p}_i^{\\mathcal{C}}[y_i^{\\mathcal{C}}]$ . Otherwise, it is zero.",
+        "bbox": [
+            112,
+            854,
+            487,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Construction of the F2C mapping matrix",
+        "text_level": 1,
+        "bbox": [
+            507,
+            437,
+            878,
+            453
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The F2C mapping matrix assesses the conditional probability of a coarse-grained entity type  $s \\in E^{\\mathcal{C}}$  given a fine-grained label  $\\ell \\in E^{\\mathcal{F}}$  (i.e.,  $\\mathbf{M}_{\\ell,s} = p(y^{\\mathcal{C}} = s|y^{\\mathcal{F}} = \\ell)$ ). As there are only a few identical texts with both coarse- and fine-grained annotations simultaneously, we can not directly calculate this conditional probability using the data alone. Thus, we approximate the probability using a coarse-grained NER model and fine-grained labeled data as follows:",
+        "bbox": [
+            507,
+            461,
+            884,
+            621
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {M} _ {\\ell , s} = p (y ^ {\\mathcal {C}} = s | y ^ {\\mathcal {F}} = \\ell) \\approx p (\\tilde {y} ^ {\\mathcal {C}} = s | y ^ {\\mathcal {F}} = \\ell).\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            623,
+            868,
+            642
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To generate the mapping matrix, we first train a coarse-grained NER model  $f^{\\mathcal{C}}(\\theta)$  with the coarse-grained dataset  $\\mathcal{D}^{\\mathcal{C}}$ . Then, we reannote the fine-grained dataset  $\\mathcal{D}^{\\mathcal{F}}$  by using the coarse-grained model  $f^{\\mathcal{C}}(\\theta)$ . As a result, we obtain parallel annotations for both coarse- and fine-grained types in the fine-grained data  $\\mathcal{D}^{\\mathcal{F}}$ . By using the parallel annotations, we can compute the co-occurrence matrix  $\\mathbf{C} \\in \\mathbb{N}^{|E^{\\mathcal{F}}| \\times |E^{\\mathcal{C}}|}$  where each cell  $\\mathbf{C}_{\\ell,s}$  is the number of tokens that are labeled as fine-grained type  $\\ell$  and coarse-grained type  $s$  together.",
+        "bbox": [
+            507,
+            644,
+            884,
+            820
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Because some labels generated by the coarse-grained model  $f^{\\mathcal{C}}(\\theta)$  can be inaccurate, we refine the co-occurrence matrix by retaining only the top- $k$  counts for each fine-grained type and setting the rest to zero. Our experiments show that our model performs best when  $k = 1$ . This process effic",
+        "bbox": [
+            507,
+            822,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3272",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tively retains only the most frequent coarse-grained categories for each fine-grained type, thereby improving the precision of the resulting mapping. Finally, we compute the conditional probabilities for all  $\\ell \\in E^{\\mathcal{F}}, s \\in E^{\\mathcal{C}}$  by using the co-occurrence counts as follows:",
+        "bbox": [
+            112,
+            84,
+            489,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {M} _ {\\ell , s} = p \\left(\\tilde {y} ^ {\\mathcal {C}} = s \\mid y ^ {\\mathcal {F}} = \\ell\\right) = \\frac {\\mathbf {C} _ {\\ell , s}}{\\sum_ {s ^ {\\prime} \\in E ^ {\\mathcal {C}}} \\mathbf {C} _ {\\ell , s ^ {\\prime}}}. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            187,
+            489,
+            223
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The F2C mapping matrix  $\\mathbf{M}$  is used to predict the coarse-grained labels using Equation (3).",
+        "bbox": [
+            112,
+            231,
+            487,
+            265
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            275,
+            260,
+            292
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            300,
+            228,
+            315
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/e79345dd5fa0595ceeb217925a4707dfdfb09308dbd11f93ad5f3022049f0a92.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"3\"># Sentences</td><td rowspan=\"2\"># Types</td></tr><tr><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td colspan=\"5\">Coarse-grained datasets</td></tr><tr><td>CoNLL&#x27;03</td><td>14k</td><td>3.3k</td><td>3.4k</td><td>4</td></tr><tr><td>OntoNotes</td><td>59.9k</td><td>8.5k</td><td>8.3k</td><td>18</td></tr><tr><td colspan=\"5\">Fine-grained dataset Few-NERD</td></tr><tr><td>10-Shot</td><td>0.3k</td><td></td><td></td><td></td></tr><tr><td>20-Shot</td><td>0.6k</td><td></td><td></td><td></td></tr><tr><td>40-Shot</td><td>1.1k</td><td>18.8k</td><td>37.6k</td><td>66</td></tr><tr><td>80-Shot</td><td>2.2k</td><td></td><td></td><td></td></tr><tr><td>100-Shot</td><td>2.7k</td><td></td><td></td><td></td></tr></table>",
+        "bbox": [
+            139,
+            330,
+            463,
+            486
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Statistics of each dataset.",
+        "bbox": [
+            183,
+            495,
+            415,
+            508
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments using a fine-grained NER dataset, Few-NERD (SUP) (Ding et al., 2021), as well as two coarse-grained datasets, namely, OntoNotes (Weischedel et al., 2013) and CoNLL'03 (Tjong Kim Sang, 2002). The fine-grained dataset Few-NERD comprises 66 entity types, whereas the coarse-grained datasets CoNLL'03 and OntoNotes consist of 4 and 18 entity types, respectively. The statistics for the datasets are listed in Table 1.",
+        "bbox": [
+            112,
+            529,
+            489,
+            689
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$K$ -shot sampling for the fine-grained dataset. Because we assumed a small number of examples for each label in the fine-grained dataset, we evaluated the performance in the  $K$ -shot learning setting. Although Few-NERD provides few-shot samples, they are obtained based on an  $N$ -way  $K$ -shot scenario, where  $N$  is considerably smaller than the total number of entity types. However, our goal is to identify named entities across all possible entity types. For this setting, we resampled  $K$ -shot examples to accommodate all-way  $K$ -shot scenarios.",
+        "bbox": [
+            112,
+            694,
+            489,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Since multiple entities exist in a single sentence, we cannot strictly generate exact  $K$ -shot samples for all the entity types. Therefore, we adopt the",
+        "bbox": [
+            112,
+            871,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$K\\sim (K + 5)$  -shot setting. In the  $K\\sim (K + 5)$  -shot setting, there are at least  $K$  examples and at most  $K + 5$  examples for each entity type. See Appendix A for more details. In our experiments, we sampled fine-grained training data for  $K = 10,20,40,80$  ,and 100.",
+        "bbox": [
+            507,
+            83,
+            884,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            193,
+            732,
+            209
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In experiments, we use transformer-based PLM, including  $\\mathrm{BERT}_{\\mathrm{BASE}}$ ,  $\\mathrm{RoBERTa}_{\\mathrm{BASE}}$ , and  $\\mathrm{RoBERTa}_{\\mathrm{LARGE}}$ . In CoFiNER, we follow  $\\mathrm{RoBERTa}_{\\mathrm{LARGE}}$  to build a baseline model. The maximum sequence length is set to 256 tokens. The AdamW optimizer (Loshchilov and Hutter, 2019) is used to train the model with a learning rate of  $2e - 5$  and a batch size of 16. The number of epochs is varied for each model. We train the fine-grained model, CoFiNER, over 30 epochs. To construct the F2C mapping matrix, the coarse-grained model is trained for 50 epochs using both CoNLL'03 and OntoNotes. To train the inconsistency filtering model, we set different epochs based on the number of shots: For 10, 20, 40, 80, and 100 shot settings, the epochs are 150, 150, 120, 50, and 30, respectively. We report the results using span-level F1. The dropout with a probability of 0.1 is applied. All the models were trained on NVIDIA RTX 3090 GPUs.",
+        "bbox": [
+            507,
+            216,
+            884,
+            536
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Compared Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            550,
+            714,
+            565
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this study, we compare the performance of CoFiNER with that of both the supervised and few-shot methods. We modified the existing methods for our experimental setup and re-implemented them accordingly.",
+        "bbox": [
+            507,
+            573,
+            882,
+            653
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Supervised method. We use  $\\mathrm{BERT}_{\\mathrm{BASE}}$ ,  $\\mathrm{RoBERTa}_{\\mathrm{BASE}}$ , and  $\\mathrm{RoBERTa}_{\\mathrm{LARGE}}$  as the supervised baselines, each including a fine-grained classifier on the head. In addition, PIQN (Shen et al., 2022) and PL-Marker (Ye et al., 2022) are methods that have achieved state-of-the-art performance in a supervised setting using the full Few-NERD dataset. All models are trained using only a Few-NERD dataset.",
+        "bbox": [
+            507,
+            658,
+            884,
+            800
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Few-shot method. The LSFS (Ma et al., 2022a) leverages a label encoder to utilize the semantics of label names, thereby achieving state-of-the-art results in low-resource NER settings. The LSFS applies a pre-finetuning strategy to learn prior knowledge from the coarse-grained dataset, OntoNotes. For a fair comparison, we also conducted pre",
+        "bbox": [
+            507,
+            806,
+            885,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3273",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/9e4735c022c6ffc69884b031d4dd4f93864fcb2b27c31f4668743c1ef742ef30.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Model</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>100-shot</td></tr><tr><td rowspan=\"5\">Supervised Method</td><td>BERTBASE (Devlin et al., 2019)</td><td>23.101</td><td>34.718</td><td>43.138</td><td>46.182</td><td>46.449</td></tr><tr><td>RoBERTaBASE (Liu et al., 2019)</td><td>21.073</td><td>34.157</td><td>39.226</td><td>45.735</td><td>46.942</td></tr><tr><td>RoBERTaLARGE (Liu et al., 2019)</td><td>29.137</td><td>41.425</td><td>51.137</td><td>55.057</td><td>54.172</td></tr><tr><td>PIQN (Shen et al., 2022)</td><td>21.750</td><td>22.007</td><td>28.533</td><td>29.339</td><td>38.658</td></tr><tr><td>PL-Marker (Ye et al., 2022)</td><td>40.902</td><td>48.064</td><td>52.395</td><td>53.249</td><td>53.061</td></tr><tr><td>Few-shot Method</td><td>LSFS (Ma et al., 2022a)</td><td>47.998</td><td>43.269</td><td>50.595</td><td>51.420</td><td>50.366</td></tr><tr><td>Proposed Method</td><td>CoFiNER</td><td>44.951</td><td>51.142</td><td>56.409</td><td>56.847</td><td>57.178</td></tr></table>",
+        "bbox": [
+            147,
+            80,
+            850,
+            200
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "finetuning on OntoNotes and performed fine-tuning on each shot of the Few-NERD dataset.",
+        "bbox": [
+            112,
+            250,
+            485,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Proposed method. We trained our CoFiNER model as proposed in §3. In each epoch, CoFiNER is first trained on two coarse-grained datasets: OntoNotes and CoNLL'03. Subsequently, it is trained on the fine-grained dataset Few-NERD. We used RoBERTa $_{\\text{LARGE}}$  as the pre-trained language model for CoFiNER in Equation (1).",
+        "bbox": [
+            112,
+            291,
+            485,
+            404
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            432,
+            265,
+            447
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2 reports the performance of CoFiNER and existing methods. The result shows that CoFiNER outperforms both supervised learning and few-shot learning methods. Because supervised learning typically needs a large volume of training data, these models underperform in low-resource settings. This demonstrates that our method effectively exploits coarse-grained datasets to enhance the performance of the fine-grained NER model. In other words, the proposed F2C mapping matrix significantly reduces the amount of fine-grained dataset required to train supervised NER models. In particular, CoFiNER achieves significant performance improvements compared to the state-of-the-art model PL-Marker, which also utilizes the same pre-trained language model RoBERTa $_{\\text{LARGE}}$  as CoFiNER.",
+        "bbox": [
+            112,
+            463,
+            489,
+            734
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The few-shot method LSFS yields the highest F1 score for the 10-shot case. However, this few-shot method suffers from early performance saturation, resulting in less than a 2.4 F1 improvement with an additional 90-shot. By contrast, the F1 score of CoFiNER increases by 12.2. Consequently, CoFiNER outperforms all the compared methods except for the 10-shot case. In summary, the proposed method yields promising results for a wide range of data sample sizes by explicitly leveraging the inherent hierarchical structure.",
+        "bbox": [
+            112,
+            741,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/0dd3b23c9414f1b74b790b57d2347285d7c469cf1b8d0a72c971759fcfea046d.jpg",
+        "table_caption": [
+            "Table 2: Results on few-shot NER. The best scores across all models are marked bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>100-shot</td></tr><tr><td>CoFiNER</td><td>44.95</td><td>51.14</td><td>56.41</td><td>56.85</td><td>57.18</td></tr><tr><td>w/o filtering</td><td>41.38</td><td>43.87</td><td>53.42</td><td>54.70</td><td>55.13</td></tr><tr><td>w/o OntoNotes</td><td>42.63</td><td>44.51</td><td>57.18</td><td>56.97</td><td>55.54</td></tr><tr><td>w/o CoNLL&#x27;03</td><td>43.99</td><td>50.07</td><td>57.54</td><td>56.71</td><td>56.27</td></tr><tr><td>w/o coarse</td><td>29.14</td><td>41.43</td><td>51.14</td><td>55.06</td><td>54.17</td></tr></table>",
+        "bbox": [
+            512,
+            247,
+            882,
+            332
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Performances of ablation study over different components.",
+        "bbox": [
+            507,
+            342,
+            880,
+            370
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.5 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            403,
+            672,
+            419
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "An ablation study is conducted to validate the effectiveness of each component of the proposed method. The results are presented in Table 3. First, we remove the inconsistency filtering  $(w / o$  filtering) and observe a significant decrease in the F1 score, ranging from 2.05 to 7.27. These results demonstrate the effectiveness of our filtering method, which excludes mislabeled entities. Second, we provide the results using a single coarse-grained dataset (w/o OntoNotes and w/o CoNLL'03). Even with a single coarse-grained dataset, our proposed method significantly outperforms w/o coarse, which is trained solely on the fine-grained dataset (i.e. RoBERTaLarge in Table 2).",
+        "bbox": [
+            505,
+            431,
+            884,
+            671
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This indicates the effectiveness of using a well-aligned hierarchy through the F2C mapping matrix and inconsistency filtering. Although we achieve a significant improvement even with a single coarse-grained dataset, we achieve a more stable result with two coarse-grained datasets. This implies that our approach effectively utilizes multiple coarse-grained datasets, although the datasets contain different sets of entity types.",
+        "bbox": [
+            507,
+            676,
+            884,
+            821
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.6 Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            843,
+            623,
+            859
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we experimentally investigate how the F2C mapping matrix and inconsistency filtering improve the accuracy of the proposed model.",
+        "bbox": [
+            507,
+            871,
+            880,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3274",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/6438e8c4c7287b9e5ab4a53f5e279aaa5f071c5b156d36d50978ade55f2201ac.jpg",
+        "image_caption": [
+            "Figure 2: The F2C mapping matrix between OntoNotes and 100-shot of the Few-NERD dataset. During the generation of the F2C mapping matrix, some coarse-grained types are not mapped to any fine-grained types. The 7 unmapped types are not represented: DATE, TIME, PERCENT, MONEY, QUANTITY, ORDINAL, and CARDINAL."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            142,
+            80,
+            860,
+            272
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b838be05e6f9191f82309e9b94331af3f976a4094b598b435f389ba3429b65ed.jpg",
+        "image_caption": [
+            "Figure 3: The F2C mapping matrix between CoNLL'03 and 100-shot of the Few-NERD dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            142,
+            335,
+            860,
+            480
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.6.1 F2C Mapping Matrix",
+        "text_level": 1,
+        "bbox": [
+            112,
+            529,
+            344,
+            545
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Mapping the entity types between the coarse- and fine-grained datasets directly affects the model performance. Hence, we investigate the mapping outcomes between the coarse- and fine-grained datasets. Figures 2 and 3 show the F2C matrices for FewNERD-OntoNotes and FewNERD-CoNLL'03, respectively. In both figures, the x- and y-axis represent the fine-grained and coarse-grained entity types, respectively. The colored text indicates the corresponding coarse-grained entity types in FewNERD, which were not used to find the mapping matrix. The mapping is reliable if we compare the y-axis and the colored types (coarse-grained types in Few-NERD). Even without manual annotation of the relationship between coarse- and fine-grained entities, our method successfully obtains reliable mapping from fine-grained to coarse-grained types. Our model can be effectively trained with both types of datasets using accurate F2C mapping.",
+        "bbox": [
+            112,
+            549,
+            489,
+            853
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 provides the F1 scores by varying the hyperparameter  $k$  to refine the F2C mapping matrix described in §3.3. 'all' refers to the usage of the complete frequency distribution when creating an",
+        "bbox": [
+            112,
+            854,
+            487,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c3bc998e35fe8e0f869651ee62c65766952daacd568e73e3190ad7b4ea60e9bb.jpg",
+        "image_caption": [
+            "Figure 4: Impact of top- $k$  values in the F2C mapping matrix. F1 is reported on the test data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            529,
+            820,
+            644
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "F2C mapping matrix. We observe that the highest performance is achieved when  $k$  is set to 1, and as  $k$  increases, the performance gradually decreases. Although the optimal value of  $k$  can vary depending on the quality of the coarse-grained data and the performance of the coarse-grained NER model, the results indicate that a good F2C mapping matrix can be obtained by ignoring minor co-occurrences.",
+        "bbox": [
+            507,
+            721,
+            884,
+            851
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Additionally, We conducted an experiment by setting the F2C mapping matrix to be learnable and comparing it with our non-learnable F2C matrix. The non-learnable approach showed better",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3275",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/51a7e9371ac336e5f10ad5c92d2104c0b1f716d05c28c2d80a91119497a0f85f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>#</td><td>Type</td><td>Example</td></tr><tr><td rowspan=\"2\">1</td><td>predict</td><td>... they may be offshoots of the intifadahEVENT, the Palestinian rebellionEVENT in ...</td></tr><tr><td>target</td><td>... they may be offshoots of the intifadah, the PalestinianNORP rebellion in ...</td></tr><tr><td rowspan=\"2\">2</td><td>predict</td><td>... players heading to CanadaGPE, particularly TorontoGPE for the All-starEVENT ...</td></tr><tr><td>target</td><td>... players heading to CanadaGPE, particularly TorontoGPE for the All-star ...</td></tr><tr><td rowspan=\"2\">3</td><td>predict</td><td>Judges at Flight 103 Lockerbie trialEVENT are expected ...</td></tr><tr><td>target</td><td>Judges at Flight 103 LockerbieGPE trial are expected ...</td></tr></table>",
+        "bbox": [
+            186,
+            80,
+            811,
+            198
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Inconsistent examples: The labeled sentence in the original coarse-grained dataset is the \"target\" type, while the label predicted by the model is the \"predict\" type. Consistent entity types are indicated in blue, while inconsistent entity types are indicated in red.",
+        "bbox": [
+            112,
+            206,
+            882,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "performance, hence we adopted this approach for CoFiNER. Detailed analysis and experiment results are in Appendix B.2.",
+        "bbox": [
+            112,
+            274,
+            489,
+            324
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6.2 Inconsistency Filtering",
+        "text_level": 1,
+        "bbox": [
+            112,
+            334,
+            354,
+            350
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We aim to examine whether inconsistency filtering successfully screened out inconsistent entities between OntoNotes and Few-NERD datasets. To conduct this analysis, we describe three inconsistent examples. Table 4 shows the predicted values and target labels, which correspond to the coarse-grained output of the fine-grained NER model trained as described in §3.2 and the golden labels of the coarse-grained dataset.",
+        "bbox": [
+            112,
+            354,
+            489,
+            495
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The first example illustrates the inconsistency in entity types between mislabeled entities. In the original coarse-grained dataset, \"Palestinian\" is labeled as NORP, but the model trained on the fine-grained dataset predicts \"Palestinian rebellion\" as its appropriate label, EVENT. However, annotators of the OntoNotes labeled \"Palestinian\" as NORP, whereas the fine-grained NER model correctly predicts the highly informative actual label span. The inconsistency caused by a label mismatch between the coarse-grained and fine-grained datasets can result in performance degradation.",
+        "bbox": [
+            112,
+            499,
+            489,
+            690
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In the second example, both \"Canada\" and \"Toronto\" are consistently labeled as GPE; thus, the model is not confused when training on these two entities. However, in the case of \"All-star\", we can observe a mismatch. This example in the coarse-grained dataset is labeled 0 instead of the correct entity type EVENT, indicating a mismatch. Through inconsistency filtering, unlabeled \"All-star\" is masked out of the training process.",
+        "bbox": [
+            112,
+            693,
+            489,
+            837
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in the examples, inconsistency filtering is necessary to mitigate the potential noise arising from mismatched entities. We analyze the filtering results for each coarse-grained label to assess its impact on model performance. Figure 5",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/da7155def09c630523ce54071fbd3f1ebb60dd871854b671c40764ccb6be0504.jpg",
+        "image_caption": [
+            "Figure 5: Performance with increasing filtering proportion between Few-NERD and OntoNotes datasets. The correlation coefficient between filtering proportion and performance improvement is 0.29."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            274,
+            878,
+            479
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "illustrates the correlation between the filtering proportion and the performance improvement for each coarse-grained label mapped to the fine-grained labels. In this figure, a higher filtering proportion indicates a greater inconsistency between the two datasets. F1 improvements indicate the difference in performance when filtering is applied and when it is not. Each data point refers to a fine-grained label mapped onto a coarse-grained label. As the proportion of the filtered entities increases, the F1 scores also increase. These improvements indicate that inconsistency filtering effectively eliminates noisy entities, enabling the model to be well-trained on consistent data only.",
+        "bbox": [
+            507,
+            573,
+            882,
+            797
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            812,
+            640,
+            827
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We proposed CoFiNER, which explicitly leverages the hierarchical structure between coarse- and fine-grained types to alleviate the low-resource problem of fine-grained NER. We devised the F2C mapping matrix that allows for fine-grained NER model",
+        "bbox": [
+            507,
+            839,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3276",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "training using additional coarse-grained datasets. However, because not all coarse-grained entities are beneficial for the fine-grained NER model, the proposed inconsistency filtering method is used to mask out noisy entities from being used in the model training process. We found through experiments that using a smaller amount of consistent data is better than using a larger amount of data without filtering, thus demonstrating the crucial role of inconsistency filtering. Empirical results confirmed the superiority of CoFiNER over both the supervised and few-shot methods.",
+        "bbox": [
+            112,
+            84,
+            490,
+            277
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            290,
+            220,
+            305
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Despite the promising empirical results of this study, there is still a limitation. The main drawback of CoFiNER is that the token-level F2C mapping matrix and inconsistency filtering may not be directly applicable to nested NER tasks. Nested NER involves identifying and classifying certain overlapping entities that exceed the token-level scope. Because CoFiNER addresses fine-grained NER at the token level, it may not accurately capture entities with nested structures. Therefore, applying our token-level approach to nested NER could pose challenges and may require further adaptations or other modeling techniques to effectively handle the hierarchical relations between nested entities.",
+        "bbox": [
+            112,
+            317,
+            489,
+            541
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            555,
+            278,
+            571
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF2022R1G1A1013549). This work was supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No. RS2023-00261068, Development of Lightweight Multimodal Anti-Phishing Models and Split-Learning Techniques for Privacy-Preserving Anti-Phishing) and (No.RS-2022-00155885, Artificial Intelligence Convergence Innovation Human Resources Development (Hanyang University ERICA)). This research was also supported by the Ministry of Trade, Industry, and Energy (MOTIE), Korea, under the \"Project for Research and Development with Middle Markets Enterprises and DNA(Data, Network, AI) Universities\" (Smart Home Based LifeCare Service & Solution Development)(P0024555) supervised by the Korea Institute for Advancement of Technology (KIAT).",
+        "bbox": [
+            112,
+            581,
+            490,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            83,
+            608,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Armen Aghajanyan, Anchit Gupta, Akshat Shrivastava, Xilun Chen, Luke Zettlemoyer, and Sonal Gupta. 2021. Muppet: Massive multi-task representations with pre-finetuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5799-5811, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Sarkar Snigdha Sarathi Das, Arzoo Katiyar, Rebecca J Passonneau, and Rui Zhang. 2021. Container: Few-shot named entity recognition via contrastive learning. arXiv preprint arXiv:2109.07589.",
+            "Leon Derczynski, Eric Nichols, Marieke van Erp, and Nut Limsopatham. 2017. Results of the WNUT2017 shared task on novel and emerging entity recognition. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 140-147, Copenhagen, Denmark. Association for Computational Linguistics.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Ning Ding, Guangwei Xu, Yulin Chen, Xiaobin Wang, Xu Han, Pengjun Xie, Haitao Zheng, and Zhiyuan Liu. 2021. Few-NERD: A few-shot named entity recognition dataset. 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 3198-3213, Online. Association for Computational Linguistics.",
+            "Maximilian Hofer, Andrey Kormilitzin, Paul Goldberg, and Alejo Nevada-Holgado. 2018. Few-shot learning for named entity recognition in medical text.",
+            "Yucheng Huang, Kai He, Yige Wang, Xianli Zhang, Tieliang Gong, Rui Mao, and Chen Li. 2022. Copner: Contrastive learning with prompt guiding for few-shot named entity recognition. In Proceedings of the 29th International conference on computational linguistics, pages 2515-2527.",
+            "Haoming Jiang, Danqing Zhang, Tianyu Cao, Bing Yin, and Tuo Zhao. 2021. Named entity recognition with small strongly labeled and large weakly labeled data.",
+            "Woohwan Jung and Kyuseok Shim. 2020. Dual supervision framework for relation extraction with distant supervision and human annotation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6411-6423, Barcelona, Spain (Online). International Committee on Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            105,
+            885,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3277",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "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.",
+            "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.",
+            "Lefteris Loukas, Manos Fergadiotis, Ilias Chalkidis, Eirini Spyropoulou, Prodromos Malakasiotis, Ion Androutsopoulos, and Georgios Paliouras. 2022. Finer: Financial numeric entity recognition for xbrl tagging. arXiv preprint arXiv:2203.06482.",
+            "Jie Ma, Miguel Ballesteros, Srikanth Doss, Rishita Anubhai, Sunil Mallya, Yaser Al-Onaizan, and Dan Roth. 2022a. Label semantics for few shot named entity recognition. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1956-1971, Dublin, Ireland. Association for Computational Linguistics.",
+            "Tingting Ma, Huiqiang Jiang, Qianhui Wu, Tiejun Zhao, and Chin-Yew Lin. 2022b. Decomposed meta-learning for few-shot named entity recognition. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 1584–1596.",
+            "Alexis Mitchell, Stephanie Strassel, Shudong Huang, and Ramez Zakhary. 2005. Ace 2004 multilingual training corpus. Linguistic Data Consortium, Philadelphia, 1:1-1.",
+            "Seongwoong Oh, Woohwan Jung, and Kyuseok Shim. 2023. Thunder: Named entity recognition using a teacher-student model with dual classifiers for strong and weak supervisions. In Proceedings of 26th European Conference on Artificial Intelligence.",
+            "Alan Ritter, Sam Clark, Mausam, and Oren Etzioni. 2011. Named entity recognition in tweets: An experimental study. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1524-1534, Edinburgh, Scotland, UK. Association for Computational Linguistics.",
+            "Yongliang Shen, Xiaobin Wang, Zeqi Tan, Guangwei Xu, Pengjun Xie, Fei Huang, Weiming Lu, and Yueting Zhuang. 2022. Parallel instance query network for named entity recognition. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 947-961, Dublin, Ireland. Association for Computational Linguistics.",
+            "Mujeen Sung, Minbyul Jeong, Yonghwa Choi, Donghyeon Kim, Jinhyuk Lee, and Jaewoo Kang. 2022. Bern2: an advanced neural biomedical named entity recognition and normalization tool. Bioinformatics, 38(20):4837-4839.",
+            "Erik F. Tjong Kim Sang. 2002. Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition. In COLING-02: The 6th"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Conference on Natural Language Learning 2002 (CoNLL-2002).",
+            "Christopher Walker, Stephanie Strassel, Julie Medero, and Kazuaki Maeda. 2006. Ace 2005 multilingual training corpus. Linguistic Data Consortium, Philadelphia, 57:45.",
+            "Peiyi Wang, Runxin Xu, Tianyu Liu, Qingyu Zhou, Yunbo Cao, Baobao Chang, and Zhifang Sui. 2022. An enhanced span-based decomposition method for few-shot sequence labeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5012-5024.",
+            "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Edward Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, et al. 2013. Ontonotes release 5.0 ldc2013t19. Linguistic Data Consortium, Philadelphia, PA, 23.",
+            "Deming Ye, Yankai Lin, Peng Li, and Maosong Sun. 2022. Packed levitated marker for entity and relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4904-4917, Dublin, Ireland. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            441
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "text_level": 1,
+        "bbox": [
+            510,
+            453,
+            596,
+            470
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Sampling Algorithm",
+        "text_level": 1,
+        "bbox": [
+            510,
+            478,
+            724,
+            495
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 1  $K\\sim (K + 5)$  sampling algorithm",
+        "bbox": [
+            510,
+            518,
+            847,
+            533
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Input: Dataset  $\\mathbf{X}$ , labeled set  $\\mathbf{Y}$ ,  $K$",
+        "bbox": [
+            542,
+            536,
+            808,
+            550
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Output: Fine-grained dataset  $D^{\\mathcal{F}}$",
+        "bbox": [
+            542,
+            552,
+            796,
+            567
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1: Sort classes in  $\\mathbf{Y}$  based on their freq. in  $\\mathbf{X}$",
+            "2:  $D^{\\mathcal{F}}\\gets \\emptyset$  Init the train dataset  $\\triangleright$  Init the count of entity classes in  $D^{\\mathcal{F}}$",
+            "3: for  $i = 1$  to  $|\\mathbf{Y}|$  do",
+            "4: Count  $[i] = 0$",
+            "5: for  $i = 1$  to  $|\\mathbf{Y}|$  do",
+            "6: if  $\\forall \\operatorname{Count}[i'] \\geq K$  then",
+            "7: break;",
+            "8: Randomly sample  $(x,y)\\in \\mathbf{X}$  s.t.  $\\mathbf{Y}_i\\in y$",
+            "9: for  $j = 1$  to  $|y|$  do",
+            "10:  $\\mathbf{Count}^{\\mathbf{Y}}[j] + = 1;$",
+            "11: if  $\\exists \\operatorname{Count}[i] + \\operatorname{Count}^{\\mathbf{Y}}[i] > K + 5$  then",
+            "12: Continue;",
+            "13: else",
+            "14:  $D^{\\mathcal{F}}\\gets D^{\\mathcal{F}}\\cup \\{(x,y)\\} ;$",
+            "15: Count  $\\leftarrow$  Count + CountY;",
+            "16: return  $D^{\\mathcal{F}}$"
+        ],
+        "bbox": [
+            515,
+            570,
+            880,
+            852
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Unlike other tasks, NER involves multiple entity occurrences within a sentence, making it too restrictive to sample an exact count. We adopt a",
+        "bbox": [
+            510,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3278",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "$K \\sim (K + 5)$ -shot setting to minimize differences in the number of entity types. Additionally, we sample the entity types, starting from those with fewer occurrences to ensure a balanced distribution of multiple entity types within the sentences. Algorithm 1 presents the  $K$ -shot sampling algorithm used in this study.",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Additional Experiments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            209,
+            361,
+            225
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B.1 Generalization of Inconsistency filtering in diverse dataset settings",
+        "text_level": 1,
+        "bbox": [
+            114,
+            235,
+            478,
+            267
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/7a4145e368ac7852c027cf322464a0a93e7771f0e77b3c1a1cf50b62933dae81.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Setting</td><td colspan=\"2\">Coarse-grained datasets</td></tr><tr><td>Few-NERD_coarse</td><td>OntoNotes + CoNLL&#x27;03</td></tr><tr><td>w/ filtering</td><td>56.63</td><td>57.18</td></tr><tr><td>w/o filtering</td><td>56.46</td><td>55.13</td></tr></table>",
+        "bbox": [
+            117,
+            282,
+            487,
+            349
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "To assess the generalization of the proposed method, we conducted experiments under various coarse- and fine-grained dataset settings.",
+        "bbox": [
+            112,
+            403,
+            485,
+            450
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "First, we verified the robustness of inconsistency filtering through experiments using different coarse-grained datasets. The fine-grained dataset, Few-NERD, remains unchanged. Since Few-NERD has both coarse- and fine-grained labels, we constructed a coarse-grained dataset Few-NERD_coarse using the coarse-grained labels from Few-NERD. When compared to Few-NERD_coarse, entities in OntoNotes and CoNLL'03 are inconsistently labeled with Few-NERD because they were independently created. In Table 5, Few-NERD_coarse exhibits a higher F1 score in the w/o filtering setting due to its consistency with fine-grained labels of Few-NERD. However, the performance improvement achieved with filtering is more substantial in the inconsistent datasets, when compared to the consistent dataset. This result indicates that inconsistency filtering improves performance by filtering out the mismatching labels. Therefore, we have demonstrated the importance of using inconsistency filtering to filter out noise when working with datasets that employ different labeling schemes. Furthermore, by achieving effective performance improvements across various coarse-grained datasets, we have provided evidence of the robustness of the filtering method.",
+        "bbox": [
+            112,
+            451,
+            489,
+            868
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Second, we validate the generalization performance through experiments conducted in different coarse- and fine-grained dataset settings. We",
+        "bbox": [
+            112,
+            871,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/ee54a497bf605476149f669d975a10e6cf010a3803e0de5738371b1ab7f1f5e9.jpg",
+        "table_caption": [
+            "Table 5: Results on different coarse-grained datasets. A fine-grained dataset is 100-shot of the Few-NERD."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>F1</td></tr><tr><td>RoBERTaLARGE</td><td>75.15</td></tr><tr><td>PL-Marker</td><td>74.03</td></tr><tr><td>CoFiNER</td><td>80.44</td></tr><tr><td>w/o filtering</td><td>78.20</td></tr><tr><td>w/o coarse</td><td>75.15</td></tr></table>",
+        "bbox": [
+            598,
+            80,
+            794,
+            187
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "set up the CoNLL'03, which has 4 entity types, as the coarse-grained dataset, and the 100-shot OntoNotes, which has a finer label with 19 entity types, as the fine-grained dataset. In Table 6, when compared to the two top-performing models in the main results, RoBERTa $_{\\text{LARGE}}$  and the state-of-the-art PL-Marker, CoFiNER show consistently higher performance. Furthermore, as shown in the ablation study that was conducted following the same methodology in §4.5, CoFiNER exhibited superior performance.",
+        "bbox": [
+            507,
+            265,
+            884,
+            439
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In conclusion, through above the two experiments, our method has been demonstrated to work robustly across different coarse- and fine-grained dataset settings.",
+        "bbox": [
+            507,
+            442,
+            882,
+            505
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B.2 Comparison with learnable F2C mapping matrix",
+        "text_level": 1,
+        "bbox": [
+            507,
+            516,
+            882,
+            546
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/dd66ebc9a4a39ca0079e3e7a86273abaff2de0a1ab0da5529e25869d341c299f.jpg",
+        "table_caption": [
+            "Table 6: Performances of different models and ablation studies on our model. RoBERTa $_{\\text{LARGE}}$  and w/o coarse are identical."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Matrix Type</td><td>F1</td></tr><tr><td>non-learnable</td><td>56.27</td></tr><tr><td>learnable</td><td>53.25</td></tr></table>",
+        "bbox": [
+            603,
+            561,
+            789,
+            618
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Results on learnable and non-learnable F2C mapping matrix on 100-shot of Few-NERD.",
+        "bbox": [
+            507,
+            627,
+            882,
+            657
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "To find the optimal F2C mapping matrix, we conducted experiments to explore the impact of making the F2C mapping matrix learnable. We use FewNERD as a fine-grained dataset and OntoNotes as a coarse-grained dataset. Table 7 shows no performance gains when the F2C mapping matrix was set to be learnable. We found that the learnable matrix tends to form a pattern similar to what is shown in Figure 4 with  $k =$  all. This result suggests that taking minor co-occurrences into account leads to an overall decrease in performance. Based on this analysis, the non-learnable mapping matrix is used in our experiments.",
+        "bbox": [
+            507,
+            675,
+            882,
+            884
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "3279",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2023/Enhancing Low-resource Fine-grained Named Entity Recognition by Leveraging Coarse-grained Datasets/dff73b45-756f-4d7a-a8f6-789c5d1784de_model.json

@@ -0,0 +1,2257 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.145,
+                0.09,
+                0.856,
+                0.131
+            ],
+            "angle": 0,
+            "content": "Enhancing Low-resource Fine-grained Named Entity Recognition by Leveraging Coarse-grained Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.286,
+                0.149,
+                0.717,
+                0.168
+            ],
+            "angle": 0,
+            "content": "Su Ah Lee\\(^{1\\dagger}\\), Seokjin Oh\\(^{1\\dagger}\\), and Woohwan Jung\\(^{1,2}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.23,
+                0.168,
+                0.774,
+                0.185
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)Department of Applied Artificial Intelligence, Hanyang University"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.447,
+                0.185,
+                0.555,
+                0.201
+            ],
+            "angle": 0,
+            "content": "2Ramply Inc."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.295,
+                0.202,
+                0.71,
+                0.218
+            ],
+            "angle": 0,
+            "content": "{sue991, seokjinoh, whjung}@hanyang.ac.kr"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.268
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.282,
+                0.461,
+                0.678
+            ],
+            "angle": 0,
+            "content": "Named Entity Recognition (NER) frequently suffers from the problem of insufficient labeled data, particularly in fine-grained NER scenarios. Although \\( K \\)-shot learning techniques can be applied, their performance tends to saturate when the number of annotations exceeds several tens of labels. To overcome this problem, we utilize existing coarse-grained datasets that offer a large number of annotations. A straightforward approach to address this problem is prefinetuning, which employs coarse-grained data for representation learning. However, it cannot directly utilize the relationships between fine-grained and coarse-grained entities, although a fine-grained entity type is likely to be a subcategory of a coarse-grained entity type. We propose a fine-grained NER model with a Fineto-Coarse(F2C) mapping matrix to leverage the hierarchical structure explicitly. In addition, we present an inconsistency filtering method to eliminate coarse-grained entities that are inconsistent with fine-grained entity types to avoid performance degradation. Our experimental results show that our method outperforms both \\( K \\)-shot learning and supervised learning methods when dealing with a small number of fine-grained annotations. Code is available at https://github.com/sue991/CoFiNER."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.693,
+                0.26,
+                0.708
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.719,
+                0.491,
+                0.897
+            ],
+            "angle": 0,
+            "content": "Named Entity Recognition (NER) is a fundamental task in locating and categorizing named entities in unstructured texts. Most research on NER has been conducted on coarse-grained datasets, including CoNLL'03 (Tjong Kim Sang, 2002), ACE04 (Mitchell et al., 2005), ACE05 (Walker et al., 2006), and OntoNotes (Weischedel et al., 2013), each of which has less than 18 categories. As the applications of NLP broaden across diverse fields, there is increasing demand for fine-grained NER that can provide more precise and"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.254,
+                0.886,
+                0.638
+            ],
+            "angle": 0,
+            "content": "detailed information extraction. Nonetheless, detailed labeling required for large datasets in the context of fine-grained NER presents several significant challenges. It is more cost-intensive and time-consuming than coarse-grained NER. In addition, it requires a high degree of expertise because domain-specific NER tasks such as financial NER (Loukas et al., 2022) and biomedical NER (Sung et al., 2022) require fine-grained labels. Thus, fine-grained NER tasks typically suffer from the data scarcity problem. Few-shot NER approaches (Ding et al., 2021) can be applied to conduct fine-grained NER with scarce fine-grained data. However, these methods do not exploit existing coarse-grained datasets that can be leveraged to improve fine-grained NER because fine-grained entities are usually subtypes of coarse-grained entities. For instance, if a model knows what Organization is, it could be easier for it to understand the concept of Government or Company. Furthermore, these methods often experience early performance saturation, necessitating the training of a new supervised learning model if the annotation extends beyond several tens of labels."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.643,
+                0.886,
+                0.803
+            ],
+            "angle": 0,
+            "content": "A pre-finetuning strategy (Aghajanyan et al., 2021; Ma et al., 2022a) was proposed to overcome the aforementioned problem. This strategy first learns the feature representations using a coarse-grained dataset before training the fine-grained model on a fine-grained dataset. In this method, coarse-grained data are solely utilized for representation learning; thus, it still does not explicitly utilize the relationships between the coarse- and fine-grained entities."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.886,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Our intuition for fully leveraging coarse-grained datasets comes mainly from the hierarchy between coarse- and fine-grained entity types. Because a coarse-grained entity type typically comprises multiple fine-grained entity types, we can enhance low-resource fine-grained NER with abundant coarse-grained data. To jointly utilize both datasets, we"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.148,
+                0.904,
+                0.372,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Major in Bio Artificial Intelligence"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3269"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.16,
+                0.946,
+                0.838,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3269-3279 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.18
+            ],
+            "angle": 0,
+            "content": "devise a method to build a mapping matrix called F2C (short for 'Fine-to-Coarse'), which connects fine-grained entity types to their corresponding coarse-grained entity types and propose a novel approach to train a fine-grained NER model with both datasets."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.182,
+                0.49,
+                0.406
+            ],
+            "angle": 0,
+            "content": "Some coarse-grained entities improperly match a fine-grained entity type because datasets may be created by different annotators for different purposes. These mismatched entities can reduce the performance of the model during training. To mitigate this problem, coarse-grained entities that can degrade the performance of fine-grained NER must be eliminated. Therefore, we introduce a filtering method called 'Inconsistency Filtering'. This approach is designed to identify and exclude any inconsistent coarse-to-fine entity mappings, ensuring a higher-quality training process, and ultimately, better model performance. The main contributions of our study are as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.419,
+                0.489,
+                0.466
+            ],
+            "angle": 0,
+            "content": "- We propose an F2C mapping matrix to directly leverage the intimate relation between coarse- and fine-grained types."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.478,
+                0.489,
+                0.526
+            ],
+            "angle": 0,
+            "content": "- We present an inconsistency filtering method to screen out coarse-grained data that are inconsistent with the fine-grained types."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.538,
+                0.489,
+                0.601
+            ],
+            "angle": 0,
+            "content": "- The empirical results show that our method achieves state-of-the-art performance by utilizing the proposed F2C mapping matrix and the inconsistency filtering method."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.419,
+                0.489,
+                0.601
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.613,
+                0.271,
+                0.629
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.643,
+                0.49,
+                0.834
+            ],
+            "angle": 0,
+            "content": "Fine-grained NER. NER is a key task in information extraction and has been extensively studied. Traditional NER datasets (Tjong Kim Sang, 2002; Ritter et al., 2011; Weischedel et al., 2013; Derczynski et al., 2017) address coarse-grained entity types in the general domain. Recently, domain-specific NER has been investigated in various fields (Hofer et al., 2018; Loukas et al., 2022; Sung et al., 2022). The domain-specific NER tasks typically employ fine-grained entity types. In addition, Ding et al. (2021) proposed a general domain fine-grained NER."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.839,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "\\(N\\)-way \\(K\\)-shot learning for NER. Since labeling domain-specific data is an expensive process, few-shot NER has gained attention. Most few-shot NER studies are conducted using an \\(N\\)-way \\(K\\)-shot episode learning (Das et al., 2021; Huang"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.326
+            ],
+            "angle": 0,
+            "content": "et al., 2022; Ma et al., 2022b; Wang et al., 2022). The objective of this approach is to train a model that can correctly classify new examples into one of \\(N\\) classes, using only \\(K\\) examples per class. To achieve this generalization performance, a large number of episodes need to be generated. Furthermore, to sample these episodes, the training data must contain a significantly larger number of classes than \\(N\\). In contrast, in our problem context, the number of fine-grained classes we aim to identify is substantially larger than the number of coarse-grained classes in the existing dataset. Consequently, the episode-based few-shot learning approaches are unsuitable for addressing this problem."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.33,
+                0.885,
+                0.588
+            ],
+            "angle": 0,
+            "content": "Leveraging auxiliary data for information extraction. Several studies have employed auxiliary data to overcome the data scarcity problem. Jiang et al. (2021); Oh et al. (2023) propose NER models trained with small, strongly labeled, and large weakly labeled data. Jung and Shim (2020) used strong and weak labels for relation extraction. However, in the previous studies, the main data and auxiliary data shared the same set of classes, which is not the case for fine-grained NER with coarse-grained labels. While the approaches of Aghajanyan et al. (2021) and Ma et al. (2022a) can be applied to our problem setting, they utilize auxiliary data for representation learning rather than explicitly utilizing the relationship between the two types of data."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.599,
+                0.698,
+                0.615
+            ],
+            "angle": 0,
+            "content": "3 Proposed Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.624,
+                0.884,
+                0.705
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce the notations and define the problem of fine-grained NER using coarse-grained data. Then, we introduce the proposed CoFiNER model, including the creation of the F2C mapping matrix."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.715,
+                0.704,
+                0.729
+            ],
+            "angle": 0,
+            "content": "3.1 Problem definition"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.736,
+                0.885,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Given a sequence of \\(n\\) tokens \\(\\mathbf{X} = \\{x_{1},x_{2},\\ldots ,x_{n}\\}\\), the NER task involves assigning type \\(y_{i}\\in E\\) to each token \\(x_{i}\\) where \\(E\\) is a predefined set of entity types. In our problem setting, we used a fine-grained dataset \\(\\mathcal{D}^{\\mathcal{F}} = \\{(\\mathbf{X}_1^{\\mathcal{F}},\\mathbf{Y}_1^{\\mathcal{F}}),\\dots,(\\mathbf{X}_{|\\mathcal{D}^{\\mathcal{F}}|}^{\\mathcal{F}},\\mathbf{X}_{|\\mathcal{D}^{\\mathcal{F}}|})\\}\\) with a predefined entity type set \\(E^{\\mathcal{F}}\\). Additionally, we possess a coarse-grained dataset \\(\\mathcal{D}^{\\mathcal{C}} = \\{(\\mathbf{X}_1^{\\mathcal{C}},\\mathbf{Y}_1^{\\mathcal{C}}),\\dots,(\\mathbf{X}_{|\\mathcal{D}^{\\mathcal{C}}|}^{\\mathcal{C}},\\mathbf{X}_{|\\mathcal{D}^{\\mathcal{C}}|})\\}\\) characterized by a coarse-grained entity set \\(E^{\\mathcal{C}}\\) (i.e. \\(|E^{\\mathcal{F}}| > |E^{\\mathcal{C}}|\\)). A coarse-grained dataset typically"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3270"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.23
+            ],
+            "angle": 0,
+            "content": "has a smaller number of types than a fine-grained dataset. Throughout this study, we use \\(\\mathcal{F}\\) and \\(\\mathcal{C}\\) to distinguish between these datasets. It should be noted that our method can be readily extended to accommodate multiple coarse-grained datasets, incorporating an intrinsic multi-level hierarchy. However, our primary discussion revolves around a single coarse-grained dataset for simplicity and readability."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.24,
+                0.313,
+                0.256
+            ],
+            "angle": 0,
+            "content": "3.2 Training CoFiNER"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.26,
+                0.489,
+                0.357
+            ],
+            "angle": 0,
+            "content": "We aim to utilize both coarse- and fine-grained datasets directly in a single model training. Figure 1 illustrates an overview of the fine-grained NER model training process using both coarse- and fine-grained datasets. The CoFiNER training process consists of the following four steps:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.361,
+                0.49,
+                0.489
+            ],
+            "angle": 0,
+            "content": "Step 1- Training a fine-grained model. In the first step, we train a fine-grained model \\( f^{\\mathcal{F}}(\\theta) \\) with the low-resource fine-grained dataset \\( \\mathcal{D}^{\\mathcal{F}} \\). This process follows a typical supervised learning approach for NER. For a training example \\( (\\mathbf{X}^{\\mathcal{F}},\\mathbf{Y}^{\\mathcal{F}}) \\), \\( \\mathbf{X}^{\\mathcal{F}} \\) is fed into a PLM (Pre-trained Language Model), such as BERT, RoBERTa, to generate a contextual representations \\( \\mathbf{h}_i\\in \\mathbb{R}^d \\) of each token \\( x_{i} \\)."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.157,
+                0.499,
+                0.488,
+                0.518
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {H} = \\left[ \\mathbf {h} _ {1}, \\dots , \\mathbf {h} _ {n} \\right] = \\mathsf {P L M} \\left(\\left[ x _ {1} ^ {\\mathcal {F}}, \\dots , x _ {n} ^ {\\mathcal {F}} \\right]\\right). \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.528,
+                0.487,
+                0.559
+            ],
+            "angle": 0,
+            "content": "Then, we apply a softmax layer to obtain the label probability distribution:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.201,
+                0.569,
+                0.401,
+                0.589
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {p} _ {i} ^ {\\mathcal {F}} = \\operatorname {s o f t m a x} (\\mathbf {W h} _ {i} + \\mathbf {b})\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.598,
+                0.49,
+                0.68
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathbf{W} \\in \\mathbb{R}^{|E^{\\mathcal{F}}| \\times d}\\) and \\(\\mathbf{b} \\in \\mathbb{R}^{|E^{\\mathcal{F}}|}\\) represent the weights and bias of the classification head, respectively. To train the model using a fine-grained dataset, we optimize the cross-entropy loss function:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.202,
+                0.688,
+                0.488,
+                0.728
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {\\mathcal {F}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\mathbf {p} _ {i} ^ {\\mathcal {F}} \\left[ y _ {i} ^ {\\mathcal {F}} \\right] \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.739,
+                0.487,
+                0.771
+            ],
+            "angle": 0,
+            "content": "where \\(y_{i}\\in E^{\\mathcal{F}}\\) is the fine-grained label for the token \\(x_{i}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.775,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Step 2 - Generating an F2C matrix. To fully leverage the hierarchy between coarse- and fine-grained entity types, we avoid training separate NER models for each dataset. Instead, we utilize a single model that incorporates an F2C mapping matrix that transforms a fine-grained output into a corresponding coarse-grained output. The F2C mapping matrix assesses the conditional probability of a coarse-grained entity type \\( s \\in E^{\\mathcal{C}} \\) given"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.083,
+                0.882,
+                0.116
+            ],
+            "angle": 0,
+            "content": "a fine-grained label \\(\\ell \\in E^{\\mathcal{F}}\\) (i.e., \\(\\mathbf{M}_{\\ell,s} = p(y^{\\mathcal{C}} = s|y^{\\mathcal{F}} = \\ell)\\))."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.117,
+                0.884,
+                0.181
+            ],
+            "angle": 0,
+            "content": "Given a fine-grained probability distribution \\(\\mathbf{p}_i^{\\mathcal{F}}\\) computed using the proposed model, the marginal probability of a coarse-grained type \\(s\\) can be computed as"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.549,
+                0.192,
+                0.842,
+                0.226
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {p} _ {i} ^ {\\mathcal {C}} [ s ] = \\sum_ {\\ell \\in E ^ {\\mathcal {F}}} p (y ^ {\\mathcal {C}} = s | y ^ {\\mathcal {F}} = \\ell) \\cdot \\mathbf {p} _ {i} ^ {\\mathcal {F}} [ \\ell ].\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.237,
+                0.882,
+                0.268
+            ],
+            "angle": 0,
+            "content": "Thus, the coarse-grained output probabilities are simply computed as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.642,
+                0.278,
+                0.884,
+                0.298
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {p} _ {i} ^ {\\mathcal {C}} = \\mathbf {p} _ {i} ^ {\\mathcal {F}} \\cdot \\mathbf {M} \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.31,
+                0.882,
+                0.376
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathbf{M} \\in \\mathbb{R}^{|E^{\\mathcal{F}}| \\times |E^{\\mathcal{C}}|}\\) is the F2C mapping matrix whose row-wise sum is 1. By introducing this F2C mapping matrix, we can train a single model using multiple datasets with different granularity levels."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.377,
+                0.884,
+                0.504
+            ],
+            "angle": 0,
+            "content": "Manual annotation is a straightforward approach that can be used when hierarchical information is unavailable. However, it is not only cost-intensive but also noisy and subjective, especially when there are multiple coarse-grained datasets or a large number of fine-grained entity types. We introduce an efficient method for automatically generating an F2C matrix in §3.3."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.509,
+                0.884,
+                0.653
+            ],
+            "angle": 0,
+            "content": "Step 3 - Filtering inconsistent coarse labels. Although a fine-grained entity type is usually a subtype of a coarse-grained type, there can be some misalignments between the coarse- and fine-grained entity types. For example, an entity \"Microsoft\" in a financial document can either be tagged as Company or Stock which are not hierarchical. This inconsistency can significantly degrade the model's performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.654,
+                0.885,
+                0.812
+            ],
+            "angle": 0,
+            "content": "To mitigate the effect of inconsistent labeling, we devise an inconsistency filtering method aimed at masking less relevant coarse labels. By automatically filtering out the inconsistent labels from the coarse-grained dataset, we investigate the coarse-grained labels using the fine-grained NER model trained in Step 1. For each token in the coarse-grained dataset, we predict the coarse-grained label using the fine-grained model and the mapping matrix as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.625,
+                0.824,
+                0.883,
+                0.844
+            ],
+            "angle": 0,
+            "content": "\\[\n\\tilde {y} _ {i} ^ {\\mathcal {C}} = \\arg \\max  \\mathbf {p} _ {i} ^ {\\mathcal {C}}. \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.855,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "If the predicted label is the same as the coarse-grained label (i.e., \\( y_{i}^{\\mathcal{C}} = \\tilde{y}_{i}^{\\mathcal{C}} \\)), we assume that the coarse-grained label is consistent with the fine-grained one and can benefit the model. Otherwise,"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3271"
+        }
+    ],
+    [
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.117,
+                0.085,
+                0.427,
+                0.096
+            ],
+            "angle": 0,
+            "content": "Step 1) training process using fine-grained dataset"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.1,
+                0.874,
+                0.384
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.326,
+                0.398,
+                0.671,
+                0.413
+            ],
+            "angle": 0,
+            "content": "Figure 1: Overall process of our proposed method."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.438,
+                0.489,
+                0.502
+            ],
+            "angle": 0,
+            "content": "we regard the label as inconsistent with fine-grained types and do not utilize the coarse-grained label in Step 4. Note that the fine-grained NER model is frozen during this phase."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.509,
+                0.49,
+                0.604
+            ],
+            "angle": 0,
+            "content": "Step 4 - Jointly training CoFiNER with both datasets. The model is trained by alternating between epochs using coarse- and fine-grained data. This is an effective learning strategy for training models using heterogeneous datasets (Jung and Shim, 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.608,
+                0.49,
+                0.771
+            ],
+            "angle": 0,
+            "content": "For the fine-grained batches, CoFiNER is trained by minimizing the loss function defined in Equation (2), as described in Step 1. Meanwhile, we use the F2C mapping matrix to generate coarse-grained outputs and utilize inconsistency filtering for coarse-grained batches. Thus, we compute the cross-entropy loss between the coarse-grained label \\( y_{i}^{\\mathcal{C}} \\) and the predicted probabilities \\( \\mathbf{p}_i^{\\mathcal{C}} \\) when the coarse-grained label is consistent with our model (i.e., \\( y_{i}^{\\mathcal{C}} = \\tilde{y}_{i}^{\\mathcal{C}} \\)) as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.157,
+                0.792,
+                0.488,
+                0.832
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {\\mathcal {C}} = - \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\mathbf {p} _ {i} ^ {\\mathcal {C}} \\left[ y _ {i} ^ {\\mathcal {C}} \\right] \\cdot \\mathbb {I} \\left[ y _ {i} ^ {\\mathcal {C}} = \\tilde {y} _ {i} ^ {\\mathcal {C}} \\right] \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.855,
+                0.489,
+                0.921
+            ],
+            "angle": 0,
+            "content": "where \\(m\\) is a length of \\(\\mathbf{X}^{\\mathcal{C}}\\). For example, suppose that the coarse label of token \\(x_{i}\\) is \\(y_{i}^{\\mathcal{C}} = 0\\mathrm{RG}\\). If the estimated coarse label \\(\\tilde{y}_i^{\\mathcal{C}}\\) is 0RG, the loss for the token is \\(-\\log \\mathbf{p}_i^{\\mathcal{C}}[y_i^{\\mathcal{C}}]\\). Otherwise, it is zero."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.438,
+                0.879,
+                0.454
+            ],
+            "angle": 0,
+            "content": "3.3 Construction of the F2C mapping matrix"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.462,
+                0.885,
+                0.622
+            ],
+            "angle": 0,
+            "content": "The F2C mapping matrix assesses the conditional probability of a coarse-grained entity type \\( s \\in E^{\\mathcal{C}} \\) given a fine-grained label \\( \\ell \\in E^{\\mathcal{F}} \\) (i.e., \\( \\mathbf{M}_{\\ell,s} = p(y^{\\mathcal{C}} = s|y^{\\mathcal{F}} = \\ell) \\)). As there are only a few identical texts with both coarse- and fine-grained annotations simultaneously, we can not directly calculate this conditional probability using the data alone. Thus, we approximate the probability using a coarse-grained NER model and fine-grained labeled data as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.527,
+                0.624,
+                0.869,
+                0.643
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {M} _ {\\ell , s} = p (y ^ {\\mathcal {C}} = s | y ^ {\\mathcal {F}} = \\ell) \\approx p (\\tilde {y} ^ {\\mathcal {C}} = s | y ^ {\\mathcal {F}} = \\ell).\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.645,
+                0.885,
+                0.821
+            ],
+            "angle": 0,
+            "content": "To generate the mapping matrix, we first train a coarse-grained NER model \\( f^{\\mathcal{C}}(\\theta) \\) with the coarse-grained dataset \\( \\mathcal{D}^{\\mathcal{C}} \\). Then, we reannote the fine-grained dataset \\( \\mathcal{D}^{\\mathcal{F}} \\) by using the coarse-grained model \\( f^{\\mathcal{C}}(\\theta) \\). As a result, we obtain parallel annotations for both coarse- and fine-grained types in the fine-grained data \\( \\mathcal{D}^{\\mathcal{F}} \\). By using the parallel annotations, we can compute the co-occurrence matrix \\( \\mathbf{C} \\in \\mathbb{N}^{|E^{\\mathcal{F}}| \\times |E^{\\mathcal{C}}|} \\) where each cell \\( \\mathbf{C}_{\\ell,s} \\) is the number of tokens that are labeled as fine-grained type \\( \\ell \\) and coarse-grained type \\( s \\) together."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.824,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Because some labels generated by the coarse-grained model \\( f^{\\mathcal{C}}(\\theta) \\) can be inaccurate, we refine the co-occurrence matrix by retaining only the top-\\( k \\) counts for each fine-grained type and setting the rest to zero. Our experiments show that our model performs best when \\( k = 1 \\). This process effic"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3272"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.18
+            ],
+            "angle": 0,
+            "content": "tively retains only the most frequent coarse-grained categories for each fine-grained type, thereby improving the precision of the resulting mapping. Finally, we compute the conditional probabilities for all \\(\\ell \\in E^{\\mathcal{F}}, s \\in E^{\\mathcal{C}}\\) by using the co-occurrence counts as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.129,
+                0.188,
+                0.49,
+                0.224
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {M} _ {\\ell , s} = p \\left(\\tilde {y} ^ {\\mathcal {C}} = s \\mid y ^ {\\mathcal {F}} = \\ell\\right) = \\frac {\\mathbf {C} _ {\\ell , s}}{\\sum_ {s ^ {\\prime} \\in E ^ {\\mathcal {C}}} \\mathbf {C} _ {\\ell , s ^ {\\prime}}}. \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.233,
+                0.489,
+                0.266
+            ],
+            "angle": 0,
+            "content": "The F2C mapping matrix \\(\\mathbf{M}\\) is used to predict the coarse-grained labels using Equation (3)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.276,
+                0.262,
+                0.293
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.302,
+                0.229,
+                0.316
+            ],
+            "angle": 0,
+            "content": "4.1 Datasets"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.14,
+                0.331,
+                0.465,
+                0.487
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"3\"># Sentences</td><td rowspan=\"2\"># Types</td></tr><tr><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td colspan=\"5\">Coarse-grained datasets</td></tr><tr><td>CoNLL&#x27;03</td><td>14k</td><td>3.3k</td><td>3.4k</td><td>4</td></tr><tr><td>OntoNotes</td><td>59.9k</td><td>8.5k</td><td>8.3k</td><td>18</td></tr><tr><td colspan=\"5\">Fine-grained dataset Few-NERD</td></tr><tr><td>10-Shot</td><td>0.3k</td><td></td><td></td><td></td></tr><tr><td>20-Shot</td><td>0.6k</td><td></td><td></td><td></td></tr><tr><td>40-Shot</td><td>1.1k</td><td>18.8k</td><td>37.6k</td><td>66</td></tr><tr><td>80-Shot</td><td>2.2k</td><td></td><td></td><td></td></tr><tr><td>100-Shot</td><td>2.7k</td><td></td><td></td><td></td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.184,
+                0.497,
+                0.416,
+                0.51
+            ],
+            "angle": 0,
+            "content": "Table 1: Statistics of each dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.53,
+                0.49,
+                0.69
+            ],
+            "angle": 0,
+            "content": "We conduct experiments using a fine-grained NER dataset, Few-NERD (SUP) (Ding et al., 2021), as well as two coarse-grained datasets, namely, OntoNotes (Weischedel et al., 2013) and CoNLL'03 (Tjong Kim Sang, 2002). The fine-grained dataset Few-NERD comprises 66 entity types, whereas the coarse-grained datasets CoNLL'03 and OntoNotes consist of 4 and 18 entity types, respectively. The statistics for the datasets are listed in Table 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.695,
+                0.49,
+                0.87
+            ],
+            "angle": 0,
+            "content": "\\(K\\)-shot sampling for the fine-grained dataset. Because we assumed a small number of examples for each label in the fine-grained dataset, we evaluated the performance in the \\(K\\)-shot learning setting. Although Few-NERD provides few-shot samples, they are obtained based on an \\(N\\)-way \\(K\\)-shot scenario, where \\(N\\) is considerably smaller than the total number of entity types. However, our goal is to identify named entities across all possible entity types. For this setting, we resampled \\(K\\)-shot examples to accommodate all-way \\(K\\)-shot scenarios."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.872,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Since multiple entities exist in a single sentence, we cannot strictly generate exact \\(K\\)-shot samples for all the entity types. Therefore, we adopt the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.084,
+                0.885,
+                0.182
+            ],
+            "angle": 0,
+            "content": "\\(K\\sim (K + 5)\\) -shot setting. In the \\(K\\sim (K + 5)\\) -shot setting, there are at least \\(K\\) examples and at most \\(K + 5\\) examples for each entity type. See Appendix A for more details. In our experiments, we sampled fine-grained training data for \\(K = 10,20,40,80\\) ,and 100."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.195,
+                0.733,
+                0.21
+            ],
+            "angle": 0,
+            "content": "4.2 Experimental Settings"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.217,
+                0.885,
+                0.537
+            ],
+            "angle": 0,
+            "content": "In experiments, we use transformer-based PLM, including \\(\\mathrm{BERT}_{\\mathrm{BASE}}\\), \\(\\mathrm{RoBERTa}_{\\mathrm{BASE}}\\), and \\(\\mathrm{RoBERTa}_{\\mathrm{LARGE}}\\). In CoFiNER, we follow \\(\\mathrm{RoBERTa}_{\\mathrm{LARGE}}\\) to build a baseline model. The maximum sequence length is set to 256 tokens. The AdamW optimizer (Loshchilov and Hutter, 2019) is used to train the model with a learning rate of \\(2e - 5\\) and a batch size of 16. The number of epochs is varied for each model. We train the fine-grained model, CoFiNER, over 30 epochs. To construct the F2C mapping matrix, the coarse-grained model is trained for 50 epochs using both CoNLL'03 and OntoNotes. To train the inconsistency filtering model, we set different epochs based on the number of shots: For 10, 20, 40, 80, and 100 shot settings, the epochs are 150, 150, 120, 50, and 30, respectively. We report the results using span-level F1. The dropout with a probability of 0.1 is applied. All the models were trained on NVIDIA RTX 3090 GPUs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.551,
+                0.715,
+                0.567
+            ],
+            "angle": 0,
+            "content": "4.3 Compared Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.574,
+                0.884,
+                0.655
+            ],
+            "angle": 0,
+            "content": "In this study, we compare the performance of CoFiNER with that of both the supervised and few-shot methods. We modified the existing methods for our experimental setup and re-implemented them accordingly."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.659,
+                0.885,
+                0.801
+            ],
+            "angle": 0,
+            "content": "Supervised method. We use \\(\\mathrm{BERT}_{\\mathrm{BASE}}\\), \\(\\mathrm{RoBERTa}_{\\mathrm{BASE}}\\), and \\(\\mathrm{RoBERTa}_{\\mathrm{LARGE}}\\) as the supervised baselines, each including a fine-grained classifier on the head. In addition, PIQN (Shen et al., 2022) and PL-Marker (Ye et al., 2022) are methods that have achieved state-of-the-art performance in a supervised setting using the full Few-NERD dataset. All models are trained using only a Few-NERD dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Few-shot method. The LSFS (Ma et al., 2022a) leverages a label encoder to utilize the semantics of label names, thereby achieving state-of-the-art results in low-resource NER settings. The LSFS applies a pre-finetuning strategy to learn prior knowledge from the coarse-grained dataset, OntoNotes. For a fair comparison, we also conducted pre"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3273"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.148,
+                0.081,
+                0.852,
+                0.202
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Method</td><td>Model</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>100-shot</td></tr><tr><td rowspan=\"5\">Supervised Method</td><td>BERTBASE (Devlin et al., 2019)</td><td>23.101</td><td>34.718</td><td>43.138</td><td>46.182</td><td>46.449</td></tr><tr><td>RoBERTaBASE (Liu et al., 2019)</td><td>21.073</td><td>34.157</td><td>39.226</td><td>45.735</td><td>46.942</td></tr><tr><td>RoBERTaLARGE (Liu et al., 2019)</td><td>29.137</td><td>41.425</td><td>51.137</td><td>55.057</td><td>54.172</td></tr><tr><td>PIQN (Shen et al., 2022)</td><td>21.750</td><td>22.007</td><td>28.533</td><td>29.339</td><td>38.658</td></tr><tr><td>PL-Marker (Ye et al., 2022)</td><td>40.902</td><td>48.064</td><td>52.395</td><td>53.249</td><td>53.061</td></tr><tr><td>Few-shot Method</td><td>LSFS (Ma et al., 2022a)</td><td>47.998</td><td>43.269</td><td>50.595</td><td>51.420</td><td>50.366</td></tr><tr><td>Proposed Method</td><td>CoFiNER</td><td>44.951</td><td>51.142</td><td>56.409</td><td>56.847</td><td>57.178</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.205,
+                0.211,
+                0.79,
+                0.226
+            ],
+            "angle": 0,
+            "content": "Table 2: Results on few-shot NER. The best scores across all models are marked bold."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.252,
+                0.486,
+                0.283
+            ],
+            "angle": 0,
+            "content": "finetuning on OntoNotes and performed fine-tuning on each shot of the Few-NERD dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.292,
+                0.487,
+                0.405
+            ],
+            "angle": 0,
+            "content": "Proposed method. We trained our CoFiNER model as proposed in §3. In each epoch, CoFiNER is first trained on two coarse-grained datasets: OntoNotes and CoNLL'03. Subsequently, it is trained on the fine-grained dataset Few-NERD. We used RoBERTa\\(_{\\text{LARGE}}\\) as the pre-trained language model for CoFiNER in Equation (1)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.433,
+                0.266,
+                0.448
+            ],
+            "angle": 0,
+            "content": "4.4 Main Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.464,
+                0.49,
+                0.736
+            ],
+            "angle": 0,
+            "content": "Table 2 reports the performance of CoFiNER and existing methods. The result shows that CoFiNER outperforms both supervised learning and few-shot learning methods. Because supervised learning typically needs a large volume of training data, these models underperform in low-resource settings. This demonstrates that our method effectively exploits coarse-grained datasets to enhance the performance of the fine-grained NER model. In other words, the proposed F2C mapping matrix significantly reduces the amount of fine-grained dataset required to train supervised NER models. In particular, CoFiNER achieves significant performance improvements compared to the state-of-the-art model PL-Marker, which also utilizes the same pre-trained language model RoBERTa\\(_{\\text{LARGE}}\\) as CoFiNER."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.743,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The few-shot method LSFS yields the highest F1 score for the 10-shot case. However, this few-shot method suffers from early performance saturation, resulting in less than a 2.4 F1 improvement with an additional 90-shot. By contrast, the F1 score of CoFiNER increases by 12.2. Consequently, CoFiNER outperforms all the compared methods except for the 10-shot case. In summary, the proposed method yields promising results for a wide range of data sample sizes by explicitly leveraging the inherent hierarchical structure."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.513,
+                0.248,
+                0.883,
+                0.334
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>100-shot</td></tr><tr><td>CoFiNER</td><td>44.95</td><td>51.14</td><td>56.41</td><td>56.85</td><td>57.18</td></tr><tr><td>w/o filtering</td><td>41.38</td><td>43.87</td><td>53.42</td><td>54.70</td><td>55.13</td></tr><tr><td>w/o OntoNotes</td><td>42.63</td><td>44.51</td><td>57.18</td><td>56.97</td><td>55.54</td></tr><tr><td>w/o CoNLL&#x27;03</td><td>43.99</td><td>50.07</td><td>57.54</td><td>56.71</td><td>56.27</td></tr><tr><td>w/o coarse</td><td>29.14</td><td>41.43</td><td>51.14</td><td>55.06</td><td>54.17</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.343,
+                0.882,
+                0.372
+            ],
+            "angle": 0,
+            "content": "Table 3: Performances of ablation study over different components."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.404,
+                0.673,
+                0.42
+            ],
+            "angle": 0,
+            "content": "4.5 Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.506,
+                0.432,
+                0.885,
+                0.673
+            ],
+            "angle": 0,
+            "content": "An ablation study is conducted to validate the effectiveness of each component of the proposed method. The results are presented in Table 3. First, we remove the inconsistency filtering \\((w / o\\) filtering) and observe a significant decrease in the F1 score, ranging from 2.05 to 7.27. These results demonstrate the effectiveness of our filtering method, which excludes mislabeled entities. Second, we provide the results using a single coarse-grained dataset (w/o OntoNotes and w/o CoNLL'03). Even with a single coarse-grained dataset, our proposed method significantly outperforms w/o coarse, which is trained solely on the fine-grained dataset (i.e. RoBERTaLarge in Table 2)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.677,
+                0.885,
+                0.822
+            ],
+            "angle": 0,
+            "content": "This indicates the effectiveness of using a well-aligned hierarchy through the F2C mapping matrix and inconsistency filtering. Although we achieve a significant improvement even with a single coarse-grained dataset, we achieve a more stable result with two coarse-grained datasets. This implies that our approach effectively utilizes multiple coarse-grained datasets, although the datasets contain different sets of entity types."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.844,
+                0.625,
+                0.86
+            ],
+            "angle": 0,
+            "content": "4.6 Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.872,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this section, we experimentally investigate how the F2C mapping matrix and inconsistency filtering improve the accuracy of the proposed model."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3274"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.144,
+                0.082,
+                0.862,
+                0.273
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.28,
+                0.885,
+                0.325
+            ],
+            "angle": 0,
+            "content": "Figure 2: The F2C mapping matrix between OntoNotes and 100-shot of the Few-NERD dataset. During the generation of the F2C mapping matrix, some coarse-grained types are not mapped to any fine-grained types. The 7 unmapped types are not represented: DATE, TIME, PERCENT, MONEY, QUANTITY, ORDINAL, and CARDINAL."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.143,
+                0.336,
+                0.862,
+                0.481
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.172,
+                0.489,
+                0.825,
+                0.505
+            ],
+            "angle": 0,
+            "content": "Figure 3: The F2C mapping matrix between CoNLL'03 and 100-shot of the Few-NERD dataset."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.53,
+                0.345,
+                0.546
+            ],
+            "angle": 0,
+            "content": "4.6.1 F2C Mapping Matrix"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.55,
+                0.49,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Mapping the entity types between the coarse- and fine-grained datasets directly affects the model performance. Hence, we investigate the mapping outcomes between the coarse- and fine-grained datasets. Figures 2 and 3 show the F2C matrices for FewNERD-OntoNotes and FewNERD-CoNLL'03, respectively. In both figures, the x- and y-axis represent the fine-grained and coarse-grained entity types, respectively. The colored text indicates the corresponding coarse-grained entity types in FewNERD, which were not used to find the mapping matrix. The mapping is reliable if we compare the y-axis and the colored types (coarse-grained types in Few-NERD). Even without manual annotation of the relationship between coarse- and fine-grained entities, our method successfully obtains reliable mapping from fine-grained to coarse-grained types. Our model can be effectively trained with both types of datasets using accurate F2C mapping."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.856,
+                0.489,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Figure 4 provides the F1 scores by varying the hyperparameter \\( k \\) to refine the F2C mapping matrix described in §3.3. 'all' refers to the usage of the complete frequency distribution when creating an"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.578,
+                0.53,
+                0.821,
+                0.645
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.659,
+                0.883,
+                0.688
+            ],
+            "angle": 0,
+            "content": "Figure 4: Impact of top-\\(k\\) values in the F2C mapping matrix. F1 is reported on the test data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.722,
+                0.885,
+                0.852
+            ],
+            "angle": 0,
+            "content": "F2C mapping matrix. We observe that the highest performance is achieved when \\( k \\) is set to 1, and as \\( k \\) increases, the performance gradually decreases. Although the optimal value of \\( k \\) can vary depending on the quality of the coarse-grained data and the performance of the coarse-grained NER model, the results indicate that a good F2C mapping matrix can be obtained by ignoring minor co-occurrences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.855,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Additionally, We conducted an experiment by setting the F2C mapping matrix to be learnable and comparing it with our non-learnable F2C matrix. The non-learnable approach showed better"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3275"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.188,
+                0.081,
+                0.812,
+                0.199
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>#</td><td>Type</td><td>Example</td></tr><tr><td rowspan=\"2\">1</td><td>predict</td><td>... they may be offshoots of the intifadahEVENT, the Palestinian rebellionEVENT in ...</td></tr><tr><td>target</td><td>... they may be offshoots of the intifadah, the PalestinianNORP rebellion in ...</td></tr><tr><td rowspan=\"2\">2</td><td>predict</td><td>... players heading to CanadaGPE, particularly TorontoGPE for the All-starEVENT ...</td></tr><tr><td>target</td><td>... players heading to CanadaGPE, particularly TorontoGPE for the All-star ...</td></tr><tr><td rowspan=\"2\">3</td><td>predict</td><td>Judges at Flight 103 Lockerbie trialEVENT are expected ...</td></tr><tr><td>target</td><td>Judges at Flight 103 LockerbieGPE trial are expected ...</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.207,
+                0.884,
+                0.251
+            ],
+            "angle": 0,
+            "content": "Table 4: Inconsistent examples: The labeled sentence in the original coarse-grained dataset is the \"target\" type, while the label predicted by the model is the \"predict\" type. Consistent entity types are indicated in blue, while inconsistent entity types are indicated in red."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.275,
+                0.49,
+                0.325
+            ],
+            "angle": 0,
+            "content": "performance, hence we adopted this approach for CoFiNER. Detailed analysis and experiment results are in Appendix B.2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.335,
+                0.356,
+                0.351
+            ],
+            "angle": 0,
+            "content": "4.6.2 Inconsistency Filtering"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.355,
+                0.49,
+                0.497
+            ],
+            "angle": 0,
+            "content": "We aim to examine whether inconsistency filtering successfully screened out inconsistent entities between OntoNotes and Few-NERD datasets. To conduct this analysis, we describe three inconsistent examples. Table 4 shows the predicted values and target labels, which correspond to the coarse-grained output of the fine-grained NER model trained as described in §3.2 and the golden labels of the coarse-grained dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.5,
+                0.49,
+                0.692
+            ],
+            "angle": 0,
+            "content": "The first example illustrates the inconsistency in entity types between mislabeled entities. In the original coarse-grained dataset, \"Palestinian\" is labeled as NORP, but the model trained on the fine-grained dataset predicts \"Palestinian rebellion\" as its appropriate label, EVENT. However, annotators of the OntoNotes labeled \"Palestinian\" as NORP, whereas the fine-grained NER model correctly predicts the highly informative actual label span. The inconsistency caused by a label mismatch between the coarse-grained and fine-grained datasets can result in performance degradation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.694,
+                0.49,
+                0.838
+            ],
+            "angle": 0,
+            "content": "In the second example, both \"Canada\" and \"Toronto\" are consistently labeled as GPE; thus, the model is not confused when training on these two entities. However, in the case of \"All-star\", we can observe a mismatch. This example in the coarse-grained dataset is labeled 0 instead of the correct entity type EVENT, indicating a mismatch. Through inconsistency filtering, unlabeled \"All-star\" is masked out of the training process."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "As shown in the examples, inconsistency filtering is necessary to mitigate the potential noise arising from mismatched entities. We analyze the filtering results for each coarse-grained label to assess its impact on model performance. Figure 5"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.514,
+                0.275,
+                0.88,
+                0.48
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.491,
+                0.884,
+                0.548
+            ],
+            "angle": 0,
+            "content": "Figure 5: Performance with increasing filtering proportion between Few-NERD and OntoNotes datasets. The correlation coefficient between filtering proportion and performance improvement is 0.29."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.574,
+                0.884,
+                0.799
+            ],
+            "angle": 0,
+            "content": "illustrates the correlation between the filtering proportion and the performance improvement for each coarse-grained label mapped to the fine-grained labels. In this figure, a higher filtering proportion indicates a greater inconsistency between the two datasets. F1 improvements indicate the difference in performance when filtering is applied and when it is not. Each data point refers to a fine-grained label mapped onto a coarse-grained label. As the proportion of the filtered entities increases, the F1 scores also increase. These improvements indicate that inconsistency filtering effectively eliminates noisy entities, enabling the model to be well-trained on consistent data only."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.813,
+                0.642,
+                0.828
+            ],
+            "angle": 0,
+            "content": "5 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We proposed CoFiNER, which explicitly leverages the hierarchical structure between coarse- and fine-grained types to alleviate the low-resource problem of fine-grained NER. We devised the F2C mapping matrix that allows for fine-grained NER model"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3276"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.491,
+                0.278
+            ],
+            "angle": 0,
+            "content": "training using additional coarse-grained datasets. However, because not all coarse-grained entities are beneficial for the fine-grained NER model, the proposed inconsistency filtering method is used to mask out noisy entities from being used in the model training process. We found through experiments that using a smaller amount of consistent data is better than using a larger amount of data without filtering, thus demonstrating the crucial role of inconsistency filtering. Empirical results confirmed the superiority of CoFiNER over both the supervised and few-shot methods."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.291,
+                0.221,
+                0.306
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.318,
+                0.49,
+                0.542
+            ],
+            "angle": 0,
+            "content": "Despite the promising empirical results of this study, there is still a limitation. The main drawback of CoFiNER is that the token-level F2C mapping matrix and inconsistency filtering may not be directly applicable to nested NER tasks. Nested NER involves identifying and classifying certain overlapping entities that exceed the token-level scope. Because CoFiNER addresses fine-grained NER at the token level, it may not accurately capture entities with nested structures. Therefore, applying our token-level approach to nested NER could pose challenges and may require further adaptations or other modeling techniques to effectively handle the hierarchical relations between nested entities."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.556,
+                0.279,
+                0.572
+            ],
+            "angle": 0,
+            "content": "Acknowledgments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.582,
+                0.492,
+                0.92
+            ],
+            "angle": 0,
+            "content": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF2022R1G1A1013549). This work was supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No. RS2023-00261068, Development of Lightweight Multimodal Anti-Phishing Models and Split-Learning Techniques for Privacy-Preserving Anti-Phishing) and (No.RS-2022-00155885, Artificial Intelligence Convergence Innovation Human Resources Development (Hanyang University ERICA)). This research was also supported by the Ministry of Trade, Industry, and Energy (MOTIE), Korea, under the \"Project for Research and Development with Middle Markets Enterprises and DNA(Data, Network, AI) Universities\" (Smart Home Based LifeCare Service & Solution Development)(P0024555) supervised by the Korea Institute for Advancement of Technology (KIAT)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.084,
+                0.61,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.107,
+                0.886,
+                0.213
+            ],
+            "angle": 0,
+            "content": "Armen Aghajanyan, Anchit Gupta, Akshat Shrivastava, Xilun Chen, Luke Zettlemoyer, and Sonal Gupta. 2021. Muppet: Massive multi-task representations with pre-finetuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5799-5811, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.221,
+                0.886,
+                0.275
+            ],
+            "angle": 0,
+            "content": "Sarkar Snigdha Sarathi Das, Arzoo Katiyar, Rebecca J Passonneau, and Rui Zhang. 2021. Container: Few-shot named entity recognition via contrastive learning. arXiv preprint arXiv:2109.07589."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.283,
+                0.886,
+                0.375
+            ],
+            "angle": 0,
+            "content": "Leon Derczynski, Eric Nichols, Marieke van Erp, and Nut Limsopatham. 2017. Results of the WNUT2017 shared task on novel and emerging entity recognition. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 140-147, Copenhagen, Denmark. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.385,
+                0.886,
+                0.504
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.512,
+                0.886,
+                0.631
+            ],
+            "angle": 0,
+            "content": "Ning Ding, Guangwei Xu, Yulin Chen, Xiaobin Wang, Xu Han, Pengjun Xie, Haitao Zheng, and Zhiyuan Liu. 2021. Few-NERD: A few-shot named entity recognition dataset. 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 3198-3213, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.64,
+                0.886,
+                0.68
+            ],
+            "angle": 0,
+            "content": "Maximilian Hofer, Andrey Kormilitzin, Paul Goldberg, and Alejo Nevada-Holgado. 2018. Few-shot learning for named entity recognition in medical text."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.689,
+                0.886,
+                0.769
+            ],
+            "angle": 0,
+            "content": "Yucheng Huang, Kai He, Yige Wang, Xianli Zhang, Tieliang Gong, Rui Mao, and Chen Li. 2022. Copner: Contrastive learning with prompt guiding for few-shot named entity recognition. In Proceedings of the 29th International conference on computational linguistics, pages 2515-2527."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.777,
+                0.886,
+                0.818
+            ],
+            "angle": 0,
+            "content": "Haoming Jiang, Danqing Zhang, Tianyu Cao, Bing Yin, and Tuo Zhao. 2021. Named entity recognition with small strongly labeled and large weakly labeled data."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.826,
+                0.886,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Woohwan Jung and Kyuseok Shim. 2020. Dual supervision framework for relation extraction with distant supervision and human annotation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6411-6423, Barcelona, Spain (Online). International Committee on Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.107,
+                0.886,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3277"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.153
+            ],
+            "angle": 0,
+            "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.162,
+                0.489,
+                0.202
+            ],
+            "angle": 0,
+            "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.211,
+                0.489,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Lefteris Loukas, Manos Fergadiotis, Ilias Chalkidis, Eirini Spyropoulou, Prodromos Malakasiotis, Ion Androutsopoulos, and Georgios Paliouras. 2022. Finer: Financial numeric entity recognition for xbrl tagging. arXiv preprint arXiv:2203.06482."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.286,
+                0.489,
+                0.378
+            ],
+            "angle": 0,
+            "content": "Jie Ma, Miguel Ballesteros, Srikanth Doss, Rishita Anubhai, Sunil Mallya, Yaser Al-Onaizan, and Dan Roth. 2022a. Label semantics for few shot named entity recognition. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1956-1971, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.387,
+                0.489,
+                0.454
+            ],
+            "angle": 0,
+            "content": "Tingting Ma, Huiqiang Jiang, Qianhui Wu, Tiejun Zhao, and Chin-Yew Lin. 2022b. Decomposed meta-learning for few-shot named entity recognition. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 1584–1596."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.463,
+                0.489,
+                0.516
+            ],
+            "angle": 0,
+            "content": "Alexis Mitchell, Stephanie Strassel, Shudong Huang, and Ramez Zakhary. 2005. Ace 2004 multilingual training corpus. Linguistic Data Consortium, Philadelphia, 1:1-1."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.525,
+                0.489,
+                0.592
+            ],
+            "angle": 0,
+            "content": "Seongwoong Oh, Woohwan Jung, and Kyuseok Shim. 2023. Thunder: Named entity recognition using a teacher-student model with dual classifiers for strong and weak supervisions. In Proceedings of 26th European Conference on Artificial Intelligence."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.601,
+                0.489,
+                0.68
+            ],
+            "angle": 0,
+            "content": "Alan Ritter, Sam Clark, Mausam, and Oren Etzioni. 2011. Named entity recognition in tweets: An experimental study. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1524-1534, Edinburgh, Scotland, UK. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.688,
+                0.489,
+                0.794
+            ],
+            "angle": 0,
+            "content": "Yongliang Shen, Xiaobin Wang, Zeqi Tan, Guangwei Xu, Pengjun Xie, Fei Huang, Weiming Lu, and Yueting Zhuang. 2022. Parallel instance query network for named entity recognition. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 947-961, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.803,
+                0.489,
+                0.869
+            ],
+            "angle": 0,
+            "content": "Mujeen Sung, Minbyul Jeong, Yonghwa Choi, Donghyeon Kim, Jinhyuk Lee, and Jaewoo Kang. 2022. Bern2: an advanced neural biomedical named entity recognition and normalization tool. Bioinformatics, 38(20):4837-4839."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.878,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Erik F. Tjong Kim Sang. 2002. Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition. In COLING-02: The 6th"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.113
+            ],
+            "angle": 0,
+            "content": "Conference on Natural Language Learning 2002 (CoNLL-2002)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.122,
+                0.884,
+                0.176
+            ],
+            "angle": 0,
+            "content": "Christopher Walker, Stephanie Strassel, Julie Medero, and Kazuaki Maeda. 2006. Ace 2005 multilingual training corpus. Linguistic Data Consortium, Philadelphia, 57:45."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.185,
+                0.884,
+                0.278
+            ],
+            "angle": 0,
+            "content": "Peiyi Wang, Runxin Xu, Tianyu Liu, Qingyu Zhou, Yunbo Cao, Baobao Chang, and Zhifang Sui. 2022. An enhanced span-based decomposition method for few-shot sequence labeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5012-5024."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.286,
+                0.884,
+                0.353
+            ],
+            "angle": 0,
+            "content": "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Edward Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, et al. 2013. Ontonotes release 5.0 ldc2013t19. Linguistic Data Consortium, Philadelphia, PA, 23."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.363,
+                0.884,
+                0.442
+            ],
+            "angle": 0,
+            "content": "Deming Ye, Yankai Lin, Peng Li, and Maosong Sun. 2022. Packed levitated marker for entity and relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4904-4917, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.884,
+                0.442
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.512,
+                0.454,
+                0.598,
+                0.471
+            ],
+            "angle": 0,
+            "content": "Appendix"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.512,
+                0.479,
+                0.725,
+                0.496
+            ],
+            "angle": 0,
+            "content": "A Sampling Algorithm"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.519,
+                0.848,
+                0.535
+            ],
+            "angle": 0,
+            "content": "Algorithm 1 \\(K\\sim (K + 5)\\) sampling algorithm"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.544,
+                0.537,
+                0.81,
+                0.551
+            ],
+            "angle": 0,
+            "content": "Input: Dataset \\(\\mathbf{X}\\), labeled set \\(\\mathbf{Y}\\), \\(K\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.544,
+                0.553,
+                0.797,
+                0.568
+            ],
+            "angle": 0,
+            "content": "Output: Fine-grained dataset \\(D^{\\mathcal{F}}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.571,
+                0.858,
+                0.586
+            ],
+            "angle": 0,
+            "content": "1: Sort classes in \\(\\mathbf{Y}\\) based on their freq. in \\(\\mathbf{X}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.587,
+                0.882,
+                0.618
+            ],
+            "angle": 0,
+            "content": "2: \\(D^{\\mathcal{F}}\\gets \\emptyset\\) Init the train dataset \\(\\triangleright\\) Init the count of entity classes in \\(D^{\\mathcal{F}}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.619,
+                0.688,
+                0.635
+            ],
+            "angle": 0,
+            "content": "3: for \\( i = 1 \\) to \\( |\\mathbf{Y}| \\) do"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.636,
+                0.678,
+                0.651
+            ],
+            "angle": 0,
+            "content": "4: Count \\([i] = 0\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.655,
+                0.687,
+                0.669
+            ],
+            "angle": 0,
+            "content": "5: for \\( i = 1 \\) to \\( |\\mathbf{Y}| \\) do"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.671,
+                0.75,
+                0.686
+            ],
+            "angle": 0,
+            "content": "6: if \\(\\forall \\operatorname{Count}[i'] \\geq K\\) then"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.689,
+                0.646,
+                0.702
+            ],
+            "angle": 0,
+            "content": "7: break;"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.706,
+                0.882,
+                0.721
+            ],
+            "angle": 0,
+            "content": "8: Randomly sample \\((x,y)\\in \\mathbf{X}\\) s.t. \\(\\mathbf{Y}_i\\in y\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.523,
+                0.722,
+                0.712,
+                0.736
+            ],
+            "angle": 0,
+            "content": "9: for \\( j = 1 \\) to \\( |y| \\) do"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.737,
+                0.73,
+                0.753
+            ],
+            "angle": 0,
+            "content": "10: \\(\\mathbf{Count}^{\\mathbf{Y}}[j] + = 1;\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.756,
+                0.875,
+                0.772
+            ],
+            "angle": 0,
+            "content": "11: if \\(\\exists \\operatorname{Count}[i] + \\operatorname{Count}^{\\mathbf{Y}}[i] > K + 5\\) then"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.775,
+                0.672,
+                0.787
+            ],
+            "angle": 0,
+            "content": "12: Continue;"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.791,
+                0.602,
+                0.802
+            ],
+            "angle": 0,
+            "content": "13: else"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.804,
+                0.772,
+                0.82
+            ],
+            "angle": 0,
+            "content": "14: \\(D^{\\mathcal{F}}\\gets D^{\\mathcal{F}}\\cup \\{(x,y)\\} ;\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.821,
+                0.8,
+                0.835
+            ],
+            "angle": 0,
+            "content": "15: Count \\(\\leftarrow\\) Count + CountY;"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.517,
+                0.838,
+                0.628,
+                0.853
+            ],
+            "angle": 0,
+            "content": "16: return \\(D^{\\mathcal{F}}\\)"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.517,
+                0.571,
+                0.882,
+                0.853
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.872,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Unlike other tasks, NER involves multiple entity occurrences within a sentence, making it too restrictive to sample an exact count. We adopt a"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3278"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.198
+            ],
+            "angle": 0,
+            "content": "\\(K \\sim (K + 5)\\)-shot setting to minimize differences in the number of entity types. Additionally, we sample the entity types, starting from those with fewer occurrences to ensure a balanced distribution of multiple entity types within the sentences. Algorithm 1 presents the \\(K\\)-shot sampling algorithm used in this study."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.21,
+                0.362,
+                0.227
+            ],
+            "angle": 0,
+            "content": "B Additional Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.236,
+                0.479,
+                0.268
+            ],
+            "angle": 0,
+            "content": "B.1 Generalization of Inconsistency filtering in diverse dataset settings"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.119,
+                0.284,
+                0.488,
+                0.35
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Setting</td><td colspan=\"2\">Coarse-grained datasets</td></tr><tr><td>Few-NERD_coarse</td><td>OntoNotes + CoNLL&#x27;03</td></tr><tr><td>w/ filtering</td><td>56.63</td><td>57.18</td></tr><tr><td>w/o filtering</td><td>56.46</td><td>55.13</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.359,
+                0.489,
+                0.388
+            ],
+            "angle": 0,
+            "content": "Table 5: Results on different coarse-grained datasets. A fine-grained dataset is 100-shot of the Few-NERD."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.404,
+                0.487,
+                0.451
+            ],
+            "angle": 0,
+            "content": "To assess the generalization of the proposed method, we conducted experiments under various coarse- and fine-grained dataset settings."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.453,
+                0.49,
+                0.869
+            ],
+            "angle": 0,
+            "content": "First, we verified the robustness of inconsistency filtering through experiments using different coarse-grained datasets. The fine-grained dataset, Few-NERD, remains unchanged. Since Few-NERD has both coarse- and fine-grained labels, we constructed a coarse-grained dataset Few-NERD_coarse using the coarse-grained labels from Few-NERD. When compared to Few-NERD_coarse, entities in OntoNotes and CoNLL'03 are inconsistently labeled with Few-NERD because they were independently created. In Table 5, Few-NERD_coarse exhibits a higher F1 score in the w/o filtering setting due to its consistency with fine-grained labels of Few-NERD. However, the performance improvement achieved with filtering is more substantial in the inconsistent datasets, when compared to the consistent dataset. This result indicates that inconsistency filtering improves performance by filtering out the mismatching labels. Therefore, we have demonstrated the importance of using inconsistency filtering to filter out noise when working with datasets that employ different labeling schemes. Furthermore, by achieving effective performance improvements across various coarse-grained datasets, we have provided evidence of the robustness of the filtering method."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.872,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Second, we validate the generalization performance through experiments conducted in different coarse- and fine-grained dataset settings. We"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.599,
+                0.082,
+                0.795,
+                0.188
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>F1</td></tr><tr><td>RoBERTaLARGE</td><td>75.15</td></tr><tr><td>PL-Marker</td><td>74.03</td></tr><tr><td>CoFiNER</td><td>80.44</td></tr><tr><td>w/o filtering</td><td>78.20</td></tr><tr><td>w/o coarse</td><td>75.15</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.197,
+                0.883,
+                0.239
+            ],
+            "angle": 0,
+            "content": "Table 6: Performances of different models and ablation studies on our model. RoBERTa\\(_{\\text{LARGE}}\\) and w/o coarse are identical."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.266,
+                0.885,
+                0.441
+            ],
+            "angle": 0,
+            "content": "set up the CoNLL'03, which has 4 entity types, as the coarse-grained dataset, and the 100-shot OntoNotes, which has a finer label with 19 entity types, as the fine-grained dataset. In Table 6, when compared to the two top-performing models in the main results, RoBERTa\\(_{\\text{LARGE}}\\) and the state-of-the-art PL-Marker, CoFiNER show consistently higher performance. Furthermore, as shown in the ablation study that was conducted following the same methodology in §4.5, CoFiNER exhibited superior performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.443,
+                0.884,
+                0.506
+            ],
+            "angle": 0,
+            "content": "In conclusion, through above the two experiments, our method has been demonstrated to work robustly across different coarse- and fine-grained dataset settings."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.517,
+                0.884,
+                0.548
+            ],
+            "angle": 0,
+            "content": "B.2 Comparison with learnable F2C mapping matrix"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.604,
+                0.562,
+                0.791,
+                0.619
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Matrix Type</td><td>F1</td></tr><tr><td>non-learnable</td><td>56.27</td></tr><tr><td>learnable</td><td>53.25</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.628,
+                0.883,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Table 7: Results on learnable and non-learnable F2C mapping matrix on 100-shot of Few-NERD."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.676,
+                0.884,
+                0.885
+            ],
+            "angle": 0,
+            "content": "To find the optimal F2C mapping matrix, we conducted experiments to explore the impact of making the F2C mapping matrix learnable. We use FewNERD as a fine-grained dataset and OntoNotes as a coarse-grained dataset. Table 7 shows no performance gains when the F2C mapping matrix was set to be learnable. We found that the learnable matrix tends to form a pattern similar to what is shown in Figure 4 with \\( k = \\) all. This result suggests that taking minor co-occurrences into account leads to an overall decrease in performance. Based on this analysis, the non-learnable mapping matrix is used in our experiments."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3279"
+        }
+    ]
+]

2023/Enhancing Low-resource Fine-grained Named Entity Recognition by Leveraging Coarse-grained Datasets/dff73b45-756f-4d7a-a8f6-789c5d1784de_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b40fa4b7c020298140bdc62b31c0f97d4d2ede92ce33b3b071b45e52a0421cf0
+size 1983659

2023/Enhancing Low-resource Fine-grained Named Entity Recognition by Leveraging Coarse-grained Datasets/full.md

@@ -0,0 +1,342 @@
+# Enhancing Low-resource Fine-grained Named Entity Recognition by Leveraging Coarse-grained Datasets
+
+Su Ah Lee $^{1\dagger}$ , Seokjin Oh $^{1\dagger}$ , and Woohwan Jung $^{1,2}$
+
+$^{1}$ Department of Applied Artificial Intelligence, Hanyang University
+
+2Ramply Inc.
+
+{sue991, seokjinoh, whjung}@hanyang.ac.kr
+
+# Abstract
+
+Named Entity Recognition (NER) frequently suffers from the problem of insufficient labeled data, particularly in fine-grained NER scenarios. Although  $K$ -shot learning techniques can be applied, their performance tends to saturate when the number of annotations exceeds several tens of labels. To overcome this problem, we utilize existing coarse-grained datasets that offer a large number of annotations. A straightforward approach to address this problem is prefinetuning, which employs coarse-grained data for representation learning. However, it cannot directly utilize the relationships between fine-grained and coarse-grained entities, although a fine-grained entity type is likely to be a subcategory of a coarse-grained entity type. We propose a fine-grained NER model with a Fineto-Coarse(F2C) mapping matrix to leverage the hierarchical structure explicitly. In addition, we present an inconsistency filtering method to eliminate coarse-grained entities that are inconsistent with fine-grained entity types to avoid performance degradation. Our experimental results show that our method outperforms both  $K$ -shot learning and supervised learning methods when dealing with a small number of fine-grained annotations. Code is available at https://github.com/sue991/CoFiNER.
+
+# 1 Introduction
+
+Named Entity Recognition (NER) is a fundamental task in locating and categorizing named entities in unstructured texts. Most research on NER has been conducted on coarse-grained datasets, including CoNLL'03 (Tjong Kim Sang, 2002), ACE04 (Mitchell et al., 2005), ACE05 (Walker et al., 2006), and OntoNotes (Weischedel et al., 2013), each of which has less than 18 categories. As the applications of NLP broaden across diverse fields, there is increasing demand for fine-grained NER that can provide more precise and
+
+detailed information extraction. Nonetheless, detailed labeling required for large datasets in the context of fine-grained NER presents several significant challenges. It is more cost-intensive and time-consuming than coarse-grained NER. In addition, it requires a high degree of expertise because domain-specific NER tasks such as financial NER (Loukas et al., 2022) and biomedical NER (Sung et al., 2022) require fine-grained labels. Thus, fine-grained NER tasks typically suffer from the data scarcity problem. Few-shot NER approaches (Ding et al., 2021) can be applied to conduct fine-grained NER with scarce fine-grained data. However, these methods do not exploit existing coarse-grained datasets that can be leveraged to improve fine-grained NER because fine-grained entities are usually subtypes of coarse-grained entities. For instance, if a model knows what Organization is, it could be easier for it to understand the concept of Government or Company. Furthermore, these methods often experience early performance saturation, necessitating the training of a new supervised learning model if the annotation extends beyond several tens of labels.
+
+A pre-finetuning strategy (Aghajanyan et al., 2021; Ma et al., 2022a) was proposed to overcome the aforementioned problem. This strategy first learns the feature representations using a coarse-grained dataset before training the fine-grained model on a fine-grained dataset. In this method, coarse-grained data are solely utilized for representation learning; thus, it still does not explicitly utilize the relationships between the coarse- and fine-grained entities.
+
+Our intuition for fully leveraging coarse-grained datasets comes mainly from the hierarchy between coarse- and fine-grained entity types. Because a coarse-grained entity type typically comprises multiple fine-grained entity types, we can enhance low-resource fine-grained NER with abundant coarse-grained data. To jointly utilize both datasets, we
+
+devise a method to build a mapping matrix called F2C (short for 'Fine-to-Coarse'), which connects fine-grained entity types to their corresponding coarse-grained entity types and propose a novel approach to train a fine-grained NER model with both datasets.
+
+Some coarse-grained entities improperly match a fine-grained entity type because datasets may be created by different annotators for different purposes. These mismatched entities can reduce the performance of the model during training. To mitigate this problem, coarse-grained entities that can degrade the performance of fine-grained NER must be eliminated. Therefore, we introduce a filtering method called 'Inconsistency Filtering'. This approach is designed to identify and exclude any inconsistent coarse-to-fine entity mappings, ensuring a higher-quality training process, and ultimately, better model performance. The main contributions of our study are as follows:
+
+- We propose an F2C mapping matrix to directly leverage the intimate relation between coarse- and fine-grained types.  
+- We present an inconsistency filtering method to screen out coarse-grained data that are inconsistent with the fine-grained types.  
+- The empirical results show that our method achieves state-of-the-art performance by utilizing the proposed F2C mapping matrix and the inconsistency filtering method.
+
+# 2 Related Work
+
+Fine-grained NER. NER is a key task in information extraction and has been extensively studied. Traditional NER datasets (Tjong Kim Sang, 2002; Ritter et al., 2011; Weischedel et al., 2013; Derczynski et al., 2017) address coarse-grained entity types in the general domain. Recently, domain-specific NER has been investigated in various fields (Hofer et al., 2018; Loukas et al., 2022; Sung et al., 2022). The domain-specific NER tasks typically employ fine-grained entity types. In addition, Ding et al. (2021) proposed a general domain fine-grained NER.
+
+$N$ -way  $K$ -shot learning for NER. Since labeling domain-specific data is an expensive process, few-shot NER has gained attention. Most few-shot NER studies are conducted using an  $N$ -way  $K$ -shot episode learning (Das et al., 2021; Huang
+
+et al., 2022; Ma et al., 2022b; Wang et al., 2022). The objective of this approach is to train a model that can correctly classify new examples into one of  $N$  classes, using only  $K$  examples per class. To achieve this generalization performance, a large number of episodes need to be generated. Furthermore, to sample these episodes, the training data must contain a significantly larger number of classes than  $N$ . In contrast, in our problem context, the number of fine-grained classes we aim to identify is substantially larger than the number of coarse-grained classes in the existing dataset. Consequently, the episode-based few-shot learning approaches are unsuitable for addressing this problem.
+
+Leveraging auxiliary data for information extraction. Several studies have employed auxiliary data to overcome the data scarcity problem. Jiang et al. (2021); Oh et al. (2023) propose NER models trained with small, strongly labeled, and large weakly labeled data. Jung and Shim (2020) used strong and weak labels for relation extraction. However, in the previous studies, the main data and auxiliary data shared the same set of classes, which is not the case for fine-grained NER with coarse-grained labels. While the approaches of Aghajanyan et al. (2021) and Ma et al. (2022a) can be applied to our problem setting, they utilize auxiliary data for representation learning rather than explicitly utilizing the relationship between the two types of data.
+
+# 3 Proposed Method
+
+In this section, we introduce the notations and define the problem of fine-grained NER using coarse-grained data. Then, we introduce the proposed CoFiNER model, including the creation of the F2C mapping matrix.
+
+# 3.1 Problem definition
+
+Given a sequence of  $n$  tokens  $\mathbf{X} = \{x_{1},x_{2},\ldots ,x_{n}\}$ , the NER task involves assigning type  $y_{i}\in E$  to each token  $x_{i}$  where  $E$  is a predefined set of entity types. In our problem setting, we used a fine-grained dataset  $\mathcal{D}^{\mathcal{F}} = \{(\mathbf{X}_1^{\mathcal{F}},\mathbf{Y}_1^{\mathcal{F}}),\dots,(\mathbf{X}_{|\mathcal{D}^{\mathcal{F}}|}^{\mathcal{F}},\mathbf{X}_{|\mathcal{D}^{\mathcal{F}}|})\}$  with a predefined entity type set  $E^{\mathcal{F}}$ . Additionally, we possess a coarse-grained dataset  $\mathcal{D}^{\mathcal{C}} = \{(\mathbf{X}_1^{\mathcal{C}},\mathbf{Y}_1^{\mathcal{C}}),\dots,(\mathbf{X}_{|\mathcal{D}^{\mathcal{C}}|}^{\mathcal{C}},\mathbf{X}_{|\mathcal{D}^{\mathcal{C}}|})\}$  characterized by a coarse-grained entity set  $E^{\mathcal{C}}$  (i.e.  $|E^{\mathcal{F}}| > |E^{\mathcal{C}}|$ ). A coarse-grained dataset typically
+
+has a smaller number of types than a fine-grained dataset. Throughout this study, we use  $\mathcal{F}$  and  $\mathcal{C}$  to distinguish between these datasets. It should be noted that our method can be readily extended to accommodate multiple coarse-grained datasets, incorporating an intrinsic multi-level hierarchy. However, our primary discussion revolves around a single coarse-grained dataset for simplicity and readability.
+
+# 3.2 Training CoFiNER
+
+We aim to utilize both coarse- and fine-grained datasets directly in a single model training. Figure 1 illustrates an overview of the fine-grained NER model training process using both coarse- and fine-grained datasets. The CoFiNER training process consists of the following four steps:
+
+Step 1- Training a fine-grained model. In the first step, we train a fine-grained model  $f^{\mathcal{F}}(\theta)$  with the low-resource fine-grained dataset  $\mathcal{D}^{\mathcal{F}}$ . This process follows a typical supervised learning approach for NER. For a training example  $(\mathbf{X}^{\mathcal{F}},\mathbf{Y}^{\mathcal{F}})$ ,  $\mathbf{X}^{\mathcal{F}}$  is fed into a PLM (Pre-trained Language Model), such as BERT, RoBERTa, to generate a contextual representations  $\mathbf{h}_i\in \mathbb{R}^d$  of each token  $x_{i}$ .
+
+$$
+\mathbf {H} = \left[ \mathbf {h} _ {1}, \dots , \mathbf {h} _ {n} \right] = \mathsf {P L M} \left(\left[ x _ {1} ^ {\mathcal {F}}, \dots , x _ {n} ^ {\mathcal {F}} \right]\right). \tag {1}
+$$
+
+Then, we apply a softmax layer to obtain the label probability distribution:
+
+$$
+\mathbf {p} _ {i} ^ {\mathcal {F}} = \operatorname {s o f t m a x} (\mathbf {W h} _ {i} + \mathbf {b})
+$$
+
+where  $\mathbf{W} \in \mathbb{R}^{|E^{\mathcal{F}}| \times d}$  and  $\mathbf{b} \in \mathbb{R}^{|E^{\mathcal{F}}|}$  represent the weights and bias of the classification head, respectively. To train the model using a fine-grained dataset, we optimize the cross-entropy loss function:
+
+$$
+L _ {\mathcal {F}} = - \frac {1}{n} \sum_ {i = 1} ^ {n} \log \mathbf {p} _ {i} ^ {\mathcal {F}} \left[ y _ {i} ^ {\mathcal {F}} \right] \tag {2}
+$$
+
+where  $y_{i}\in E^{\mathcal{F}}$  is the fine-grained label for the token  $x_{i}$
+
+Step 2 - Generating an F2C matrix. To fully leverage the hierarchy between coarse- and fine-grained entity types, we avoid training separate NER models for each dataset. Instead, we utilize a single model that incorporates an F2C mapping matrix that transforms a fine-grained output into a corresponding coarse-grained output. The F2C mapping matrix assesses the conditional probability of a coarse-grained entity type  $s \in E^{\mathcal{C}}$  given
+
+a fine-grained label  $\ell \in E^{\mathcal{F}}$  (i.e.,  $\mathbf{M}_{\ell,s} = p(y^{\mathcal{C}} = s|y^{\mathcal{F}} = \ell)$ ).
+
+Given a fine-grained probability distribution  $\mathbf{p}_i^{\mathcal{F}}$  computed using the proposed model, the marginal probability of a coarse-grained type  $s$  can be computed as
+
+$$
+\mathbf {p} _ {i} ^ {\mathcal {C}} [ s ] = \sum_ {\ell \in E ^ {\mathcal {F}}} p (y ^ {\mathcal {C}} = s | y ^ {\mathcal {F}} = \ell) \cdot \mathbf {p} _ {i} ^ {\mathcal {F}} [ \ell ].
+$$
+
+Thus, the coarse-grained output probabilities are simply computed as follows:
+
+$$
+\mathbf {p} _ {i} ^ {\mathcal {C}} = \mathbf {p} _ {i} ^ {\mathcal {F}} \cdot \mathbf {M} \tag {3}
+$$
+
+where  $\mathbf{M} \in \mathbb{R}^{|E^{\mathcal{F}}| \times |E^{\mathcal{C}}|}$  is the F2C mapping matrix whose row-wise sum is 1. By introducing this F2C mapping matrix, we can train a single model using multiple datasets with different granularity levels.
+
+Manual annotation is a straightforward approach that can be used when hierarchical information is unavailable. However, it is not only cost-intensive but also noisy and subjective, especially when there are multiple coarse-grained datasets or a large number of fine-grained entity types. We introduce an efficient method for automatically generating an F2C matrix in §3.3.
+
+Step 3 - Filtering inconsistent coarse labels. Although a fine-grained entity type is usually a subtype of a coarse-grained type, there can be some misalignments between the coarse- and fine-grained entity types. For example, an entity "Microsoft" in a financial document can either be tagged as Company or Stock which are not hierarchical. This inconsistency can significantly degrade the model's performance.
+
+To mitigate the effect of inconsistent labeling, we devise an inconsistency filtering method aimed at masking less relevant coarse labels. By automatically filtering out the inconsistent labels from the coarse-grained dataset, we investigate the coarse-grained labels using the fine-grained NER model trained in Step 1. For each token in the coarse-grained dataset, we predict the coarse-grained label using the fine-grained model and the mapping matrix as follows:
+
+$$
+\tilde {y} _ {i} ^ {\mathcal {C}} = \arg \max  \mathbf {p} _ {i} ^ {\mathcal {C}}. \tag {4}
+$$
+
+If the predicted label is the same as the coarse-grained label (i.e.,  $y_{i}^{\mathcal{C}} = \tilde{y}_{i}^{\mathcal{C}}$ ), we assume that the coarse-grained label is consistent with the fine-grained one and can benefit the model. Otherwise,
+
+![](images/1dc02e5b809b045ff1d5d338ca4cd39fe360119f51bd2c351449be365b81bec2.jpg)  
+Step 1) training process using fine-grained dataset  
+Figure 1: Overall process of our proposed method.
+
+we regard the label as inconsistent with fine-grained types and do not utilize the coarse-grained label in Step 4. Note that the fine-grained NER model is frozen during this phase.
+
+Step 4 - Jointly training CoFiNER with both datasets. The model is trained by alternating between epochs using coarse- and fine-grained data. This is an effective learning strategy for training models using heterogeneous datasets (Jung and Shim, 2020).
+
+For the fine-grained batches, CoFiNER is trained by minimizing the loss function defined in Equation (2), as described in Step 1. Meanwhile, we use the F2C mapping matrix to generate coarse-grained outputs and utilize inconsistency filtering for coarse-grained batches. Thus, we compute the cross-entropy loss between the coarse-grained label  $y_{i}^{\mathcal{C}}$  and the predicted probabilities  $\mathbf{p}_i^{\mathcal{C}}$  when the coarse-grained label is consistent with our model (i.e.,  $y_{i}^{\mathcal{C}} = \tilde{y}_{i}^{\mathcal{C}}$ ) as follows:
+
+$$
+L _ {\mathcal {C}} = - \frac {1}{m} \sum_ {i = 1} ^ {m} \log \mathbf {p} _ {i} ^ {\mathcal {C}} \left[ y _ {i} ^ {\mathcal {C}} \right] \cdot \mathbb {I} \left[ y _ {i} ^ {\mathcal {C}} = \tilde {y} _ {i} ^ {\mathcal {C}} \right] \tag {5}
+$$
+
+where  $m$  is a length of  $\mathbf{X}^{\mathcal{C}}$ . For example, suppose that the coarse label of token  $x_{i}$  is  $y_{i}^{\mathcal{C}} = 0\mathrm{RG}$ . If the estimated coarse label  $\tilde{y}_i^{\mathcal{C}}$  is 0RG, the loss for the token is  $-\log \mathbf{p}_i^{\mathcal{C}}[y_i^{\mathcal{C}}]$ . Otherwise, it is zero.
+
+# 3.3 Construction of the F2C mapping matrix
+
+The F2C mapping matrix assesses the conditional probability of a coarse-grained entity type  $s \in E^{\mathcal{C}}$  given a fine-grained label  $\ell \in E^{\mathcal{F}}$  (i.e.,  $\mathbf{M}_{\ell,s} = p(y^{\mathcal{C}} = s|y^{\mathcal{F}} = \ell)$ ). As there are only a few identical texts with both coarse- and fine-grained annotations simultaneously, we can not directly calculate this conditional probability using the data alone. Thus, we approximate the probability using a coarse-grained NER model and fine-grained labeled data as follows:
+
+$$
+\mathbf {M} _ {\ell , s} = p (y ^ {\mathcal {C}} = s | y ^ {\mathcal {F}} = \ell) \approx p (\tilde {y} ^ {\mathcal {C}} = s | y ^ {\mathcal {F}} = \ell).
+$$
+
+To generate the mapping matrix, we first train a coarse-grained NER model  $f^{\mathcal{C}}(\theta)$  with the coarse-grained dataset  $\mathcal{D}^{\mathcal{C}}$ . Then, we reannote the fine-grained dataset  $\mathcal{D}^{\mathcal{F}}$  by using the coarse-grained model  $f^{\mathcal{C}}(\theta)$ . As a result, we obtain parallel annotations for both coarse- and fine-grained types in the fine-grained data  $\mathcal{D}^{\mathcal{F}}$ . By using the parallel annotations, we can compute the co-occurrence matrix  $\mathbf{C} \in \mathbb{N}^{|E^{\mathcal{F}}| \times |E^{\mathcal{C}}|}$  where each cell  $\mathbf{C}_{\ell,s}$  is the number of tokens that are labeled as fine-grained type  $\ell$  and coarse-grained type  $s$  together.
+
+Because some labels generated by the coarse-grained model  $f^{\mathcal{C}}(\theta)$  can be inaccurate, we refine the co-occurrence matrix by retaining only the top- $k$  counts for each fine-grained type and setting the rest to zero. Our experiments show that our model performs best when  $k = 1$ . This process effic
+
+tively retains only the most frequent coarse-grained categories for each fine-grained type, thereby improving the precision of the resulting mapping. Finally, we compute the conditional probabilities for all  $\ell \in E^{\mathcal{F}}, s \in E^{\mathcal{C}}$  by using the co-occurrence counts as follows:
+
+$$
+\mathbf {M} _ {\ell , s} = p \left(\tilde {y} ^ {\mathcal {C}} = s \mid y ^ {\mathcal {F}} = \ell\right) = \frac {\mathbf {C} _ {\ell , s}}{\sum_ {s ^ {\prime} \in E ^ {\mathcal {C}}} \mathbf {C} _ {\ell , s ^ {\prime}}}. \tag {6}
+$$
+
+The F2C mapping matrix  $\mathbf{M}$  is used to predict the coarse-grained labels using Equation (3).
+
+# 4 Experiments
+
+# 4.1 Datasets
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="3"># Sentences</td><td rowspan="2"># Types</td></tr><tr><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td colspan="5">Coarse-grained datasets</td></tr><tr><td>CoNLL&#x27;03</td><td>14k</td><td>3.3k</td><td>3.4k</td><td>4</td></tr><tr><td>OntoNotes</td><td>59.9k</td><td>8.5k</td><td>8.3k</td><td>18</td></tr><tr><td colspan="5">Fine-grained dataset Few-NERD</td></tr><tr><td>10-Shot</td><td>0.3k</td><td></td><td></td><td></td></tr><tr><td>20-Shot</td><td>0.6k</td><td></td><td></td><td></td></tr><tr><td>40-Shot</td><td>1.1k</td><td>18.8k</td><td>37.6k</td><td>66</td></tr><tr><td>80-Shot</td><td>2.2k</td><td></td><td></td><td></td></tr><tr><td>100-Shot</td><td>2.7k</td><td></td><td></td><td></td></tr></table>
+
+Table 1: Statistics of each dataset.
+
+We conduct experiments using a fine-grained NER dataset, Few-NERD (SUP) (Ding et al., 2021), as well as two coarse-grained datasets, namely, OntoNotes (Weischedel et al., 2013) and CoNLL'03 (Tjong Kim Sang, 2002). The fine-grained dataset Few-NERD comprises 66 entity types, whereas the coarse-grained datasets CoNLL'03 and OntoNotes consist of 4 and 18 entity types, respectively. The statistics for the datasets are listed in Table 1.
+
+$K$ -shot sampling for the fine-grained dataset. Because we assumed a small number of examples for each label in the fine-grained dataset, we evaluated the performance in the  $K$ -shot learning setting. Although Few-NERD provides few-shot samples, they are obtained based on an  $N$ -way  $K$ -shot scenario, where  $N$  is considerably smaller than the total number of entity types. However, our goal is to identify named entities across all possible entity types. For this setting, we resampled  $K$ -shot examples to accommodate all-way  $K$ -shot scenarios.
+
+Since multiple entities exist in a single sentence, we cannot strictly generate exact  $K$ -shot samples for all the entity types. Therefore, we adopt the
+
+$K\sim (K + 5)$  -shot setting. In the  $K\sim (K + 5)$  -shot setting, there are at least  $K$  examples and at most  $K + 5$  examples for each entity type. See Appendix A for more details. In our experiments, we sampled fine-grained training data for  $K = 10,20,40,80$  ,and 100.
+
+# 4.2 Experimental Settings
+
+In experiments, we use transformer-based PLM, including  $\mathrm{BERT}_{\mathrm{BASE}}$ ,  $\mathrm{RoBERTa}_{\mathrm{BASE}}$ , and  $\mathrm{RoBERTa}_{\mathrm{LARGE}}$ . In CoFiNER, we follow  $\mathrm{RoBERTa}_{\mathrm{LARGE}}$  to build a baseline model. The maximum sequence length is set to 256 tokens. The AdamW optimizer (Loshchilov and Hutter, 2019) is used to train the model with a learning rate of  $2e - 5$  and a batch size of 16. The number of epochs is varied for each model. We train the fine-grained model, CoFiNER, over 30 epochs. To construct the F2C mapping matrix, the coarse-grained model is trained for 50 epochs using both CoNLL'03 and OntoNotes. To train the inconsistency filtering model, we set different epochs based on the number of shots: For 10, 20, 40, 80, and 100 shot settings, the epochs are 150, 150, 120, 50, and 30, respectively. We report the results using span-level F1. The dropout with a probability of 0.1 is applied. All the models were trained on NVIDIA RTX 3090 GPUs.
+
+# 4.3 Compared Methods
+
+In this study, we compare the performance of CoFiNER with that of both the supervised and few-shot methods. We modified the existing methods for our experimental setup and re-implemented them accordingly.
+
+Supervised method. We use  $\mathrm{BERT}_{\mathrm{BASE}}$ ,  $\mathrm{RoBERTa}_{\mathrm{BASE}}$ , and  $\mathrm{RoBERTa}_{\mathrm{LARGE}}$  as the supervised baselines, each including a fine-grained classifier on the head. In addition, PIQN (Shen et al., 2022) and PL-Marker (Ye et al., 2022) are methods that have achieved state-of-the-art performance in a supervised setting using the full Few-NERD dataset. All models are trained using only a Few-NERD dataset.
+
+Few-shot method. The LSFS (Ma et al., 2022a) leverages a label encoder to utilize the semantics of label names, thereby achieving state-of-the-art results in low-resource NER settings. The LSFS applies a pre-finetuning strategy to learn prior knowledge from the coarse-grained dataset, OntoNotes. For a fair comparison, we also conducted pre
+
+<table><tr><td>Method</td><td>Model</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>100-shot</td></tr><tr><td rowspan="5">Supervised Method</td><td>BERTBASE (Devlin et al., 2019)</td><td>23.101</td><td>34.718</td><td>43.138</td><td>46.182</td><td>46.449</td></tr><tr><td>RoBERTaBASE (Liu et al., 2019)</td><td>21.073</td><td>34.157</td><td>39.226</td><td>45.735</td><td>46.942</td></tr><tr><td>RoBERTaLARGE (Liu et al., 2019)</td><td>29.137</td><td>41.425</td><td>51.137</td><td>55.057</td><td>54.172</td></tr><tr><td>PIQN (Shen et al., 2022)</td><td>21.750</td><td>22.007</td><td>28.533</td><td>29.339</td><td>38.658</td></tr><tr><td>PL-Marker (Ye et al., 2022)</td><td>40.902</td><td>48.064</td><td>52.395</td><td>53.249</td><td>53.061</td></tr><tr><td>Few-shot Method</td><td>LSFS (Ma et al., 2022a)</td><td>47.998</td><td>43.269</td><td>50.595</td><td>51.420</td><td>50.366</td></tr><tr><td>Proposed Method</td><td>CoFiNER</td><td>44.951</td><td>51.142</td><td>56.409</td><td>56.847</td><td>57.178</td></tr></table>
+
+finetuning on OntoNotes and performed fine-tuning on each shot of the Few-NERD dataset.
+
+Proposed method. We trained our CoFiNER model as proposed in §3. In each epoch, CoFiNER is first trained on two coarse-grained datasets: OntoNotes and CoNLL'03. Subsequently, it is trained on the fine-grained dataset Few-NERD. We used RoBERTa $_{\text{LARGE}}$  as the pre-trained language model for CoFiNER in Equation (1).
+
+# 4.4 Main Results
+
+Table 2 reports the performance of CoFiNER and existing methods. The result shows that CoFiNER outperforms both supervised learning and few-shot learning methods. Because supervised learning typically needs a large volume of training data, these models underperform in low-resource settings. This demonstrates that our method effectively exploits coarse-grained datasets to enhance the performance of the fine-grained NER model. In other words, the proposed F2C mapping matrix significantly reduces the amount of fine-grained dataset required to train supervised NER models. In particular, CoFiNER achieves significant performance improvements compared to the state-of-the-art model PL-Marker, which also utilizes the same pre-trained language model RoBERTa $_{\text{LARGE}}$  as CoFiNER.
+
+The few-shot method LSFS yields the highest F1 score for the 10-shot case. However, this few-shot method suffers from early performance saturation, resulting in less than a 2.4 F1 improvement with an additional 90-shot. By contrast, the F1 score of CoFiNER increases by 12.2. Consequently, CoFiNER outperforms all the compared methods except for the 10-shot case. In summary, the proposed method yields promising results for a wide range of data sample sizes by explicitly leveraging the inherent hierarchical structure.
+
+Table 2: Results on few-shot NER. The best scores across all models are marked bold.  
+
+<table><tr><td>Model</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>100-shot</td></tr><tr><td>CoFiNER</td><td>44.95</td><td>51.14</td><td>56.41</td><td>56.85</td><td>57.18</td></tr><tr><td>w/o filtering</td><td>41.38</td><td>43.87</td><td>53.42</td><td>54.70</td><td>55.13</td></tr><tr><td>w/o OntoNotes</td><td>42.63</td><td>44.51</td><td>57.18</td><td>56.97</td><td>55.54</td></tr><tr><td>w/o CoNLL&#x27;03</td><td>43.99</td><td>50.07</td><td>57.54</td><td>56.71</td><td>56.27</td></tr><tr><td>w/o coarse</td><td>29.14</td><td>41.43</td><td>51.14</td><td>55.06</td><td>54.17</td></tr></table>
+
+Table 3: Performances of ablation study over different components.
+
+# 4.5 Ablation Study
+
+An ablation study is conducted to validate the effectiveness of each component of the proposed method. The results are presented in Table 3. First, we remove the inconsistency filtering  $(w / o$  filtering) and observe a significant decrease in the F1 score, ranging from 2.05 to 7.27. These results demonstrate the effectiveness of our filtering method, which excludes mislabeled entities. Second, we provide the results using a single coarse-grained dataset (w/o OntoNotes and w/o CoNLL'03). Even with a single coarse-grained dataset, our proposed method significantly outperforms w/o coarse, which is trained solely on the fine-grained dataset (i.e. RoBERTaLarge in Table 2).
+
+This indicates the effectiveness of using a well-aligned hierarchy through the F2C mapping matrix and inconsistency filtering. Although we achieve a significant improvement even with a single coarse-grained dataset, we achieve a more stable result with two coarse-grained datasets. This implies that our approach effectively utilizes multiple coarse-grained datasets, although the datasets contain different sets of entity types.
+
+# 4.6 Analysis
+
+In this section, we experimentally investigate how the F2C mapping matrix and inconsistency filtering improve the accuracy of the proposed model.
+
+![](images/6438e8c4c7287b9e5ab4a53f5e279aaa5f071c5b156d36d50978ade55f2201ac.jpg)  
+Figure 2: The F2C mapping matrix between OntoNotes and 100-shot of the Few-NERD dataset. During the generation of the F2C mapping matrix, some coarse-grained types are not mapped to any fine-grained types. The 7 unmapped types are not represented: DATE, TIME, PERCENT, MONEY, QUANTITY, ORDINAL, and CARDINAL.
+
+![](images/b838be05e6f9191f82309e9b94331af3f976a4094b598b435f389ba3429b65ed.jpg)  
+Figure 3: The F2C mapping matrix between CoNLL'03 and 100-shot of the Few-NERD dataset.
+
+# 4.6.1 F2C Mapping Matrix
+
+Mapping the entity types between the coarse- and fine-grained datasets directly affects the model performance. Hence, we investigate the mapping outcomes between the coarse- and fine-grained datasets. Figures 2 and 3 show the F2C matrices for FewNERD-OntoNotes and FewNERD-CoNLL'03, respectively. In both figures, the x- and y-axis represent the fine-grained and coarse-grained entity types, respectively. The colored text indicates the corresponding coarse-grained entity types in FewNERD, which were not used to find the mapping matrix. The mapping is reliable if we compare the y-axis and the colored types (coarse-grained types in Few-NERD). Even without manual annotation of the relationship between coarse- and fine-grained entities, our method successfully obtains reliable mapping from fine-grained to coarse-grained types. Our model can be effectively trained with both types of datasets using accurate F2C mapping.
+
+Figure 4 provides the F1 scores by varying the hyperparameter  $k$  to refine the F2C mapping matrix described in §3.3. 'all' refers to the usage of the complete frequency distribution when creating an
+
+![](images/c3bc998e35fe8e0f869651ee62c65766952daacd568e73e3190ad7b4ea60e9bb.jpg)  
+Figure 4: Impact of top- $k$  values in the F2C mapping matrix. F1 is reported on the test data.
+
+F2C mapping matrix. We observe that the highest performance is achieved when  $k$  is set to 1, and as  $k$  increases, the performance gradually decreases. Although the optimal value of  $k$  can vary depending on the quality of the coarse-grained data and the performance of the coarse-grained NER model, the results indicate that a good F2C mapping matrix can be obtained by ignoring minor co-occurrences.
+
+Additionally, We conducted an experiment by setting the F2C mapping matrix to be learnable and comparing it with our non-learnable F2C matrix. The non-learnable approach showed better
+
+<table><tr><td>#</td><td>Type</td><td>Example</td></tr><tr><td rowspan="2">1</td><td>predict</td><td>... they may be offshoots of the intifadahEVENT, the Palestinian rebellionEVENT in ...</td></tr><tr><td>target</td><td>... they may be offshoots of the intifadah, the PalestinianNORP rebellion in ...</td></tr><tr><td rowspan="2">2</td><td>predict</td><td>... players heading to CanadaGPE, particularly TorontoGPE for the All-starEVENT ...</td></tr><tr><td>target</td><td>... players heading to CanadaGPE, particularly TorontoGPE for the All-star ...</td></tr><tr><td rowspan="2">3</td><td>predict</td><td>Judges at Flight 103 Lockerbie trialEVENT are expected ...</td></tr><tr><td>target</td><td>Judges at Flight 103 LockerbieGPE trial are expected ...</td></tr></table>
+
+Table 4: Inconsistent examples: The labeled sentence in the original coarse-grained dataset is the "target" type, while the label predicted by the model is the "predict" type. Consistent entity types are indicated in blue, while inconsistent entity types are indicated in red.
+
+performance, hence we adopted this approach for CoFiNER. Detailed analysis and experiment results are in Appendix B.2.
+
+# 4.6.2 Inconsistency Filtering
+
+We aim to examine whether inconsistency filtering successfully screened out inconsistent entities between OntoNotes and Few-NERD datasets. To conduct this analysis, we describe three inconsistent examples. Table 4 shows the predicted values and target labels, which correspond to the coarse-grained output of the fine-grained NER model trained as described in §3.2 and the golden labels of the coarse-grained dataset.
+
+The first example illustrates the inconsistency in entity types between mislabeled entities. In the original coarse-grained dataset, "Palestinian" is labeled as NORP, but the model trained on the fine-grained dataset predicts "Palestinian rebellion" as its appropriate label, EVENT. However, annotators of the OntoNotes labeled "Palestinian" as NORP, whereas the fine-grained NER model correctly predicts the highly informative actual label span. The inconsistency caused by a label mismatch between the coarse-grained and fine-grained datasets can result in performance degradation.
+
+In the second example, both "Canada" and "Toronto" are consistently labeled as GPE; thus, the model is not confused when training on these two entities. However, in the case of "All-star", we can observe a mismatch. This example in the coarse-grained dataset is labeled 0 instead of the correct entity type EVENT, indicating a mismatch. Through inconsistency filtering, unlabeled "All-star" is masked out of the training process.
+
+As shown in the examples, inconsistency filtering is necessary to mitigate the potential noise arising from mismatched entities. We analyze the filtering results for each coarse-grained label to assess its impact on model performance. Figure 5
+
+![](images/da7155def09c630523ce54071fbd3f1ebb60dd871854b671c40764ccb6be0504.jpg)  
+Figure 5: Performance with increasing filtering proportion between Few-NERD and OntoNotes datasets. The correlation coefficient between filtering proportion and performance improvement is 0.29.
+
+illustrates the correlation between the filtering proportion and the performance improvement for each coarse-grained label mapped to the fine-grained labels. In this figure, a higher filtering proportion indicates a greater inconsistency between the two datasets. F1 improvements indicate the difference in performance when filtering is applied and when it is not. Each data point refers to a fine-grained label mapped onto a coarse-grained label. As the proportion of the filtered entities increases, the F1 scores also increase. These improvements indicate that inconsistency filtering effectively eliminates noisy entities, enabling the model to be well-trained on consistent data only.
+
+# 5 Conclusion
+
+We proposed CoFiNER, which explicitly leverages the hierarchical structure between coarse- and fine-grained types to alleviate the low-resource problem of fine-grained NER. We devised the F2C mapping matrix that allows for fine-grained NER model
+
+training using additional coarse-grained datasets. However, because not all coarse-grained entities are beneficial for the fine-grained NER model, the proposed inconsistency filtering method is used to mask out noisy entities from being used in the model training process. We found through experiments that using a smaller amount of consistent data is better than using a larger amount of data without filtering, thus demonstrating the crucial role of inconsistency filtering. Empirical results confirmed the superiority of CoFiNER over both the supervised and few-shot methods.
+
+# Limitations
+
+Despite the promising empirical results of this study, there is still a limitation. The main drawback of CoFiNER is that the token-level F2C mapping matrix and inconsistency filtering may not be directly applicable to nested NER tasks. Nested NER involves identifying and classifying certain overlapping entities that exceed the token-level scope. Because CoFiNER addresses fine-grained NER at the token level, it may not accurately capture entities with nested structures. Therefore, applying our token-level approach to nested NER could pose challenges and may require further adaptations or other modeling techniques to effectively handle the hierarchical relations between nested entities.
+
+# Acknowledgments
+
+This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF2022R1G1A1013549). This work was supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No. RS2023-00261068, Development of Lightweight Multimodal Anti-Phishing Models and Split-Learning Techniques for Privacy-Preserving Anti-Phishing) and (No.RS-2022-00155885, Artificial Intelligence Convergence Innovation Human Resources Development (Hanyang University ERICA)). This research was also supported by the Ministry of Trade, Industry, and Energy (MOTIE), Korea, under the "Project for Research and Development with Middle Markets Enterprises and DNA(Data, Network, AI) Universities" (Smart Home Based LifeCare Service & Solution Development)(P0024555) supervised by the Korea Institute for Advancement of Technology (KIAT).
+
+# References
+
+Armen Aghajanyan, Anchit Gupta, Akshat Shrivastava, Xilun Chen, Luke Zettlemoyer, and Sonal Gupta. 2021. Muppet: Massive multi-task representations with pre-finetuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5799-5811, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Sarkar Snigdha Sarathi Das, Arzoo Katiyar, Rebecca J Passonneau, and Rui Zhang. 2021. Container: Few-shot named entity recognition via contrastive learning. arXiv preprint arXiv:2109.07589.  
+Leon Derczynski, Eric Nichols, Marieke van Erp, and Nut Limsopatham. 2017. Results of the WNUT2017 shared task on novel and emerging entity recognition. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 140-147, Copenhagen, Denmark. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Ning Ding, Guangwei Xu, Yulin Chen, Xiaobin Wang, Xu Han, Pengjun Xie, Haitao Zheng, and Zhiyuan Liu. 2021. Few-NERD: A few-shot named entity recognition dataset. 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 3198-3213, Online. Association for Computational Linguistics.  
+Maximilian Hofer, Andrey Kormilitzin, Paul Goldberg, and Alejo Nevada-Holgado. 2018. Few-shot learning for named entity recognition in medical text.  
+Yucheng Huang, Kai He, Yige Wang, Xianli Zhang, Tieliang Gong, Rui Mao, and Chen Li. 2022. Copner: Contrastive learning with prompt guiding for few-shot named entity recognition. In Proceedings of the 29th International conference on computational linguistics, pages 2515-2527.  
+Haoming Jiang, Danqing Zhang, Tianyu Cao, Bing Yin, and Tuo Zhao. 2021. Named entity recognition with small strongly labeled and large weakly labeled data.  
+Woohwan Jung and Kyuseok Shim. 2020. Dual supervision framework for relation extraction with distant supervision and human annotation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6411-6423, Barcelona, Spain (Online). International Committee on Computational Linguistics.
+
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Lefteris Loukas, Manos Fergadiotis, Ilias Chalkidis, Eirini Spyropoulou, Prodromos Malakasiotis, Ion Androutsopoulos, and Georgios Paliouras. 2022. Finer: Financial numeric entity recognition for xbrl tagging. arXiv preprint arXiv:2203.06482.  
+Jie Ma, Miguel Ballesteros, Srikanth Doss, Rishita Anubhai, Sunil Mallya, Yaser Al-Onaizan, and Dan Roth. 2022a. Label semantics for few shot named entity recognition. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1956-1971, Dublin, Ireland. Association for Computational Linguistics.  
+Tingting Ma, Huiqiang Jiang, Qianhui Wu, Tiejun Zhao, and Chin-Yew Lin. 2022b. Decomposed meta-learning for few-shot named entity recognition. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 1584–1596.  
+Alexis Mitchell, Stephanie Strassel, Shudong Huang, and Ramez Zakhary. 2005. Ace 2004 multilingual training corpus. Linguistic Data Consortium, Philadelphia, 1:1-1.  
+Seongwoong Oh, Woohwan Jung, and Kyuseok Shim. 2023. Thunder: Named entity recognition using a teacher-student model with dual classifiers for strong and weak supervisions. In Proceedings of 26th European Conference on Artificial Intelligence.  
+Alan Ritter, Sam Clark, Mausam, and Oren Etzioni. 2011. Named entity recognition in tweets: An experimental study. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1524-1534, Edinburgh, Scotland, UK. Association for Computational Linguistics.  
+Yongliang Shen, Xiaobin Wang, Zeqi Tan, Guangwei Xu, Pengjun Xie, Fei Huang, Weiming Lu, and Yueting Zhuang. 2022. Parallel instance query network for named entity recognition. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 947-961, Dublin, Ireland. Association for Computational Linguistics.  
+Mujeen Sung, Minbyul Jeong, Yonghwa Choi, Donghyeon Kim, Jinhyuk Lee, and Jaewoo Kang. 2022. Bern2: an advanced neural biomedical named entity recognition and normalization tool. Bioinformatics, 38(20):4837-4839.  
+Erik F. Tjong Kim Sang. 2002. Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition. In COLING-02: The 6th
+
+Conference on Natural Language Learning 2002 (CoNLL-2002).  
+Christopher Walker, Stephanie Strassel, Julie Medero, and Kazuaki Maeda. 2006. Ace 2005 multilingual training corpus. Linguistic Data Consortium, Philadelphia, 57:45.  
+Peiyi Wang, Runxin Xu, Tianyu Liu, Qingyu Zhou, Yunbo Cao, Baobao Chang, and Zhifang Sui. 2022. An enhanced span-based decomposition method for few-shot sequence labeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5012-5024.  
+Ralph Weischedel, Martha Palmer, Mitchell Marcus, Edward Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, et al. 2013. Ontonotes release 5.0 ldc2013t19. Linguistic Data Consortium, Philadelphia, PA, 23.  
+Deming Ye, Yankai Lin, Peng Li, and Maosong Sun. 2022. Packed levitated marker for entity and relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4904-4917, Dublin, Ireland. Association for Computational Linguistics.
+
+# Appendix
+
+# A Sampling Algorithm
+
+Algorithm 1  $K\sim (K + 5)$  sampling algorithm
+
+Input: Dataset  $\mathbf{X}$ , labeled set  $\mathbf{Y}$ ,  $K$
+
+Output: Fine-grained dataset  $D^{\mathcal{F}}$
+
+1: Sort classes in  $\mathbf{Y}$  based on their freq. in  $\mathbf{X}$  
+2:  $D^{\mathcal{F}}\gets \emptyset$  Init the train dataset  $\triangleright$  Init the count of entity classes in  $D^{\mathcal{F}}$  
+3: for  $i = 1$  to  $|\mathbf{Y}|$  do  
+4: Count  $[i] = 0$  
+5: for  $i = 1$  to  $|\mathbf{Y}|$  do  
+6: if  $\forall \operatorname{Count}[i'] \geq K$  then  
+7: break;  
+8: Randomly sample  $(x,y)\in \mathbf{X}$  s.t.  $\mathbf{Y}_i\in y$  
+9: for  $j = 1$  to  $|y|$  do  
+10:  $\mathbf{Count}^{\mathbf{Y}}[j] + = 1;$  
+11: if  $\exists \operatorname{Count}[i] + \operatorname{Count}^{\mathbf{Y}}[i] > K + 5$  then  
+12: Continue;  
+13: else  
+14:  $D^{\mathcal{F}}\gets D^{\mathcal{F}}\cup \{(x,y)\} ;$  
+15: Count  $\leftarrow$  Count + CountY;  
+16: return  $D^{\mathcal{F}}$
+
+Unlike other tasks, NER involves multiple entity occurrences within a sentence, making it too restrictive to sample an exact count. We adopt a
+
+$K \sim (K + 5)$ -shot setting to minimize differences in the number of entity types. Additionally, we sample the entity types, starting from those with fewer occurrences to ensure a balanced distribution of multiple entity types within the sentences. Algorithm 1 presents the  $K$ -shot sampling algorithm used in this study.
+
+# B Additional Experiments
+
+# B.1 Generalization of Inconsistency filtering in diverse dataset settings
+
+<table><tr><td rowspan="2">Setting</td><td colspan="2">Coarse-grained datasets</td></tr><tr><td>Few-NERD_coarse</td><td>OntoNotes + CoNLL&#x27;03</td></tr><tr><td>w/ filtering</td><td>56.63</td><td>57.18</td></tr><tr><td>w/o filtering</td><td>56.46</td><td>55.13</td></tr></table>
+
+To assess the generalization of the proposed method, we conducted experiments under various coarse- and fine-grained dataset settings.
+
+First, we verified the robustness of inconsistency filtering through experiments using different coarse-grained datasets. The fine-grained dataset, Few-NERD, remains unchanged. Since Few-NERD has both coarse- and fine-grained labels, we constructed a coarse-grained dataset Few-NERD_coarse using the coarse-grained labels from Few-NERD. When compared to Few-NERD_coarse, entities in OntoNotes and CoNLL'03 are inconsistently labeled with Few-NERD because they were independently created. In Table 5, Few-NERD_coarse exhibits a higher F1 score in the w/o filtering setting due to its consistency with fine-grained labels of Few-NERD. However, the performance improvement achieved with filtering is more substantial in the inconsistent datasets, when compared to the consistent dataset. This result indicates that inconsistency filtering improves performance by filtering out the mismatching labels. Therefore, we have demonstrated the importance of using inconsistency filtering to filter out noise when working with datasets that employ different labeling schemes. Furthermore, by achieving effective performance improvements across various coarse-grained datasets, we have provided evidence of the robustness of the filtering method.
+
+Second, we validate the generalization performance through experiments conducted in different coarse- and fine-grained dataset settings. We
+
+Table 5: Results on different coarse-grained datasets. A fine-grained dataset is 100-shot of the Few-NERD.  
+
+<table><tr><td>Model</td><td>F1</td></tr><tr><td>RoBERTaLARGE</td><td>75.15</td></tr><tr><td>PL-Marker</td><td>74.03</td></tr><tr><td>CoFiNER</td><td>80.44</td></tr><tr><td>w/o filtering</td><td>78.20</td></tr><tr><td>w/o coarse</td><td>75.15</td></tr></table>
+
+set up the CoNLL'03, which has 4 entity types, as the coarse-grained dataset, and the 100-shot OntoNotes, which has a finer label with 19 entity types, as the fine-grained dataset. In Table 6, when compared to the two top-performing models in the main results, RoBERTa $_{\text{LARGE}}$  and the state-of-the-art PL-Marker, CoFiNER show consistently higher performance. Furthermore, as shown in the ablation study that was conducted following the same methodology in §4.5, CoFiNER exhibited superior performance.
+
+In conclusion, through above the two experiments, our method has been demonstrated to work robustly across different coarse- and fine-grained dataset settings.
+
+# B.2 Comparison with learnable F2C mapping matrix
+
+Table 6: Performances of different models and ablation studies on our model. RoBERTa $_{\text{LARGE}}$  and w/o coarse are identical.  
+
+<table><tr><td>Matrix Type</td><td>F1</td></tr><tr><td>non-learnable</td><td>56.27</td></tr><tr><td>learnable</td><td>53.25</td></tr></table>
+
+Table 7: Results on learnable and non-learnable F2C mapping matrix on 100-shot of Few-NERD.
+
+To find the optimal F2C mapping matrix, we conducted experiments to explore the impact of making the F2C mapping matrix learnable. We use FewNERD as a fine-grained dataset and OntoNotes as a coarse-grained dataset. Table 7 shows no performance gains when the F2C mapping matrix was set to be learnable. We found that the learnable matrix tends to form a pattern similar to what is shown in Figure 4 with  $k =$  all. This result suggests that taking minor co-occurrences into account leads to an overall decrease in performance. Based on this analysis, the non-learnable mapping matrix is used in our experiments.

2023/Enhancing Low-resource Fine-grained Named Entity Recognition by Leveraging Coarse-grained Datasets/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd813dcc58d25670c7121547d6f6cc1b998594826e15bf914ac9219f3aca8542
+size 470688

2023/Enhancing Structured Evidence Extraction for Fact Verification/ebbbd673-d241-414d-91ff-1168a052ced6_content_list.json

@@ -0,0 +1,1767 @@
+[
+    {
+        "type": "text",
+        "text": "Enhancing Structured Evidence Extraction for Fact Verification",
+        "text_level": 1,
+        "bbox": [
+            166,
+            90,
+            831,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zirui Wu, Nan Hu, Yansong Feng*",
+        "bbox": [
+            346,
+            142,
+            655,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Wangxuan Institute of Computer Technology, Peking University, China",
+        "bbox": [
+            211,
+            159,
+            791,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{ziruiwu,hunan,fengyansong}@pku.edu.cn",
+        "bbox": [
+            307,
+            177,
+            694,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            341,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Open-domain fact verification is the task of verifying claims in natural language texts against extracted evidence. FEVERIOUS is a benchmark that requires extracting and integrating both unstructured and structured evidence to verify a given claim. Previous models suffer from low recall of structured evidence extraction, i.e., table extraction and cell selection. In this paper, we propose a simple but effective method to enhance the extraction of structured evidence by leveraging the row and column semantics of tables. Our method comprises two components: (i) a coarse-grained table extraction module that selects tables based on rows and columns relevant to the claim and (ii) a fine-grained cell selection graph that combines both formats of evidence and enables multihop and numerical reasoning. We evaluate our method on FEVERIOUS and achieve an evidence recall of  $60.01\\%$  on the test set, which is  $6.14\\%$  higher than the previous state-of-the-art performance. Our results demonstrate that our method can extract tables and select cells effectively, and provide better evidence sets for verdict prediction. Our code is released at https://github.com/WilliamZR/see-st",
+        "bbox": [
+            141,
+            281,
+            460,
+            650
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            677,
+            258,
+            692
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Open-domain fact verification is the task of verifying a factual claim based on evidence extracted from a knowledge base(Guo et al., 2022; Hardalov et al., 2022). A complete evidence set provides necessary information for a verification module to support or refute the claim. In the real world, fact-checkers could use both unstructured and structured data as evidence.",
+        "bbox": [
+            112,
+            703,
+            489,
+            829
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Tables are a structured form of evidence that provides an organized representation of information that facilitates effortless comparison and cross-referencing. By presenting information in cells,",
+        "bbox": [
+            112,
+            832,
+            489,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a30dfb3cb74cadfded92f601f88388672a67962304357d8f8be3d53e2d883569.jpg",
+        "image_caption": [
+            "Figure 1: Example of FEVERIOUS with gold evidence. The gold cell evidence is highlighted in orange."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            252,
+            875,
+            514
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tables enable efficient look-up of individual values, while the arrangement of rows and columns allows for the extraction of high-order semantics through various operations, such as comparisons, filtering, arithmetic calculations, and the identification of minimum and maximum values. The two-dimension structure and high-order semantics of tables make them more difficult to extract compared with texts. Regarding the previous method to extract structured evidence. Aly et al. (2021) treats tables as text sequences neglecting the table structure itself and extracts them only based on lexical matching. This approach is also used in many following works (Bouziane et al., 2021; Saeed et al., 2021; Hu et al., 2022, 2023). Consequently, these methods struggle to extract tables with little word overlap with the claim. Alternatively, Bouziane et al. (2021) and Gi et al. (2021) convert each table cell into a text sequence and perform cell retrieval directly. Kotonya et al. (2021) construct a reason-",
+        "bbox": [
+            507,
+            596,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding Author.",
+        "bbox": [
+            141,
+            904,
+            294,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "6631",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6631-6641 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            159,
+            945,
+            835,
+            973
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/66785a60fe362381abd20e3d7bd3ce868abb47c5bb5f99bdf937adc6e20bffaa.jpg",
+        "image_caption": [
+            "Figure 2: Overview of our proposed approach SEE-ST. It extracts structured evidence based on retrieved pages and extracted sentences. SEE-ST includes two modules: coarse-grained table extraction (§2.1) and fined-grained cell selection based on an evidence graph(§2.2)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            87,
+            884,
+            340
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ing graph of sentences and cells to extract evidence. Yet these methods overlook the context within the table and fail to capture high-order semantics from rows and columns.",
+        "bbox": [
+            112,
+            423,
+            487,
+            486
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Figure 1 presents an example from FEVEROUS (Aly et al., 2021), a benchmark on open-domain fact verification using both sentences and table cells as evidence. The verification model is expected to extract both unstructured and structured evidence to verify the claim. The structured table can be interpreted in several ways for the model: a complete table, a combination of rows or columns, or a composition of isolated cells. The entire table encompasses a substantial amount of information. During table extraction, the Site and Attendance columns are irrelevant and only introduce confusion for both lexical and semantic methods (Herzig et al., 2021). Cells typically contain short phrases and numbers, which are often insufficient for conveying comprehensive information. In contrast, rows and columns possess a similar volume of information to the claim, thereby exhibiting comparable semantic granularity. The extraction process should prioritize the information from the 3rd row and the Result column. During the cell selection step, the selection module must capture the high-order semantics in the Result column to verify the statement ... lost all its games except for one.... Moreover, it needs to identify the cells October 7 and Illinois State by integrating informa",
+        "bbox": [
+            110,
+            500,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tion from the Date and Opponent columns and the 3rd row. Overall rows and columns have comparable semantic granularity with the claim and are the most suitable fundamental unit for both table extraction and cell selection.",
+        "bbox": [
+            507,
+            423,
+            882,
+            502
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose Structured Evidence Extraction with Row and Column Semantics of Tables (SEE-ST), a novel approach to enhance both table extraction and cell selection modules for fact verification by leveraging the semantics of rows and columns. This method aligns the semantic granularity of structured evidence with claim and sentence evidence, enabling more precise evidence extraction. SEE-ST starts by extracting tables from retrieved Wikipedia pages and focuses on identifying the most relevant rows and columns for the claim, thereby reducing confusion arising from irrelevant cells. Since a cell represents the intersection of a row and a column, its selection can be conducted by analyzing both dimensions. We integrate rows and columns within a graph neural network to facilitate fine-grained cell selection while preserving high-order semantics. Additionally, we incorporate extracted sentence evidence into the graph to merge information from evidence of both formats at a similar semantic granularity.",
+        "bbox": [
+            507,
+            510,
+            884,
+            848
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions can be summarized as follows: (i) We propose to leverage the row and column semantics to enhance table extraction and cell selection. (ii) We design an evidence graph for cell",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "6632",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "selection. It enables rows and columns to interact with sentences at comparable semantic granularity while still maintaining high-order semantics from table evidence. (iii) Our proposed method SEE-ST achieves state-of-the-art performance in evidence extraction on the FEVEROUS dataset, with better evidence sets contributing to the increased accuracy of the verification module.",
+        "bbox": [
+            112,
+            84,
+            487,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2 Our Model",
+        "text_level": 1,
+        "bbox": [
+            112,
+            227,
+            245,
+            242
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "FEVEROUS is a benchmark on open-domain fact verification (Aly et al., 2021). The task is to verify a claim  $c$  based on Wikipedia. We follow the three-step pipeline in open-domain fact verification (Thorne et al., 2018; Aly et al., 2021). The three steps are document retrieval, evidence extraction, and verdict prediction. The first step retrieves  $n_p$  relevant pages from the Wikipedia dump. Then the second step extracts sentence evidence  $S$  and cell evidence  $C$  from the retrieved pages. At last, the model predicts the veracity label  $\\hat{y}$  of the claim based on sentence and cell evidence.",
+        "bbox": [
+            112,
+            254,
+            487,
+            445
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "SEE-ST enhances the extraction of structured evidence in the second step of the pipeline. It operates on the same conceptual framework, utilizing row and column semantics to enhance evidence extraction (Figure 2). SEE-ST performs structured evidence extraction in two stages: (i) Coarse-grained table extraction that retrieves top  $n_t$  tables for the second step. It scores relevance between rows or columns with the claim and ranks tables accordingly and (ii) Fine-grained cell selection that retrieves cell evidence from complex tables. It constructs a graph neural network of rows, columns, and sentences at a comparable semantic granularity. It utilizes a graph attention network to enable delicate information integration between different pieces of evidence. The cells are selected as the intersection of rows and columns.",
+        "bbox": [
+            115,
+            449,
+            489,
+            720
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "At the verification step, we utilize DCUF (Hu et al., 2022), a method that converts evidence into dual-channel encodings to verify the claim.",
+        "bbox": [
+            112,
+            722,
+            487,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.1 Coarse-grained Table Extraction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            784,
+            418,
+            799
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce our table extraction module. First, the claim and table pair are fed to TAPAS, a pre-trained table model aware of table structures(Herzig et al., 2020). Instead of using embedding of [CLS] token to predict the relevance between the claim and table. our module predicts the relevance based on row and column semantics.",
+        "bbox": [
+            112,
+            806,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Each row and column in the evidence table is represented as the average of all the token embeddings in the same row or column. Then we predict a probability score for each row and column separately.",
+        "bbox": [
+            507,
+            84,
+            882,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The selective score for the table is computed as the multiplication of the maximum probability score for the rows and columns.",
+        "bbox": [
+            507,
+            149,
+            882,
+            195
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {P} (t) = \\max \\left(\\mathrm {P} _ {r o w}\\right) \\times \\max \\left(\\mathrm {P} _ {c o l}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            208,
+            823,
+            225
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The training set is constructed with all the tables that contain gold cell evidence as positive examples. As for negative examples, we use top tables retrieved by DrQA (Chen et al., 2017) which do not contain cell evidence but have lexical overlap with the claim. The ratio of positive and negative examples is 1 at every mini-batch in training. During training, the model is trained to select table rows and columns that contain gold cell evidence. The loss function is computed with a cross-entropy loss on each row and column.",
+        "bbox": [
+            505,
+            237,
+            882,
+            412
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL = \\alpha_ {t} L ^ {R o w} + \\beta_ {t} L ^ {C o l}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            606,
+            423,
+            784,
+            441
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "$L^{Row}$ , and  $L^{Col}$  are cross-entropy loss functions of rows and columns. These rows and columns that have gold cell evidence are labeled as select while rows/columns in the same table but do not contain cell evidence and rows/columns in negative examples are labeled as not select.",
+        "bbox": [
+            507,
+            451,
+            882,
+            548
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "At inference, a selective score is computed for each table in retrieved pages. Top  $n_t$  tables are selected for the following fine-grained cell selection.",
+        "bbox": [
+            507,
+            550,
+            882,
+            598
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Fine-grained Cell Selection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            625,
+            766,
+            639
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following the idea of leveraging row and column semantics, we build an evidence graph  $\\mathcal{G}(\\mathcal{V},\\mathcal{E})$  to select cell evidence. The first two types of nodes  $v_{0},\\ldots,v_{M}$  and  $v_{M+1},\\ldots,v_{M+N}$  encode table rows and columns. Shown as yellow nodes in Figure 2.  $M$  and  $N$  are the total number of table rows and columns. The third type of node is sentences  $v_{M+N},\\ldots,v_{K}$ , shown as blue nodes in Figure 2. Overall, the evidence graph has  $K$  nodes of unstructured and structured evidence at a comparable semantic granularity.",
+        "bbox": [
+            505,
+            646,
+            882,
+            821
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Graph edges  $\\mathcal{E}$  are constructed between related evidence following three rules: (i) Cross-Format Related:  $v_{i}$  and  $v_{j}$  are different format evidence of sentence and table, and in the same Wikipedia page. (ii) Structure-Based:  $v_{i}$  and  $v_{j}$  are nodes from the same table and (iii) Entity-Based:  $v_{i}$  and  $v_{j}$  have",
+        "bbox": [
+            507,
+            822,
+            882,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "6633",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "common entities and hyperlinks or sentences on the same Wikipedia page.",
+        "bbox": [
+            112,
+            84,
+            485,
+            116
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We feed each claim-sentence pair to a RoBERTa model and use the embedding for [CLS] token to initialize the sentence nodes in our graph. And nodes of table rows and columns are initialized using the same method as in §2.1.",
+        "bbox": [
+            112,
+            116,
+            485,
+            196
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Then we apply a layer of Graph Attention Network (GAT) (Velicković et al., 2018) with a residual connection to update representations of each node. It dynamically allocates weights between different nodes of evidence, allowing the graph to capture semantic connections between evidence.",
+        "bbox": [
+            112,
+            197,
+            487,
+            292
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, we use a two-layer feed-forward network to predict the retrieved probability for each node separately. The selective score for cell evidence is the multiplication of the probability of the row node and column node that contains the cell. Top  $n_c$  cells are extracted for verdict prediction.",
+        "bbox": [
+            112,
+            294,
+            487,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The loss function is computed as the weighted sum of the loss on nodes and cells:",
+        "bbox": [
+            112,
+            390,
+            485,
+            420
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL = L ^ {S} + \\alpha_ {c} \\cdot L ^ {C o l} + \\beta_ {c} \\cdot L ^ {R o w} + \\gamma_ {c} \\cdot L ^ {C e}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            134,
+            432,
+            467,
+            451
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$L^{S}, L^{Col}, L^{Row}$ , and  $L^{Ce}$  are cross-entropy loss computed on sentences, table columns, table rows, and cells.  $\\alpha_{c}, \\beta_{c}$  and  $\\gamma_{c}$  are coefficients to adjust weights between nodes and cells. Nodes of sentences in the gold evidence set are labeled as positive. Nodes of rows and columns are labeled as positive if they contain the gold cell evidence.",
+        "bbox": [
+            112,
+            461,
+            487,
+            576
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            588,
+            258,
+            605
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We utilize FEVERIOUS as the test bed for our approach, as it is, to our knowledge, the only open-domain fact verification benchmark that combines both unstructured and structured evidence. FEVEROUS aims to accomplish two objectives: first, to extract sentence and table cell evidence from English Wikipedia, which contains over 95.6 million sentences and 11.8 million tables; and second, to predict the veracity label  $\\hat{y}$  of a given claim. The possible veracity labels include SUPPORTS, REFUTES, and NOT ENOUGH INFO (NEI). Some claims can be verified using either sentences or tables alone, while approximately  $59\\%$  of instances need structured evidence for claim verification.  $27\\%$  of the instances need to combine evidence from unstructured and structured evidence to verify. The quantitative characteristics of the benchmark are presented in Table 1. Detailed experiment settings can be found in Appendix A.",
+        "bbox": [
+            112,
+            613,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/56fbddd06d907b0b87015d6711b9f794c03fd5e17d4edb36ab99450c71ae02db.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Train</td><td>Dev</td><td>Test</td><td>Total</td></tr><tr><td>Supported</td><td>41,835</td><td>3,908</td><td>3,372</td><td>49,115</td></tr><tr><td>Refuted</td><td>27,215</td><td>3,481</td><td>2,973</td><td>33,669</td></tr><tr><td>NEI</td><td>2,241</td><td>501</td><td>1,500</td><td>4,242</td></tr><tr><td>Sentences</td><td>31,607</td><td>3,745</td><td>3,589</td><td>38,941</td></tr><tr><td>Cells</td><td>25,020</td><td>2,738</td><td>2,816</td><td>30,574</td></tr><tr><td>Sentences+Cells</td><td>20,865</td><td>2,468</td><td>2,062</td><td>25,395</td></tr><tr><td>Total</td><td>71,291</td><td>7,890</td><td>7,845</td><td>87,026</td></tr></table>",
+        "bbox": [
+            534,
+            80,
+            857,
+            206
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Quantitative characteristics of FEVEROUS.",
+        "bbox": [
+            515,
+            217,
+            873,
+            231
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            258,
+            640,
+            272
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Evidence extraction is crucial in fact verification, it provides the verdict prediction module with the necessary evidence to verify the claim. FEVEROUS uses precision, recall and F1 score to measure the quality of retrieved evidence.  $\\mathbb{E}$  is the collection of gold evidence sets and  $\\hat{E}$  is the retrieved evidence set. Recall of a specific type of evidence is computed for an instance as:",
+        "bbox": [
+            505,
+            280,
+            882,
+            407
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {R e c a l l} (\\mathbb {E}, \\hat {E}) = \\frac {| \\hat {E} \\cap \\cup \\mathbb {E} | _ {i}}{| \\cup \\mathbb {E} | _ {i}}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            420,
+            800,
+            458
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$|E|_i$  is the number of  $i$  type evidence in set  $E$ . Evidence type  $i$  could be sentence, table or cell.",
+        "bbox": [
+            507,
+            469,
+            882,
+            501
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A retrieved evidence set is considered complete iff at least one gold evidence set is included. The recall of evidence set is defined for an instance as follows:",
+        "bbox": [
+            507,
+            502,
+            880,
+            565
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {R e c a l l} (\\mathbb {E}, \\hat {E}) = \\left\\{ \\begin{array}{l l} 1 & \\exists E \\in \\mathbb {E}: E \\subseteq \\hat {E} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            537,
+            576,
+            843,
+            611
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The number of retrieved evidence is limited to 5 sentences and 25 cells at maximum (Aly et al., 2021). In the following discussion, recall refers to evidence recall unless specified for an evidence type.",
+        "bbox": [
+            507,
+            623,
+            880,
+            703
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The evaluation for verdict prediction uses label accuracy(Acc.) and FEVERIOUS score(F.S) as its metrics. They are defined for an instance as:",
+        "bbox": [
+            507,
+            703,
+            880,
+            752
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\operatorname {A c c}. (y, \\hat {y}) = \\left\\{ \\begin{array}{l l} 1 & y = \\hat {y} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\\\ \\mathrm {F . S} = \\operatorname {A c c}. (y, \\hat {y}) \\times \\operatorname {R e c a l l} (\\mathbb {E}, \\hat {\\mathrm {E}}) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            542,
+            765,
+            845,
+            821
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$y$  is the gold label and  $\\hat{y}$  is the predicted label. FEVEROUS score considers both the correct prediction of the veracity label and the completeness of the retrieved evidence set.",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "6634",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/4af302f739f3a56a2382c0b13f22383a284cb7f7d5924d9e4d67bedc97f7963c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Models</td><td>Table</td><td>Sentence</td><td>Cell</td><td>Evidence</td></tr><tr><td>Baseline</td><td>56</td><td>53</td><td>29</td><td>30</td></tr><tr><td>FaBULOUS</td><td>-</td><td>56.6</td><td>34.2</td><td>40.4</td></tr><tr><td>DCUF</td><td>75.59</td><td>62.54</td><td>58.41</td><td>43.22</td></tr><tr><td>UnifEE</td><td>75.59</td><td>75.36</td><td>67.44</td><td>55.08</td></tr><tr><td>Our Model</td><td>80.86</td><td>75.50</td><td>77.16</td><td>61.43</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            482,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Recall of different formats of evidence on the development set.",
+        "bbox": [
+            112,
+            187,
+            485,
+            217
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2 Main Results",
+        "text_level": 1,
+        "bbox": [
+            114,
+            246,
+            265,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evidence Extraction Results Table2 shows the extraction results of our model on the development set for different categories of evidence. All the previous methods restrict their evidence set to at most 5 sentences, 3 tables, and 25 cells. Official baseline (Aly et al., 2021) and FaBULOUS (Bouziane et al., 2021) use a weaker document retrieval module which results in error propagation and low evidence recall. Our model uses the same document retrieval module as DCUF (Hu et al., 2022) and UnifEE (Hu et al., 2023), which obtains a recall of  $85.20\\%$  for Wikipedia pages. Therefore recall of each format of evidence can be used to compare the performance of evidence extraction directly. SEE-ST leverages the row and column semantics of tables and increases the recall of tables by  $5.27\\%$ . SEE-ST also increases the recall of cells by  $11.65\\%$  and  $9.27\\%$  compared with DCUF and UnifEE. This suggests the two stages of SEE-ST, table extraction and cell selection can both improve the performance of structured evidence extraction.",
+        "bbox": [
+            115,
+            271,
+            489,
+            608
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Overall Results Table 3 shows overall performance on the dev and test set based on our evidence extraction results. We achieve  $60.01\\%$  evidence recall on the blind test set, which is  $6.14\\%$  higher than the previous state-of-the-art UnifEE (Hu et al., 2023). Following the verification method proposed in Hu et al. (2022), we obtain accuracy  $74.68\\% / 65.16\\%$  and FEVEROUS score  $49.73\\% / 44.75\\%$  on dev/test set. This demonstrates our enhancement on structured evidence extraction provides verification with more accurate evidence so that it could make the right predictions.",
+        "bbox": [
+            112,
+            611,
+            489,
+            804
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Since SEE-ST obtains the maximum number of evidence in extraction, we also experiment on UnifEE without its threshold selection mechanism to compare its performance with SEE-ST more fairly. SEE-ST still improves the evidence recall by a large margin. Our choice of discarding threshold selection is further analyzed in Appendix B.",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The accuracy on the test set is generally about  $10\\%$  lower than the accuracy on dev set. The main reason is the unbalanced instances of NEI claims in different splits. Our analysis of verdict prediction results shows that DCUF performs poorly in NEI instances. This explains the accuracy gap between the dev and test set.",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            214,
+            672,
+            229
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table Extraction We assess the effectiveness of our table extraction model through the following ablation experiments: (i) w/o Column Semantics: We train a model that only considers the relevance between rows and the claim for table extraction, scoring tables based on the criterion max  $(\\mathrm{P}_{row})$ . (ii) w/o Row Semantics: Analogously, we score tables using max  $(\\mathrm{P}_{col})$ . (iii) w/o Row and Column: We eliminate the design of table extraction and instead utilize the embedding of the [CLS] token in claim-table encodings for table extraction. The recall of the top 3 tables is presented in Table 4.",
+        "bbox": [
+            507,
+            237,
+            884,
+            430
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the setups without column or row semantics, the recall of the top 3 tables decreases by  $1.03\\%$  and  $1.76\\%$ , respectively. In the w/o Row and Column setting, table recall drops significantly by  $17.75\\%$ , resulting in a recall of  $63.11\\%$ , which is even lower than the recall achieved by DrQA (Table 2). These results suggest that our table extraction module in SEE-ST, which combines semantics from both rows and columns, effectively captures the coarse-grained relevance between claims and tables.",
+        "bbox": [
+            507,
+            432,
+            885,
+            593
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Cell Selection To evaluate the effectiveness of our evidence graph for cell selection, we design ablation experiments with an evidence graph constructed with the top 3 tables. The experimental settings are as follows: (i) w/o Structured Edges: We introduce additional edges between rows and columns from the same Wikipedia pages but not the same table. (ii) w/o Edge Pattern: We directly connect every node in the evidence graph to each other. (iii) w/o Row Semantics: Following Hu et al. (2023), we build an evidence graph of sentences and columns, and separately, a graph of table cells, using columns as intermediates to select cells. (iv) w/o Evidence Graph: We follow the table extraction pipeline with an added term for the cell loss function, selecting each cell based solely on the context within the same table.",
+        "bbox": [
+            507,
+            595,
+            885,
+            868
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The results of the ablation experiments are presented in Table 5. When increasing the number of tables in the evidence graph from 3 to 5, cell and",
+        "bbox": [
+            507,
+            871,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "6635",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/1542e0105cb3bdca1c2ed98a28b3e9785da06fffda66c6572170215727e28972.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"5\">Dev</td><td colspan=\"5\">Test</td></tr><tr><td>F.S</td><td>Acc.</td><td>E-P</td><td>E-R</td><td>E-F1</td><td>F.S</td><td>Acc.</td><td>E-P</td><td>E-R</td><td>E-F1</td></tr><tr><td>Official Baseline</td><td>19</td><td>53</td><td>12</td><td>30</td><td>17</td><td>17.73</td><td>48.48</td><td>10.17</td><td>28.78</td><td>15.03</td></tr><tr><td>EURECOM</td><td>19</td><td>53</td><td>12</td><td>29</td><td>17</td><td>20.01</td><td>47.79</td><td>13.73</td><td>33.73</td><td>19.52</td></tr><tr><td>Z team</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>22.51</td><td>49.01</td><td>7.76</td><td>42.64</td><td>13.12</td></tr><tr><td>CARE</td><td>26</td><td>63</td><td>7</td><td>37</td><td>12</td><td>23</td><td>53</td><td>7</td><td>37</td><td>11</td></tr><tr><td>NCU</td><td>29</td><td>60</td><td>10</td><td>42</td><td>17</td><td>25.14</td><td>52.29</td><td>9.91</td><td>39.07</td><td>15.81</td></tr><tr><td>Papelo</td><td>28</td><td>66</td><td>-</td><td>-</td><td>-</td><td>25.92</td><td>57.57</td><td>7.16</td><td>34.60</td><td>11.87</td></tr><tr><td>FaBULOUS</td><td>30</td><td>65</td><td>8</td><td>43</td><td>14</td><td>27.01</td><td>56.07</td><td>7.73</td><td>42.58</td><td>13.08</td></tr><tr><td>DCUF</td><td>35.77</td><td>72.91</td><td>15.06</td><td>43.22</td><td>22.34</td><td>33.97</td><td>63.21</td><td>14.79</td><td>44.10</td><td>22.15</td></tr><tr><td>UnifEE</td><td>44.86</td><td>73.67</td><td>19.04</td><td>55.08</td><td>28.30</td><td>41.50</td><td>65.04</td><td>18.35</td><td>53.87</td><td>27.37</td></tr><tr><td>UnifEE*</td><td>46.13</td><td>73.14</td><td>12.48</td><td>56.22</td><td>20.42</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Our Model</td><td>49.73</td><td>74.68</td><td>10.60</td><td>61.43</td><td>18.07</td><td>44.75</td><td>65.16</td><td>9.81</td><td>60.01</td><td>16.89</td></tr></table>",
+        "bbox": [
+            157,
+            80,
+            840,
+            266
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f8abb76ad5aadc3832c14b12b390afb189b4fab8c0a63262e7db9f79841df353.jpg",
+        "table_caption": [
+            "Table 3: Model performance on the development set and test set. F.S is FEVEROUS score and Acc. is the accuracy of veracity labels. E-R, E-P and E-F1 are recall, precision and F1 computed based on the evidence set. All metrics are averaged on all instances. UnifEE* uses the maximum number of evidence as the constraint for the evidence set."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Models</td><td>Recall@3</td></tr><tr><td>SEE-ST</td><td>80.86</td></tr><tr><td>w/o Column Semantics</td><td>79.83</td></tr><tr><td>w/o Row Semantics</td><td>79.10</td></tr><tr><td>w/o Row and Column</td><td>63.11</td></tr></table>",
+        "bbox": [
+            164,
+            341,
+            435,
+            420
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f6444b31fdbc67a9691640796fcc0c7b2bb9433ad23ff508aeb9a45202c16066.jpg",
+        "table_caption": [
+            "Table 4: Ablation study of coarse-grained table extraction. Recall@3 is the recall of top 3 extracted tables."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Recall</td><td>Cell</td><td>Evidence</td></tr><tr><td colspan=\"3\">Top5 Tables</td></tr><tr><td>SEE-ST</td><td>77.16</td><td>61.43</td></tr><tr><td colspan=\"3\">Top3 Tables</td></tr><tr><td>SEE-ST</td><td>74.20</td><td>60.16</td></tr><tr><td>w/o Structured Edges</td><td>72.30</td><td>58.61</td></tr><tr><td>w/o Edge Pattern</td><td>72.22</td><td>58.56</td></tr><tr><td>w/o Row Semantics</td><td>73.16</td><td>59.24</td></tr><tr><td>w/o Evidence Graph</td><td>68.37</td><td>56.08</td></tr></table>",
+        "bbox": [
+            124,
+            476,
+            478,
+            640
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Ablation study of fine-grained cell selection.",
+        "bbox": [
+            117,
+            650,
+            480,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "evidence recall rise by  $2.96\\%$  and  $1.38\\%$ , respectively, achieving a recall rate of  $77.16\\%$  for cell evidence and  $61.43\\%$  for evidence recall. In the w/o Structured Evidence setting, rows and columns from different tables are connected if they are on the same Wikipedia page, resulting in a decrease of  $1.98\\%$  in cell recall and  $1.53\\%$  in evidence recall. This demonstrates that the Structured Edges designed in  $\\S 2.2$  are beneficial for selecting cells as intersections of rows and columns. Connecting all cells in the graph further compromises the performance of structured evidence. In the w/o Row Semantics setting, we build two evidence graphs following Hu et al. (2023): one containing sentences",
+        "bbox": [
+            112,
+            693,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "and columns, and another with cells. Cell recall and evidence recall decline by  $1.04\\%$  and  $0.92\\%$  in this setting, indicating that SEE-ST's utilization of row and column representations captures cell information more precisely. Moreover, compared to the two graphs of sentences and cells, SEE-ST is more computationally efficient with only one evidence graph. Finally, the w/o Evidence Graph setting loses all connections between sentences and tables, as well as shared entity information among evidence. Each node is predicted based on its own information, resulting in a significant drop in cell and evidence recall. This highlights the importance of the design of evidence graph that enables interactions between texts and tables.",
+        "bbox": [
+            507,
+            344,
+            884,
+            586
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            598,
+            618,
+            614
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Error Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            624,
+            672,
+            639
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Here we perform a detailed error analysis for evidence extraction.",
+        "bbox": [
+            507,
+            645,
+            882,
+            675
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Error Source Analysis The error source for evidence sets is defined as the format of evidence that is not extracted completely in the three-step pipeline (§2). We use Page to denote that document retrieval fails to retrieve all the pages containing evidence. For those instances that have a complete document set, the error source is defined according to the format of evidence fail to extract: Unstructured (sentences), Structured (tables or cells) and Both. The instances with a complete evidence set are denoted as Complete. About 1% of instances have other evidence types such as lists and table captions. These instances are excluded during error analysis. The impact of ignorance of other evidence types is discussed in the Limitations section.",
+        "bbox": [
+            507,
+            678,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6636",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d3704b25906302a66669dd3ac694ce93de22446bfcfd8372c7c34c483df24d64.jpg",
+        "image_caption": [
+            "Figure 3: Error analysis of extracted evidence set for the dev set. Left: Overall error source analysis. Right: Error source proportions of claims with different reasoning challenges."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            80,
+            484,
+            282
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/aadca6d93ba0869925c95909195a3c89c10ba7abc39b08e38109597c771fde35.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            510,
+            80,
+            884,
+            282
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The result of the error source analysis is shown in Figure 3. The document retriever fails to retrieve complete documents for  $15.8\\%$  instances in the dev set. Our proposed method decreases the proportion of cases missing structured evidence from  $14.28\\%$  to  $9.34\\%$ . It shows the effectiveness of SEE-ST in structured evidence extraction.",
+        "bbox": [
+            112,
+            346,
+            487,
+            458
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Analysis based on challenge types A fact-checker system would face a specific reasoning challenge when extracting evidence and verifying claims for FEVERIOUS. The challenges include multi-hop reasoning, entity disambiguation, search term not in claim, combining tables and texts and numerical reasoning. The challenges that do not belong to those five types are categorized as other.",
+        "bbox": [
+            112,
+            464,
+            489,
+            592
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We further evaluate the performance of SEE-ST in each category to demonstrate its ability to retrieve evidence for claims with different reasoning challenges. SEE-ST achieves great extraction performance over claims that need to combine tables and texts with a recall rate of  $69.27\\%$  showing that the unified graph neural network can effectively integrate information from both formats of evidence. The recall rates of claims that need multi-hop reasoning, entity disambiguation or search terms are lower than  $50\\%$ . But the analysis in Figure 3 shows that the main error source of incomplete evidence extraction is document retrieval. It fails to retrieve complete document sets for each challenge type listed above with an error rate of  $38.6\\%$ ,  $28.4\\%$  and  $54.6\\%$  while only about  $7\\%$  instances in these challenges lose complete evidence sets because of unsuccessful structured evidence extraction. This demonstrates the ability of our model to retrieve evidence in more challenging scenarios. Our model",
+        "bbox": [
+            115,
+            596,
+            489,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "also improves the evidence extraction performance of numerical reasoning. The rate of Structured errors is  $34.82\\%$  for UnifEE (Hu et al., 2023). SEEST increases the recall rate by  $8.25\\%$  and decreases the rate of Structured errors by  $7.35\\%$ . Yet numerical reasoning remains the main challenge in structured evidence extraction.",
+        "bbox": [
+            507,
+            347,
+            884,
+            458
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 Case Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            473,
+            643,
+            489
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we demonstrate two cases for evidence extraction in Figure 4. For the claim on 1991 Waterford City Council election, our table extraction module successfully extracts the table of election statistics from the Wikipedia dump. The main challenge for this case is numerical reasoning. Verifying ... six parties all of which have at least two seats requires high-order semantics in Seats column. Such semantics is maintained in coarse-grained table extraction and fine-grained cell selection. SEE-ST extracts cells in Party and Seats columns as evidence. It shows the numerical reasoning ability of SEE-ST.",
+        "bbox": [
+            505,
+            500,
+            882,
+            708
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As for the evidence extraction for the claim of The Castle of Iron, it demonstrates the ability of SEE-ST to perform multi-hop reasoning. The claim states that ...was published by a large-scale publishing company in New York, and is refuted by sentence evidence Gnome Press was an American small-press publishing company.... However, the name of the publisher is not mentioned in the claim. It is the structured evidence that links the claim with the sentence evidence. The design of our evidence graph could enable interaction between sentence and table cells to achieve multi-hop reasoning over different formats of evidence.",
+        "bbox": [
+            507,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "6637",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Claim: 1991 Waterford City Council election was planned to take place on 27 June 1991 with six parties all of which have at least two seats.",
+        "bbox": [
+            119,
+            87,
+            480,
+            107
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Label: SUPPORTS Challenge: Numerical Reasoning",
+        "bbox": [
+            122,
+            109,
+            431,
+            118
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/20bff8e45837b39b4c347689066e254e8055456a21f21911cb0507abc672e8e1.jpg",
+        "table_caption": [
+            "Extracted Evidence",
+            "Structured:"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Party</td><td>Seats</td><td>±</td><td>First Pref. Votes</td><td>FPv%</td></tr><tr><td>Fianna Fáil</td><td>3</td><td>-2</td><td>3,165</td><td>19.01%</td></tr><tr><td>Fine Gael</td><td>2</td><td>-2</td><td>2,220</td><td>13.34%</td></tr><tr><td>Labour</td><td>3</td><td>+1</td><td>3,415</td><td>20.5%</td></tr><tr><td>Workers&#x27;s Party</td><td>3</td><td>+1</td><td>3,359</td><td>20.18%</td></tr><tr><td>Progressive Democrats</td><td>2</td><td>+2</td><td>1,666</td><td>10.01%</td></tr><tr><td>Independent</td><td>2</td><td>-</td><td>2,360</td><td>14.2%</td></tr><tr><td>Totals</td><td>15</td><td>-</td><td>16,645</td><td>100%</td></tr></table>",
+        "bbox": [
+            122,
+            143,
+            478,
+            307
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Unstructured:",
+        "text_level": 1,
+        "bbox": [
+            129,
+            313,
+            200,
+            323
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "An election to Waterford City Council took place on 27 June 1991 as part of that year's Irish local elections.",
+        "bbox": [
+            129,
+            325,
+            477,
+            343
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Waterford City Council Comhairle Cathrach Phort Láirge) was the authority responsible for local governments in the city of waterford.",
+        "bbox": [
+            129,
+            343,
+            467,
+            362
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "…",
+        "bbox": [
+            129,
+            362,
+            142,
+            370
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Claim: The Castle of Iron is a fantasy novel written by L. Sprague de Camp and Fletcher Pratt and was published by a large-scale publishing company in New York.",
+        "bbox": [
+            497,
+            87,
+            850,
+            116
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Label: REFUTES",
+        "bbox": [
+            502,
+            118,
+            589,
+            128
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Challenge: Multi-hop Reasoning",
+        "bbox": [
+            633,
+            118,
+            800,
+            128
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/5d75a30a7e6d684acc603badba0ed92138aabee13360435923444d9d36b527c4.jpg",
+        "table_caption": [
+            "Extracted Evidence",
+            "Structured:"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Author</td><td>L. Sprague de Camp and Fletcher Pratt</td><td>Publisher</td><td>Gnome Press</td></tr><tr><td>Cover artist</td><td>Hannes Bok</td><td>Publication date</td><td>1941, 1950</td></tr><tr><td>Country</td><td>United States</td><td>Media type</td><td>Print</td></tr><tr><td>Language</td><td>English</td><td>Pages</td><td>224</td></tr><tr><td>Series</td><td>Harold Shea</td><td>Preceded by</td><td>The Incomplete Enchanter</td></tr><tr><td>Genre</td><td>Fantasy</td><td>Followed by</td><td>Wall of Serpents</td></tr></table>",
+        "bbox": [
+            502,
+            156,
+            868,
+            290
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Unstructured:",
+        "text_level": 1,
+        "bbox": [
+            502,
+            302,
+            573,
+            311
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Gnome Press was an American small-press publishing company primarily known for publishing many science fiction classics.",
+        "bbox": [
+            502,
+            312,
+            873,
+            331
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The Castle of Iron is the title of a fantasy novella by American authors L. Sprague de Camp and Fletcher Pratt, and of the novel into which it was later expanded by the same authors.",
+        "bbox": [
+            502,
+            331,
+            863,
+            359
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "…",
+        "bbox": [
+            502,
+            360,
+            514,
+            368
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: Demonstrations of extracted evidence set for claims in FEVERIOUS. Gold evidence is highlighted in orange. The underlines indicate the cells selected by SEE-ST and table headers are marked with bold texts.",
+        "bbox": [
+            112,
+            382,
+            880,
+            411
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            436,
+            268,
+            451
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Open-domain fact verification extracts evidence from a knowledge base (KB) to verify the claim. In the real world, unstructured and structured evidence can be used to verify the claim. FEVER (Thorne et al., 2018) and HoVer (Jiang et al., 2020) focus on extracting unstructured evidence i.e. sentences from the Wikipedia dump. And the research of structured evidence mainly focuses on tables. TabFact (Chen et al., 2019) and InfoTabs (Gupta et al., 2020) verify claims with tables and infoboxes from Wikipedia. SEM-TABFACT (Wang et al., 2021) and SciTab (Lu et al., 2023) are constructed on claims and tables from scientific papers. Yet these datasets all provide the table evidence for verification as in a closed-domain setting. To our knowledge, FEVEROUS is the only open-domain fact verification dataset that considers table as evidence (Aly et al., 2021). Therefore, our work uses FEVEROUS as a testbed for table extraction and cell selection.",
+        "bbox": [
+            112,
+            466,
+            489,
+            788
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As for table extraction, the lexical matching approach is widely used in fact verification (Aly et al., 2021; Hu et al., 2022, 2023). Chen et al. (2020) encodes tables with BERT and concatenate embedding with structured features to rank tables. TAPAS enhances table encoding with additional table-aware positional embeddings (Herzig et al., 2021). (Pan et al., 2021) also leverages represent",
+        "bbox": [
+            112,
+            790,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tations of row and column to rank tables but overlooks the context of other cells. TaBERT (Yin et al., 2020) utilizes a content-snapshot mechanism to only encode the relevant rows thus may lose the high-order semantics of columns.",
+        "bbox": [
+            507,
+            437,
+            884,
+            517
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Gi et al. (2021) converts each cell into a text sequence to select cell evidence, bypassing the table extraction step. Acharya (2021) selects cells through dependency parsing. Jindal et al. (2021) combines single-cell NLI task with cell-wise relevance to extract evidence at cell-level semantic granularity. These methods are performed at the semantic granularity of cells. Approaches in question answering have proposed to select cells as the intersection of rows and columns (Glass et al., 2021; Pan et al., 2021). All these above methods do not consider the interaction between evidence.",
+        "bbox": [
+            507,
+            533,
+            884,
+            726
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The idea of evidence graph has been explored in previous works. Kotonya et al. (2021) linearizes cells into sentences to construct a fully connected evidence graph with texts. UnifEE (Hu et al., 2023) proposes to use column nodes as intermediates between sentence-level and cell-level evidence graphs. Our method is more computationally efficient with only one evidence graph. Our design enables rows and columns to interact with sentences at comparable semantic granularity while still maintaining high-order semantics from tables.",
+        "bbox": [
+            507,
+            741,
+            885,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "6638",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusions",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            253,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper presents SEE-ST, a novel method for enhancing the extraction of structured evidence. By leveraging both row and column semantics, SEE-ST can operate with a comparable evidence granularity, enabling both coarse-grained table extraction and fine-grained cell selection. Experimental results demonstrate the effectiveness of SEE-ST, as it significantly improves the recall of structured evidence compared to existing methods in fact verification. Moreover, SEE-ST outperforms state-of-the-art methods in evidence recall and FEVERIOUS score on FEVEROUS benchmark. These findings underline the impact of SEE-ST in enhancing structured evidence extraction.",
+        "bbox": [
+            112,
+            110,
+            492,
+            335
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            347,
+            220,
+            362
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our work explores the extraction of structured evidence in enhancing table extraction and cell selection. In the real world, fact checkers also use other structured evidence as knowledge graphs and databases to verify the claim. FactKG is a fact verification dataset on knowledge graphs (Kim et al., 2023). The verification model is required to retrieve subgraphs as evidence for verdict prediction. SEE-ST can not be directly applied to this task. We will try to build a unified evidence extraction module that could retrieve both tables and subgraphs for fact verification in our future work.",
+        "bbox": [
+            112,
+            374,
+            489,
+            565
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our work only enhances the extraction of structured evidence, while the whole evidence extraction process still suffers from error propagation from document retrieval. The error propagation in the three-step pipeline makes it extremely difficult for the model to retrieve complete evidence sets and predict the right verdict label. A new pipeline for open-domain fact verification is needed to solve this issue.",
+        "bbox": [
+            112,
+            567,
+            489,
+            709
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            724,
+            285,
+            739
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by NSFC (62161160339). We would like to thank the anonymous reviewers for their helpful comments and suggestions. For any correspondence, please contact Yansong Feng.",
+        "bbox": [
+            112,
+            750,
+            489,
+            814
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            841,
+            213,
+            857
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "ref_text",
+        "text": "Kaushik Acharya. 2021. KaushikAcharya at SemEval-2021 task 9: Candidate generation for fact verification over tables. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval",
+        "bbox": [
+            115,
+            865,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "2021), pages 1271-1275, Online. Association for Computational Linguistics.",
+            "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.",
+            "Mostafa Bouziane, Hugo Perrin, Amine Sadeq, Thanh Nguyen, Aurélien Cluzeau, and Julien Mardas. 2021. Fabulous: Fact-checking based on understanding of language over unstructured and structured information. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 31-39.",
+            "Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer open-domain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870-1879.",
+            "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. arXiv preprint arXiv:1909.02164.",
+            "Zhiyu Chen, Mohamed Trabelsi, Jeff Heflin, Yinan Xu, and Brian D. Davison. 2020. Table search using a deep contextualized language model. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020, pages 589-598. ACM.",
+            "In-Zu Gi, Ting-Yu Fang, and Richard Tzong-Han Tsai. 2021. Verdict inference with claim and retrieved elements using roberta. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 60-65.",
+            "Michael Glass, Mustafa Canim, Alfio Gliozzo, Saneem Chemmengath, Vishwajeet Kumar, Rishav Chakravarti, Avi Sil, Feifei Pan, Samarth Bharadwaj, and Nicolas Rodolfo Fauceglia. 2021. Capturing row and column semantics in transformer based question answering over tables. In Annual Conference of the North American Chapter of the Association for Computational Linguistics.",
+            "Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. A survey on automated fact-checking. Transactions of the Association for Computational Linguistics, 10:178-206.",
+            "Vivek Gupta, Maitrey Mehta, Pegah Nokhiz, and Vivek Srikumar. 2020. INFOTABS: Inference on tables as semi-structured data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2309–2324, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "6639",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Momchil Hardalov, Arnav Arora, Preslav Nakov, and Isabelle Augenstein. 2022. A survey on stance detection for mis-and disinformation identification. In *Findings of the Association for Computational Linguistics: NAACL* 2022, pages 1259-1277.",
+            "Jonathan Herzig, Thomas Müller, 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, pages 512-519, Online. Association for Computational Linguistics.",
+            "Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Mueller, Francesco Piccinno, and Julian Eisenschlos. 2020. Tapas: Weakly supervised table parsing via pre-training. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4320-4333.",
+            "Nan Hu, Zirui Wu, Yuxuan Lai, Xiao Liu, and Yansong Feng. 2022. Dual-channel evidence fusion for fact verification over texts and tables. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5232-5242.",
+            "Nan Hu, Zirui Wu, Yuxuan Lai, Chen Zhang, and Yansong Feng. 2023. Unifee: Unified evidence extraction for fact verification. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1142-1152.",
+            "Yichen Jiang, Shikha Bordia, Zheng Zhong, Charles Dognin, Maneesh Singh, and Mohit Bansal. 2020. HoVer: A dataset for many-hop fact extraction and claim verification. In *Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)*.",
+            "Aditya Jindal, Ankur Gupta, Jaya Srivastava, Preeti Menghwani, Vijit Malik, Vishesh Kaushik, and Ashutosh Modi. 2021. BreakingBERT@IITK at SemEval-2021 task 9: Statement verification and evidence finding with tables. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 327-337, Online. Association for Computational Linguistics.",
+            "Jiho Kim, Sungjin Park, Yeonsu Kwon, Yohan Jo, James Thorne, and Edward Choi. 2023. FactKG: Fact verification via reasoning on knowledge graphs. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16190-16206, Toronto, Canada. Association for Computational Linguistics.",
+            "Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Neema Kotonya, Thomas Spooner, Daniele Magazzeni, and Francesca Toni. 2021. Graph reasoning with context-aware linearization for interpretable fact extraction and verification. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 21-30.",
+            "Xinyuan Lu, Liangming Pan, Qian Liu, Preslav Nakov, and Min-Yen Kan. 2023. Scitab: A challenging benchmark for compositional reasoning and claim verification on scientific tables. arXiv preprint arXiv:2305.13186.",
+            "Feifei Pan, Mustafa Canim, Michael Glass, Alfio Gliozzo, and Peter Fox. 2021. Cltr: An end-to-end, transformer-based system for cell-level table retrieval and table question answering. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations, pages 202-209.",
+            "Mohammed Saeed, Giulio Alfarano, Khai Nguyen, Duc Pham, Raphael Troncy, and Paolo Papotti. 2021. Neural re-rankers for evidence retrieval in the FEVERIOUS task. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 108-112, Dominican Republic. Association for Computational Linguistics.",
+            "James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics.",
+            "Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. In International Conference on Learning Representations.",
+            "Nancy X. R. Wang, Diwakar Mahajan, Marina Danilevsky, and Sara Rosenthal. 2021. SemEval-2021 task 9: Fact verification and evidence finding for tabular data in scientific documents (SEM-TABFACTS). In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 317-326, Online. 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."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            834
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            510,
+            846,
+            749,
+            860
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "In our approach, we retrieve  $n_p = 5$  pages from Wikipedia for each claim following Hu et al. (2022). And we extract  $n_s = 5$  sentences and  $n_t = 5$  tables",
+        "bbox": [
+            510,
+            872,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "6640",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "from the pages to construct our evidence graph. As for the cell selection, we limit the selected cells to a maximum of  $n_c = 25$  cells. For a fair comparison with baselines, the source tables of cells are constrained to the top 3 tables. The tables are re-ranked by the highest cell selective score at cell selection.",
+        "bbox": [
+            112,
+            84,
+            487,
+            195
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "For table extraction, we utilize TAPAS-base<sup>1</sup> to encode the claim-table pair. Our model is trained with Adam optimizer (Kingma and Ba, 2014) with a batch size of 8. The learning rate is  $10^{-7}$  for TAPAS and  $10^{-5}$  for the classifier.  $\\alpha_{t}$  and  $\\beta_{t}$  are assigned a value of 1.",
+        "bbox": [
+            110,
+            197,
+            487,
+            294
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Regarding our evidence graph, we employ Spacy $^2$  to extract common entities from evidence to connect relevant nodes. RoBERTa-base $^3$  and TAPAS-base are chosen as the encoders for node initialization. The model is optimized using the Adam optimizer (Kingma and Ba, 2014) with a batch size of 4 and a peak learning rate of  $10^{-6}$ . We set the weights for different coefficients of loss functions as follows:  $\\alpha_{c} = \\beta_{c} = 2$  and  $\\gamma_{c} = 1$ . The model takes 14 hours to train for 3 epochs and 34 minutes to select cells in the dev set on a single NVIDIA A100 GPU.",
+        "bbox": [
+            115,
+            294,
+            489,
+            487
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Evidence Precision Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            502,
+            394,
+            519
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "With the constraint of a maximum number of sentences and cells in the evidence set, there are two ways to construct evidence sets. The previous SOTA method UnifEE (Hu et al., 2023) uses a threshold to construct an evidence set with high precision at the expense of a slight decrease in evidence recall. Other baselines use evidence as much as fit the constraints to construct the evidence sets resulting in higher evidence recall. In this section, we aim to analyze which strategy achieves higher accuracy and FEVERIOUS score at the verdict prediction step. We filter out unnecessary evidence with different thresholds and verify claims based on the filtered evidence sets. The results are presented in Table 6.",
+        "bbox": [
+            110,
+            529,
+            489,
+            769
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The results indicate that as the evidence threshold increases, evidence precision improves at the cost of reduced evidence recall. However, label accuracy remains largely unaffected by changes in the evidence set. Threshold selection leads to a decrease in the FEVERIOUS score. Overall, the",
+        "bbox": [
+            112,
+            771,
+            487,
+            868
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "verification model used in this work is robust to threshold selection within the same extracted evidence set, and verdict prediction achieves the highest FEVERIOUS score without threshold selection. The annotations of the gold evidence set may not be exhaustive of all possible evidence, which could result in an underestimation of precision due to the presence of correct evidence that is not annotated as gold evidence. In summary, SEE-ST employs the maximum number of sentences and cells as constrain for evidence sets for optimal verdict prediction performance.",
+        "bbox": [
+            507,
+            84,
+            884,
+            277
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/64eed44385341bfad2f88643ffcfe6896874a1a3faeefae3a6df0aec54b8d04e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Threshold</td><td>E-P</td><td>E-R</td><td>E-F1</td><td>F.S</td><td>Acc.</td></tr><tr><td>0</td><td>10.60</td><td>61.43</td><td>18.07</td><td>49.73</td><td>74.68</td></tr><tr><td>0.001</td><td>12.99</td><td>60.17</td><td>21.36</td><td>48.38</td><td>74.25</td></tr><tr><td>0.005</td><td>15.85</td><td>59.60</td><td>25.04</td><td>48.04</td><td>74.03</td></tr><tr><td>0.01</td><td>17.88</td><td>58.54</td><td>27.49</td><td>47.28</td><td>74.22</td></tr><tr><td>0.05</td><td>23.39</td><td>50.78</td><td>32.04</td><td>42.56</td><td>73.99</td></tr></table>",
+        "bbox": [
+            524,
+            288,
+            867,
+            380
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/e43875bbf156e9d5c91e232bbd1eebadadd92643034f2a8d8a4f8badef8010dc.jpg",
+        "table_caption": [
+            "Table 6: Effects of evidence extraction on verdict prediction."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Verdict</td><td>Recall</td></tr><tr><td>NEI</td><td>41.72</td></tr><tr><td>SUPPORTS</td><td>61.38</td></tr><tr><td>REFUTES</td><td>61.84</td></tr></table>",
+        "bbox": [
+            591,
+            445,
+            801,
+            527
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 7: The performance of evidence extraction for claims with different veracity labels.",
+        "bbox": [
+            507,
+            537,
+            882,
+            567
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Performance on Different Verdict",
+        "text_level": 1,
+        "bbox": [
+            507,
+            594,
+            836,
+            609
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We further analyze the performance of evidence extraction and verdict prediction for claims with different veracity labels, as shown in Table 7. Generally, extracting evidence for REFUTES claims is more challenging, as the information contained in the evidence is not always explicitly included in the claim, making it difficult to retrieve based on relevance alone. However, our model achieves comparable extraction performance for both SUPPORTS and REFUTES types, with recall rates of  $61.38\\%$  and  $61.84\\%$ , respectively. The recall for NEI instances is approximately  $20\\%$  lower than for the other types. NEI claims cannot be verified based on evidence from Wikipedia alone; only the most relevant information is listed in the gold evidence set, making it difficult to extract a complete evidence set for NEI claims.",
+        "bbox": [
+            505,
+            619,
+            884,
+            892
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_footnote",
+        "text": "1https://huggingface.co/google/tapas-base",
+        "bbox": [
+            134,
+            878,
+            448,
+            891
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_footnote",
+        "text": "2https://spacy.io/",
+        "bbox": [
+            136,
+            891,
+            275,
+            904
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://huggingface.co/roberta-base",
+        "bbox": [
+            136,
+            904,
+            411,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "6641",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2023/Enhancing Structured Evidence Extraction for Fact Verification/ebbbd673-d241-414d-91ff-1168a052ced6_model.json

@@ -0,0 +1,2092 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.168,
+                0.091,
+                0.833,
+                0.112
+            ],
+            "angle": 0,
+            "content": "Enhancing Structured Evidence Extraction for Fact Verification"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.347,
+                0.143,
+                0.656,
+                0.159
+            ],
+            "angle": 0,
+            "content": "Zirui Wu, Nan Hu, Yansong Feng*"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.213,
+                0.16,
+                0.792,
+                0.177
+            ],
+            "angle": 0,
+            "content": "Wangxuan Institute of Computer Technology, Peking University, China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.308,
+                0.178,
+                0.695,
+                0.193
+            ],
+            "angle": 0,
+            "content": "{ziruiwu,hunan,fengyansong}@pku.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.342,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.282,
+                0.461,
+                0.651
+            ],
+            "angle": 0,
+            "content": "Open-domain fact verification is the task of verifying claims in natural language texts against extracted evidence. FEVERIOUS is a benchmark that requires extracting and integrating both unstructured and structured evidence to verify a given claim. Previous models suffer from low recall of structured evidence extraction, i.e., table extraction and cell selection. In this paper, we propose a simple but effective method to enhance the extraction of structured evidence by leveraging the row and column semantics of tables. Our method comprises two components: (i) a coarse-grained table extraction module that selects tables based on rows and columns relevant to the claim and (ii) a fine-grained cell selection graph that combines both formats of evidence and enables multihop and numerical reasoning. We evaluate our method on FEVERIOUS and achieve an evidence recall of \\(60.01\\%\\) on the test set, which is \\(6.14\\%\\) higher than the previous state-of-the-art performance. Our results demonstrate that our method can extract tables and select cells effectively, and provide better evidence sets for verdict prediction. Our code is released at https://github.com/WilliamZR/see-st"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.678,
+                0.26,
+                0.693
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.704,
+                0.49,
+                0.831
+            ],
+            "angle": 0,
+            "content": "Open-domain fact verification is the task of verifying a factual claim based on evidence extracted from a knowledge base(Guo et al., 2022; Hardalov et al., 2022). A complete evidence set provides necessary information for a verification module to support or refute the claim. In the real world, fact-checkers could use both unstructured and structured data as evidence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.833,
+                0.49,
+                0.897
+            ],
+            "angle": 0,
+            "content": "Tables are a structured form of evidence that provides an organized representation of information that facilitates effortless comparison and cross-referencing. By presenting information in cells,"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.526,
+                0.253,
+                0.877,
+                0.516
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.529,
+                0.886,
+                0.559
+            ],
+            "angle": 0,
+            "content": "Figure 1: Example of FEVERIOUS with gold evidence. The gold cell evidence is highlighted in orange."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.598,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "tables enable efficient look-up of individual values, while the arrangement of rows and columns allows for the extraction of high-order semantics through various operations, such as comparisons, filtering, arithmetic calculations, and the identification of minimum and maximum values. The two-dimension structure and high-order semantics of tables make them more difficult to extract compared with texts. Regarding the previous method to extract structured evidence. Aly et al. (2021) treats tables as text sequences neglecting the table structure itself and extracts them only based on lexical matching. This approach is also used in many following works (Bouziane et al., 2021; Saeed et al., 2021; Hu et al., 2022, 2023). Consequently, these methods struggle to extract tables with little word overlap with the claim. Alternatively, Bouziane et al. (2021) and Gi et al. (2021) convert each table cell into a text sequence and perform cell retrieval directly. Kotonya et al. (2021) construct a reason-"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.142,
+                0.905,
+                0.295,
+                0.919
+            ],
+            "angle": 0,
+            "content": "* Corresponding Author."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6631"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.16,
+                0.946,
+                0.836,
+                0.974
+            ],
+            "angle": 0,
+            "content": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6631-6641 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.129,
+                0.089,
+                0.885,
+                0.341
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.355,
+                0.884,
+                0.4
+            ],
+            "angle": 0,
+            "content": "Figure 2: Overview of our proposed approach SEE-ST. It extracts structured evidence based on retrieved pages and extracted sentences. SEE-ST includes two modules: coarse-grained table extraction (§2.1) and fined-grained cell selection based on an evidence graph(§2.2)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.424,
+                0.489,
+                0.487
+            ],
+            "angle": 0,
+            "content": "ing graph of sentences and cells to extract evidence. Yet these methods overlook the context within the table and fail to capture high-order semantics from rows and columns."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.501,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Figure 1 presents an example from FEVEROUS (Aly et al., 2021), a benchmark on open-domain fact verification using both sentences and table cells as evidence. The verification model is expected to extract both unstructured and structured evidence to verify the claim. The structured table can be interpreted in several ways for the model: a complete table, a combination of rows or columns, or a composition of isolated cells. The entire table encompasses a substantial amount of information. During table extraction, the Site and Attendance columns are irrelevant and only introduce confusion for both lexical and semantic methods (Herzig et al., 2021). Cells typically contain short phrases and numbers, which are often insufficient for conveying comprehensive information. In contrast, rows and columns possess a similar volume of information to the claim, thereby exhibiting comparable semantic granularity. The extraction process should prioritize the information from the 3rd row and the Result column. During the cell selection step, the selection module must capture the high-order semantics in the Result column to verify the statement ... lost all its games except for one.... Moreover, it needs to identify the cells October 7 and Illinois State by integrating informa"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.424,
+                0.884,
+                0.504
+            ],
+            "angle": 0,
+            "content": "tion from the Date and Opponent columns and the 3rd row. Overall rows and columns have comparable semantic granularity with the claim and are the most suitable fundamental unit for both table extraction and cell selection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.511,
+                0.885,
+                0.849
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose Structured Evidence Extraction with Row and Column Semantics of Tables (SEE-ST), a novel approach to enhance both table extraction and cell selection modules for fact verification by leveraging the semantics of rows and columns. This method aligns the semantic granularity of structured evidence with claim and sentence evidence, enabling more precise evidence extraction. SEE-ST starts by extracting tables from retrieved Wikipedia pages and focuses on identifying the most relevant rows and columns for the claim, thereby reducing confusion arising from irrelevant cells. Since a cell represents the intersection of a row and a column, its selection can be conducted by analyzing both dimensions. We integrate rows and columns within a graph neural network to facilitate fine-grained cell selection while preserving high-order semantics. Additionally, we incorporate extracted sentence evidence into the graph to merge information from evidence of both formats at a similar semantic granularity."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.856,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Our contributions can be summarized as follows: (i) We propose to leverage the row and column semantics to enhance table extraction and cell selection. (ii) We design an evidence graph for cell"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6632"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.213
+            ],
+            "angle": 0,
+            "content": "selection. It enables rows and columns to interact with sentences at comparable semantic granularity while still maintaining high-order semantics from table evidence. (iii) Our proposed method SEE-ST achieves state-of-the-art performance in evidence extraction on the FEVEROUS dataset, with better evidence sets contributing to the increased accuracy of the verification module."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.228,
+                0.247,
+                0.243
+            ],
+            "angle": 0,
+            "content": "2 Our Model"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.255,
+                0.489,
+                0.446
+            ],
+            "angle": 0,
+            "content": "FEVEROUS is a benchmark on open-domain fact verification (Aly et al., 2021). The task is to verify a claim \\( c \\) based on Wikipedia. We follow the three-step pipeline in open-domain fact verification (Thorne et al., 2018; Aly et al., 2021). The three steps are document retrieval, evidence extraction, and verdict prediction. The first step retrieves \\( n_p \\) relevant pages from the Wikipedia dump. Then the second step extracts sentence evidence \\( S \\) and cell evidence \\( C \\) from the retrieved pages. At last, the model predicts the veracity label \\( \\hat{y} \\) of the claim based on sentence and cell evidence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.45,
+                0.49,
+                0.721
+            ],
+            "angle": 0,
+            "content": "SEE-ST enhances the extraction of structured evidence in the second step of the pipeline. It operates on the same conceptual framework, utilizing row and column semantics to enhance evidence extraction (Figure 2). SEE-ST performs structured evidence extraction in two stages: (i) Coarse-grained table extraction that retrieves top \\( n_t \\) tables for the second step. It scores relevance between rows or columns with the claim and ranks tables accordingly and (ii) Fine-grained cell selection that retrieves cell evidence from complex tables. It constructs a graph neural network of rows, columns, and sentences at a comparable semantic granularity. It utilizes a graph attention network to enable delicate information integration between different pieces of evidence. The cells are selected as the intersection of rows and columns."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.724,
+                0.488,
+                0.772
+            ],
+            "angle": 0,
+            "content": "At the verification step, we utilize DCUF (Hu et al., 2022), a method that converts evidence into dual-channel encodings to verify the claim."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.785,
+                0.419,
+                0.8
+            ],
+            "angle": 0,
+            "content": "2.1 Coarse-grained Table Extraction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.807,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce our table extraction module. First, the claim and table pair are fed to TAPAS, a pre-trained table model aware of table structures(Herzig et al., 2020). Instead of using embedding of [CLS] token to predict the relevance between the claim and table. our module predicts the relevance based on row and column semantics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.884,
+                0.15
+            ],
+            "angle": 0,
+            "content": "Each row and column in the evidence table is represented as the average of all the token embeddings in the same row or column. Then we predict a probability score for each row and column separately."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.15,
+                0.883,
+                0.196
+            ],
+            "angle": 0,
+            "content": "The selective score for the table is computed as the multiplication of the maximum probability score for the rows and columns."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.567,
+                0.209,
+                0.825,
+                0.227
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathrm {P} (t) = \\max \\left(\\mathrm {P} _ {r o w}\\right) \\times \\max \\left(\\mathrm {P} _ {c o l}\\right)\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.238,
+                0.884,
+                0.413
+            ],
+            "angle": 0,
+            "content": "The training set is constructed with all the tables that contain gold cell evidence as positive examples. As for negative examples, we use top tables retrieved by DrQA (Chen et al., 2017) which do not contain cell evidence but have lexical overlap with the claim. The ratio of positive and negative examples is 1 at every mini-batch in training. During training, the model is trained to select table rows and columns that contain gold cell evidence. The loss function is computed with a cross-entropy loss on each row and column."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.608,
+                0.424,
+                0.785,
+                0.442
+            ],
+            "angle": 0,
+            "content": "\\[\nL = \\alpha_ {t} L ^ {R o w} + \\beta_ {t} L ^ {C o l}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.453,
+                0.883,
+                0.549
+            ],
+            "angle": 0,
+            "content": "\\(L^{Row}\\), and \\(L^{Col}\\) are cross-entropy loss functions of rows and columns. These rows and columns that have gold cell evidence are labeled as select while rows/columns in the same table but do not contain cell evidence and rows/columns in negative examples are labeled as not select."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.551,
+                0.884,
+                0.599
+            ],
+            "angle": 0,
+            "content": "At inference, a selective score is computed for each table in retrieved pages. Top \\( n_t \\) tables are selected for the following fine-grained cell selection."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.626,
+                0.768,
+                0.64
+            ],
+            "angle": 0,
+            "content": "2.2 Fine-grained Cell Selection"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.647,
+                0.884,
+                0.822
+            ],
+            "angle": 0,
+            "content": "Following the idea of leveraging row and column semantics, we build an evidence graph \\(\\mathcal{G}(\\mathcal{V},\\mathcal{E})\\) to select cell evidence. The first two types of nodes \\(v_{0},\\ldots,v_{M}\\) and \\(v_{M+1},\\ldots,v_{M+N}\\) encode table rows and columns. Shown as yellow nodes in Figure 2. \\(M\\) and \\(N\\) are the total number of table rows and columns. The third type of node is sentences \\(v_{M+N},\\ldots,v_{K}\\), shown as blue nodes in Figure 2. Overall, the evidence graph has \\(K\\) nodes of unstructured and structured evidence at a comparable semantic granularity."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.824,
+                0.884,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Graph edges \\(\\mathcal{E}\\) are constructed between related evidence following three rules: (i) Cross-Format Related: \\(v_{i}\\) and \\(v_{j}\\) are different format evidence of sentence and table, and in the same Wikipedia page. (ii) Structure-Based: \\(v_{i}\\) and \\(v_{j}\\) are nodes from the same table and (iii) Entity-Based: \\(v_{i}\\) and \\(v_{j}\\) have"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6633"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.486,
+                0.117
+            ],
+            "angle": 0,
+            "content": "common entities and hyperlinks or sentences on the same Wikipedia page."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.117,
+                0.486,
+                0.197
+            ],
+            "angle": 0,
+            "content": "We feed each claim-sentence pair to a RoBERTa model and use the embedding for [CLS] token to initialize the sentence nodes in our graph. And nodes of table rows and columns are initialized using the same method as in §2.1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.198,
+                0.488,
+                0.293
+            ],
+            "angle": 0,
+            "content": "Then we apply a layer of Graph Attention Network (GAT) (Velicković et al., 2018) with a residual connection to update representations of each node. It dynamically allocates weights between different nodes of evidence, allowing the graph to capture semantic connections between evidence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.295,
+                0.488,
+                0.39
+            ],
+            "angle": 0,
+            "content": "Finally, we use a two-layer feed-forward network to predict the retrieved probability for each node separately. The selective score for cell evidence is the multiplication of the probability of the row node and column node that contains the cell. Top \\( n_c \\) cells are extracted for verdict prediction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.391,
+                0.487,
+                0.422
+            ],
+            "angle": 0,
+            "content": "The loss function is computed as the weighted sum of the loss on nodes and cells:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.135,
+                0.433,
+                0.468,
+                0.452
+            ],
+            "angle": 0,
+            "content": "\\[\nL = L ^ {S} + \\alpha_ {c} \\cdot L ^ {C o l} + \\beta_ {c} \\cdot L ^ {R o w} + \\gamma_ {c} \\cdot L ^ {C e}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.462,
+                0.488,
+                0.577
+            ],
+            "angle": 0,
+            "content": "\\(L^{S}, L^{Col}, L^{Row}\\), and \\(L^{Ce}\\) are cross-entropy loss computed on sentences, table columns, table rows, and cells. \\(\\alpha_{c}, \\beta_{c}\\) and \\(\\gamma_{c}\\) are coefficients to adjust weights between nodes and cells. Nodes of sentences in the gold evidence set are labeled as positive. Nodes of rows and columns are labeled as positive if they contain the gold cell evidence."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.589,
+                0.26,
+                0.606
+            ],
+            "angle": 0,
+            "content": "3 Experiments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.614,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We utilize FEVERIOUS as the test bed for our approach, as it is, to our knowledge, the only open-domain fact verification benchmark that combines both unstructured and structured evidence. FEVEROUS aims to accomplish two objectives: first, to extract sentence and table cell evidence from English Wikipedia, which contains over 95.6 million sentences and 11.8 million tables; and second, to predict the veracity label \\(\\hat{y}\\) of a given claim. The possible veracity labels include SUPPORTS, REFUTES, and NOT ENOUGH INFO (NEI). Some claims can be verified using either sentences or tables alone, while approximately \\(59\\%\\) of instances need structured evidence for claim verification. \\(27\\%\\) of the instances need to combine evidence from unstructured and structured evidence to verify. The quantitative characteristics of the benchmark are presented in Table 1. Detailed experiment settings can be found in Appendix A."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.536,
+                0.082,
+                0.858,
+                0.208
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Train</td><td>Dev</td><td>Test</td><td>Total</td></tr><tr><td>Supported</td><td>41,835</td><td>3,908</td><td>3,372</td><td>49,115</td></tr><tr><td>Refuted</td><td>27,215</td><td>3,481</td><td>2,973</td><td>33,669</td></tr><tr><td>NEI</td><td>2,241</td><td>501</td><td>1,500</td><td>4,242</td></tr><tr><td>Sentences</td><td>31,607</td><td>3,745</td><td>3,589</td><td>38,941</td></tr><tr><td>Cells</td><td>25,020</td><td>2,738</td><td>2,816</td><td>30,574</td></tr><tr><td>Sentences+Cells</td><td>20,865</td><td>2,468</td><td>2,062</td><td>25,395</td></tr><tr><td>Total</td><td>71,291</td><td>7,890</td><td>7,845</td><td>87,026</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.516,
+                0.218,
+                0.874,
+                0.232
+            ],
+            "angle": 0,
+            "content": "Table 1: Quantitative characteristics of FEVEROUS."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.259,
+                0.642,
+                0.273
+            ],
+            "angle": 0,
+            "content": "3.1 Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.281,
+                0.884,
+                0.409
+            ],
+            "angle": 0,
+            "content": "Evidence extraction is crucial in fact verification, it provides the verdict prediction module with the necessary evidence to verify the claim. FEVEROUS uses precision, recall and F1 score to measure the quality of retrieved evidence. \\(\\mathbb{E}\\) is the collection of gold evidence sets and \\(\\hat{E}\\) is the retrieved evidence set. Recall of a specific type of evidence is computed for an instance as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.592,
+                0.421,
+                0.8,
+                0.459
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {R e c a l l} (\\mathbb {E}, \\hat {E}) = \\frac {| \\hat {E} \\cap \\cup \\mathbb {E} | _ {i}}{| \\cup \\mathbb {E} | _ {i}}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.47,
+                0.884,
+                0.502
+            ],
+            "angle": 0,
+            "content": "\\(|E|_i\\) is the number of \\(i\\) type evidence in set \\(E\\). Evidence type \\(i\\) could be sentence, table or cell."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.503,
+                0.882,
+                0.566
+            ],
+            "angle": 0,
+            "content": "A retrieved evidence set is considered complete iff at least one gold evidence set is included. The recall of evidence set is defined for an instance as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.538,
+                0.577,
+                0.844,
+                0.612
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {R e c a l l} (\\mathbb {E}, \\hat {E}) = \\left\\{ \\begin{array}{l l} 1 & \\exists E \\in \\mathbb {E}: E \\subseteq \\hat {E} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.624,
+                0.882,
+                0.704
+            ],
+            "angle": 0,
+            "content": "The number of retrieved evidence is limited to 5 sentences and 25 cells at maximum (Aly et al., 2021). In the following discussion, recall refers to evidence recall unless specified for an evidence type."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.705,
+                0.882,
+                0.753
+            ],
+            "angle": 0,
+            "content": "The evaluation for verdict prediction uses label accuracy(Acc.) and FEVERIOUS score(F.S) as its metrics. They are defined for an instance as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.544,
+                0.766,
+                0.846,
+                0.822
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\operatorname {A c c}. (y, \\hat {y}) = \\left\\{ \\begin{array}{l l} 1 & y = \\hat {y} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\\\ \\mathrm {F . S} = \\operatorname {A c c}. (y, \\hat {y}) \\times \\operatorname {R e c a l l} (\\mathbb {E}, \\hat {\\mathrm {E}}) \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.855,
+                0.884,
+                0.918
+            ],
+            "angle": 0,
+            "content": "\\(y\\) is the gold label and \\(\\hat{y}\\) is the predicted label. FEVEROUS score considers both the correct prediction of the veracity label and the completeness of the retrieved evidence set."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6634"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.119,
+                0.082,
+                0.483,
+                0.18
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Models</td><td>Table</td><td>Sentence</td><td>Cell</td><td>Evidence</td></tr><tr><td>Baseline</td><td>56</td><td>53</td><td>29</td><td>30</td></tr><tr><td>FaBULOUS</td><td>-</td><td>56.6</td><td>34.2</td><td>40.4</td></tr><tr><td>DCUF</td><td>75.59</td><td>62.54</td><td>58.41</td><td>43.22</td></tr><tr><td>UnifEE</td><td>75.59</td><td>75.36</td><td>67.44</td><td>55.08</td></tr><tr><td>Our Model</td><td>80.86</td><td>75.50</td><td>77.16</td><td>61.43</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.189,
+                0.487,
+                0.218
+            ],
+            "angle": 0,
+            "content": "Table 2: Recall of different formats of evidence on the development set."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.247,
+                0.266,
+                0.262
+            ],
+            "angle": 0,
+            "content": "3.2 Main Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.272,
+                0.49,
+                0.609
+            ],
+            "angle": 0,
+            "content": "Evidence Extraction Results Table2 shows the extraction results of our model on the development set for different categories of evidence. All the previous methods restrict their evidence set to at most 5 sentences, 3 tables, and 25 cells. Official baseline (Aly et al., 2021) and FaBULOUS (Bouziane et al., 2021) use a weaker document retrieval module which results in error propagation and low evidence recall. Our model uses the same document retrieval module as DCUF (Hu et al., 2022) and UnifEE (Hu et al., 2023), which obtains a recall of \\(85.20\\%\\) for Wikipedia pages. Therefore recall of each format of evidence can be used to compare the performance of evidence extraction directly. SEE-ST leverages the row and column semantics of tables and increases the recall of tables by \\(5.27\\%\\). SEE-ST also increases the recall of cells by \\(11.65\\%\\) and \\(9.27\\%\\) compared with DCUF and UnifEE. This suggests the two stages of SEE-ST, table extraction and cell selection can both improve the performance of structured evidence extraction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.612,
+                0.49,
+                0.805
+            ],
+            "angle": 0,
+            "content": "Overall Results Table 3 shows overall performance on the dev and test set based on our evidence extraction results. We achieve \\(60.01\\%\\) evidence recall on the blind test set, which is \\(6.14\\%\\) higher than the previous state-of-the-art UnifEE (Hu et al., 2023). Following the verification method proposed in Hu et al. (2022), we obtain accuracy \\(74.68\\% / 65.16\\%\\) and FEVEROUS score \\(49.73\\% / 44.75\\%\\) on dev/test set. This demonstrates our enhancement on structured evidence extraction provides verification with more accurate evidence so that it could make the right predictions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.807,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Since SEE-ST obtains the maximum number of evidence in extraction, we also experiment on UnifEE without its threshold selection mechanism to compare its performance with SEE-ST more fairly. SEE-ST still improves the evidence recall by a large margin. Our choice of discarding threshold selection is further analyzed in Appendix B."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.198
+            ],
+            "angle": 0,
+            "content": "The accuracy on the test set is generally about \\(10\\%\\) lower than the accuracy on dev set. The main reason is the unbalanced instances of NEI claims in different splits. Our analysis of verdict prediction results shows that DCUF performs poorly in NEI instances. This explains the accuracy gap between the dev and test set."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.215,
+                0.673,
+                0.23
+            ],
+            "angle": 0,
+            "content": "3.3 Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.238,
+                0.885,
+                0.431
+            ],
+            "angle": 0,
+            "content": "Table Extraction We assess the effectiveness of our table extraction model through the following ablation experiments: (i) w/o Column Semantics: We train a model that only considers the relevance between rows and the claim for table extraction, scoring tables based on the criterion max \\((\\mathrm{P}_{row})\\). (ii) w/o Row Semantics: Analogously, we score tables using max \\((\\mathrm{P}_{col})\\). (iii) w/o Row and Column: We eliminate the design of table extraction and instead utilize the embedding of the [CLS] token in claim-table encodings for table extraction. The recall of the top 3 tables is presented in Table 4."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.433,
+                0.886,
+                0.594
+            ],
+            "angle": 0,
+            "content": "In the setups without column or row semantics, the recall of the top 3 tables decreases by \\(1.03\\%\\) and \\(1.76\\%\\), respectively. In the w/o Row and Column setting, table recall drops significantly by \\(17.75\\%\\), resulting in a recall of \\(63.11\\%\\), which is even lower than the recall achieved by DrQA (Table 2). These results suggest that our table extraction module in SEE-ST, which combines semantics from both rows and columns, effectively captures the coarse-grained relevance between claims and tables."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.596,
+                0.886,
+                0.869
+            ],
+            "angle": 0,
+            "content": "Cell Selection To evaluate the effectiveness of our evidence graph for cell selection, we design ablation experiments with an evidence graph constructed with the top 3 tables. The experimental settings are as follows: (i) w/o Structured Edges: We introduce additional edges between rows and columns from the same Wikipedia pages but not the same table. (ii) w/o Edge Pattern: We directly connect every node in the evidence graph to each other. (iii) w/o Row Semantics: Following Hu et al. (2023), we build an evidence graph of sentences and columns, and separately, a graph of table cells, using columns as intermediates to select cells. (iv) w/o Evidence Graph: We follow the table extraction pipeline with an added term for the cell loss function, selecting each cell based solely on the context within the same table."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.872,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The results of the ablation experiments are presented in Table 5. When increasing the number of tables in the evidence graph from 3 to 5, cell and"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6635"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.158,
+                0.082,
+                0.842,
+                0.267
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"5\">Dev</td><td colspan=\"5\">Test</td></tr><tr><td>F.S</td><td>Acc.</td><td>E-P</td><td>E-R</td><td>E-F1</td><td>F.S</td><td>Acc.</td><td>E-P</td><td>E-R</td><td>E-F1</td></tr><tr><td>Official Baseline</td><td>19</td><td>53</td><td>12</td><td>30</td><td>17</td><td>17.73</td><td>48.48</td><td>10.17</td><td>28.78</td><td>15.03</td></tr><tr><td>EURECOM</td><td>19</td><td>53</td><td>12</td><td>29</td><td>17</td><td>20.01</td><td>47.79</td><td>13.73</td><td>33.73</td><td>19.52</td></tr><tr><td>Z team</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>22.51</td><td>49.01</td><td>7.76</td><td>42.64</td><td>13.12</td></tr><tr><td>CARE</td><td>26</td><td>63</td><td>7</td><td>37</td><td>12</td><td>23</td><td>53</td><td>7</td><td>37</td><td>11</td></tr><tr><td>NCU</td><td>29</td><td>60</td><td>10</td><td>42</td><td>17</td><td>25.14</td><td>52.29</td><td>9.91</td><td>39.07</td><td>15.81</td></tr><tr><td>Papelo</td><td>28</td><td>66</td><td>-</td><td>-</td><td>-</td><td>25.92</td><td>57.57</td><td>7.16</td><td>34.60</td><td>11.87</td></tr><tr><td>FaBULOUS</td><td>30</td><td>65</td><td>8</td><td>43</td><td>14</td><td>27.01</td><td>56.07</td><td>7.73</td><td>42.58</td><td>13.08</td></tr><tr><td>DCUF</td><td>35.77</td><td>72.91</td><td>15.06</td><td>43.22</td><td>22.34</td><td>33.97</td><td>63.21</td><td>14.79</td><td>44.10</td><td>22.15</td></tr><tr><td>UnifEE</td><td>44.86</td><td>73.67</td><td>19.04</td><td>55.08</td><td>28.30</td><td>41.50</td><td>65.04</td><td>18.35</td><td>53.87</td><td>27.37</td></tr><tr><td>UnifEE*</td><td>46.13</td><td>73.14</td><td>12.48</td><td>56.22</td><td>20.42</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Our Model</td><td>49.73</td><td>74.68</td><td>10.60</td><td>61.43</td><td>18.07</td><td>44.75</td><td>65.16</td><td>9.81</td><td>60.01</td><td>16.89</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.277,
+                0.883,
+                0.32
+            ],
+            "angle": 0,
+            "content": "Table 3: Model performance on the development set and test set. F.S is FEVEROUS score and Acc. is the accuracy of veracity labels. E-R, E-P and E-F1 are recall, precision and F1 computed based on the evidence set. All metrics are averaged on all instances. UnifEE* uses the maximum number of evidence as the constraint for the evidence set."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.166,
+                0.342,
+                0.436,
+                0.421
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Models</td><td>Recall@3</td></tr><tr><td>SEE-ST</td><td>80.86</td></tr><tr><td>w/o Column Semantics</td><td>79.83</td></tr><tr><td>w/o Row Semantics</td><td>79.10</td></tr><tr><td>w/o Row and Column</td><td>63.11</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.431,
+                0.489,
+                0.46
+            ],
+            "angle": 0,
+            "content": "Table 4: Ablation study of coarse-grained table extraction. Recall@3 is the recall of top 3 extracted tables."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.125,
+                0.477,
+                0.479,
+                0.641
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Recall</td><td>Cell</td><td>Evidence</td></tr><tr><td colspan=\"3\">Top5 Tables</td></tr><tr><td>SEE-ST</td><td>77.16</td><td>61.43</td></tr><tr><td colspan=\"3\">Top3 Tables</td></tr><tr><td>SEE-ST</td><td>74.20</td><td>60.16</td></tr><tr><td>w/o Structured Edges</td><td>72.30</td><td>58.61</td></tr><tr><td>w/o Edge Pattern</td><td>72.22</td><td>58.56</td></tr><tr><td>w/o Row Semantics</td><td>73.16</td><td>59.24</td></tr><tr><td>w/o Evidence Graph</td><td>68.37</td><td>56.08</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.118,
+                0.651,
+                0.481,
+                0.666
+            ],
+            "angle": 0,
+            "content": "Table 5: Ablation study of fine-grained cell selection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.694,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "evidence recall rise by \\(2.96\\%\\) and \\(1.38\\%\\), respectively, achieving a recall rate of \\(77.16\\%\\) for cell evidence and \\(61.43\\%\\) for evidence recall. In the w/o Structured Evidence setting, rows and columns from different tables are connected if they are on the same Wikipedia page, resulting in a decrease of \\(1.98\\%\\) in cell recall and \\(1.53\\%\\) in evidence recall. This demonstrates that the Structured Edges designed in \\(\\S 2.2\\) are beneficial for selecting cells as intersections of rows and columns. Connecting all cells in the graph further compromises the performance of structured evidence. In the w/o Row Semantics setting, we build two evidence graphs following Hu et al. (2023): one containing sentences"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.345,
+                0.885,
+                0.587
+            ],
+            "angle": 0,
+            "content": "and columns, and another with cells. Cell recall and evidence recall decline by \\(1.04\\%\\) and \\(0.92\\%\\) in this setting, indicating that SEE-ST's utilization of row and column representations captures cell information more precisely. Moreover, compared to the two graphs of sentences and cells, SEE-ST is more computationally efficient with only one evidence graph. Finally, the w/o Evidence Graph setting loses all connections between sentences and tables, as well as shared entity information among evidence. Each node is predicted based on its own information, resulting in a significant drop in cell and evidence recall. This highlights the importance of the design of evidence graph that enables interactions between texts and tables."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.599,
+                0.619,
+                0.615
+            ],
+            "angle": 0,
+            "content": "4 Analysis"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.625,
+                0.673,
+                0.64
+            ],
+            "angle": 0,
+            "content": "4.1 Error Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.646,
+                0.884,
+                0.676
+            ],
+            "angle": 0,
+            "content": "Here we perform a detailed error analysis for evidence extraction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.679,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Error Source Analysis The error source for evidence sets is defined as the format of evidence that is not extracted completely in the three-step pipeline (§2). We use Page to denote that document retrieval fails to retrieve all the pages containing evidence. For those instances that have a complete document set, the error source is defined according to the format of evidence fail to extract: Unstructured (sentences), Structured (tables or cells) and Both. The instances with a complete evidence set are denoted as Complete. About 1% of instances have other evidence types such as lists and table captions. These instances are excluded during error analysis. The impact of ignorance of other evidence types is discussed in the Limitations section."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6636"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.115,
+                0.081,
+                0.485,
+                0.284
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.511,
+                0.081,
+                0.885,
+                0.283
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.293,
+                0.884,
+                0.323
+            ],
+            "angle": 0,
+            "content": "Figure 3: Error analysis of extracted evidence set for the dev set. Left: Overall error source analysis. Right: Error source proportions of claims with different reasoning challenges."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.347,
+                0.489,
+                0.459
+            ],
+            "angle": 0,
+            "content": "The result of the error source analysis is shown in Figure 3. The document retriever fails to retrieve complete documents for \\(15.8\\%\\) instances in the dev set. Our proposed method decreases the proportion of cases missing structured evidence from \\(14.28\\%\\) to \\(9.34\\%\\). It shows the effectiveness of SEE-ST in structured evidence extraction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.465,
+                0.49,
+                0.593
+            ],
+            "angle": 0,
+            "content": "Analysis based on challenge types A fact-checker system would face a specific reasoning challenge when extracting evidence and verifying claims for FEVERIOUS. The challenges include multi-hop reasoning, entity disambiguation, search term not in claim, combining tables and texts and numerical reasoning. The challenges that do not belong to those five types are categorized as other."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.598,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "We further evaluate the performance of SEE-ST in each category to demonstrate its ability to retrieve evidence for claims with different reasoning challenges. SEE-ST achieves great extraction performance over claims that need to combine tables and texts with a recall rate of \\(69.27\\%\\) showing that the unified graph neural network can effectively integrate information from both formats of evidence. The recall rates of claims that need multi-hop reasoning, entity disambiguation or search terms are lower than \\(50\\%\\). But the analysis in Figure 3 shows that the main error source of incomplete evidence extraction is document retrieval. It fails to retrieve complete document sets for each challenge type listed above with an error rate of \\(38.6\\%\\), \\(28.4\\%\\) and \\(54.6\\%\\) while only about \\(7\\%\\) instances in these challenges lose complete evidence sets because of unsuccessful structured evidence extraction. This demonstrates the ability of our model to retrieve evidence in more challenging scenarios. Our model"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.348,
+                0.885,
+                0.459
+            ],
+            "angle": 0,
+            "content": "also improves the evidence extraction performance of numerical reasoning. The rate of Structured errors is \\(34.82\\%\\) for UnifEE (Hu et al., 2023). SEEST increases the recall rate by \\(8.25\\%\\) and decreases the rate of Structured errors by \\(7.35\\%\\). Yet numerical reasoning remains the main challenge in structured evidence extraction."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.474,
+                0.644,
+                0.491
+            ],
+            "angle": 0,
+            "content": "5 Case Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.501,
+                0.884,
+                0.709
+            ],
+            "angle": 0,
+            "content": "In this section, we demonstrate two cases for evidence extraction in Figure 4. For the claim on 1991 Waterford City Council election, our table extraction module successfully extracts the table of election statistics from the Wikipedia dump. The main challenge for this case is numerical reasoning. Verifying ... six parties all of which have at least two seats requires high-order semantics in Seats column. Such semantics is maintained in coarse-grained table extraction and fine-grained cell selection. SEE-ST extracts cells in Party and Seats columns as evidence. It shows the numerical reasoning ability of SEE-ST."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.71,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "As for the evidence extraction for the claim of The Castle of Iron, it demonstrates the ability of SEE-ST to perform multi-hop reasoning. The claim states that ...was published by a large-scale publishing company in New York, and is refuted by sentence evidence Gnome Press was an American small-press publishing company.... However, the name of the publisher is not mentioned in the claim. It is the structured evidence that links the claim with the sentence evidence. The design of our evidence graph could enable interaction between sentence and table cells to achieve multi-hop reasoning over different formats of evidence."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6637"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.121,
+                0.089,
+                0.482,
+                0.108
+            ],
+            "angle": 0,
+            "content": "Claim: 1991 Waterford City Council election was planned to take place on 27 June 1991 with six parties all of which have at least two seats."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.123,
+                0.11,
+                0.432,
+                0.12
+            ],
+            "angle": 0,
+            "content": "Label: SUPPORTS Challenge: Numerical Reasoning"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.252,
+                0.125,
+                0.358,
+                0.134
+            ],
+            "angle": 0,
+            "content": "Extracted Evidence"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.124,
+                0.133,
+                0.186,
+                0.142
+            ],
+            "angle": 0,
+            "content": "Structured:"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.124,
+                0.144,
+                0.48,
+                0.308
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Party</td><td>Seats</td><td>±</td><td>First Pref. Votes</td><td>FPv%</td></tr><tr><td>Fianna Fáil</td><td>3</td><td>-2</td><td>3,165</td><td>19.01%</td></tr><tr><td>Fine Gael</td><td>2</td><td>-2</td><td>2,220</td><td>13.34%</td></tr><tr><td>Labour</td><td>3</td><td>+1</td><td>3,415</td><td>20.5%</td></tr><tr><td>Workers&#x27;s Party</td><td>3</td><td>+1</td><td>3,359</td><td>20.18%</td></tr><tr><td>Progressive Democrats</td><td>2</td><td>+2</td><td>1,666</td><td>10.01%</td></tr><tr><td>Independent</td><td>2</td><td>-</td><td>2,360</td><td>14.2%</td></tr><tr><td>Totals</td><td>15</td><td>-</td><td>16,645</td><td>100%</td></tr></table>"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.131,
+                0.315,
+                0.201,
+                0.324
+            ],
+            "angle": 0,
+            "content": "Unstructured:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.326,
+                0.478,
+                0.344
+            ],
+            "angle": 0,
+            "content": "An election to Waterford City Council took place on 27 June 1991 as part of that year's Irish local elections."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.344,
+                0.468,
+                0.363
+            ],
+            "angle": 0,
+            "content": "Waterford City Council Comhairle Cathrach Phort Láirge) was the authority responsible for local governments in the city of waterford."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.363,
+                0.144,
+                0.372
+            ],
+            "angle": 0,
+            "content": "…"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.089,
+                0.852,
+                0.117
+            ],
+            "angle": 0,
+            "content": "Claim: The Castle of Iron is a fantasy novel written by L. Sprague de Camp and Fletcher Pratt and was published by a large-scale publishing company in New York."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.503,
+                0.119,
+                0.59,
+                0.129
+            ],
+            "angle": 0,
+            "content": "Label: REFUTES"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.635,
+                0.119,
+                0.801,
+                0.129
+            ],
+            "angle": 0,
+            "content": "Challenge: Multi-hop Reasoning"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.642,
+                0.134,
+                0.748,
+                0.143
+            ],
+            "angle": 0,
+            "content": "Extracted Evidence"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.503,
+                0.145,
+                0.566,
+                0.153
+            ],
+            "angle": 0,
+            "content": "Structured:"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.504,
+                0.158,
+                0.87,
+                0.291
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Author</td><td>L. Sprague de Camp and Fletcher Pratt</td><td>Publisher</td><td>Gnome Press</td></tr><tr><td>Cover artist</td><td>Hannes Bok</td><td>Publication date</td><td>1941, 1950</td></tr><tr><td>Country</td><td>United States</td><td>Media type</td><td>Print</td></tr><tr><td>Language</td><td>English</td><td>Pages</td><td>224</td></tr><tr><td>Series</td><td>Harold Shea</td><td>Preceded by</td><td>The Incomplete Enchanter</td></tr><tr><td>Genre</td><td>Fantasy</td><td>Followed by</td><td>Wall of Serpents</td></tr></table>"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.503,
+                0.303,
+                0.574,
+                0.312
+            ],
+            "angle": 0,
+            "content": "Unstructured:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.503,
+                0.313,
+                0.875,
+                0.332
+            ],
+            "angle": 0,
+            "content": "Gnome Press was an American small-press publishing company primarily known for publishing many science fiction classics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.503,
+                0.332,
+                0.865,
+                0.36
+            ],
+            "angle": 0,
+            "content": "The Castle of Iron is the title of a fantasy novella by American authors L. Sprague de Camp and Fletcher Pratt, and of the novel into which it was later expanded by the same authors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.503,
+                0.361,
+                0.515,
+                0.369
+            ],
+            "angle": 0,
+            "content": "…"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.384,
+                0.882,
+                0.412
+            ],
+            "angle": 0,
+            "content": "Figure 4: Demonstrations of extracted evidence set for claims in FEVERIOUS. Gold evidence is highlighted in orange. The underlines indicate the cells selected by SEE-ST and table headers are marked with bold texts."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.437,
+                0.269,
+                0.452
+            ],
+            "angle": 0,
+            "content": "6 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.467,
+                0.49,
+                0.789
+            ],
+            "angle": 0,
+            "content": "Open-domain fact verification extracts evidence from a knowledge base (KB) to verify the claim. In the real world, unstructured and structured evidence can be used to verify the claim. FEVER (Thorne et al., 2018) and HoVer (Jiang et al., 2020) focus on extracting unstructured evidence i.e. sentences from the Wikipedia dump. And the research of structured evidence mainly focuses on tables. TabFact (Chen et al., 2019) and InfoTabs (Gupta et al., 2020) verify claims with tables and infoboxes from Wikipedia. SEM-TABFACT (Wang et al., 2021) and SciTab (Lu et al., 2023) are constructed on claims and tables from scientific papers. Yet these datasets all provide the table evidence for verification as in a closed-domain setting. To our knowledge, FEVEROUS is the only open-domain fact verification dataset that considers table as evidence (Aly et al., 2021). Therefore, our work uses FEVEROUS as a testbed for table extraction and cell selection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.791,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "As for table extraction, the lexical matching approach is widely used in fact verification (Aly et al., 2021; Hu et al., 2022, 2023). Chen et al. (2020) encodes tables with BERT and concatenate embedding with structured features to rank tables. TAPAS enhances table encoding with additional table-aware positional embeddings (Herzig et al., 2021). (Pan et al., 2021) also leverages represent"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.438,
+                0.885,
+                0.518
+            ],
+            "angle": 0,
+            "content": "tations of row and column to rank tables but overlooks the context of other cells. TaBERT (Yin et al., 2020) utilizes a content-snapshot mechanism to only encode the relevant rows thus may lose the high-order semantics of columns."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.534,
+                0.885,
+                0.727
+            ],
+            "angle": 0,
+            "content": "Gi et al. (2021) converts each cell into a text sequence to select cell evidence, bypassing the table extraction step. Acharya (2021) selects cells through dependency parsing. Jindal et al. (2021) combines single-cell NLI task with cell-wise relevance to extract evidence at cell-level semantic granularity. These methods are performed at the semantic granularity of cells. Approaches in question answering have proposed to select cells as the intersection of rows and columns (Glass et al., 2021; Pan et al., 2021). All these above methods do not consider the interaction between evidence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.743,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "The idea of evidence graph has been explored in previous works. Kotonya et al. (2021) linearizes cells into sentences to construct a fully connected evidence graph with texts. UnifEE (Hu et al., 2023) proposes to use column nodes as intermediates between sentence-level and cell-level evidence graphs. Our method is more computationally efficient with only one evidence graph. Our design enables rows and columns to interact with sentences at comparable semantic granularity while still maintaining high-order semantics from tables."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6638"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.254,
+                0.099
+            ],
+            "angle": 0,
+            "content": "7 Conclusions"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.111,
+                0.493,
+                0.336
+            ],
+            "angle": 0,
+            "content": "This paper presents SEE-ST, a novel method for enhancing the extraction of structured evidence. By leveraging both row and column semantics, SEE-ST can operate with a comparable evidence granularity, enabling both coarse-grained table extraction and fine-grained cell selection. Experimental results demonstrate the effectiveness of SEE-ST, as it significantly improves the recall of structured evidence compared to existing methods in fact verification. Moreover, SEE-ST outperforms state-of-the-art methods in evidence recall and FEVERIOUS score on FEVEROUS benchmark. These findings underline the impact of SEE-ST in enhancing structured evidence extraction."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.348,
+                0.221,
+                0.363
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.375,
+                0.49,
+                0.566
+            ],
+            "angle": 0,
+            "content": "Our work explores the extraction of structured evidence in enhancing table extraction and cell selection. In the real world, fact checkers also use other structured evidence as knowledge graphs and databases to verify the claim. FactKG is a fact verification dataset on knowledge graphs (Kim et al., 2023). The verification model is required to retrieve subgraphs as evidence for verdict prediction. SEE-ST can not be directly applied to this task. We will try to build a unified evidence extraction module that could retrieve both tables and subgraphs for fact verification in our future work."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.568,
+                0.49,
+                0.711
+            ],
+            "angle": 0,
+            "content": "Our work only enhances the extraction of structured evidence, while the whole evidence extraction process still suffers from error propagation from document retrieval. The error propagation in the three-step pipeline makes it extremely difficult for the model to retrieve complete evidence sets and predict the right verdict label. A new pipeline for open-domain fact verification is needed to solve this issue."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.725,
+                0.287,
+                0.74
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.751,
+                0.49,
+                0.815
+            ],
+            "angle": 0,
+            "content": "This work is supported by NSFC (62161160339). We would like to thank the anonymous reviewers for their helpful comments and suggestions. For any correspondence, please contact Yansong Feng."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.843,
+                0.214,
+                0.858
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.866,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Kaushik Acharya. 2021. KaushikAcharya at SemEval-2021 task 9: Candidate generation for fact verification over tables. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.114
+            ],
+            "angle": 0,
+            "content": "2021), pages 1271-1275, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.122,
+                0.885,
+                0.241
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.25,
+                0.885,
+                0.33
+            ],
+            "angle": 0,
+            "content": "Mostafa Bouziane, Hugo Perrin, Amine Sadeq, Thanh Nguyen, Aurélien Cluzeau, and Julien Mardas. 2021. Fabulous: Fact-checking based on understanding of language over unstructured and structured information. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 31-39."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.338,
+                0.885,
+                0.404
+            ],
+            "angle": 0,
+            "content": "Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer open-domain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870-1879."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.413,
+                0.885,
+                0.478
+            ],
+            "angle": 0,
+            "content": "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. arXiv preprint arXiv:1909.02164."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.488,
+                0.885,
+                0.58
+            ],
+            "angle": 0,
+            "content": "Zhiyu Chen, Mohamed Trabelsi, Jeff Heflin, Yinan Xu, and Brian D. Davison. 2020. Table search using a deep contextualized language model. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020, pages 589-598. ACM."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.589,
+                0.885,
+                0.655
+            ],
+            "angle": 0,
+            "content": "In-Zu Gi, Ting-Yu Fang, and Richard Tzong-Han Tsai. 2021. Verdict inference with claim and retrieved elements using roberta. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 60-65."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.664,
+                0.885,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Michael Glass, Mustafa Canim, Alfio Gliozzo, Saneem Chemmengath, Vishwajeet Kumar, Rishav Chakravarti, Avi Sil, Feifei Pan, Samarth Bharadwaj, and Nicolas Rodolfo Fauceglia. 2021. Capturing row and column semantics in transformer based question answering over tables. In Annual Conference of the North American Chapter of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.778,
+                0.885,
+                0.831
+            ],
+            "angle": 0,
+            "content": "Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. A survey on automated fact-checking. Transactions of the Association for Computational Linguistics, 10:178-206."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.84,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Vivek Gupta, Maitrey Mehta, Pegah Nokhiz, and Vivek Srikumar. 2020. INFOTABS: Inference on tables as semi-structured data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2309–2324, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6639"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.153
+            ],
+            "angle": 0,
+            "content": "Momchil Hardalov, Arnav Arora, Preslav Nakov, and Isabelle Augenstein. 2022. A survey on stance detection for mis-and disinformation identification. In *Findings of the Association for Computational Linguistics: NAACL* 2022, pages 1259-1277."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.165,
+                0.49,
+                0.271
+            ],
+            "angle": 0,
+            "content": "Jonathan Herzig, Thomas Müller, 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, pages 512-519, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.284,
+                0.49,
+                0.363
+            ],
+            "angle": 0,
+            "content": "Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Mueller, Francesco Piccinno, and Julian Eisenschlos. 2020. Tapas: Weakly supervised table parsing via pre-training. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4320-4333."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.376,
+                0.49,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Nan Hu, Zirui Wu, Yuxuan Lai, Xiao Liu, and Yansong Feng. 2022. Dual-channel evidence fusion for fact verification over texts and tables. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5232-5242."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.469,
+                0.49,
+                0.548
+            ],
+            "angle": 0,
+            "content": "Nan Hu, Zirui Wu, Yuxuan Lai, Chen Zhang, and Yansong Feng. 2023. Unifee: Unified evidence extraction for fact verification. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1142-1152."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.562,
+                0.49,
+                0.641
+            ],
+            "angle": 0,
+            "content": "Yichen Jiang, Shikha Bordia, Zheng Zhong, Charles Dognin, Maneesh Singh, and Mohit Bansal. 2020. HoVer: A dataset for many-hop fact extraction and claim verification. In *Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)*."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.655,
+                0.49,
+                0.76
+            ],
+            "angle": 0,
+            "content": "Aditya Jindal, Ankur Gupta, Jaya Srivastava, Preeti Menghwani, Vijit Malik, Vishesh Kaushik, and Ashutosh Modi. 2021. BreakingBERT@IITK at SemEval-2021 task 9: Statement verification and evidence finding with tables. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 327-337, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.773,
+                0.49,
+                0.865
+            ],
+            "angle": 0,
+            "content": "Jiho Kim, Sungjin Park, Yeonsu Kwon, Yohan Jo, James Thorne, and Edward Choi. 2023. FactKG: Fact verification via reasoning on knowledge graphs. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16190-16206, Toronto, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.878,
+                0.489,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Neema Kotonya, Thomas Spooner, Daniele Magazzeni, and Francesca Toni. 2021. Graph reasoning with context-aware linearization for interpretable fact extraction and verification. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 21-30."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.174,
+                0.885,
+                0.239
+            ],
+            "angle": 0,
+            "content": "Xinyuan Lu, Liangming Pan, Qian Liu, Preslav Nakov, and Min-Yen Kan. 2023. Scitab: A challenging benchmark for compositional reasoning and claim verification on scientific tables. arXiv preprint arXiv:2305.13186."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.25,
+                0.885,
+                0.355
+            ],
+            "angle": 0,
+            "content": "Feifei Pan, Mustafa Canim, Michael Glass, Alfio Gliozzo, and Peter Fox. 2021. Cltr: An end-to-end, transformer-based system for cell-level table retrieval and table question answering. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations, pages 202-209."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.364,
+                0.885,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Mohammed Saeed, Giulio Alfarano, Khai Nguyen, Duc Pham, Raphael Troncy, and Paolo Papotti. 2021. Neural re-rankers for evidence retrieval in the FEVERIOUS task. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 108-112, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.466,
+                0.885,
+                0.583
+            ],
+            "angle": 0,
+            "content": "James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.593,
+                0.885,
+                0.645
+            ],
+            "angle": 0,
+            "content": "Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. In International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.655,
+                0.885,
+                0.759
+            ],
+            "angle": 0,
+            "content": "Nancy X. R. Wang, Diwakar Mahajan, Marina Danilevsky, and Sara Rosenthal. 2021. SemEval-2021 task 9: Fact verification and evidence finding for tabular data in scientific documents (SEM-TABFACTS). In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 317-326, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.769,
+                0.885,
+                0.835
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.835
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.512,
+                0.847,
+                0.751,
+                0.862
+            ],
+            "angle": 0,
+            "content": "A Implementation Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.873,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": "In our approach, we retrieve \\( n_p = 5 \\) pages from Wikipedia for each claim following Hu et al. (2022). And we extract \\( n_s = 5 \\) sentences and \\( n_t = 5 \\) tables"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6640"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.489,
+                0.196
+            ],
+            "angle": 0,
+            "content": "from the pages to construct our evidence graph. As for the cell selection, we limit the selected cells to a maximum of \\( n_c = 25 \\) cells. For a fair comparison with baselines, the source tables of cells are constrained to the top 3 tables. The tables are re-ranked by the highest cell selective score at cell selection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.198,
+                0.489,
+                0.295
+            ],
+            "angle": 0,
+            "content": "For table extraction, we utilize TAPAS-base<sup>1</sup> to encode the claim-table pair. Our model is trained with Adam optimizer (Kingma and Ba, 2014) with a batch size of 8. The learning rate is \\(10^{-7}\\) for TAPAS and \\(10^{-5}\\) for the classifier. \\(\\alpha_{t}\\) and \\(\\beta_{t}\\) are assigned a value of 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.296,
+                0.49,
+                0.488
+            ],
+            "angle": 0,
+            "content": "Regarding our evidence graph, we employ Spacy\\(^2\\) to extract common entities from evidence to connect relevant nodes. RoBERTa-base\\(^3\\) and TAPAS-base are chosen as the encoders for node initialization. The model is optimized using the Adam optimizer (Kingma and Ba, 2014) with a batch size of 4 and a peak learning rate of \\(10^{-6}\\). We set the weights for different coefficients of loss functions as follows: \\(\\alpha_{c} = \\beta_{c} = 2\\) and \\(\\gamma_{c} = 1\\). The model takes 14 hours to train for 3 epochs and 34 minutes to select cells in the dev set on a single NVIDIA A100 GPU."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.503,
+                0.395,
+                0.52
+            ],
+            "angle": 0,
+            "content": "B Evidence Precision Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.53,
+                0.49,
+                0.77
+            ],
+            "angle": 0,
+            "content": "With the constraint of a maximum number of sentences and cells in the evidence set, there are two ways to construct evidence sets. The previous SOTA method UnifEE (Hu et al., 2023) uses a threshold to construct an evidence set with high precision at the expense of a slight decrease in evidence recall. Other baselines use evidence as much as fit the constraints to construct the evidence sets resulting in higher evidence recall. In this section, we aim to analyze which strategy achieves higher accuracy and FEVERIOUS score at the verdict prediction step. We filter out unnecessary evidence with different thresholds and verify claims based on the filtered evidence sets. The results are presented in Table 6."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.772,
+                0.489,
+                0.869
+            ],
+            "angle": 0,
+            "content": "The results indicate that as the evidence threshold increases, evidence precision improves at the cost of reduced evidence recall. However, label accuracy remains largely unaffected by changes in the evidence set. Threshold selection leads to a decrease in the FEVERIOUS score. Overall, the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.278
+            ],
+            "angle": 0,
+            "content": "verification model used in this work is robust to threshold selection within the same extracted evidence set, and verdict prediction achieves the highest FEVERIOUS score without threshold selection. The annotations of the gold evidence set may not be exhaustive of all possible evidence, which could result in an underestimation of precision due to the presence of correct evidence that is not annotated as gold evidence. In summary, SEE-ST employs the maximum number of sentences and cells as constrain for evidence sets for optimal verdict prediction performance."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.525,
+                0.29,
+                0.868,
+                0.381
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Threshold</td><td>E-P</td><td>E-R</td><td>E-F1</td><td>F.S</td><td>Acc.</td></tr><tr><td>0</td><td>10.60</td><td>61.43</td><td>18.07</td><td>49.73</td><td>74.68</td></tr><tr><td>0.001</td><td>12.99</td><td>60.17</td><td>21.36</td><td>48.38</td><td>74.25</td></tr><tr><td>0.005</td><td>15.85</td><td>59.60</td><td>25.04</td><td>48.04</td><td>74.03</td></tr><tr><td>0.01</td><td>17.88</td><td>58.54</td><td>27.49</td><td>47.28</td><td>74.22</td></tr><tr><td>0.05</td><td>23.39</td><td>50.78</td><td>32.04</td><td>42.56</td><td>73.99</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.39,
+                0.884,
+                0.417
+            ],
+            "angle": 0,
+            "content": "Table 6: Effects of evidence extraction on verdict prediction."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.592,
+                0.446,
+                0.803,
+                0.528
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Verdict</td><td>Recall</td></tr><tr><td>NEI</td><td>41.72</td></tr><tr><td>SUPPORTS</td><td>61.38</td></tr><tr><td>REFUTES</td><td>61.84</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.538,
+                0.883,
+                0.568
+            ],
+            "angle": 0,
+            "content": "Table 7: The performance of evidence extraction for claims with different veracity labels."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.595,
+                0.838,
+                0.61
+            ],
+            "angle": 0,
+            "content": "C Performance on Different Verdict"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.62,
+                0.885,
+                0.893
+            ],
+            "angle": 0,
+            "content": "We further analyze the performance of evidence extraction and verdict prediction for claims with different veracity labels, as shown in Table 7. Generally, extracting evidence for REFUTES claims is more challenging, as the information contained in the evidence is not always explicitly included in the claim, making it difficult to retrieve based on relevance alone. However, our model achieves comparable extraction performance for both SUPPORTS and REFUTES types, with recall rates of \\(61.38\\%\\) and \\(61.84\\%\\), respectively. The recall for NEI instances is approximately \\(20\\%\\) lower than for the other types. NEI claims cannot be verified based on evidence from Wikipedia alone; only the most relevant information is listed in the gold evidence set, making it difficult to extract a complete evidence set for NEI claims."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.136,
+                0.879,
+                0.449,
+                0.892
+            ],
+            "angle": 0,
+            "content": "1https://huggingface.co/google/tapas-base"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.137,
+                0.892,
+                0.276,
+                0.905
+            ],
+            "angle": 0,
+            "content": "2https://spacy.io/"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.137,
+                0.905,
+                0.412,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>3</sup>https://huggingface.co/roberta-base"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.136,
+                0.879,
+                0.449,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6641"
+        }
+    ]
+]

2023/Enhancing Structured Evidence Extraction for Fact Verification/ebbbd673-d241-414d-91ff-1168a052ced6_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1c8c143722358937b30fde2157e01697860df62b8768383d55481c0894b4dfd
+size 774535

2023/Enhancing Structured Evidence Extraction for Fact Verification/full.md

@@ -0,0 +1,318 @@
+# Enhancing Structured Evidence Extraction for Fact Verification
+
+Zirui Wu, Nan Hu, Yansong Feng*
+
+Wangxuan Institute of Computer Technology, Peking University, China
+
+{ziruiwu,hunan,fengyansong}@pku.edu.cn
+
+# Abstract
+
+Open-domain fact verification is the task of verifying claims in natural language texts against extracted evidence. FEVERIOUS is a benchmark that requires extracting and integrating both unstructured and structured evidence to verify a given claim. Previous models suffer from low recall of structured evidence extraction, i.e., table extraction and cell selection. In this paper, we propose a simple but effective method to enhance the extraction of structured evidence by leveraging the row and column semantics of tables. Our method comprises two components: (i) a coarse-grained table extraction module that selects tables based on rows and columns relevant to the claim and (ii) a fine-grained cell selection graph that combines both formats of evidence and enables multihop and numerical reasoning. We evaluate our method on FEVERIOUS and achieve an evidence recall of  $60.01\%$  on the test set, which is  $6.14\%$  higher than the previous state-of-the-art performance. Our results demonstrate that our method can extract tables and select cells effectively, and provide better evidence sets for verdict prediction. Our code is released at https://github.com/WilliamZR/see-st
+
+# 1 Introduction
+
+Open-domain fact verification is the task of verifying a factual claim based on evidence extracted from a knowledge base(Guo et al., 2022; Hardalov et al., 2022). A complete evidence set provides necessary information for a verification module to support or refute the claim. In the real world, fact-checkers could use both unstructured and structured data as evidence.
+
+Tables are a structured form of evidence that provides an organized representation of information that facilitates effortless comparison and cross-referencing. By presenting information in cells,
+
+![](images/a30dfb3cb74cadfded92f601f88388672a67962304357d8f8be3d53e2d883569.jpg)  
+Figure 1: Example of FEVERIOUS with gold evidence. The gold cell evidence is highlighted in orange.
+
+tables enable efficient look-up of individual values, while the arrangement of rows and columns allows for the extraction of high-order semantics through various operations, such as comparisons, filtering, arithmetic calculations, and the identification of minimum and maximum values. The two-dimension structure and high-order semantics of tables make them more difficult to extract compared with texts. Regarding the previous method to extract structured evidence. Aly et al. (2021) treats tables as text sequences neglecting the table structure itself and extracts them only based on lexical matching. This approach is also used in many following works (Bouziane et al., 2021; Saeed et al., 2021; Hu et al., 2022, 2023). Consequently, these methods struggle to extract tables with little word overlap with the claim. Alternatively, Bouziane et al. (2021) and Gi et al. (2021) convert each table cell into a text sequence and perform cell retrieval directly. Kotonya et al. (2021) construct a reason-
+
+![](images/66785a60fe362381abd20e3d7bd3ce868abb47c5bb5f99bdf937adc6e20bffaa.jpg)  
+Figure 2: Overview of our proposed approach SEE-ST. It extracts structured evidence based on retrieved pages and extracted sentences. SEE-ST includes two modules: coarse-grained table extraction (§2.1) and fined-grained cell selection based on an evidence graph(§2.2).
+
+ing graph of sentences and cells to extract evidence. Yet these methods overlook the context within the table and fail to capture high-order semantics from rows and columns.
+
+Figure 1 presents an example from FEVEROUS (Aly et al., 2021), a benchmark on open-domain fact verification using both sentences and table cells as evidence. The verification model is expected to extract both unstructured and structured evidence to verify the claim. The structured table can be interpreted in several ways for the model: a complete table, a combination of rows or columns, or a composition of isolated cells. The entire table encompasses a substantial amount of information. During table extraction, the Site and Attendance columns are irrelevant and only introduce confusion for both lexical and semantic methods (Herzig et al., 2021). Cells typically contain short phrases and numbers, which are often insufficient for conveying comprehensive information. In contrast, rows and columns possess a similar volume of information to the claim, thereby exhibiting comparable semantic granularity. The extraction process should prioritize the information from the 3rd row and the Result column. During the cell selection step, the selection module must capture the high-order semantics in the Result column to verify the statement ... lost all its games except for one.... Moreover, it needs to identify the cells October 7 and Illinois State by integrating informa
+
+tion from the Date and Opponent columns and the 3rd row. Overall rows and columns have comparable semantic granularity with the claim and are the most suitable fundamental unit for both table extraction and cell selection.
+
+In this paper, we propose Structured Evidence Extraction with Row and Column Semantics of Tables (SEE-ST), a novel approach to enhance both table extraction and cell selection modules for fact verification by leveraging the semantics of rows and columns. This method aligns the semantic granularity of structured evidence with claim and sentence evidence, enabling more precise evidence extraction. SEE-ST starts by extracting tables from retrieved Wikipedia pages and focuses on identifying the most relevant rows and columns for the claim, thereby reducing confusion arising from irrelevant cells. Since a cell represents the intersection of a row and a column, its selection can be conducted by analyzing both dimensions. We integrate rows and columns within a graph neural network to facilitate fine-grained cell selection while preserving high-order semantics. Additionally, we incorporate extracted sentence evidence into the graph to merge information from evidence of both formats at a similar semantic granularity.
+
+Our contributions can be summarized as follows: (i) We propose to leverage the row and column semantics to enhance table extraction and cell selection. (ii) We design an evidence graph for cell
+
+selection. It enables rows and columns to interact with sentences at comparable semantic granularity while still maintaining high-order semantics from table evidence. (iii) Our proposed method SEE-ST achieves state-of-the-art performance in evidence extraction on the FEVEROUS dataset, with better evidence sets contributing to the increased accuracy of the verification module.
+
+# 2 Our Model
+
+FEVEROUS is a benchmark on open-domain fact verification (Aly et al., 2021). The task is to verify a claim  $c$  based on Wikipedia. We follow the three-step pipeline in open-domain fact verification (Thorne et al., 2018; Aly et al., 2021). The three steps are document retrieval, evidence extraction, and verdict prediction. The first step retrieves  $n_p$  relevant pages from the Wikipedia dump. Then the second step extracts sentence evidence  $S$  and cell evidence  $C$  from the retrieved pages. At last, the model predicts the veracity label  $\hat{y}$  of the claim based on sentence and cell evidence.
+
+SEE-ST enhances the extraction of structured evidence in the second step of the pipeline. It operates on the same conceptual framework, utilizing row and column semantics to enhance evidence extraction (Figure 2). SEE-ST performs structured evidence extraction in two stages: (i) Coarse-grained table extraction that retrieves top  $n_t$  tables for the second step. It scores relevance between rows or columns with the claim and ranks tables accordingly and (ii) Fine-grained cell selection that retrieves cell evidence from complex tables. It constructs a graph neural network of rows, columns, and sentences at a comparable semantic granularity. It utilizes a graph attention network to enable delicate information integration between different pieces of evidence. The cells are selected as the intersection of rows and columns.
+
+At the verification step, we utilize DCUF (Hu et al., 2022), a method that converts evidence into dual-channel encodings to verify the claim.
+
+# 2.1 Coarse-grained Table Extraction
+
+In this section, we introduce our table extraction module. First, the claim and table pair are fed to TAPAS, a pre-trained table model aware of table structures(Herzig et al., 2020). Instead of using embedding of [CLS] token to predict the relevance between the claim and table. our module predicts the relevance based on row and column semantics.
+
+Each row and column in the evidence table is represented as the average of all the token embeddings in the same row or column. Then we predict a probability score for each row and column separately.
+
+The selective score for the table is computed as the multiplication of the maximum probability score for the rows and columns.
+
+$$
+\mathrm {P} (t) = \max \left(\mathrm {P} _ {r o w}\right) \times \max \left(\mathrm {P} _ {c o l}\right)
+$$
+
+The training set is constructed with all the tables that contain gold cell evidence as positive examples. As for negative examples, we use top tables retrieved by DrQA (Chen et al., 2017) which do not contain cell evidence but have lexical overlap with the claim. The ratio of positive and negative examples is 1 at every mini-batch in training. During training, the model is trained to select table rows and columns that contain gold cell evidence. The loss function is computed with a cross-entropy loss on each row and column.
+
+$$
+L = \alpha_ {t} L ^ {R o w} + \beta_ {t} L ^ {C o l}
+$$
+
+$L^{Row}$ , and  $L^{Col}$  are cross-entropy loss functions of rows and columns. These rows and columns that have gold cell evidence are labeled as select while rows/columns in the same table but do not contain cell evidence and rows/columns in negative examples are labeled as not select.
+
+At inference, a selective score is computed for each table in retrieved pages. Top  $n_t$  tables are selected for the following fine-grained cell selection.
+
+# 2.2 Fine-grained Cell Selection
+
+Following the idea of leveraging row and column semantics, we build an evidence graph  $\mathcal{G}(\mathcal{V},\mathcal{E})$  to select cell evidence. The first two types of nodes  $v_{0},\ldots,v_{M}$  and  $v_{M+1},\ldots,v_{M+N}$  encode table rows and columns. Shown as yellow nodes in Figure 2.  $M$  and  $N$  are the total number of table rows and columns. The third type of node is sentences  $v_{M+N},\ldots,v_{K}$ , shown as blue nodes in Figure 2. Overall, the evidence graph has  $K$  nodes of unstructured and structured evidence at a comparable semantic granularity.
+
+Graph edges  $\mathcal{E}$  are constructed between related evidence following three rules: (i) Cross-Format Related:  $v_{i}$  and  $v_{j}$  are different format evidence of sentence and table, and in the same Wikipedia page. (ii) Structure-Based:  $v_{i}$  and  $v_{j}$  are nodes from the same table and (iii) Entity-Based:  $v_{i}$  and  $v_{j}$  have
+
+common entities and hyperlinks or sentences on the same Wikipedia page.
+
+We feed each claim-sentence pair to a RoBERTa model and use the embedding for [CLS] token to initialize the sentence nodes in our graph. And nodes of table rows and columns are initialized using the same method as in §2.1.
+
+Then we apply a layer of Graph Attention Network (GAT) (Velicković et al., 2018) with a residual connection to update representations of each node. It dynamically allocates weights between different nodes of evidence, allowing the graph to capture semantic connections between evidence.
+
+Finally, we use a two-layer feed-forward network to predict the retrieved probability for each node separately. The selective score for cell evidence is the multiplication of the probability of the row node and column node that contains the cell. Top  $n_c$  cells are extracted for verdict prediction.
+
+The loss function is computed as the weighted sum of the loss on nodes and cells:
+
+$$
+L = L ^ {S} + \alpha_ {c} \cdot L ^ {C o l} + \beta_ {c} \cdot L ^ {R o w} + \gamma_ {c} \cdot L ^ {C e}
+$$
+
+$L^{S}, L^{Col}, L^{Row}$ , and  $L^{Ce}$  are cross-entropy loss computed on sentences, table columns, table rows, and cells.  $\alpha_{c}, \beta_{c}$  and  $\gamma_{c}$  are coefficients to adjust weights between nodes and cells. Nodes of sentences in the gold evidence set are labeled as positive. Nodes of rows and columns are labeled as positive if they contain the gold cell evidence.
+
+# 3 Experiments
+
+We utilize FEVERIOUS as the test bed for our approach, as it is, to our knowledge, the only open-domain fact verification benchmark that combines both unstructured and structured evidence. FEVEROUS aims to accomplish two objectives: first, to extract sentence and table cell evidence from English Wikipedia, which contains over 95.6 million sentences and 11.8 million tables; and second, to predict the veracity label  $\hat{y}$  of a given claim. The possible veracity labels include SUPPORTS, REFUTES, and NOT ENOUGH INFO (NEI). Some claims can be verified using either sentences or tables alone, while approximately  $59\%$  of instances need structured evidence for claim verification.  $27\%$  of the instances need to combine evidence from unstructured and structured evidence to verify. The quantitative characteristics of the benchmark are presented in Table 1. Detailed experiment settings can be found in Appendix A.
+
+<table><tr><td></td><td>Train</td><td>Dev</td><td>Test</td><td>Total</td></tr><tr><td>Supported</td><td>41,835</td><td>3,908</td><td>3,372</td><td>49,115</td></tr><tr><td>Refuted</td><td>27,215</td><td>3,481</td><td>2,973</td><td>33,669</td></tr><tr><td>NEI</td><td>2,241</td><td>501</td><td>1,500</td><td>4,242</td></tr><tr><td>Sentences</td><td>31,607</td><td>3,745</td><td>3,589</td><td>38,941</td></tr><tr><td>Cells</td><td>25,020</td><td>2,738</td><td>2,816</td><td>30,574</td></tr><tr><td>Sentences+Cells</td><td>20,865</td><td>2,468</td><td>2,062</td><td>25,395</td></tr><tr><td>Total</td><td>71,291</td><td>7,890</td><td>7,845</td><td>87,026</td></tr></table>
+
+Table 1: Quantitative characteristics of FEVEROUS.
+
+# 3.1 Evaluation
+
+Evidence extraction is crucial in fact verification, it provides the verdict prediction module with the necessary evidence to verify the claim. FEVEROUS uses precision, recall and F1 score to measure the quality of retrieved evidence.  $\mathbb{E}$  is the collection of gold evidence sets and  $\hat{E}$  is the retrieved evidence set. Recall of a specific type of evidence is computed for an instance as:
+
+$$
+\operatorname {R e c a l l} (\mathbb {E}, \hat {E}) = \frac {| \hat {E} \cap \cup \mathbb {E} | _ {i}}{| \cup \mathbb {E} | _ {i}}
+$$
+
+$|E|_i$  is the number of  $i$  type evidence in set  $E$ . Evidence type  $i$  could be sentence, table or cell.
+
+A retrieved evidence set is considered complete iff at least one gold evidence set is included. The recall of evidence set is defined for an instance as follows:
+
+$$
+\operatorname {R e c a l l} (\mathbb {E}, \hat {E}) = \left\{ \begin{array}{l l} 1 & \exists E \in \mathbb {E}: E \subseteq \hat {E} \\ 0 & \text {o t h e r w i s e} \end{array} \right.
+$$
+
+The number of retrieved evidence is limited to 5 sentences and 25 cells at maximum (Aly et al., 2021). In the following discussion, recall refers to evidence recall unless specified for an evidence type.
+
+The evaluation for verdict prediction uses label accuracy(Acc.) and FEVERIOUS score(F.S) as its metrics. They are defined for an instance as:
+
+$$
+\begin{array}{l} \operatorname {A c c}. (y, \hat {y}) = \left\{ \begin{array}{l l} 1 & y = \hat {y} \\ 0 & \text {o t h e r w i s e} \end{array} \right. \\ \mathrm {F . S} = \operatorname {A c c}. (y, \hat {y}) \times \operatorname {R e c a l l} (\mathbb {E}, \hat {\mathrm {E}}) \\ \end{array}
+$$
+
+$y$  is the gold label and  $\hat{y}$  is the predicted label. FEVEROUS score considers both the correct prediction of the veracity label and the completeness of the retrieved evidence set.
+
+<table><tr><td>Models</td><td>Table</td><td>Sentence</td><td>Cell</td><td>Evidence</td></tr><tr><td>Baseline</td><td>56</td><td>53</td><td>29</td><td>30</td></tr><tr><td>FaBULOUS</td><td>-</td><td>56.6</td><td>34.2</td><td>40.4</td></tr><tr><td>DCUF</td><td>75.59</td><td>62.54</td><td>58.41</td><td>43.22</td></tr><tr><td>UnifEE</td><td>75.59</td><td>75.36</td><td>67.44</td><td>55.08</td></tr><tr><td>Our Model</td><td>80.86</td><td>75.50</td><td>77.16</td><td>61.43</td></tr></table>
+
+Table 2: Recall of different formats of evidence on the development set.
+
+# 3.2 Main Results
+
+Evidence Extraction Results Table2 shows the extraction results of our model on the development set for different categories of evidence. All the previous methods restrict their evidence set to at most 5 sentences, 3 tables, and 25 cells. Official baseline (Aly et al., 2021) and FaBULOUS (Bouziane et al., 2021) use a weaker document retrieval module which results in error propagation and low evidence recall. Our model uses the same document retrieval module as DCUF (Hu et al., 2022) and UnifEE (Hu et al., 2023), which obtains a recall of  $85.20\%$  for Wikipedia pages. Therefore recall of each format of evidence can be used to compare the performance of evidence extraction directly. SEE-ST leverages the row and column semantics of tables and increases the recall of tables by  $5.27\%$ . SEE-ST also increases the recall of cells by  $11.65\%$  and  $9.27\%$  compared with DCUF and UnifEE. This suggests the two stages of SEE-ST, table extraction and cell selection can both improve the performance of structured evidence extraction.
+
+Overall Results Table 3 shows overall performance on the dev and test set based on our evidence extraction results. We achieve  $60.01\%$  evidence recall on the blind test set, which is  $6.14\%$  higher than the previous state-of-the-art UnifEE (Hu et al., 2023). Following the verification method proposed in Hu et al. (2022), we obtain accuracy  $74.68\% / 65.16\%$  and FEVEROUS score  $49.73\% / 44.75\%$  on dev/test set. This demonstrates our enhancement on structured evidence extraction provides verification with more accurate evidence so that it could make the right predictions.
+
+Since SEE-ST obtains the maximum number of evidence in extraction, we also experiment on UnifEE without its threshold selection mechanism to compare its performance with SEE-ST more fairly. SEE-ST still improves the evidence recall by a large margin. Our choice of discarding threshold selection is further analyzed in Appendix B.
+
+The accuracy on the test set is generally about  $10\%$  lower than the accuracy on dev set. The main reason is the unbalanced instances of NEI claims in different splits. Our analysis of verdict prediction results shows that DCUF performs poorly in NEI instances. This explains the accuracy gap between the dev and test set.
+
+# 3.3 Ablation Study
+
+Table Extraction We assess the effectiveness of our table extraction model through the following ablation experiments: (i) w/o Column Semantics: We train a model that only considers the relevance between rows and the claim for table extraction, scoring tables based on the criterion max  $(\mathrm{P}_{row})$ . (ii) w/o Row Semantics: Analogously, we score tables using max  $(\mathrm{P}_{col})$ . (iii) w/o Row and Column: We eliminate the design of table extraction and instead utilize the embedding of the [CLS] token in claim-table encodings for table extraction. The recall of the top 3 tables is presented in Table 4.
+
+In the setups without column or row semantics, the recall of the top 3 tables decreases by  $1.03\%$  and  $1.76\%$ , respectively. In the w/o Row and Column setting, table recall drops significantly by  $17.75\%$ , resulting in a recall of  $63.11\%$ , which is even lower than the recall achieved by DrQA (Table 2). These results suggest that our table extraction module in SEE-ST, which combines semantics from both rows and columns, effectively captures the coarse-grained relevance between claims and tables.
+
+Cell Selection To evaluate the effectiveness of our evidence graph for cell selection, we design ablation experiments with an evidence graph constructed with the top 3 tables. The experimental settings are as follows: (i) w/o Structured Edges: We introduce additional edges between rows and columns from the same Wikipedia pages but not the same table. (ii) w/o Edge Pattern: We directly connect every node in the evidence graph to each other. (iii) w/o Row Semantics: Following Hu et al. (2023), we build an evidence graph of sentences and columns, and separately, a graph of table cells, using columns as intermediates to select cells. (iv) w/o Evidence Graph: We follow the table extraction pipeline with an added term for the cell loss function, selecting each cell based solely on the context within the same table.
+
+The results of the ablation experiments are presented in Table 5. When increasing the number of tables in the evidence graph from 3 to 5, cell and
+
+<table><tr><td rowspan="2">Models</td><td colspan="5">Dev</td><td colspan="5">Test</td></tr><tr><td>F.S</td><td>Acc.</td><td>E-P</td><td>E-R</td><td>E-F1</td><td>F.S</td><td>Acc.</td><td>E-P</td><td>E-R</td><td>E-F1</td></tr><tr><td>Official Baseline</td><td>19</td><td>53</td><td>12</td><td>30</td><td>17</td><td>17.73</td><td>48.48</td><td>10.17</td><td>28.78</td><td>15.03</td></tr><tr><td>EURECOM</td><td>19</td><td>53</td><td>12</td><td>29</td><td>17</td><td>20.01</td><td>47.79</td><td>13.73</td><td>33.73</td><td>19.52</td></tr><tr><td>Z team</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>22.51</td><td>49.01</td><td>7.76</td><td>42.64</td><td>13.12</td></tr><tr><td>CARE</td><td>26</td><td>63</td><td>7</td><td>37</td><td>12</td><td>23</td><td>53</td><td>7</td><td>37</td><td>11</td></tr><tr><td>NCU</td><td>29</td><td>60</td><td>10</td><td>42</td><td>17</td><td>25.14</td><td>52.29</td><td>9.91</td><td>39.07</td><td>15.81</td></tr><tr><td>Papelo</td><td>28</td><td>66</td><td>-</td><td>-</td><td>-</td><td>25.92</td><td>57.57</td><td>7.16</td><td>34.60</td><td>11.87</td></tr><tr><td>FaBULOUS</td><td>30</td><td>65</td><td>8</td><td>43</td><td>14</td><td>27.01</td><td>56.07</td><td>7.73</td><td>42.58</td><td>13.08</td></tr><tr><td>DCUF</td><td>35.77</td><td>72.91</td><td>15.06</td><td>43.22</td><td>22.34</td><td>33.97</td><td>63.21</td><td>14.79</td><td>44.10</td><td>22.15</td></tr><tr><td>UnifEE</td><td>44.86</td><td>73.67</td><td>19.04</td><td>55.08</td><td>28.30</td><td>41.50</td><td>65.04</td><td>18.35</td><td>53.87</td><td>27.37</td></tr><tr><td>UnifEE*</td><td>46.13</td><td>73.14</td><td>12.48</td><td>56.22</td><td>20.42</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Our Model</td><td>49.73</td><td>74.68</td><td>10.60</td><td>61.43</td><td>18.07</td><td>44.75</td><td>65.16</td><td>9.81</td><td>60.01</td><td>16.89</td></tr></table>
+
+Table 3: Model performance on the development set and test set. F.S is FEVEROUS score and Acc. is the accuracy of veracity labels. E-R, E-P and E-F1 are recall, precision and F1 computed based on the evidence set. All metrics are averaged on all instances. UnifEE* uses the maximum number of evidence as the constraint for the evidence set.  
+
+<table><tr><td>Models</td><td>Recall@3</td></tr><tr><td>SEE-ST</td><td>80.86</td></tr><tr><td>w/o Column Semantics</td><td>79.83</td></tr><tr><td>w/o Row Semantics</td><td>79.10</td></tr><tr><td>w/o Row and Column</td><td>63.11</td></tr></table>
+
+Table 4: Ablation study of coarse-grained table extraction. Recall@3 is the recall of top 3 extracted tables.  
+
+<table><tr><td>Recall</td><td>Cell</td><td>Evidence</td></tr><tr><td colspan="3">Top5 Tables</td></tr><tr><td>SEE-ST</td><td>77.16</td><td>61.43</td></tr><tr><td colspan="3">Top3 Tables</td></tr><tr><td>SEE-ST</td><td>74.20</td><td>60.16</td></tr><tr><td>w/o Structured Edges</td><td>72.30</td><td>58.61</td></tr><tr><td>w/o Edge Pattern</td><td>72.22</td><td>58.56</td></tr><tr><td>w/o Row Semantics</td><td>73.16</td><td>59.24</td></tr><tr><td>w/o Evidence Graph</td><td>68.37</td><td>56.08</td></tr></table>
+
+Table 5: Ablation study of fine-grained cell selection.
+
+evidence recall rise by  $2.96\%$  and  $1.38\%$ , respectively, achieving a recall rate of  $77.16\%$  for cell evidence and  $61.43\%$  for evidence recall. In the w/o Structured Evidence setting, rows and columns from different tables are connected if they are on the same Wikipedia page, resulting in a decrease of  $1.98\%$  in cell recall and  $1.53\%$  in evidence recall. This demonstrates that the Structured Edges designed in  $\S 2.2$  are beneficial for selecting cells as intersections of rows and columns. Connecting all cells in the graph further compromises the performance of structured evidence. In the w/o Row Semantics setting, we build two evidence graphs following Hu et al. (2023): one containing sentences
+
+and columns, and another with cells. Cell recall and evidence recall decline by  $1.04\%$  and  $0.92\%$  in this setting, indicating that SEE-ST's utilization of row and column representations captures cell information more precisely. Moreover, compared to the two graphs of sentences and cells, SEE-ST is more computationally efficient with only one evidence graph. Finally, the w/o Evidence Graph setting loses all connections between sentences and tables, as well as shared entity information among evidence. Each node is predicted based on its own information, resulting in a significant drop in cell and evidence recall. This highlights the importance of the design of evidence graph that enables interactions between texts and tables.
+
+# 4 Analysis
+
+# 4.1 Error Analysis
+
+Here we perform a detailed error analysis for evidence extraction.
+
+Error Source Analysis The error source for evidence sets is defined as the format of evidence that is not extracted completely in the three-step pipeline (§2). We use Page to denote that document retrieval fails to retrieve all the pages containing evidence. For those instances that have a complete document set, the error source is defined according to the format of evidence fail to extract: Unstructured (sentences), Structured (tables or cells) and Both. The instances with a complete evidence set are denoted as Complete. About 1% of instances have other evidence types such as lists and table captions. These instances are excluded during error analysis. The impact of ignorance of other evidence types is discussed in the Limitations section.
+
+![](images/d3704b25906302a66669dd3ac694ce93de22446bfcfd8372c7c34c483df24d64.jpg)  
+Figure 3: Error analysis of extracted evidence set for the dev set. Left: Overall error source analysis. Right: Error source proportions of claims with different reasoning challenges.
+
+![](images/aadca6d93ba0869925c95909195a3c89c10ba7abc39b08e38109597c771fde35.jpg)
+
+The result of the error source analysis is shown in Figure 3. The document retriever fails to retrieve complete documents for  $15.8\%$  instances in the dev set. Our proposed method decreases the proportion of cases missing structured evidence from  $14.28\%$  to  $9.34\%$ . It shows the effectiveness of SEE-ST in structured evidence extraction.
+
+Analysis based on challenge types A fact-checker system would face a specific reasoning challenge when extracting evidence and verifying claims for FEVERIOUS. The challenges include multi-hop reasoning, entity disambiguation, search term not in claim, combining tables and texts and numerical reasoning. The challenges that do not belong to those five types are categorized as other.
+
+We further evaluate the performance of SEE-ST in each category to demonstrate its ability to retrieve evidence for claims with different reasoning challenges. SEE-ST achieves great extraction performance over claims that need to combine tables and texts with a recall rate of  $69.27\%$  showing that the unified graph neural network can effectively integrate information from both formats of evidence. The recall rates of claims that need multi-hop reasoning, entity disambiguation or search terms are lower than  $50\%$ . But the analysis in Figure 3 shows that the main error source of incomplete evidence extraction is document retrieval. It fails to retrieve complete document sets for each challenge type listed above with an error rate of  $38.6\%$ ,  $28.4\%$  and  $54.6\%$  while only about  $7\%$  instances in these challenges lose complete evidence sets because of unsuccessful structured evidence extraction. This demonstrates the ability of our model to retrieve evidence in more challenging scenarios. Our model
+
+also improves the evidence extraction performance of numerical reasoning. The rate of Structured errors is  $34.82\%$  for UnifEE (Hu et al., 2023). SEEST increases the recall rate by  $8.25\%$  and decreases the rate of Structured errors by  $7.35\%$ . Yet numerical reasoning remains the main challenge in structured evidence extraction.
+
+# 5 Case Study
+
+In this section, we demonstrate two cases for evidence extraction in Figure 4. For the claim on 1991 Waterford City Council election, our table extraction module successfully extracts the table of election statistics from the Wikipedia dump. The main challenge for this case is numerical reasoning. Verifying ... six parties all of which have at least two seats requires high-order semantics in Seats column. Such semantics is maintained in coarse-grained table extraction and fine-grained cell selection. SEE-ST extracts cells in Party and Seats columns as evidence. It shows the numerical reasoning ability of SEE-ST.
+
+As for the evidence extraction for the claim of The Castle of Iron, it demonstrates the ability of SEE-ST to perform multi-hop reasoning. The claim states that ...was published by a large-scale publishing company in New York, and is refuted by sentence evidence Gnome Press was an American small-press publishing company.... However, the name of the publisher is not mentioned in the claim. It is the structured evidence that links the claim with the sentence evidence. The design of our evidence graph could enable interaction between sentence and table cells to achieve multi-hop reasoning over different formats of evidence.
+
+Claim: 1991 Waterford City Council election was planned to take place on 27 June 1991 with six parties all of which have at least two seats.
+
+Label: SUPPORTS Challenge: Numerical Reasoning
+
+Extracted Evidence  
+Structured:  
+
+<table><tr><td>Party</td><td>Seats</td><td>±</td><td>First Pref. Votes</td><td>FPv%</td></tr><tr><td>Fianna Fáil</td><td>3</td><td>-2</td><td>3,165</td><td>19.01%</td></tr><tr><td>Fine Gael</td><td>2</td><td>-2</td><td>2,220</td><td>13.34%</td></tr><tr><td>Labour</td><td>3</td><td>+1</td><td>3,415</td><td>20.5%</td></tr><tr><td>Workers&#x27;s Party</td><td>3</td><td>+1</td><td>3,359</td><td>20.18%</td></tr><tr><td>Progressive Democrats</td><td>2</td><td>+2</td><td>1,666</td><td>10.01%</td></tr><tr><td>Independent</td><td>2</td><td>-</td><td>2,360</td><td>14.2%</td></tr><tr><td>Totals</td><td>15</td><td>-</td><td>16,645</td><td>100%</td></tr></table>
+
+# Unstructured:
+
+An election to Waterford City Council took place on 27 June 1991 as part of that year's Irish local elections.
+
+Waterford City Council Comhairle Cathrach Phort Láirge) was the authority responsible for local governments in the city of waterford.
+
+…
+
+Claim: The Castle of Iron is a fantasy novel written by L. Sprague de Camp and Fletcher Pratt and was published by a large-scale publishing company in New York.
+
+Label: REFUTES
+
+Challenge: Multi-hop Reasoning
+
+Extracted Evidence  
+Structured:  
+
+<table><tr><td>Author</td><td>L. Sprague de Camp and Fletcher Pratt</td><td>Publisher</td><td>Gnome Press</td></tr><tr><td>Cover artist</td><td>Hannes Bok</td><td>Publication date</td><td>1941, 1950</td></tr><tr><td>Country</td><td>United States</td><td>Media type</td><td>Print</td></tr><tr><td>Language</td><td>English</td><td>Pages</td><td>224</td></tr><tr><td>Series</td><td>Harold Shea</td><td>Preceded by</td><td>The Incomplete Enchanter</td></tr><tr><td>Genre</td><td>Fantasy</td><td>Followed by</td><td>Wall of Serpents</td></tr></table>
+
+# Unstructured:
+
+Gnome Press was an American small-press publishing company primarily known for publishing many science fiction classics.
+
+The Castle of Iron is the title of a fantasy novella by American authors L. Sprague de Camp and Fletcher Pratt, and of the novel into which it was later expanded by the same authors.
+
+…
+
+Figure 4: Demonstrations of extracted evidence set for claims in FEVERIOUS. Gold evidence is highlighted in orange. The underlines indicate the cells selected by SEE-ST and table headers are marked with bold texts.
+
+# 6 Related Work
+
+Open-domain fact verification extracts evidence from a knowledge base (KB) to verify the claim. In the real world, unstructured and structured evidence can be used to verify the claim. FEVER (Thorne et al., 2018) and HoVer (Jiang et al., 2020) focus on extracting unstructured evidence i.e. sentences from the Wikipedia dump. And the research of structured evidence mainly focuses on tables. TabFact (Chen et al., 2019) and InfoTabs (Gupta et al., 2020) verify claims with tables and infoboxes from Wikipedia. SEM-TABFACT (Wang et al., 2021) and SciTab (Lu et al., 2023) are constructed on claims and tables from scientific papers. Yet these datasets all provide the table evidence for verification as in a closed-domain setting. To our knowledge, FEVEROUS is the only open-domain fact verification dataset that considers table as evidence (Aly et al., 2021). Therefore, our work uses FEVEROUS as a testbed for table extraction and cell selection.
+
+As for table extraction, the lexical matching approach is widely used in fact verification (Aly et al., 2021; Hu et al., 2022, 2023). Chen et al. (2020) encodes tables with BERT and concatenate embedding with structured features to rank tables. TAPAS enhances table encoding with additional table-aware positional embeddings (Herzig et al., 2021). (Pan et al., 2021) also leverages represent
+
+tations of row and column to rank tables but overlooks the context of other cells. TaBERT (Yin et al., 2020) utilizes a content-snapshot mechanism to only encode the relevant rows thus may lose the high-order semantics of columns.
+
+Gi et al. (2021) converts each cell into a text sequence to select cell evidence, bypassing the table extraction step. Acharya (2021) selects cells through dependency parsing. Jindal et al. (2021) combines single-cell NLI task with cell-wise relevance to extract evidence at cell-level semantic granularity. These methods are performed at the semantic granularity of cells. Approaches in question answering have proposed to select cells as the intersection of rows and columns (Glass et al., 2021; Pan et al., 2021). All these above methods do not consider the interaction between evidence.
+
+The idea of evidence graph has been explored in previous works. Kotonya et al. (2021) linearizes cells into sentences to construct a fully connected evidence graph with texts. UnifEE (Hu et al., 2023) proposes to use column nodes as intermediates between sentence-level and cell-level evidence graphs. Our method is more computationally efficient with only one evidence graph. Our design enables rows and columns to interact with sentences at comparable semantic granularity while still maintaining high-order semantics from tables.
+
+# 7 Conclusions
+
+This paper presents SEE-ST, a novel method for enhancing the extraction of structured evidence. By leveraging both row and column semantics, SEE-ST can operate with a comparable evidence granularity, enabling both coarse-grained table extraction and fine-grained cell selection. Experimental results demonstrate the effectiveness of SEE-ST, as it significantly improves the recall of structured evidence compared to existing methods in fact verification. Moreover, SEE-ST outperforms state-of-the-art methods in evidence recall and FEVERIOUS score on FEVEROUS benchmark. These findings underline the impact of SEE-ST in enhancing structured evidence extraction.
+
+# Limitations
+
+Our work explores the extraction of structured evidence in enhancing table extraction and cell selection. In the real world, fact checkers also use other structured evidence as knowledge graphs and databases to verify the claim. FactKG is a fact verification dataset on knowledge graphs (Kim et al., 2023). The verification model is required to retrieve subgraphs as evidence for verdict prediction. SEE-ST can not be directly applied to this task. We will try to build a unified evidence extraction module that could retrieve both tables and subgraphs for fact verification in our future work.
+
+Our work only enhances the extraction of structured evidence, while the whole evidence extraction process still suffers from error propagation from document retrieval. The error propagation in the three-step pipeline makes it extremely difficult for the model to retrieve complete evidence sets and predict the right verdict label. A new pipeline for open-domain fact verification is needed to solve this issue.
+
+# Acknowledgements
+
+This work is supported by NSFC (62161160339). We would like to thank the anonymous reviewers for their helpful comments and suggestions. For any correspondence, please contact Yansong Feng.
+
+# References
+
+Kaushik Acharya. 2021. KaushikAcharya at SemEval-2021 task 9: Candidate generation for fact verification over tables. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval
+
+2021), pages 1271-1275, Online. Association for Computational Linguistics.  
+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.  
+Mostafa Bouziane, Hugo Perrin, Amine Sadeq, Thanh Nguyen, Aurélien Cluzeau, and Julien Mardas. 2021. Fabulous: Fact-checking based on understanding of language over unstructured and structured information. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 31-39.  
+Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer open-domain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870-1879.  
+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. arXiv preprint arXiv:1909.02164.  
+Zhiyu Chen, Mohamed Trabelsi, Jeff Heflin, Yinan Xu, and Brian D. Davison. 2020. Table search using a deep contextualized language model. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020, pages 589-598. ACM.  
+In-Zu Gi, Ting-Yu Fang, and Richard Tzong-Han Tsai. 2021. Verdict inference with claim and retrieved elements using roberta. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 60-65.  
+Michael Glass, Mustafa Canim, Alfio Gliozzo, Saneem Chemmengath, Vishwajeet Kumar, Rishav Chakravarti, Avi Sil, Feifei Pan, Samarth Bharadwaj, and Nicolas Rodolfo Fauceglia. 2021. Capturing row and column semantics in transformer based question answering over tables. In Annual Conference of the North American Chapter of the Association for Computational Linguistics.  
+Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. A survey on automated fact-checking. Transactions of the Association for Computational Linguistics, 10:178-206.  
+Vivek Gupta, Maitrey Mehta, Pegah Nokhiz, and Vivek Srikumar. 2020. INFOTABS: Inference on tables as semi-structured data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2309–2324, Online. Association for Computational Linguistics.
+
+Momchil Hardalov, Arnav Arora, Preslav Nakov, and Isabelle Augenstein. 2022. A survey on stance detection for mis-and disinformation identification. In *Findings of the Association for Computational Linguistics: NAACL* 2022, pages 1259-1277.  
+Jonathan Herzig, Thomas Müller, 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, pages 512-519, Online. Association for Computational Linguistics.  
+Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Mueller, Francesco Piccinno, and Julian Eisenschlos. 2020. Tapas: Weakly supervised table parsing via pre-training. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4320-4333.  
+Nan Hu, Zirui Wu, Yuxuan Lai, Xiao Liu, and Yansong Feng. 2022. Dual-channel evidence fusion for fact verification over texts and tables. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5232-5242.  
+Nan Hu, Zirui Wu, Yuxuan Lai, Chen Zhang, and Yansong Feng. 2023. Unifee: Unified evidence extraction for fact verification. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1142-1152.  
+Yichen Jiang, Shikha Bordia, Zheng Zhong, Charles Dognin, Maneesh Singh, and Mohit Bansal. 2020. HoVer: A dataset for many-hop fact extraction and claim verification. In *Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)*.  
+Aditya Jindal, Ankur Gupta, Jaya Srivastava, Preeti Menghwani, Vijit Malik, Vishesh Kaushik, and Ashutosh Modi. 2021. BreakingBERT@IITK at SemEval-2021 task 9: Statement verification and evidence finding with tables. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 327-337, Online. Association for Computational Linguistics.  
+Jiho Kim, Sungjin Park, Yeonsu Kwon, Yohan Jo, James Thorne, and Edward Choi. 2023. FactKG: Fact verification via reasoning on knowledge graphs. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16190-16206, Toronto, Canada. Association for Computational Linguistics.  
+Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
+
+Neema Kotonya, Thomas Spooner, Daniele Magazzeni, and Francesca Toni. 2021. Graph reasoning with context-aware linearization for interpretable fact extraction and verification. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 21-30.  
+Xinyuan Lu, Liangming Pan, Qian Liu, Preslav Nakov, and Min-Yen Kan. 2023. Scitab: A challenging benchmark for compositional reasoning and claim verification on scientific tables. arXiv preprint arXiv:2305.13186.  
+Feifei Pan, Mustafa Canim, Michael Glass, Alfio Gliozzo, and Peter Fox. 2021. Cltr: An end-to-end, transformer-based system for cell-level table retrieval and table question answering. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations, pages 202-209.  
+Mohammed Saeed, Giulio Alfarano, Khai Nguyen, Duc Pham, Raphael Troncy, and Paolo Papotti. 2021. Neural re-rankers for evidence retrieval in the FEVERIOUS task. In Proceedings of the Fourth Workshop on Fact Extraction and VERIFICATION (FEVER), pages 108-112, Dominican Republic. Association for Computational Linguistics.  
+James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics.  
+Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. In International Conference on Learning Representations.  
+Nancy X. R. Wang, Diwakar Mahajan, Marina Danilevsky, and Sara Rosenthal. 2021. SemEval-2021 task 9: Fact verification and evidence finding for tabular data in scientific documents (SEM-TABFACTS). In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 317-326, Online. 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.
+
+# A Implementation Details
+
+In our approach, we retrieve  $n_p = 5$  pages from Wikipedia for each claim following Hu et al. (2022). And we extract  $n_s = 5$  sentences and  $n_t = 5$  tables
+
+from the pages to construct our evidence graph. As for the cell selection, we limit the selected cells to a maximum of  $n_c = 25$  cells. For a fair comparison with baselines, the source tables of cells are constrained to the top 3 tables. The tables are re-ranked by the highest cell selective score at cell selection.
+
+For table extraction, we utilize TAPAS-base<sup>1</sup> to encode the claim-table pair. Our model is trained with Adam optimizer (Kingma and Ba, 2014) with a batch size of 8. The learning rate is  $10^{-7}$  for TAPAS and  $10^{-5}$  for the classifier.  $\alpha_{t}$  and  $\beta_{t}$  are assigned a value of 1.
+
+Regarding our evidence graph, we employ Spacy $^2$  to extract common entities from evidence to connect relevant nodes. RoBERTa-base $^3$  and TAPAS-base are chosen as the encoders for node initialization. The model is optimized using the Adam optimizer (Kingma and Ba, 2014) with a batch size of 4 and a peak learning rate of  $10^{-6}$ . We set the weights for different coefficients of loss functions as follows:  $\alpha_{c} = \beta_{c} = 2$  and  $\gamma_{c} = 1$ . The model takes 14 hours to train for 3 epochs and 34 minutes to select cells in the dev set on a single NVIDIA A100 GPU.
+
+# B Evidence Precision Analysis
+
+With the constraint of a maximum number of sentences and cells in the evidence set, there are two ways to construct evidence sets. The previous SOTA method UnifEE (Hu et al., 2023) uses a threshold to construct an evidence set with high precision at the expense of a slight decrease in evidence recall. Other baselines use evidence as much as fit the constraints to construct the evidence sets resulting in higher evidence recall. In this section, we aim to analyze which strategy achieves higher accuracy and FEVERIOUS score at the verdict prediction step. We filter out unnecessary evidence with different thresholds and verify claims based on the filtered evidence sets. The results are presented in Table 6.
+
+The results indicate that as the evidence threshold increases, evidence precision improves at the cost of reduced evidence recall. However, label accuracy remains largely unaffected by changes in the evidence set. Threshold selection leads to a decrease in the FEVERIOUS score. Overall, the
+
+verification model used in this work is robust to threshold selection within the same extracted evidence set, and verdict prediction achieves the highest FEVERIOUS score without threshold selection. The annotations of the gold evidence set may not be exhaustive of all possible evidence, which could result in an underestimation of precision due to the presence of correct evidence that is not annotated as gold evidence. In summary, SEE-ST employs the maximum number of sentences and cells as constrain for evidence sets for optimal verdict prediction performance.
+
+<table><tr><td>Threshold</td><td>E-P</td><td>E-R</td><td>E-F1</td><td>F.S</td><td>Acc.</td></tr><tr><td>0</td><td>10.60</td><td>61.43</td><td>18.07</td><td>49.73</td><td>74.68</td></tr><tr><td>0.001</td><td>12.99</td><td>60.17</td><td>21.36</td><td>48.38</td><td>74.25</td></tr><tr><td>0.005</td><td>15.85</td><td>59.60</td><td>25.04</td><td>48.04</td><td>74.03</td></tr><tr><td>0.01</td><td>17.88</td><td>58.54</td><td>27.49</td><td>47.28</td><td>74.22</td></tr><tr><td>0.05</td><td>23.39</td><td>50.78</td><td>32.04</td><td>42.56</td><td>73.99</td></tr></table>
+
+Table 6: Effects of evidence extraction on verdict prediction.  
+
+<table><tr><td>Verdict</td><td>Recall</td></tr><tr><td>NEI</td><td>41.72</td></tr><tr><td>SUPPORTS</td><td>61.38</td></tr><tr><td>REFUTES</td><td>61.84</td></tr></table>
+
+Table 7: The performance of evidence extraction for claims with different veracity labels.
+
+# C Performance on Different Verdict
+
+We further analyze the performance of evidence extraction and verdict prediction for claims with different veracity labels, as shown in Table 7. Generally, extracting evidence for REFUTES claims is more challenging, as the information contained in the evidence is not always explicitly included in the claim, making it difficult to retrieve based on relevance alone. However, our model achieves comparable extraction performance for both SUPPORTS and REFUTES types, with recall rates of  $61.38\%$  and  $61.84\%$ , respectively. The recall for NEI instances is approximately  $20\%$  lower than for the other types. NEI claims cannot be verified based on evidence from Wikipedia alone; only the most relevant information is listed in the gold evidence set, making it difficult to extract a complete evidence set for NEI claims.

2023/Enhancing Structured Evidence Extraction for Fact Verification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2a59015368349038a357753a35290d4e07a10e7e3bff24ed8b048574b717723
+size 449764

2023/Enhancing Task-oriented Dialogue Systems with Generative Post-processing Networks/ff4f198b-74ec-4f81-bac0-82b9a8b9efc4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:994720f0e95d39c839d56fc007f98b499349fe7c72f61573c271e04b266d3297
+size 749463

2023/Enhancing Task-oriented Dialogue Systems with Generative Post-processing Networks/full.md

@@ -0,0 +1,465 @@
+# Enhancing Task-oriented Dialogue Systems with Generative Post-processing Networks
+
+Atsumoto Ohashi Ryuichiro Higashinaka
+
+Graduate School of Informatics, Nagoya University
+
+ohashi.atsumoto.c@s.mail.nagoya-u.ac.jp
+
+higashinaka@i.nagoya-u.ac.jp
+
+# Abstract
+
+Recently, post-processing networks (PPNs), which modify the outputs of arbitrary modules including non-differentiable ones in task-oriented dialogue systems, have been proposed. PPNs have successfully improved the dialogue performance by post-processing natural language understanding (NLU), dialogue state tracking (DST), and dialogue policy (Policy) modules with a classification-based approach. However, they cannot be applied to natural language generation (NLG) modules because the post-processing of utterances output by NLG modules requires a generative approach. In this study, we propose a new post-processing component for NLG, generative post-processing networks (GenPPNs). For optimizing GenPPNs via reinforcement learning, the reward function incorporates dialogue act contribution, a new measure to evaluate the contribution of GenPPN-generated utterances with regard to task completion in dialogue. Through simulation and human evaluation experiments based on the MultiWOZ dataset, we confirmed that GenPPNs improve the task completion performance of task-oriented dialogue systems<sup>1</sup>.
+
+# 1 Introduction
+
+A typical task-oriented dialogue system has a pipelined structure consisting of four modules (Young et al., 2013; Zhang et al., 2020): natural language understanding (NLU), dialogue state tracking (DST), dialogue policy (Policy), and natural language generation (NLG). Many studies have used reinforcement learning (RL) to improve the task completion performance of an entire pipelined dialogue system by fine-tuning modules directly (Lee et al., 2021; Lin et al., 2021; Chen et al., 2023).
+
+Recently, Ohashi and Higashinaka (2022b) proposed a novel method using Post-Processing Networks (PPNs) that can optimize pipelined dialogue
+
+![](images/acdf4dc7c58f23a3db9f9820e93dcccc98f0dc6f1b36edbc94c0e73e03eb974e.jpg)  
+Figure 1: Diagrams of MLP-based PPN for NLU, DST, and Policy modules proposed by Ohashi and Higashinaka (2022b) and PLM-based GenPPN for NLG proposed in this study.
+
+systems that consist of arbitrary modules including non-differentiable ones (e.g., rule-based and Web API-based). Their method uses RL to optimize a neural-based component called a PPN, which modifies the output of each module. PPNs have been applied to three modules, namely NLU, DST, and Policy, and have been shown to improve the task completion performance of various pipelined systems. In PPNs, the post-processing of module outputs is treated as a binary classification task, i.e., adding or removing slots in the output of each module, and this classification task is modeled with a multi-layer perceptron (MLP). This means that the current PPNs could not be applied to NLGs, where the unit of output is not slots but a sequence of tokens, i.e., natural language.
+
+To overcome this limitation, we propose a Pre-trained Language Model (PLM)-based Generative Post-processing Network (GenPPN) that can postprocess the output of NLGs in a manner similar to conventional PPNs (Figure 1). To optimize GenPPN via commonly used RL frameworks for PLMs (Ziegler et al., 2019; Stiennon et al., 2020), a reward for each utterance at each turn (an utterance-level reward) is required. In a task-oriented dialogue, however, a reward indicating success or failure is obtained only at the end of a multi-turn interaction (a dialogue-level reward). Therefore, we introduce a dialogue act (DA) contribution for distributing the dialogue-level reward to the utterance
+
+level reward. Here, a DA is the meaning representation of the information that NLG converts into an utterance, and DA contribution is a measure of how much the DA of each utterance contributes to the final task completion of the dialogue.
+
+Experiments on the MultiWOZ dataset (Budzianowski et al., 2018) confirm that GenPPNs can improve the task performance of the entire dialogue system, regardless of the architecture of the NLG module. Furthermore, an ablation study reveals that the introduction of DA contribution is effective for learning GenPPNs to improve task completion. The contributions of this study are threefold:
+
+- We propose the generative post-processing network (GenPPN) to modify the output utterances of the NLG module, which was impossible with conventional PPNs.  
+- We introduce DA contribution to optimize the GenPPN by evaluating the impact of each utterance on dialogue task completion.  
+- Simulation experiments on the MultiWOZ dataset confirm that the proposed GenPPN improves the task completion performance of the entire dialogue system, regardless of the architecture of the NLG module. We also validated that a GenPPN optimized using simulation is effective in human evaluation experiments.
+
+# 2 Related Work
+
+# 2.1 Optimization of Pipelined Task-oriented Dialogue Systems
+
+Methods have been proposed to optimize the task completion performance of an entire pipelined system using RL. Zhao and Eskenazi (2016) and Li et al. (2017) optimized a Policy network implemented in MLP using the Deep Q-Network algorithm (Mnih et al., 2013) to achieve robustness against errors that occur in real dialogues. Mehri et al. (2019) proposed a method to add additional parameters to NLU, Policy, and NLG and optimize them using RL. By expressing a dialogue state output by a DST as a probability distribution, Lee et al. (2021) made the entire system differentiable and jointly optimized it via RL. A method of fine-tuning an NLU while optimizing a Policy in dialogue simulations was proposed (Lin et al., 2021). Tseng et al. (2021) proposed a domain-adaptive learning
+
+framework that simultaneously optimizes the Policy module of the dialogue system and the user simulator. An RL framework for online DST optimization was also proposed to improve dialogue management performance (Chen et al., 2023).
+
+Instead of training each module, Ohashi and Higashinaka (2022b) proposed a generalized method that optimizes PPNs, classification-based models that modify the outputs of NLU, DST, and Policy, to enhance the task completion performance. In this paper, we propose a new generative-based PPN for post-processing NLG modules, which has not been supported by conventional PPNs.
+
+# 2.2 Natural Language Generation Module for Task-oriented Dialogues
+
+Conventional NLGs for task-oriented dialogues used template-based or rule-based methods (Walker et al., 2002; Stent et al., 2004). Later, data-driven methods using machine learning were proposed (Oh and Rudnicky, 2002; Angeli et al., 2010; Mairesse and Young, 2014) that do not require the cost of template and rule creation.
+
+In recent years, many generative models based on deep learning have been proposed (Wen et al., 2016; Tran and Nguyen, 2017; Su et al., 2018). Wen et al. (2015) proposed an SC-LSTM that controls utterance generation using DA feature vectors and reading gates. SC-GPT (Peng et al., 2020) is the best NLG model of MultiWOZ. It achieves high performance by fine-tuning GPT-2 (Radford et al., 2019) on many task-oriented dialogue datasets such as MultiWOZ (Budzianowski et al., 2018) and the Schema-Guided Dialogue Dataset (Rastogi et al., 2020).
+
+For task-oriented dialogue, it is crucial not only to generate natural utterances via maximum likelihood estimation (MLE) but also to accurately reflect the input DA's content. To achieve this, Balakrishnan et al. (2019) introduced a conditional decoding approach utilizing a tree-shaped semantic representation, enhancing the slot content in generated utterances. Furthermore, Li et al. (2020) offered a method to lower the slot error rate in utterances using an iterative RL framework for slot consistency. Ohashi and Higashinaka (2022a) presented a fine-tuning method for utterance generation that uses a user's NLU model so that the NLU can understand the DA accurately.
+
+In all of the above studies, the optimization was performed at the utterance level using a fixed cor
+
+![](images/0d1d334e480d9916b4062228667747770b3953b8427b0d06d60ff213ce9fb776.jpg)  
+Figure 2: Schematic diagram of GenPPN and how it is optimized. GenPPN  $\mathcal{M}$  rewrites utterance  $u$  generated by NLG of dialogue system into  $u^{\prime}$ . This rewriting of  $\mathcal{M}$  is optimized using reward  $r(u^{\prime})$  calculated on the basis of DA contribution  $c_{(I,s)}$ , so system's ability to complete task improves.
+
+pus of utterances. In this study, we aim to optimize post-processing not only on the utterance level but also on the dialogue level to improve the task completion performance for multi-turn dialogues.
+
+# 3 Method
+
+# 3.1 Overview
+
+Figure 2 shows a schematic diagram of GenPPN and how it is optimized. At the beginning of each interaction, the user is given a user goal consisting of a list of constraints to be informed to the system and information to be obtained from the system through the dialogue. During the dialogue, from the user's utterance at a certain turn  $t$ , the NLU and DST modules of the system first estimate the current belief state of the user. Then, on the basis of the belief state, the Policy module determines the next actions to be taken by the system as DAs  $a_{t}$ . The  $a$  takes a structure containing one or more triples of intent  $I$ , slot  $s$ , and value  $v$ :
+
+$$
+a = \left\{\left(I _ {i}, s _ {i}, v _ {i}\right) \mid i = 1, \dots , | a | \right\} \tag {1}
+$$
+
+The NLG module maps  $a_{t}$  to the system's utterance  $u_{t}$ . Here, GenPPN  $\mathcal{M}$  post-processes, i.e., rewrites  $u_{t}$ . A prompt  $x_{t} = \mathrm{Prompt}(h_{t},a_{t},u_{t})$  is created from the previous dialogue context  $h_t$ ,  $a_{t}$ , and  $u_{t}$ , and  $\mathcal{M}$  generates a new utterance  $u_{t}^{\prime}$ :  $u_{t}^{\prime}\sim \mathcal{M}(x_{t})$  (see Section 3.2 for details on GenPPN and the prompt format). The user then receives  $u_{t}^{\prime}$ , estimates the system DAs  $\hat{a}_t$ , and the user's utterance is output for the next turn  $t + 1$  on the basis of the  $\hat{a}_{t}$  and current status of the user goal. This exchange is repeated until the user goal is achieved or until the maximum number of turns is reached.
+
+Let  $T$  be the number of turns at the end of a dialogue. In addition, the dialogue result  $e \in \{0,1\}$  of whether the user goal was achieved is determined,
+
+i.e., 0 means that the task failed, and 1 means that the task succeeded. The objective is to obtain the optimal parameters  $\theta^{*}$  for the GenPPN  $\mathcal{M}_{\theta}$  such that the expected value of  $e$  is maximized:
+
+$$
+\theta^ {*} = \arg \max  _ {\theta} \max  _ {U ^ {\prime} \sim \mathcal {M} _ {\theta} (X)} \mathbb {E} [ e (U ^ {\prime}) ] \tag {2}
+$$
+
+where  $U^{\prime} = \{u_{1}^{\prime},u_{2}^{\prime},\dots,u_{T}^{\prime}\}$  are the system utterances sampled by  $\mathcal{M}$  from prompts  $X = \{x_{1},x_{2},\ldots ,x_{T}\}$  at each turn.
+
+Since the component to be trained is not the NLG itself but the GenPPN, our optimization is independent of the architecture and differentiability of NLG modules. In the following subsections, we describe the GenPPN, reward design for optimizing Eq. (2), and RL algorithm used.
+
+# 3.2 Generative Post-processing with PLM
+
+We assume that the model  $\mathcal{M}$  of GenPPN is a language model based on the Transformer architecture (Vaswani et al., 2017). At turn  $t$ , the input prompt  $x_{t}$  is created from the previous dialogue context  $h_t$ , DA  $a_t$ , and system utterances  $u_{t}$  by using a hand-crafted prompt (specific prompts are shown in Section A.2 in the appendix).
+
+Through RL,  $\mathcal{M}$  learns to generate a rewritten system utterance  $u^{\prime}$  such that the dialogue result  $e$  is maximized. However, even in the early stages of learning, if the system cannot produce at least reasonable system utterances, the dialogue always breaks down, and the training does not progress properly. For this reason, we adopt an instruction-tuned PLM (Wei et al., 2022; Chung et al., 2022; Taori et al., 2023), which has shown high performance on many NLP benchmarks including task-oriented dialogue modeling in few- and zero-shot settings (Pan et al., 2023; Hudeček and Dusek,
+
+2023). Since the general PLM has a massive number of parameters, we use Low-Rank Adaptation (LoRA) (Hu et al., 2022), which has good parameter efficiency. That is, a small number of parameters  $\theta$  is added to each self-attention module in the PLM, and only  $\theta$  is optimized during training.
+
+# 3.3 Reward with DA Contribution
+
+Eq. (2) implies that only  $e$  or a dialogue-level reward at the end of a multi-turn dialogue should be used to evaluate the generated utterances  $U^{\prime}$  for all  $T$  turns. However, since this reward is too sparse and it is not feasible to optimize  $\mathcal{M}$ , we approximate Eq. (2) by using the utterance-level reward  $r$ , which evaluates how much of an effect each utterance has on the final task completion:
+
+$$
+\theta^ {*} = \arg \max  _ {\theta} \underset {U ^ {\prime} \sim \mathcal {M} _ {\theta} (X)} {\mathbb {E}} \left[ \sum_ {u _ {t} ^ {\prime} \in U ^ {\prime}} ^ {T} r \left(u _ {t} ^ {\prime}\right) \right] \tag {3}
+$$
+
+The role of NLG is to accurately convey DAs to the user, i.e., to generate utterances such that  $a_{t} = \hat{a}_{t}$ , but this kind of evaluation based on the consistency of  $a_{t}$  and  $\hat{a}_{t}$  does not take into account  $e$ , so it cannot be used for  $r$  directly.
+
+With this in mind, we design  $r$  by combining  $e$  with the consistency evaluation for  $a_{t}$  and  $\hat{a}_{t}$ . Specifically, we statistically measure the contribution of correctly conveying  $(I,s,v) \in a_{t}$  to the user in terms of the impact on  $e$  on the basis of the dialogues we have sampled. Then, we give a contribution-weighted reward to the utterance  $u_{t}^{\prime}$  that succeeds in conveying  $(I,s,v)$ .
+
+To realize this, first, at the end of each dialogue, triples  $\{(a_t,\hat{a}_t,e)|t\in T\}$  consisting of  $a_{t}$  and  $\hat{a}_{t}$  at each turn together with  $e$  are added to the DA history  $\mathcal{D}$ . Here,  $e$  is determined retrospectively on the basis of the outcome at the end of the dialogue;  $e$  of all turns in a successful dialogue is 1,  $\{(a_t,\hat{a}_t,1)|t\in T\}$ , and  $e$  of all turns in a failed dialogue is 0,  $\{(a_t,\hat{a}_t,0)|t\in T\}$ . Here, all  $(a,\hat{a})$  pairs in  $\mathcal{D}$  sampled so far are split into  $\mathcal{S}$  and  $\mathcal{F}$  depending on whether the task was successful:
+
+$$
+\mathcal {S} = \left\{\left(a, \hat {a}\right) \mid \left(a, \hat {a}, e\right) \in D, e = 1 \right\}
+$$
+
+$$
+\mathcal {F} = \left\{\left(a, \hat {a}\right) \mid \left(a, \hat {a}, e\right) \in D, e = 0 \right\}
+$$
+
+Then, for each  $(I,s,v)\in a_{t}$ , the contribution  $c_{(I,s)}$  of  $(I,s)$  is calculated as follows:
+
+$$
+c _ {(I, s)} = \frac {n _ {(I , s)} ^ {\operatorname {R e c} , \mathcal {S}} + n _ {(I , s)} ^ {\operatorname {U n r} , \mathcal {F}}}{n _ {(I , s)} ^ {\operatorname {R e c} , \mathcal {S}} + n _ {(I , s)} ^ {\operatorname {R e c} , \mathcal {F}} + n _ {(I , s)} ^ {\operatorname {U n r} , \mathcal {S}} + n _ {(I , s)} ^ {\operatorname {U n r} , \mathcal {F}}} \tag {4}
+$$
+
+where  $n_{(I,s)}^{\mathrm{Rec},S}$  is the number of times that  $(I,s)$  has been correctly recognized by the user (subscript "Rec" for "recognized") in a successful dialogue  $S$ :
+
+$$
+n _ {(I, s)} ^ {\mathrm {R e c}, \mathcal {S}} = \sum_ {(a, \hat {a}) \in \mathcal {S}} \sum_ {(I ^ {\prime}, s ^ {\prime}, v ^ {\prime}) \in a \cap \hat {a}} [ (I, s) = (I ^ {\prime}, s ^ {\prime}) ]
+$$
+
+Similarly,  $n_{(I,s)}^{\mathrm{Unr},S}$  indicates the number of times in  $S$  that  $(I,s)$  was not recognized by the user (subscript "Unr" for "unrecognized"),  $\mathfrak{n}_{(I,s)}^{\mathrm{Rec},\mathcal{F}}$  indicates the number of times in  $\mathcal{F}$  that  $(I,s)$  was recognized by the user, and  $n_{(I,s)}^{\mathrm{Unr},\mathcal{F}}$  indicates the number of times in  $\mathcal{F}$  that  $(I,s)$  was not recognized by the user.
+
+Since  $c_{(I,s)}$  is the co-occurrence probability of  $(I,s)$  being recognized (or unrecognized) and the task being a success (or a failure), this quantifies how accurately conveying  $(I,s)$  leads to task success. Also,  $c_{(I,s)}$  is a value that is updated as the number of dialogue samples increases during training. That is,  $c_{(I,s)}$  can be computed adaptively to the current performance of the GenPPN at each learning step. Note that  $v$  is not taken into account for each count because the possible values of  $v$  are so large (e.g., instances of phone number, address, etc.) that counting them separately would yield unreliable statistics.
+
+Finally, the reward  $r(u_t^{\prime})$  for the utterance generated by GenPPN at turn  $t$  is calculated using this DA contribution  $c_{(I,s)}$  as follows:
+
+$$
+r \left(u _ {t} ^ {\prime}\right) = \text {A g g r e g a t e} \left(w \left(a _ {t}, \hat {a} _ {t}\right)\right) \tag {5}
+$$
+
+$$
+\begin{array}{l} w \left(a _ {t}, \hat {a} _ {t}\right) = \left\{\tau \cdot c _ {(I, s)} \mid (I, s, v) \in a _ {t} \cap \hat {a} _ {t} \right\} \tag {6} \\ \cup \left\{- \tau \cdot c _ {(I, s)} \mid (I, s, v) \in a _ {t} \cap \bar {\hat {a}} _ {t} \right\} \\ \end{array}
+$$
+
+Here, the DAs correctly recognized by the user  $(a_{t}\cap \hat{a}_{t})$  are given a default score  $\tau$ , and conversely, the DAs not recognized by the user  $(a\cap \bar{\hat{a}}_t)$  are given a negative score  $-\tau$ , and  $w(a_{t},\hat{a}_{t})$  is the set of these scores weighted by each DA contribution  $c_{(I,s)}$ . The constant  $\tau$  is a hyperparameter, and "Aggregate" is a function for aggregating  $w(a_{t},\hat{a}_{t})$  into a scalar value for the reward. We design the following two types of Aggregate functions and empirically determine which one is better:
+
+mean Output the mean of  $w(a_{t},\hat{a}_{t})$
+
+absmax Output the value with the highest absolute value among  $w(a_{t},\hat{a}_{t})$ . This is to emphasize the weighted score of  $(I,s,v)$  with the highest contribution.
+
+At the beginning of learning, the number of DAs recorded in  $\mathcal{D}$  is small, so it is expected that an appropriate  $c_{(I,s)}$  cannot be calculated. Therefore, prior to learning,  $(a,\hat{a})$  is sampled and recorded in  $\mathcal{D}$  by conducting multiple dialogues between the dialogue system and the user simulator without using GenPPN.
+
+# 3.4 Optimization via RL
+
+Following Stiennon et al. (2020), the RL algorithm used in this method is Proximal Policy Optimization (PPO) (Schulman et al., 2017). As a value network, a linear layer that outputs a scalar value randomly initialized with the parameter  $\phi$  is added, and the overall trainable parameters are set to  $\psi = [\theta ;\phi ]$ . Also, following (Ziegler et al., 2019), to prevent the probability distribution of  $\mathcal{M}_{\psi}$  from deviating too far from that of the original  $\mathcal{M}$  due to parameter updates and thus losing its naturalness, a Kullback-Leibler (KL) divergence penalty is added to  $r(u_t^{\prime})$  as the final reward  $R_{t}$  for utterance  $u_{t}^{\prime}$  at turn  $t$ :
+
+$$
+R _ {t} = r \left(u _ {t} ^ {\prime}\right) - \beta \log \frac {\mathcal {M} _ {\psi} \left(u _ {t} ^ {\prime} \mid x _ {t}\right)}{\mathcal {M} \left(u _ {t} ^ {\prime} \mid x _ {t}\right)} \tag {7}
+$$
+
+The clipped surrogate objective  $\mathcal{L}(\psi)$  (Schulman et al., 2017) is used to optimize  $\psi$  with the advantage estimated from the reward and value network. Since we need the user's subjective understanding results  $\hat{a}$ , this study uses a user simulator to optimize the GenPPN. The learning algorithm is summarized in Algorithm 1 in the appendix.
+
+# 4 Experiments
+
+In our experiments, we first evaluated the effectiveness of GenPPNs using a user simulator. Then, using the optimized GenPPN, we conducted a dialogue evaluation experiment using human subjects.
+
+# 4.1 Dataset and Platform
+
+We evaluated the effectiveness of our GenPPN using a dialogue system and a user simulator implemented on the basis of the MultiWOZ dataset (Budzianowski et al., 2018). MultiWOZ is a task-oriented dialogue dataset between a clerk and a customer at a travel information center, collected in Wizard-of-OZ style. It contains a variety of tasks across a total of seven domains (attraction, hotel, hospital, restaurant, taxi, train, and police).
+
+We used ConvLab-2 (Zhu et al., 2020), a platform for evaluating task-oriented dialogue systems
+
+that provide various modules for dialogue systems, a user simulator, and an evaluation tool. The following describes the dialogue system, user simulator, and user goals used in this experiment.
+
+Dialogue System To make it easier to assess changes in the performance of utterance generation, the other modules (that is, NLU, DST, and Policy) should have stable performance. Therefore, we used the best-performing BERT (Devlin et al., 2019)-based NLU (Chen et al., 2019), rule-based DST, and rule-based Policy available in ConvLab-2. BERT-based NLU is a model that uses representations embedded by BERT to classify user intentions in user utterances and extract slots by sequence labeling. Rule-based DST and Policy are modules implemented using hand-crafted rules. Three models were selected as NLG modules (see Section 4.2 for the NLG models used), and GenPPN was applied to each of them to verify its generality.
+
+User Simulator For the user simulator, we used a combination of BERT-based NLU, agenda-based Policy (Schatzmann et al., 2007), and template-based NLG. The agenda-based Policy models a user's behavior in MultiWOZ by using a stack-like agenda created using hand-crafted rules.
+
+User Goal A user goal for each dialogue is randomly generated; the domains are randomly selected from one to three domains (out of all seven domains). The slots are also randomly selected on the basis of the slots' frequency in MultiWOZ.
+
+# 4.2 NLG Baselines
+
+We applied GenPPN to each of the three NLGs available in ConvLab-2 with different architectures in order to demonstrate that it works for a variety of NLGs. In addition, one NLG optimized with only utterance-level rewards, without considering task success, was also evaluated for comparison.
+
+Template NLG An NLG model that uses the template utterances representing each DA. Because each utterance is carefully designed by hand, this model has significantly higher performance than other NLG baselines (Takanobu et al., 2020).
+
+SC-LSTM (Wen et al., 2015) An LSTM-based model with a reading gate mechanism. This model takes binary feature vectors representing DAs as context and decodes utterances.
+
+<table><tr><td rowspan="2">NLG</td><td rowspan="2">Task Success</td><td colspan="3">Inform</td><td rowspan="2">Book Rate</td><td rowspan="2">Turn ↓</td><td rowspan="2">DA F1</td></tr><tr><td>F1</td><td>Precision</td><td>Recall</td></tr><tr><td>Template NLG</td><td>77.25</td><td>78.44</td><td>74.19</td><td>89.13</td><td>83.91</td><td>7.67</td><td>71.73</td></tr><tr><td>+ GenPPNmean</td><td>77.93</td><td>79.75</td><td>75.62</td><td>89.41</td><td>84.33</td><td>7.63</td><td>76.98</td></tr><tr><td>+ GenPPNabsmax</td><td>78.91*</td><td>79.86</td><td>75.58</td><td>89.93</td><td>85.19</td><td>7.02</td><td>78.23</td></tr><tr><td>SC-LSTM</td><td>54.00</td><td>67.45</td><td>68.07</td><td>72.48</td><td>67.69</td><td>11.65</td><td>60.56</td></tr><tr><td>+ GenPPNmean</td><td>60.64**</td><td>75.40</td><td>75.42</td><td>81.01</td><td>78.74</td><td>9.42</td><td>79.80</td></tr><tr><td>+ GenPPNabsmax</td><td>72.95**</td><td>79.46</td><td>77.38</td><td>86.16</td><td>78.46</td><td>7.21</td><td>79.08</td></tr><tr><td>SC-GPT</td><td>64.94</td><td>78.06</td><td>73.60</td><td>88.51</td><td>56.94</td><td>7.80</td><td>71.53</td></tr><tr><td>+ GenPPNmean</td><td>73.63**</td><td>76.54</td><td>71.81</td><td>87.79</td><td>82.08</td><td>8.03</td><td>73.07</td></tr><tr><td>+ GenPPNabsmax</td><td>73.34**</td><td>77.34</td><td>72.29</td><td>89.03</td><td>80.79</td><td>7.50</td><td>73.72</td></tr><tr><td>GPT-2 + RL</td><td>72.36</td><td>76.70</td><td>73.50</td><td>85.99</td><td>76.81</td><td>7.47</td><td>81.17</td></tr><tr><td>+ GenPPNmean</td><td>74.02</td><td>77.10</td><td>73.87</td><td>86.82</td><td>79.19</td><td>7.54</td><td>80.98</td></tr><tr><td>+ GenPPNabsmax</td><td>75.20**</td><td>78.79</td><td>75.58</td><td>88.02</td><td>79.80</td><td>7.15</td><td>80.08</td></tr></table>
+
+Table 1: Performance of dialogue system with each NLG and with GenPPN applied to them. Subscripts in GenPPN indicate Aggregate function used in reward calculation. Highest score for each NLG is in bold. * and ** indicate Task Success for GenPPN and were significantly better than original NLGs at  $p < 0.05$  and  $< 0.01$ , respectively, in McNemar test with Bonferroni correction.
+
+SC-GPT (Peng et al., 2020) A GPT-2 based model that generates utterances from DA text sequences. This has been trained on seven task-oriented dialogue corpora including MultiWOZ and the Schema-Guided Dialogue Dataset (Rastogi et al., 2020), and it is a SOTA on the MultiWOZ NLG benchmark.
+
+# GPT-2 + RL (Ohashi and Higashinaka, 2022a)
+
+A GPT-2-based NLG model optimized with an utterance-level reward. This model was trained to maximize the accuracy of  $a$  and  $\hat{a}$  (as measured by F1) using the DA-system utterance pairs in the MultiWOZ corpus in an offline fashion.
+
+Note that, since the purpose of our experiment is to verify whether GenPPN can enhance the performance of NLG irrespective of its architecture or base performance, the verification of methods that fine-tune NLG models themselves is outside the scope of this study.
+
+# 4.3 Evaluation Metrics
+
+In the evaluation of the dialogue system, we used common metrics for task-oriented dialogues: Turn, Inform F1/Precision/Recall, Book Rate, Task Success. Turn indicates the number of turns required for each dialogue; the smaller it is, the more efficiently the dialogue can be conducted. Inform F1/Precision/Recall indicate whether the system responded appropriately to the user's requests. Book Rate indicates whether the entity booked by the system correctly matched that of the user's goal.
+
+Task Success is evaluated by calculating whether Inform Recall and Match Rate both became one within the maximum number of turns. To evaluate the utterance-level performance of NLG, we additionally used  $DA$ $F1$ , which measures the match rate of  $a$  and  $\hat{a}$  at each turn by F1 (Ohashi and Higashinaka, 2022a; Guo et al., 2023).
+
+# 4.4 Training GenPPN
+
+Stanford Alpaca 7B (Taori et al., 2023) was used as the instruction-tuned PLM for the GenPPN. This model is a fine-tuned version of LLaMA-7B (Touvron et al., 2023) with a 52K instruction dataset. The input prompts include the dialogue history, the previous DA, the system utterances generated by NLG from that DA, and the instructions for rewriting that system utterance (specific prompts are shown in Section A.2 in the appendix). Throughout the experiment, we first sampled 1K dialogues to initialize the DA history and then trained 200 iterations. See Section A.3 for further training details.
+
+For the evaluation, the GenPPN at the highest-reward iteration was used, and dialogue simulations were performed using 1,024 user goals prepared specifically for the test. We report the average scores in this paper.
+
+# 4.5 Main Results
+
+Table 1 shows the evaluation results<sup>2</sup>. For Template and SC-LSTM, GenPPN improved all evalua
+
+<table><tr><td>NLG</td><td>Success</td><td>Inf. F1</td><td>Book</td><td>DA F1</td></tr><tr><td>SC-LSTM + GenPPN</td><td>72.95</td><td>79.46</td><td>78.46</td><td>79.08</td></tr><tr><td>w/o adaptive c(I,s)</td><td>62.11</td><td>74.80</td><td>78.15</td><td>70.12</td></tr><tr><td>w/o e</td><td>59.47</td><td>77.02</td><td>71.06</td><td>76.00</td></tr><tr><td>w/o r(u&#x27;)</td><td>63.67</td><td>77.58</td><td>80.83</td><td>62.69</td></tr><tr><td>w/o context</td><td>68.46</td><td>76.25</td><td>74.17</td><td>80.28</td></tr><tr><td>w/o response</td><td>47.36</td><td>68.96</td><td>74.67</td><td>65.21</td></tr></table>
+
+tion metrics. The absmax aggregation function improved the final performance more than the mean. For SC-LSTM, Task Success was nearly 19 points better than the original SC-LSTM, especially when using absmax. These results indicate that a better strategy for evaluating each utterance is to prioritize learning to convey the DA with the highest contribution rather than to take the average. Considering that the improvement in Template NLG was only about  $1.7\%$ , the lower the performance of the original NLG, the more room there is for improvement by GenPPN.
+
+For SC-GPT, Task Success, Book Rate, and Turn steadily improved. However, Inform Precision slightly decreased, unlike other cases where GenPPN was applied to other NLG modules. This suggests that GenPPN learned that improving Book Rate is more important for improving Task Success, even at the expense of Inform Precision.
+
+GPT-2 + RL has a higher DA F1 than any of the GenPPNs. This is reasonable, given that GPT-2 + RL is optimized using DA F1 as a reward. GenPPN was able to further improve GPT-2 + RL's Task Success while the utterance-level metric DA F1 slightly decreased from the original score. This suggests that not only utterance-level rewards but dialogue-level rewards must be introduced to improve the task completion performance of the entire system.
+
+Note that the performance of SC-LSTM + GenPPN and SC-GPT + GenPPN is lower than that of Template NLG baseline. However, the primary goal of our study is not to obtain an NLG model with SOTA performance but rather to enhance the performance of NLG irrespective of its architecture or its base performance. In our experiments, the dialogue performance of NLG models like SC-LSTM and SC-GPT, in addition to Template NLG, were improved using GenPPN, showing that our primary goal has been achieved.
+
+Table 2: Ablation results when applying GenPPN to SC-LSTM. "absumax" Aggregate function is used in calculation of rewards using DA contributions.  
+
+<table><tr><td>NLG</td><td>N</td><td>Success</td><td>Turns</td><td>Und.</td><td>App.</td><td>Sat.</td></tr><tr><td>SC-LSTM</td><td>54</td><td>33.33</td><td>17.72</td><td>4.09</td><td>4.17</td><td>4.04</td></tr><tr><td>+ GenPPN</td><td>50</td><td>52.00*</td><td>15.08**</td><td>4.04</td><td>4.14</td><td>3.98</td></tr></table>
+
+Table 3: Human evaluation results.  $N$  is number of subjects who interacted with each system. "Und," "App," and "Sat." represent user's subjective evaluation of system's understanding, appropriateness of system utterance, and satisfaction with dialogue, respectively. Scores of SC-LSTM and our GenPPN were compared using the McNemar test; significant differences are indicated with  $* (p < 0.05)$  and  $\*\* (p < 0.01)$ .
+
+# 4.6 Ablation Study
+
+We analyzed the impact of each factor in the GenPPN optimization on the final performance. For the NLG model in this ablation study, we used SC-LSTM, which showed the greatest improvement in performance over the other NLGs in Table 1. Table 2 shows the results.
+
+Adaptive DA Contribution "w/o adaptive  $c_{(I,s)}$ " is a reward design that does not update the DA history, and in all steps, uses the DA contribution calculated only from the 1,000 dialogues sampled prior to training. This resulted in a 10 points decrease in Task Success. Therefore, we found that it is important to constantly update the DA contributions as the GenPPN changes.
+
+Dialogue-level Reward "w/o e" does not use DA contribution, and only an utterance-level reward based on the accuracy (F1) between  $a$  and  $\hat{a}$  is used. That is, the reward design does not take into account the dialogue-level reward  $e$  related to task completion. As a result, Task Success decreased significantly, while DA F1 did not decrease much. Therefore, it was shown that the DA contribution is a useful factor in improving the task success rate.
+
+Utterance-level Reward "w/o  $r(u')$ " does not use the utterance-level reward  $r(u')$  but only the last dialogue evaluation result as a reward. That is, all utterances were given the dialogue result  $e \in \{0,1\}$  equally. Although Book Rate improved, Inform F1 did not, indicating that it is difficult to optimize GenPPN with only a sparse  $e$  reward, as described in Section 3.3.
+
+Dialogue Context "w/o context" indicates that the input of GenPPN does not include the dialogue history. The slight decrease in Task Success, Inform F1, and Book Rate indicates that GenPPN should make use of the previous history to improve
+
+<table><tr><td>Context</td><td colspan="2">User: Hello, can you help me find a restaurant for my upcoming trip to Cambridge? Hmm, I &#x27;ll try Asian oriental food. I&#x27;m looking for an expensive restaurant. I also would like information on a place to eat in the centre.
+System: kymmoy serves asian oriental food.
+User: May I have the address? I would like their phone number.</td></tr><tr><td>System DA</td><td colspan="2">{ (Inform, Attraction-Addr, pool way, whitehill road), (Inform, Attraction-Phone, 01223902088), (Inform, Restaurant-Addr, 52 Mill Road City Centre), (Inform, Restaurant-Phone, 01223311911) }</td></tr><tr><td>System Response</td><td>SC-LSTM:
+Their address is</td><td>+ GenPPN:
+Their address is 52 Mill Road City Centre and phone number is 01223311911.</td></tr><tr><td>DA predicted by user</td><td>{}</td><td>{ (Inform, Restaurant-Addr, 52 Mill Road City Centre), (Inform, Restaurant-Phone, 01223311911) }</td></tr></table>
+
+Table 4: Response examples of systems using SC-LSTM and SC-LSTM + GenPPN in same context.
+
+dialogue performance. Meanwhile, the improvement in DA F1 suggests that dialogue history is not necessary if we only want to optimize at the utterance level.
+
+System Response "w/o response" does not rewrite the system utterance, but GenPPN generates it directly from the dialogue history and DA. It can be seen that the performance of all the methods significantly degraded. The reason for this is that the untrained GenPPN, i.e., Alpaca-7B, has never seen the mapping from DA to system utterance in MultiWOZ and thus has difficulty generating system utterances in a zero-shot setting. Therefore, it is considered more reasonable to learn to rewrite utterances rather than to generate them from scratch.
+
+# 4.7 Human Evaluation
+
+We tested whether GenPPN optimized by dialogue simulation is also effective for humans. In this experiment, we also used the system using SC-LSTM, which showed the best performance improvement over the other NLGs in Table 1. Over 50 crowd workers were recruited via Amazon Mechanical Turk, and each worker interacted once with one of the systems using SC-LSTM or SC-LSTM + GenPPN for up to 20 turns after being instructed about their user goal. Each worker had 20 turns to judge whether or not the task was successful; after 20 turns, the task was forced to fail. After the dialogue was completed, each worker was asked to rate the system's ability to understand the language (Und.), the accuracy of the system's responses (App.), and overall satisfaction with the dialogue (Sat.) on a 5-point Likert scale. See Section A.5 in the appendix for more details on the human evaluation settings.
+
+Table 3 shows the results of the evaluations of each system. GenPPN significantly improved in terms of both Task Success and Turn, which are
+
+measures of task completion. However, the user's subjective ratings of understanding, appropriateness, and dialogue satisfaction showed no improvement, which is reasonable given that GenPPN is optimized for task completion only. This also means that GenPPN did not generate unnatural utterances to improve Task Success and Turn, which is considered a positive result. We would like to evaluate the effectiveness of GenPPN for other NLG baselines than SC-LSTM by humans in the future.
+
+# 5 Case Study
+
+Table 4 compares the behavior of two systems using SC-LSTM and using SC-LSTM + GenPPN in a dialogue. As a context, the user requested the address and phone number of the restaurant, but the Policy module of the dialogue system mistakenly decided to answer with the address and phone number of an attraction as well. Here, SC-LSTM was not able to generate sentences because there was no example in its training data, in which DAs in different domains occur at the same time. As a result, SC-LSTM failed to convey any information to the user. In contrast, GenPPN reflected the address and phone number of the restaurant in the post-processed utterance, and as a result, the information was correctly conveyed as requested by the user, and the task was successful.
+
+It is worth noting that the utterance generated by post-processing does not contain any information about the attraction. This is probably because GenPPN judged from the dialogue context that the information about the attraction was not what the user requested, and that conveying this piece of information would interfere with the task completion. These results indicate that GenPPN can correct errors propagated from preceding modules beyond the realm of NLG so that the entire system's performance improves.
+
+# 6 Summary and Future Work
+
+Post-Processing Networks (PPNs), which modify the outputs of arbitrary modules including nondifferentiable ones in task-oriented dialogue systems, have been applied to NLU, DST, and Policy modules. In this paper, we proposed Generative Post-processing Networks (GenPPNs) for NLG modules. We optimized GenPPN toward the task completion performance of the entire system by using DA contribution-based rewards. Experiments on the MultiWOZ dataset and dialogue simulations confirmed that our GenPPN could improve the task success rate of various dialogue systems, regardless of the NLG architecture. It was also confirmed that a GenPPN optimized with a dialogue simulation was effective in a human evaluation experiment.
+
+In the future, we would like to integrate our GenPPN with the existing PPNs of NLU, DST, and Policy to realize the post-processing of all modules and further improve the systems performance. Since GenPPN is a generative model, we would like to extend it not only to text but also to the post-processing of speech recognition and speech synthesis modules in spoken dialogue systems.
+
+# Limitations
+
+Our work has several limitations, which we aim to address in our future work.
+
+First, this study implemented, trained, and evaluated the GenPPN using only the MultiWOZ dataset. Therefore, the applicability of the GenPPN to task-oriented dialogues other than MultiWOZ has not been verified. In particular, the DA-dependent reward design in this study needs to be improved for application to dialogues such as open-domain chats, where ontologies such as intent, slot, and value are not defined.
+
+Second, this study used a user simulator for training the GenPPN. When applied to a new domain, the cost of implementing a user simulator is likely to be high. Therefore, a user simulator-free learning method is needed in this respect. In addition, a GenPPN optimized by a user simulator learns user-simulator-specific utterances, which may not be the best utterances for humans. To learn the optimal GenPPN for humans, it is essential to design a reward model that takes humans into account, using a human-in-the-loop approach.
+
+Third, this study required running a PLM with a relatively large parameter size. During inference, in addition to the system's response generation, the
+
+GenPPN is required to rewrite utterances, which is expected to increase the operation cost and generation time. In addition, more computational resources are required during learning. Therefore, it is necessary to study ways to reduce learning and inference costs.
+
+# Ethics Statement
+
+We used the publicly available MultiWOZ dataset in full compliance with its terms of use. Our use of the LLaMA language model was also entirely consistent with the prescribed usage guidelines. No private or confidential data was used at any stage of the research. For our human evaluation experiment, we underwent and passed an ethical review from our institution and strictly adhered to its rules and guidelines. The anonymity, privacy, and rights of the subjects were preserved throughout. We acknowledge that there are potential ethical concerns associated with the use of large pre-trained language models, such as the risk of generating harmful or discriminatory statements.
+
+# Acknowledgments
+
+This work was supported by JSPS KAKENHI Grant Number 19H05692. We used the computational resources of the supercomputer "Flow" at the Information Technology Center, Nagoya University.
+
+# References
+
+Gabor Angeli, Percy Liang, and Dan Klein. 2010. A simple domain-independent probabilistic approach to generation. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 502-512.  
+Anusha Balakrishnan, Jinfeng Rao, Kartikeya Upasani, Michael White, and Rajen Subba. 2019. Constrained decoding for neural NLG from compositional representations in task-oriented dialogue. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 831-844.  
+Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Inigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gašić. 2018. MultiWOZ - A Large-Scale Multi-Domain Wizard-of-Oz Dataset for Task-Oriented Dialogue Modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5016-5026.  
+Qian Chen, Zhu Zhuo, and Wen Wang. 2019. BERT for Joint Intent Classification and Slot Filling. arXiv preprint arXiv:1902.10909.
+
+Zhi Chen, Lu Chen, Xiang Zhou, and Kai Yu. 2023. Deep reinforcement learning for on-line dialogue state tracking. In Proceedings of Man-Machine Speech Communication, pages 278-292.  
+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.  
+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, pages 4171-4186.  
+Ao Guo, Atsumoto Ohashi, Yuya Chiba, Yuiko Tsunomori, Ryu Hirai, and Ryuichiro Higashinaka. 2023. Personality-aware natural language generation for task-oriented dialogue using reinforcement learning. In Proceedings of the 32nd IEEE International Conference on Robot and Human Interactive Communication.  
+Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In Proceedings of International Conference on Learning Representations.  
+Vojtěch Hudeček and Ondrej Dusek. 2023. Are large language models all you need for task-oriented dialogue? In Proceedings of the 24th Meeting of the Special Interest Group on Discourse and Dialogue, pages 216-228, Prague, Czechia. Association for Computational Linguistics.  
+Diederik P Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980.  
+Hwaran Lee, Seokhwan Jo, Hyungjun Kim, Sangkeun Jung, and Tae-Yoon Kim. 2021. SUMBT+LaRL: Effective Multi-Domain End-to-End Neural Task-Oriented Dialog System. IEEE Access, pages 116133-116146.  
+Xiujun Li, Yun-Nung Chen, Lihong Li, Jianfeng Gao, and Asli Celikyilmaz. 2017. End-to-End TaskCompletion Neural Dialogue Systems. In Proceedings of the Eighth International Joint Conference on Natural Language Processing, pages 733-743.  
+Yangming Li, Kaisheng Yao, Libo Qin, Wanxiang Che, Xiaolong Li, and Ting Liu. 2020. Slot-consistent nlg for task-oriented dialogue systems with iterative rectification network. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 97-106.  
+Zichuan Lin, Jing Huang, Bowen Zhou, Xiaodong He, and Tengyu Ma. 2021. Joint System-Wise Optimization for Pipeline Goal-Oriented Dialog System. arXiv preprint arXiv:2106.04835.
+
+François Mairesse and Steve Young. 2014. Stochastic Language Generation in Dialogue using Factored Language Models. Computational Linguistics, pages 763-799.  
+Shikib Mehri, Tejas Srinivasan, and Maxine Eskenazi 2019. Structured Fusion Networks for Dialog. In Proceedings of the 20th Annual SIGdial Meeting on Discourse and Dialogue, pages 165-177.  
+Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. 2013. Playing Atari with Deep Reinforcement Learning. arXiv preprint arXiv:1312.5602.  
+Alice H Oh and Alexander I Rudnicky. 2002. Stochastic natural language generation for spoken dialog systems. Computer Speech & Language, pages 387-407.  
+Atsumoto Ohashi and Ryuichiro Higashinaka. 2022a Adaptive natural language generation for task-oriented dialogue via reinforcement learning". In Proceedings of the 29th International Conference on Computational Linguistics, pages 242-252.  
+Atsumoto Ohashi and Ryuichiro Higashinaka. 2022b Post-processing networks: Method for optimizing pipeline task-oriented dialogue systems using reinforcement learning. In Proceedings of the 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 1-13.  
+Wenbo Pan, Qiguang Chen, Xiao Xu, Wanxiang Che, and Libo Qin. 2023. A preliminary evaluation of chatgpt for zero-shot dialogue understanding. arXiv preprint arXiv:2304.04256.  
+Baolin Peng, Chenguang Zhu, Chunyuan Li, Xiujun Li, Jinchao Li, Michael Zeng, and Jianfeng Gao 2020. Few-shot Natural Language Generation for Task-Oriented Dialog. In Findings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 172-182.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan. Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Abhinav Rastogi, Xiaoxue Zang, Srinivas Sunkara, Raghav Gupta, and Pranav Khaitan. 2020. Towards Scalable Multi-Domain Conversational Agents: The Schema-Guided Dialogue Dataset. In Proceedings of the 34th AAAI Conference on Artificial Intelligence, pages 8689-8696.  
+Jost Schatzmann, Blaise Thomson, Karl Weilhammer, Hui Ye, and Steve Young. 2007. Agenda-Based User Simulation for Bootstrapping a POMDP Dialogue System. In Proceedings of Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics, pages 149-152.
+
+John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal Policy Optimization Algorithms. arXiv preprint arXiv:1707.06347.  
+Amanda Stent, Rashmi Prasad, and Marilyn Walker. 2004. Trainable Sentence Planning for Complex Information Presentations in Spoken Dialog Systems. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, pages 79-86.  
+Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. 2020. Learning to summarize with human feedback. In Proceedings of Advances in Neural Information Processing Systems, pages 3008-3021.  
+Shang-Yu Su, Kai-Ling Lo, Yi-Ting Yeh, and Yun-Nung Chen. 2018. Natural Language Generation by Hierarchical Decoding with Linguistic Patterns. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 61-66.  
+Ryuichi Takanobu, Qi Zhu, Jinchao Li, Baolin Peng, Jianfeng Gao, and Minlie Huang. 2020. Is Your Goal-Oriented Dialog Model Performing Really Well? Empirical Analysis of System-wise Evaluation. In Proceedings of the 21st Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 297-310.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford Alpaca: An Instruction-following LLaMA model. https://github.com/tatsu-lab/stanford_alpaca.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Van-Khanh Tran and Le-Minh Nguyen. 2017. Natural Language Generation for Spoken Dialogue System using RNN Encoder-Decoder Networks. In Proceedings of the 21st Conference on Computational Natural Language Learning, pages 442-451.  
+Bo-Hsiang Tseng, Yinpei Dai, Florian Kreyssig, and Bill Byrne. 2021. Transferable Dialogue Systems and User Simulators. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 152-166.  
+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 Proceedings of Advances in Neural Information Processing Systems, pages 5998-6008.  
+Marilyn A. Walker, Owen C. Rambow, and Monica Rogati. 2002. Training a sentence planner for spoken dialogue using boosting. Computer Speech & Language, 16(3):409-433.  
+Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V Le. 2022. Finetuned language models are zero-shot learners. In Proceedings of International Conference on Learning Representations.  
+Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić, Lina M. Rojas-Barahona, Pei-Hao Su, David Vandyke, and Steve Young. 2016. Multi-domain Neural Network Language Generation for Spoken Dialogue Systems. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 120-129.  
+Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić, PeiHao Su, David Vandyke, and Steve Young. 2015. Semantically Conditioned LSTM-based Natural Language Generation for Spoken Dialogue Systems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1711-1721.  
+Steve Young, Milica Gašić, Blaise Thomson, and Jason D. Williams. 2013. Pomdp-based statistical spoken dialog systems: A review. Proceedings of the IEEE, pages 1160-1179.  
+Zheng Zhang, Ryuichi Takanobu, Qi Zhu, MinLie Huang, and XiaoYan Zhu. 2020. Recent advances and challenges in task-oriented dialog systems. Science China Technological Sciences, pages 1-17.  
+Tiancheng Zhao and Maxine Eskenazi. 2016. Towards End-to-End Learning for Dialog State Tracking and Management using Deep Reinforcement Learning. In Proceedings of the 17th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 1-10.  
+Qi Zhu, Zheng Zhang, Yan Fang, Xiang Li, Ryuichi Takanobu, Jinchao Li, Baolin Peng, Jianfeng Gao, Xiaoyan Zhu, and Minlie Huang. 2020. ConvLab-2: An Open-Source Toolkit for Building, Evaluating, and Diagnosing Dialogue Systems. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 142-149.  
+Daniel M Ziegler, Nisan Stiannon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2019. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593.
+
+# A Appendix
+
+# A.1 GenPPN Optimization Algorithm
+
+# Algorithm 1 Optimization of GenPPN via PPO
+
+Require: Dialogue system  $\mathcal{A}$ , User simulator  $\mathcal{U}$
+
+Require: GenPPN  $\mathcal{M}$
+
+1: Initialize DA history  $\mathcal{D}$  by sampling several dialogues using  $\mathcal{A}$  and  $\mathcal{U}$  
+2: Prepare  $\mathcal{M}_{\psi_{\mathrm{old}}}$  with randomly initialized LoRA parameters  $\theta$  and value network parameters  $\phi$  
+3: for each training iteration do  
+4: while #turns does not reach batch size do  
+5: Sample a dialogue and  $(a_{t},\hat{a}_{t})$  for each turn  $t$  using  $\mathcal{A}$ ,  $\mathcal{M}_{\psi_{\mathrm{old}}}$ , and  $\mathcal{U}$  
+6: Obtain final evaluation result  $e$  
+7: Add  $(a_{t},\hat{a}_{t},e)$  of each  $t$  to  $\mathcal{D}$  
+8: end while  
+9: Calculate reward  $R_{t}$  using Eq. (7)  
+10: Compute advantage estimates  
+11: Optimize  $\mathcal{L}(\psi)$  with a certain epoch and mini-batch size  
+12: Update  $\psi_{\mathrm{old}}\gets \psi$  
+13: end for
+
+# A.2 Prompt for GenPPN
+
+At a given turn  $t$ , the input prompts to GenPPN include the DA  $a_{t}$  output by the system's Policy module and the original system utterance  $u_{t}$  generated by the NLG from  $a_{t}$ . As the dialogue context, we also included user utterances at  $t - 1$ , DAs  $a_{t - 1}$ , system utterances  $u_{t - 1}^{\prime}$  generated by GenPPN, and user utterances at  $t$ . The format of the instructional text in the prompts was adapted from the phrases in the Alpaca dataset (Taori et al., 2023). Examples of prompts are shown in Figure 3. In our experiments, due to the maximum input length of the model, the dialogue history was limited to the past one turn.
+
+# A.3 Details of GenPPN Training
+
+As the model for GenPPN, we used Stanford Alpaca-7B trained weights for the prompts, which are publicly available on HuggingFace Hub<sup>3</sup>. The maximum number of input tokens for a prompt was set to 512, and the maximum number of generated tokens was set to 128. When generating utterances, sampling was performed with the beam size set to 1, temperature to 1.0, and top- $p$  to 1.0. The  $\tau$  used in the reward calculation (Eq. (6)) was set to 1.
+
+<table><tr><td colspan="2">Hyperparameter Name</td><td>Value</td></tr><tr><td rowspan="3">LoRA</td><td>Target projection matrix of self-attention module</td><td>query, key value, output</td></tr><tr><td>Rank</td><td>16</td></tr><tr><td>Scaling factor α</td><td>16</td></tr><tr><td rowspan="10">PPO</td><td>Total iterations</td><td>200</td></tr><tr><td>Total batch size</td><td>512</td></tr><tr><td>Epoch</td><td>4</td></tr><tr><td>Total mini-batch size</td><td>32</td></tr><tr><td>Learning rate</td><td>1e-5</td></tr><tr><td>Optimizer</td><td>Adam</td></tr><tr><td>Discount factor γ</td><td>1.0</td></tr><tr><td>GAE factor λ</td><td>0.95</td></tr><tr><td>Clipping ε</td><td>0.2</td></tr><tr><td>Coef. of KL penalty β</td><td>0.01</td></tr></table>
+
+Table 5: Hyperparameter settings
+
+Table 5 shows the hyperparameters of LoRA and PPO during training. Throughout the experiment, we first sampled 1,000 dialogues to initialize the DA history and then trained 200 iterations, with a batch size of 512 turns per iteration (corresponding to about 50 dialogues) and a fixed learning rate of 1e-5 with Adam Optimizer (Kingma and Ba, 2014). We used  $16 \times \mathrm{V}100$  32-GB computational resources. Training of 200 iterations took about seven hours to complete.
+
+Figure 4 shows the training curves when GenPPN is applied to Template NLG, SC-LSTM, SC-GPT, and GPT-2 + RL respectively. The figure shows that both the reward and Task Success are low at the start of training (i.e., before RL is applied) and improve as RL progresses, which means that fine-tuning LLMs with RL is important for our GenPPN. We have yet to investigate the effects of instruction tuning. We plan to validate this by comparing the performance of instruction-tuned models other than Alpaca (e.g., Flan-T5 (Chung et al., 2022)) and non-instruction-tuned models (e.g., LLaMA (Touvron et al., 2023)).
+
+# A.4 Additional Ablations
+
+In addition to the study on SC-LSTM in Section 4.6, we conducted the same ablation study on two higher-performing NLGs: Template NLG and GPT-2 + RL. Table 6 shows the results. In Template NLG, the improvement with GenPPN was the greatest when no context was used. Nevertheless, for both Template NLG and GPT-2 w/ RL, there was a trend showing that adaptive DA contribution was important for performance improvement.
+
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+Instruction:
+
+Begin by reading a conversation between a customer and a chatbot about travel information.
+
+Customer: I am looking for a hotel call kirkwood house . I need a hotel as well .
+
+Chatbot Action: Inform-Hotel(Price=moderate; Area=north; Type=guesthouse; Parking=yes; Stars=4;
+
+Name=kirkwood house; Internet=yes)
+
+Chatbot: kirkwood house is a moderate 4 star guesthouse in the north.
+
+Customer: Do they have free parking?
+
+Chatbot Action: Inform-Hotel(Name=kirkwood house; Parking=yes)
+
+Chatbot: kirkwood house is available.
+
+Your task is to rephrase the chatbot's last utterance so the customer can understand it. Make sure to include the content of 'Chatbot Action:' in the utterance. If no rephrasing is necessary, repeat the original utterance.
+
+Response:
+
+Chatbot:
+
+Figure 3: Examples of prompts entered into GenPPN. "Customer:" "Chatbot Action:" and "Chatbot:" denote user utterance, system DA, and system utterance at each turn, respectively.  
+
+<table><tr><td>NLG</td><td>Success</td><td>Inf. F1</td><td>Book</td><td>DA F1</td></tr><tr><td>Template NLG + GenPPN</td><td>78.91</td><td>79.86</td><td>85.19</td><td>78.23</td></tr><tr><td>w/o adaptive c(I,s)</td><td>76.27</td><td>78.07</td><td>83.00</td><td>75.05</td></tr><tr><td>w/o e</td><td>77.15</td><td>79.29</td><td>83.30</td><td>76.89</td></tr><tr><td>w/o r(u&#x27;)</td><td>71.09</td><td>78.14</td><td>80.89</td><td>63.68</td></tr><tr><td>w/o context</td><td>79.00</td><td>79.92</td><td>85.29</td><td>78.58</td></tr></table>
+
+(a) Template NLG  
+
+<table><tr><td>NLG</td><td>Success</td><td>Inf. F1</td><td>Book</td><td>DA F1</td></tr><tr><td>GPT-2 + RL + GenPPN</td><td>75.20</td><td>78.79</td><td>79.80</td><td>80.08</td></tr><tr><td>w/o adaptive c(I,s)</td><td>67.68</td><td>76.75</td><td>79.31</td><td>81.04</td></tr><tr><td>w/o e</td><td>72.56</td><td>76.57</td><td>77.36</td><td>80.63</td></tr><tr><td>w/o r(u&#x27;)</td><td>72.85</td><td>78.52</td><td>81.82</td><td>71.15</td></tr><tr><td>w/o context</td><td>73.24</td><td>77.18</td><td>77.14</td><td>85.11</td></tr></table>
+
+(b) GPT-2 + RL
+
+Table 6: Ablation results when applying GenPPN to Template NLG and GPT-2 + RL. "absmax" Aggregate function is used in calculation of rewards using DA contributions. Bold and underlined scores indicate top and second top, respectively.
+
+# A.5 Details of Human Evaluation
+
+In collecting participants for the human evaluation via Amazon Mechanical Turk (AMT), we recruited workers who met the following four conditions: (1) had residence in an English-speaking county, (2) had at least 100 completed Human-Intelligence Tasks (HIT) on AMT, (3) had an acceptance rate of  $95\%$  or higher on the HIT on AMT, and (4) answered all five common sense questions correctly.
+
+In the experimental procedure, each worker was first instructed about a certain user goal. The user goal was randomly generated by a program as in the dialogue simulation. Considering the high comprehension ability of humans, the number of domains included in the user goal was set to 3, which is a higher difficulty level than in the simulation. After reading the user goal, the workers interacted with either the system using SC-LSTM or the system using SC-LSTM + GenPPN. The maximum number of turns in the dialogue was 20, and each worker judged whether the task was completed within the
+
+maximum number of turns. After 20 turns, the task was forced to fail. After the dialogue, the workers answered the three-question questionnaire described in Section 4.7.
+
+Considering that each HIT should take about 10 minutes to complete, the reward was set at \(2, and the time limit was 30 minutes. The number of subjects was not equal because the system was randomly selected each time, resulting in 54 evaluators for SC-LSTM and 50 evaluators for SC-LSTM + GenPPN.
+
+![](images/27af0ec18d20eb6d73896fddbc6d414a1ae68c4d69e05ead332f81107f605881.jpg)
+
+![](images/50006ed5294937be77af0f73e3a0b57b1cc6ccc7d7c85fbd08b53059e0492b83.jpg)
+
+![](images/a47f6b7f6783afc5438e465f61f723b5e9aca791218ece04252423abb0ff71c6.jpg)
+
+![](images/54c63c26aa98307511e90346d269f1d11b995146c176bc1918e2ce1dae22542b.jpg)
+
+![](images/20ab61e5d5ca004b0e5225dfb380b37d97cb370692d01b7f52e9edd76a74ff0d.jpg)  
+(a) Template NLG
+
+![](images/20434b9614c2093c1c2d08edc6954ff0d4648c543cf4b7faa58ef539b60ecd22.jpg)
+
+![](images/855f51576e139c899e67e22349a9274fdb353d37e9efbdd419b6550d1aff2a67.jpg)
+
+![](images/4d5fb91faccc3930e29d906c080447a33955b8c9bc765af40d4af1482ac14362.jpg)  
+(b) SC-LSTM
+
+![](images/da44c910a82f075c86ad1bbde95b6e28e8c038333073239df14d12eff685a4b7.jpg)
+
+![](images/9120430c02b13e6634091575c0c060b68072309f97ee268fd3015a8053d8bfbb.jpg)  
+Figure 4: Training curves when GenPPN is applied to each NLG. Either "mean" or "absmax" is used as aggregation function. "Reward" indicates value calculated by Eq. (5).
+
+![](images/55ef56697a173428c1815d8f9f598086cd7787bc02a7ceb82946d44b43d28811.jpg)  
+(c) SC-GPT  
+(d) GPT-2 + RL
+
+![](images/5e9d79fa491b9fcf9d309a80346c53f8ddce59ba00b4291ef9be9c0781252750.jpg)

2023/Enhancing Task-oriented Dialogue Systems with Generative Post-processing Networks/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7839286ef6fa5bd41a922e9ed5ba4af711bce2ae832916ce5015d261d70d69f9
+size 585042