Add Batch 01a2e52b-a81a-40cb-8ee5-e90b3194ec03 data

.gitattributes

@@ -3391,3 +3391,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/EmoTransKG_[[:space:]]An[[:space:]]Innovative[[:space:]]Emotion[[:space:]]Knowledge[[:space:]]Graph[[:space:]]to[[:space:]]Reveal[[:space:]]Emotion[[:space:]]Transformation/a4037023-fb45-4fb6-87e5-81a2977a4607_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/EmotionQueen_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]Empathy[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/04ee7f3f-b75b-4ed1-a8b2-eb274b97fde5_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/EmpathicStories++_[[:space:]]A[[:space:]]Multimodal[[:space:]]Dataset[[:space:]]for[[:space:]]Empathy[[:space:]]Towards[[:space:]]Personal[[:space:]]Experiences/c0642440-959f-4fa5-b71c-8cc638e695ae_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Empowering[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Textual[[:space:]]Data[[:space:]]Augmentation/81478ff5-f7a3-4b15-af38-b4922331bb0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Empowering[[:space:]]cross-lingual[[:space:]]abilities[[:space:]]of[[:space:]]instruction-tuned[[:space:]]large[[:space:]]language[[:space:]]models[[:space:]]by[[:space:]]translation-following[[:space:]]demonstrations/f7644dfc-cd06-4320-a14e-02a1e95a33d9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Encoding[[:space:]]Hierarchical[[:space:]]Schema[[:space:]]via[[:space:]]Concept[[:space:]]Flow[[:space:]]for[[:space:]]Multifaceted[[:space:]]Ideology[[:space:]]Detection/5e9641cb-740e-48d6-bdd3-e4eca45bddbc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/End-to-End[[:space:]]Emotion[[:space:]]Semantic[[:space:]]Parsing/4ffbaf63-28e0-4983-95b1-695face241d4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhanced[[:space:]]Language[[:space:]]Model[[:space:]]Truthfulness[[:space:]]with[[:space:]]Learnable[[:space:]]Intervention[[:space:]]and[[:space:]]Uncertainty[[:space:]]Expression/1e32c8cd-be45-4789-ad01-c4457a07e0bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhanced[[:space:]]Visual[[:space:]]Instruction[[:space:]]Tuning[[:space:]]with[[:space:]]Synthesized[[:space:]]Image-Dialogue[[:space:]]Data/5db7e691-801e-49cf-90b6-2abbaa2a383c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Adverse[[:space:]]Drug[[:space:]]Event[[:space:]]Detection[[:space:]]with[[:space:]]Multimodal[[:space:]]Dataset_[[:space:]]Corpus[[:space:]]Creation[[:space:]]and[[:space:]]Model[[:space:]]Development/5f2bd2d8-1c47-4ff0-bd6b-5d72a8900ae6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Cross[[:space:]]Text-Molecule[[:space:]]Learning[[:space:]]by[[:space:]]Self-Augmentation/206d3795-d53b-42c9-8de8-6d682eb0ac0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Distractor[[:space:]]Generation[[:space:]]for[[:space:]]Multiple-Choice[[:space:]]Questions[[:space:]]with[[:space:]]Retrieval[[:space:]]Augmented[[:space:]]Pretraining[[:space:]]and[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Integration/e2e79252-24e8-4303-bd7f-47125cbcbdea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Hallucination[[:space:]]Detection[[:space:]]through[[:space:]]Perturbation-Based[[:space:]]Synthetic[[:space:]]Data[[:space:]]Generation[[:space:]]in[[:space:]]System[[:space:]]Responses/230dbc31-6c35-4128-bdfd-69eaac066870_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Hyperbolic[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Embeddings[[:space:]]via[[:space:]]Lorentz[[:space:]]Transformations/f9607213-3e28-4ff3-b5fd-5d44688255d6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Idiomatic[[:space:]]Representation[[:space:]]in[[:space:]]Multiple[[:space:]]Languages[[:space:]]via[[:space:]]an[[:space:]]Adaptive[[:space:]]Contrastive[[:space:]]Triplet[[:space:]]Loss/c3eb5d3e-726c-4c24-b98b-e2ea581e6f42_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Semantic[[:space:]]Consistency[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Model[[:space:]]Editing_[[:space:]]An[[:space:]]Interpretability-Oriented[[:space:]]Approach/397db60d-0296-4988-b105-45201e5cbf89_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Sentence[[:space:]]Simplification[[:space:]]in[[:space:]]Portuguese_[[:space:]]Leveraging[[:space:]]Paraphrases,[[:space:]]Context,[[:space:]]and[[:space:]]Linguistic[[:space:]]Features/783c361f-fb4f-4a1c-8192-777b22269d97_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhancing[[:space:]]Text-to-SQL[[:space:]]Parsing[[:space:]]through[[:space:]]Question[[:space:]]Rewriting[[:space:]]and[[:space:]]Execution-Guided[[:space:]]Refinement/63662ef4-b176-43d7-b473-c8e7d75b1d0f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Episodic[[:space:]]Memory[[:space:]]Retrieval[[:space:]]from[[:space:]]LLMs_[[:space:]]A[[:space:]]Neuromorphic[[:space:]]Mechanism[[:space:]]to[[:space:]]Generate[[:space:]]Commonsense[[:space:]]Counterfactuals[[:space:]]for[[:space:]]Relation[[:space:]]Extraction/b2a107bb-3384-409c-9568-6de5592f8091_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Epistemology[[:space:]]of[[:space:]]Language[[:space:]]Models_[[:space:]]Do[[:space:]]Language[[:space:]]Models[[:space:]]Have[[:space:]]Holistic[[:space:]]Knowledge_/63e35d62-be10-4076-8a40-c8db5b7a15ad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Error[[:space:]]Analysis[[:space:]]Prompting[[:space:]]Enables[[:space:]]Human-Like[[:space:]]Translation[[:space:]]Evaluation[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/56585711-fb1e-4d9b-ac01-911dcd8eefbf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]LLMs’[[:space:]]Mathematical[[:space:]]Reasoning[[:space:]]in[[:space:]]Financial[[:space:]]Document[[:space:]]Question[[:space:]]Answering/6b41d5ac-9ed1-4fff-bfb8-6adcfb1c02e0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Biases[[:space:]]in[[:space:]]Persona-Steered[[:space:]]Generation/f0fe1485-df93-4135-bb68-4f764d48a454_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Health-related[[:space:]]Queries[[:space:]]with[[:space:]]Presuppositions/65ee9331-a0f0-4ef2-906e-807ba9d2cb24_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]on[[:space:]]Wikipedia-Style[[:space:]]Survey[[:space:]]Generation/d6843c47-d949-45d5-9aeb-02813ca69981_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]Mathematical[[:space:]]Reasoning[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]A[[:space:]]Focus[[:space:]]on[[:space:]]Error[[:space:]]Identification[[:space:]]and[[:space:]]Correction/a7264680-7855-4885-9910-e6d7f8fbf4b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]Robustness[[:space:]]of[[:space:]]Generative[[:space:]]Search[[:space:]]Engine[[:space:]]on[[:space:]]Adversarial[[:space:]]Factoid[[:space:]]Questions/0860db4c-4bf1-4597-a9de-6d6e9e63447a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]Structural[[:space:]]Generalization[[:space:]]in[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/d9fe2cd1-ece9-4132-b1ee-bf5fbe79ad62_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]the[[:space:]]Elementary[[:space:]]Multilingual[[:space:]]Capabilities[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]MultiQ/43fa1e3f-d092-407a-a17a-3e6a85781fa0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]the[[:space:]]Smooth[[:space:]]Control[[:space:]]of[[:space:]]Attribute[[:space:]]Intensity[[:space:]]in[[:space:]]Text[[:space:]]Generation[[:space:]]with[[:space:]]LLMs/02d779e4-2ada-4a7b-bb26-93a9f62bc8ea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evaluating[[:space:]]the[[:space:]]Validity[[:space:]]of[[:space:]]Word-level[[:space:]]Adversarial[[:space:]]Attacks[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/efc7349a-4864-465b-99ac-84c7e07e83f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Everything[[:space:]]of[[:space:]]Thoughts_[[:space:]]Defying[[:space:]]the[[:space:]]Law[[:space:]]of[[:space:]]Penrose[[:space:]]Triangle[[:space:]]for[[:space:]]Thought[[:space:]]Generation/e3195831-3fc2-473e-9453-2803da8e48fc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Evidence[[:space:]]Retrieval[[:space:]]is[[:space:]]almost[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]for[[:space:]]Fact[[:space:]]Verification/537628fc-ab9c-4658-80e4-f2bc535a40ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exciting[[:space:]]Mood[[:space:]]Changes_[[:space:]]A[[:space:]]Time-aware[[:space:]]Hierarchical[[:space:]]Transformer[[:space:]]for[[:space:]]Change[[:space:]]Detection[[:space:]]Modelling/529cdfb2-b839-48ef-9306-c1cb43b9590f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploiting[[:space:]]Positional[[:space:]]Bias[[:space:]]for[[:space:]]Query-Agnostic[[:space:]]Generative[[:space:]]Content[[:space:]]in[[:space:]]Search/703f2788-ddc8-4b7e-9131-e4cc55db3685_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploiting[[:space:]]Target[[:space:]]Language[[:space:]]Data[[:space:]]for[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation[[:space:]]Beyond[[:space:]]Back[[:space:]]Translation/8b9261e9-e149-4c43-af8b-7ca3482a3579_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploring[[:space:]]Defeasibility[[:space:]]in[[:space:]]Causal[[:space:]]Reasoning/ceb8fa8b-6abe-4cf1-ab67-64dea8ee5d1d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploring[[:space:]]Domain[[:space:]]Robust[[:space:]]Lightweight[[:space:]]Reward[[:space:]]Models[[:space:]]based[[:space:]]on[[:space:]]Router[[:space:]]Mechanism/04ece7ca-7d3b-43fb-ab08-e5dec02a2215_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploring[[:space:]]Mathematical[[:space:]]Extrapolation[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Synthetic[[:space:]]Data/13aac60b-b90e-45e1-ad9f-65ead04a5ef2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploring[[:space:]]Ordinality[[:space:]]in[[:space:]]Text[[:space:]]Classification_[[:space:]]A[[:space:]]Comparative[[:space:]]Study[[:space:]]of[[:space:]]Explicit[[:space:]]and[[:space:]]Implicit[[:space:]]Techniques/67c1ff30-92ca-4f5e-9567-e2e2e1d94ebf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploring[[:space:]]Reasoning[[:space:]]Biases[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Through[[:space:]]Syllogism_[[:space:]]Insights[[:space:]]from[[:space:]]the[[:space:]]NeuBAROCO[[:space:]]Dataset/c7954673-083e-4f27-ae6b-66436fb9a144_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploring[[:space:]]Reversal[[:space:]]Mathematical[[:space:]]Reasoning[[:space:]]Ability[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/246610bd-3940-4f09-947f-57f2f430f686_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploring[[:space:]]Spatial[[:space:]]Schema[[:space:]]Intuitions[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]and[[:space:]]Vision[[:space:]]Models/50a30c1e-5b1a-45f5-ba87-a3deee8df01b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Exploring[[:space:]]the[[:space:]]Potential[[:space:]]of[[:space:]]Dense[[:space:]]Information[[:space:]]in[[:space:]]Multimodal[[:space:]]Alignment/4a40a093-7c96-4d66-aab3-9ab31e13dff6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Extending[[:space:]]Context[[:space:]]Window[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Semantic[[:space:]]Compression/c42f6530-72ca-4b7f-b7d0-b917c0908860_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Extracting[[:space:]]Polymer[[:space:]]Nanocomposite[[:space:]]Samples[[:space:]]from[[:space:]]Full-Length[[:space:]]Documents/9f9840f2-3733-409c-a8bd-24caf9d267b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Extracting[[:space:]]and[[:space:]]Encoding_[[:space:]]Leveraging[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]and[[:space:]]Medical[[:space:]]Knowledge[[:space:]]to[[:space:]]Enhance[[:space:]]Radiological[[:space:]]Text[[:space:]]Representation/f883f859-1896-496e-8212-f391b7ca2e40_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FENICE_[[:space:]]Factuality[[:space:]]Evaluation[[:space:]]of[[:space:]]summarization[[:space:]]based[[:space:]]on[[:space:]]Natural[[:space:]]language[[:space:]]Inference[[:space:]]and[[:space:]]Claim[[:space:]]Extraction/d2c5a253-11f9-4328-b64a-1a94c4b72e58_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FOCUS_[[:space:]]Forging[[:space:]]Originality[[:space:]]through[[:space:]]Contrastive[[:space:]]Use[[:space:]]in[[:space:]]Self-Plagiarism[[:space:]]for[[:space:]]Language[[:space:]]Models/b7f4620f-8e31-4a90-9f84-240c14798a0d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FRVA_[[:space:]]Fact-Retrieval[[:space:]]and[[:space:]]Verification[[:space:]]Augmented[[:space:]]Entailment[[:space:]]Tree[[:space:]]Generation[[:space:]]for[[:space:]]Explainable[[:space:]]Question[[:space:]]Answering/dd35f37d-f5a7-425a-8cdd-00d1ace34843_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FUSE_[[:space:]]Measure-Theoretic[[:space:]]Compact[[:space:]]Fuzzy[[:space:]]Set[[:space:]]Representation[[:space:]]for[[:space:]]Taxonomy[[:space:]]Expansion/0fd7d117-40d1-4ffd-8696-517edc1ecf93_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fact-Checking[[:space:]]the[[:space:]]Output[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Token-Level[[:space:]]Uncertainty[[:space:]]Quantification/6288a5d3-64aa-4a76-9ca9-d1262dbf5a3d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fact-and-Reflection[[:space:]](FaR)[[:space:]]Improves[[:space:]]Confidence[[:space:]]Calibration[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/39dcad23-1550-4f26-a2ad-eeb74f47b230_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fair[[:space:]]Federated[[:space:]]Learning[[:space:]]with[[:space:]]Biased[[:space:]]Vision-Language[[:space:]]Models/dd744380-d2eb-495a-ae0c-c2d9e8dd5993_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Faithful[[:space:]]Persona-based[[:space:]]Conversational[[:space:]]Dataset[[:space:]]Generation[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/c2284ee5-1cae-4b1b-a99e-ae8d07fae4f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fantastic[[:space:]]Semantics[[:space:]]and[[:space:]]Where[[:space:]]to[[:space:]]Find[[:space:]]Them_[[:space:]]Investigating[[:space:]]Which[[:space:]]Layers[[:space:]]of[[:space:]]Generative[[:space:]]LLMs[[:space:]]Reflect[[:space:]]Lexical[[:space:]]Semantics/4a3fd6b8-b99c-4a6a-86a3-bca851b7f0a5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fast[[:space:]]Randomized[[:space:]]Low-Rank[[:space:]]Adaptation[[:space:]]of[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]PAC[[:space:]]Regularization/58416902-870a-4893-a4bd-34e641faa481_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FastGAS_[[:space:]]Fast[[:space:]]Graph-based[[:space:]]Annotation[[:space:]]Selection[[:space:]]for[[:space:]]In-Context[[:space:]]Learning/45075499-0df8-43cf-80d2-4de3c1994942_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Few-shot[[:space:]]Dialogue[[:space:]]Strategy[[:space:]]Learning[[:space:]]for[[:space:]]Motivational[[:space:]]Interviewing[[:space:]]via[[:space:]]Inductive[[:space:]]Reasoning/bd887f71-595b-4c48-801d-bd97a8bc9c2b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Figuratively[[:space:]]Speaking_[[:space:]]Authorship[[:space:]]Attribution[[:space:]]via[[:space:]]Multi-Task[[:space:]]Figurative[[:space:]]Language[[:space:]]Modeling/3be12eee-dc6e-41c3-9b4b-add07b8e3949_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FinTral_[[:space:]]A[[:space:]]Family[[:space:]]of[[:space:]]GPT-4[[:space:]]Level[[:space:]]Multimodal[[:space:]]Financial[[:space:]]Large[[:space:]]Language[[:space:]]Models/6eb84294-8d70-4dcd-8ea9-cce871ce9737_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Finding[[:space:]]and[[:space:]]Editing[[:space:]]Multi-Modal[[:space:]]Neurons[[:space:]]in[[:space:]]Pre-Trained[[:space:]]Transformers/5a630d08-038e-4e24-9afd-ae3a0cbeb0d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Findings[[:space:]]of[[:space:]]the[[:space:]]Association[[:space:]]for[[:space:]]Computational[[:space:]]Linguistics_[[:space:]]ACL[[:space:]]2024/e63d2d7d-f907-4861-bdb5-9124c4675e16_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fine-tuning[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Joint[[:space:]]Rewriting[[:space:]]and[[:space:]]Completion[[:space:]]of[[:space:]]Code[[:space:]]with[[:space:]]Potential[[:space:]]Bugs/bf2ed97a-c8b5-4cae-8c7d-d343714921af_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Fine-tuning[[:space:]]with[[:space:]]HED-IT_[[:space:]]The[[:space:]]impact[[:space:]]of[[:space:]]human[[:space:]]post-editing[[:space:]]for[[:space:]]dialogical[[:space:]]language[[:space:]]models/65024fa8-5479-4344-ba3d-9f19ef737348_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Flexible[[:space:]]Weight[[:space:]]Tuning[[:space:]]and[[:space:]]Weight[[:space:]]Fusion[[:space:]]Strategies[[:space:]]for[[:space:]]Continual[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition/7876569e-3c69-40a0-ae01-739fa207733b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/FlowVQA_[[:space:]]Mapping[[:space:]]Multimodal[[:space:]]Logic[[:space:]]in[[:space:]]Visual[[:space:]]Question[[:space:]]Answering[[:space:]]with[[:space:]]Flowcharts/8f1fbd29-5974-47de-834a-7c9ad846fcf3_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/Empowering Large Language Models for Textual Data Augmentation/81478ff5-f7a3-4b15-af38-b4922331bb0b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:481312db3193b11cb91d97c2bfeafa1d5c1f5e43c38498423943ed162592216c
+size 894758

2024/Empowering Large Language Models for Textual Data Augmentation/full.md

@@ -0,0 +1,414 @@
+# Empowering Large Language Models for Textual Data Augmentation
+
+Yichuan Li $^{1*}$ , Kaize Ding $^{2*}$ , Jianling Wang $^{3}$ , Kyumin Lee $^{1}$
+
+$^{1}$ Worcester Polytechnic Institute,  $^{2}$ Northwestern University  $^{3}$ Google DeepMind {yli29, kmlee}  $@$  wpi.edu, kaize.ding@northwestern.edu, jianlingw@google.com
+
+# Abstract
+
+With the capabilities of understanding and executing natural language instructions, Large language models (LLMs) can potentially act as a powerful tool for textual data augmentation. However, the quality of augmented data depends heavily on the augmentation instructions provided, and the effectiveness can fluctuate across different downstream tasks. While manually crafting and selecting instructions can offer some improvement, this approach faces scalability and consistency issues in practice due to the diversity of downstream tasks. In this work, we address these limitations by proposing a new solution, which can automatically generate a large pool of augmentation instructions and select the most suitable task-informed instructions, thereby empowering LLMs to create high-quality augmented data for different downstream tasks. Empirically, the proposed approach consistently generates augmented data with better quality compared to non-LLM and LLM-based data augmentation methods, leading to the best performance on 26 few-shot learning tasks sourced from a wide range of application domains.
+
+# 1 Introduction
+
+Large language models (LLMs) have recently demonstrated their potential in performing data augmentation on text data (Dai et al., 2023; Chung et al., 2023; Yu et al., 2023; Yoo et al., 2021). Serving as a semantic-preserving transformation function, LLMs transform original texts based on instructions to create diverse and informative data augmentations. With the augmented data, users can further train a spreadable and affordable model (e.g. OPT (Zhang et al., 2022)) to perform specific tasks. Unlike traditional heuristic-based methods such as word swapping (Wei and Zou, 2019) and model-based methods like back-translation (Fadaee et al.,
+
+![](images/0855bc8e68983859a4b93b87b261e6b7dfbef88f32a6170e7bbc2b92652059a7.jpg)  
+Figure 1: A simple demo of pronouns replacement augmentation instruction on text entailment task: GLUE-MRPC (Wang et al., 2019) and question answering task: OpenBookQA (Mihaylov et al., 2018).
+
+2017), LLMs offer great potential to produce more fluent, diverse, and semantically consistent augmentations for text data, owing to their great understanding and generalization capabilities.
+
+Despite the early success of LLMs for textual data augmentation, existing methods (Dai et al., 2023) that simply prompt LLMs with human-crafted augmentation instructions (i.e., Manual-LLMDA methods) have the following major bottlenecks: (1) Firstly, their efficacy heavily relies on the quality of the augmentation instructions, which are manually engineered by domain experts. This manual process is not only domain knowledge-intensive but also prone to inconsistencies, potentially compromising the quality of augmented data. Subtle variations in how these instructions are formulated can significantly influence the outcomes, as demonstrated by recent studies (Ishibashi et al., 2023; Zhu et al., 2023); (2) Secondly, usually text augmentation instructions are written in a task-agnostic form for a general purpose, however, the lack of context information on downstream tasks could lead to dramatic performance disparity on different downstream tasks, as shown in Fig. 1. Without considering the specific properties of the target tasks, LLM may generate low-quality augmented data (Ribeiro et al., 2020; Wei and Zou, 2019).
+
+To address the aforementioned challenges, in this paper, we introduce a new framework - Self-LLMDA that automates augmentation instruction generation and selection, facilitating LLM to generate task-specific augmented data. The initial phase of Self-LLMDA aims to broaden the span of seed augmentation strategies through the generation of diverse and effective instructions based on LLMs. Following this, Self-LLMDA employs a scoring model to identify and select the most relevant instructions that are likely to bolster the performance of target models. Such a new textual data augmentation approach ensures a balance between the generative breadth of augmentation instructions and targeted precision of task-specific guidance for downstream tasks.
+
+In our study, we conduct extensive experiments across a large collection of few-shot learning tasks used in previous studies (Min et al., 2022; Ye et al., 2021; Khashabi et al., 2020). This collection includes 26 different types of tasks across hate speech detection, question answering, natural language inference, and phrase detection datasets. Our study stands out for its extensive coverage of tasks, setting a new benchmark in the application of LLMs for textual data augmentation when compared to previous work (Dai et al., 2023; Li et al., 2023; Chung et al., 2023). The empirical results demonstrate that the proposed approach Self-LLMDA significantly outperforms various baseline methods in generating high-quality augmented textual data. To summarize, our main contributions are as follows:
+
+- We introduce a framework Self-LLMDA, which automates the generation and selection of task-specific augmentation instructions for LLMs, providing effective data augmentation for text data.  
+- Through a comprehensive set of experiments, we validate the effectiveness of Self-LLMDA, demonstrating its superior performance in enhancing data quality and model accuracy over existing text data augmentation methods.  
+- Our in-depth analyses reveal that Self-LLMDA can well generalize across various target models and previously unseen augmentation instructions, demonstrating its versatility and potential for broad applicability.
+
+# 2 Related Work
+
+# 2.1 Non-LLM Textual Data Augmentation
+
+Conventional textual data augmentation methods encompass a variety of techniques aimed at enhanc
+
+ing the diversity of textual datasets without relying on large language models (i.e., Non-LLMDA methods). Those methods range from simple heuristic-based methods to generative model-based methods. For heuristic-based approaches, such as synonym replacement (Zhang et al., 2016) and word shuffling, stand out for their computational efficiency and simplicity, making them ideal for large-scale data augmentation with minimal computational demands. Another notable example is the Easy Data Augmentation (EDA) technique introduced by Wei and Zou (2019), which employs token-level perturbations—random insertion, deletion, and swapping—to improve performance across a spectrum of text classification tasks.
+
+For model-based approaches, researchers have employed seq2seq and language models for data augmentation. Back-translation (Fadaee et al., 2017) employs translation models to preserve semantic integrity while generating paraphrases (Fadaee et al., 2017). Conditional masked language models like BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019) can also be utilized for data augmentation (Cheng et al., 2022; Wu et al., 2018). By masking words within sentences and subsequently generating replacements, these models introduce linguistic variations. Furthermore, other methods (Kumar et al., 2021; Edwards et al., 2023) leverage the capabilities of generative language models like GPT-2 (Radford et al., 2019) and BART (Lewis et al., 2019) for data augmentation. These approaches perform conditional generation based on class labels. Additionally, some studies have explored augmentation in the feature space. Mixup techniques interpolate within word or sentence embeddings (Guo et al., 2019), while others introduce random multiplicative and additive noise to the feature vectors (Kurata et al., 2016). Despite their utility, these conventional Non-LLMDA methods often come with limitations in readability and contextual consistency.
+
+# 2.2 LLM-based Textual Data Augmentation
+
+Recent advancements in LLMs have demonstrated their superiority in generating high quality and contextually relevant augmented data (Brown et al., 2020). LLMs are increasingly employed as label-preserving transformation functions, where an original example is transformed or perturbed according to manually crafted instructions (Dai et al., 2023; Yoo et al., 2021; Piedboeuf and Langlais, 2023).
+
+![](images/75782e59e34536283b8cad876fb78fe96488dcc786beca0c3f263c23dacb1c82.jpg)  
+Figure 2: The pipeline of Self-LLMDA. We first prompt the LLM to generate a diverse set of candidate augmentation instructions (§ 4.1). Then we select the instruction (§ 4.2) and apply it with the task data to LLM to get augmentations.
+
+![](images/8b20df25412b876dcbd127becab21fe77aa86463fdb5fa42b41d71f9dcfcd713.jpg)
+
+Concurrently, several studies (Chung et al., 2023; Yu et al., 2023; Li et al., 2023; Ubani et al., 2023; Meng et al., 2022) have explored the generation of conceptually similar yet semantically distinct synthetic examples. These methods, however, mostly rely on manual instruction design. In contrast, our work automatically generates label-preserving augmentation instructions by prompting LLMs, thus reducing dependency on manually crafted instructions. Furthermore, we introduce an instruction selection model that chooses appropriate instructions for arbitrary downstream tasks.
+
+# 3 Preliminary
+
+Problem Definition. Textual data augmentation involves applying a label-preserving transformation function  $T(\cdot)$  to a dataset  $\mathcal{D} = \{(\mathbf{x}_i, \mathbf{y}_i)\}_{i=1}^k$ , where each example consists of an input text  $\mathbf{x}_i$  (a sequence of tokens) and a corresponding label  $\mathbf{y}_i$  (also a sequence of tokens). The augmented dataset  $\mathcal{D}'$  is generated as follows, ensuring that the output label  $\mathbf{y}_i'$  remains unchanged:
+
+$$
+\mathbf {x} _ {i} ^ {\prime} = T \left(\mathbf {x} _ {i}\right), \mathbf {y} _ {i} ^ {\prime} = \mathbf {y} _ {i}. \tag {1}
+$$
+
+A target model  $F$  is then trained on the union of the original and augmented datasets,  $\mathcal{D} \cup \mathcal{D}'$ , with the training objective defined as:
+
+$$
+\mathcal {L} _ {(\hat {\mathbf {x}} _ {i}, \hat {\mathbf {y}} _ {i}) \in \mathcal {D} \cup \mathcal {D} ^ {\prime}} \left(F _ {\theta} (\hat {\mathbf {x}} _ {i}), \hat {\mathbf {y}} _ {i}\right). \tag {2}
+$$
+
+Therefore, designing an effective transformation function  $T(\cdot)$  that produces high-quality augmented data  $\mathcal{D}'$  is crucial for improving the downstream performance of model  $F_{\theta}$ .
+
+Manual-LLMDA. For Manual-LLMDA methods, the transformation function  $T(\cdot)$  is realized through a combination of an LLM and a manual-crafted instruction  $\mathbf{I}_{\mathrm{man}}$  (e.g., paraphrasing). The LLM is prompted to generate semantic-preserving transformations of the input text  $\mathbf{x}_i$  for the augmented
+
+dataset  $\mathcal{D}^{\prime}$
+
+$$
+\mathbf {x} _ {i} ^ {\prime} = \operatorname {L L M} \left(\mathbf {I} _ {\text {m a n}}, \mathbf {x} _ {i}\right), \mathbf {y} _ {i} ^ {\prime} = \mathbf {y} _ {i} \tag {3}
+$$
+
+# 4 Proposed Approach - Self-LLMDA
+
+To reduce the human efforts in designing augmentation instructions and selecting a task-specific instruction for a given task, we propose Self-LLMDA depicted in Fig. 2. The process begins with the LLM generating a diverse set of potential instructions  $\mathcal{I} = \{\mathbf{I}_j\}_{j=0}^n$  from a given set of seed instructions  $\mathcal{I}_{\mathrm{seed}} = \{\mathbf{I}_{\mathrm{man}}\}$ :
+
+$$
+\mathcal {I} = \operatorname {L L M} \left(\mathcal {I} _ {\text {s e e d}}\right). \tag {4}
+$$
+
+A selection model  $S$  then scores these generated instructions against the dataset  $\mathcal{D}$  to identify the most suitable instruction  $\mathbf{I}^{*}$ :
+
+$$
+\mathbf {I} ^ {*} = S (\mathcal {I}, \mathcal {D}). \tag {5}
+$$
+
+Based on the selected instruction  $\mathbf{I}^*$ , the LLM performs data augmentation on  $\mathcal{D}$ , producing an enhanced augmented dataset  $\mathcal{D}'$  for training the target model more effectively.
+
+# 4.1 Augmentation Instruction Self-Generation
+
+Inspired by the self-instruct methodology (Wang et al., 2022), this phase generates augmentation instructions from a seed set of 13 human-crafted instructions. These seed instructions act as exemplars, guiding the LLMs toward the creation of novel and diverse instructions that maintain the semantic integrity of the input text. To generate a broad and diverse set of augmentation instructions without the bias introduced by a few task examples, we exclude the task-specific data from the instruction generation. This will leverage the zero-shot learning capabilities of LLMs to produce a wide array of potential augmentation instructions. We use the following prompt to encourage LLMs to explore various augmentation techniques:
+
+![](images/8567d38d97489052241266f5e106ff1fd778b2e9ef626bf3bd0a785d46352829.jpg)  
+Figure 3: Illustration of the Instruction selection scoring model.  $F_{\theta}$  is the target model.
+
+"Come up with a series of textual data augmentation methods and you need to generate more diverse data augmentation method that can keep the semantic meaning of the input sentence.  $\{\mathbf{I}_{\mathrm{seed}}\}$ "
+
+Through iterative cycles of generation and refinement, we filter out instructions that are too similar to existing ones based on ROUGE-L (Lin, 2004). The unique generated instructions from each iteration are then incorporated back into the seed instruction pool, enriching the seed instructions for subsequent generation rounds. This process is repeated until we reach a collection of 100 augmentation instructions. To ensure diversity and eliminate redundancy, we further refine this set by removing duplicates based on their method names. This filtration results in a final set of 51 unique augmentation instructions.
+
+# 4.2 Task-Informed Instruction Selection
+
+Recognizing that augmentation instructions may not be universally applicable across different tasks, we implement a selection mechanism, tailored to the specific requirements of each task and its corresponding target model. This process involves a scoring model  $S$  to evaluate the suitability of each instruction for the task at hand. The scoring model  $S$ , as shown in Fig. 3, outputs a ranking score  $q_{j}$  indicating the instruction's effectiveness based on the pair of instruction and task dataset. Based on the notable instruction-following capabilities of FLAN-T5 (Chung et al., 2022; Raffel et al., 2023), we choose FLAN-T5-Large (Chung et al., 2022; Raffel et al., 2023) as the backbone of our scoring model. The input for scoring model  $S$  is:
+
+"Given the dataset for task  $\mathcal{T}$  and the instruction data, determine if this is a suitable instruction to address the task for model  $F$ . Task Dataset:  $\{\mathbf{x}_i\}_{i=0}^m$  Instruction:  $\mathbf{I}_j$ . Is this instruction appropriate?"
+
+where  $\mathcal{T}$  is the task name (e.g. GLUE-RTE),  $F$  is the target model name (e.g. OPT-125m). Since most of the tasks did not have a task description and manually designing the task description is time consuming, we utilize the few-shot examples  $\{\mathbf{x}_i\}_{i=0}^m$  from the dataset as the task description. Here, we calculate  $q_j$  by assessing the logit value of the "yes" token of the last position of input from FLAN-T5-Large, as shown in Fig. 3. Next, we will introduce the optimization and inference procedure of the scoring model  $S$  respectively.
+
+Model Optimization. The instruction selection model is trained to prioritize generated augmentation instructions based on their impact on downstream task performance. Its goal is to assign the highest scores to instructions that lead to the most effective data augmentation. To enhance scalability and computational efficiency, our model optimizes the selection process for a given task  $\mathcal{D}$  by sampling a subset of augmentation instructions  $\{\mathbf{I}_j\}_{j=0}^n$  (where  $n > 1$ ) from the pool of candidates. The model then computes scores  $\{q_j\}_{j=0}^n$  representing the relative effectiveness of each instruction. The optimization objective is formulated as a cross-entropy loss, designed to accurately distinguish between the effectiveness of these instructions  $\{\mathbf{I}_j\}_{j=0}^n$ . The loss function is given by:
+
+$$
+\mathcal {L} _ {S} = - \sum_ {j = 0} ^ {n} \operatorname {i s} _ {-} \max  (r _ {j}) \log \sigma (q _ {j}) \tag {6}
+$$
+
+Here, is_max serves as a binary indicator function that identifies the instruction yielding the maximum effectiveness, and  $\sigma$  is softmax that normalizes the  $q_{j}$  (a probability generating "yes" token interpreted as a ranking score associated with  $j$ th instruction) over the sampled augmentation instructions,  $r_{j}$  is the downstream task performance of a trained target model on augmented data  $\mathcal{D}_j^\prime \cup \mathcal{D}$ .
+
+Model Inference. When encountering a new task, the selection model  $S$  evaluates all potential instructions to determine the most suitable one  $\mathbf{I}^*$ , denoted by the highest score:
+
+$$
+\mathbf {I} ^ {*} = \mathbf {I} _ {\operatorname {a r g m a x} \left(\left\{q _ {j} \right\} _ {j = 0} ^ {| \tau |}\right)} \tag {7}
+$$
+
+This optimal instruction,  $\mathbf{I}^*$ , is then employed to prompt the LLMs to generate augmented data. This selection mechanism ensures the use of the most effective instruction for enhancing data utility across diverse NLP tasks.
+
+# 5 Experiment
+
+# 5.1 Experimental Setup
+
+Evaluation Datasets. In this study, we select 26 few-shot learning tasks spanning a wide range of NLP challenges, sourced from CrossFit (Ye et al., 2021), UnifiedQA (Khashabi et al., 2020), and MetaICL (Min et al., 2022). These datasets were chosen for their diversity, encompassing both classification tasks (Class)—such as natural language inference, paraphrase detection, and hate speech identification—and non-classification (Non-Class) tasks, notably question answering, to ensure a broad evaluation spectrum. The selection of tasks is significantly larger and more diverse than that in other relevant works (Dai et al., 2023; Chung et al., 2023).
+
+To investigate the generalization ability of Self-LLMDA, we split the 26 tasks into training and test tasks as for form "train  $\rightarrow$  test". We train the augmentation instruction selection methods on training tasks and evaluated it on test tasks. The task split involves four settings: Class  $\rightarrow$  Class, Class  $\rightarrow$  Non-Class, Non-Class  $\rightarrow$  Class, and Random  $\rightarrow$  Random, where "Random" represents a mixture of randomly selected tasks *. This design allows us to investigate the performance of selection models when applied across similar and disparate task types, providing insights into their generalizability and effectiveness.
+
+Evaluation Metrics. To handle all types of tasks simultaneously, we unify all downstream tasks, including classification and non-classification tasks, using a text-to-text approach (Raffel et al., 2023). For each task, we feed the input text to the target model  $F_{\theta}$  and train it to generate the corresponding target text. We choose OPT (Zhang et al., 2022) from three different sizes (e.g. 125m, 350m and  $1.3\mathrm{b}^{*}$ ) as our target models  $F_{\theta}$ . During training*,  $F_{\theta}$  takes the training example  $\mathbf{x}_i$  as the input, and is optimised to
+
+generate  $\mathbf{y}_i$  using the negative likelihood objective function:
+
+$$
+\mathcal {L} _ {F _ {\theta}} \left(\mathbf {y} _ {i}\right) = - \sum_ {t = 1} ^ {| \mathbf {y} _ {i} |} \log P _ {F _ {\theta}} \left(y _ {i} ^ {t} \mid \mathbf {x} _ {i}, \mathbf {y} _ {i} ^ {<   t}\right) \tag {8}
+$$
+
+During inference time, given the test input  $\mathbf{x}_{test}$  as well as a set of candidates  $\mathcal{C}$ , which is either a set of labels (in classification tasks) or answer options (in non-classification tasks), the  $F_{\theta}$  computes the conditional probability of each label  $\mathbf{c} \in \mathcal{C}$ , where  $\mathbf{c}$  is a sequence of tokens. The label with the maximum conditional probability is returned as a prediction:
+
+$$
+\operatorname {a r g m a x} _ {\mathbf {c} \in \mathcal {C}} \left(\sum_ {t = 1} ^ {| \mathbf {c} |} \log P _ {F _ {\theta}} \left(c ^ {t} \mid \mathbf {x} _ {\text {t e x t}}, \mathbf {c} ^ {<   t}\right)\right) \tag {9}
+$$
+
+Specifically, we use macro-F1 for classification tasks, and accuracy for non-classification tasks in our experiment. The overall performance is then quantified by computing the macro-average of these scores across all tasks, encapsulating both accuracy and macro-F1 metrics. To ensure robustness and reduce sampling bias, each experiment under each splitting setting is replicated with five different random seeds. For each few-shot task, we adopt a uniform approach by randomly selecting  $k = 16$  training examples. Following (Min et al., 2022), we did not make perfect label balance between  $k$  training examples.
+
+Baseline Methods. In this study, we compare our novel augmentation pipeline, Self-LLMDA, with two different categories of data augmentation methods as baselines: Non-LLMDA and Manual-LLMDA. For both Manual-LLMDA and Self-LLMDA, we employ GPT-3.5 Turbo as the backbone LLM. For detailed descriptions of these baseline methods, please see Appendix E. Specifically:
+
+- Non-LLMDA methods. This category includes 13 traditional augmentation techniques: Character-Level: Operations such as random swaps, OCR Errors simulation, deletions, insertions, and substitutions. Word-Level: Transformations, including word swaps, deletions, spelling errors, and embedding-based insertions. Contextual-Level: Utilization of language models for word insertions (e.g., using GPT2 (Brown et al., 2020)) and substitutions (e.g., with BERT (Devlin et al., 2018)), and back-translation (Fadaee et al., 2017).
+
+<table><tr><td>TextulDA</td><td colspan="3">Class → Class</td><td colspan="3">Class → Non-Class</td><td colspan="3">Non-Class → Class</td><td colspan="3">Random → Random</td></tr><tr><td></td><td>125m</td><td>350m</td><td>1.3b</td><td>125m</td><td>350m</td><td>1.3b</td><td>125m</td><td>350m</td><td>1.3b</td><td>125m</td><td>350m</td><td>1.3b</td></tr><tr><td>Original</td><td>44.80</td><td>41.94</td><td>42.82</td><td>38.49</td><td>42.04</td><td>42.42</td><td>46.49</td><td>44.13</td><td>44.52</td><td>42.73</td><td>42.92</td><td>44.10</td></tr><tr><td rowspan="15">Non-LLMDA</td><td>Char. Swap</td><td>44.94</td><td>42.22</td><td>39.46</td><td>40.56</td><td>42.76</td><td>47.08</td><td>44.69</td><td>44.06</td><td>43.89</td><td>42.45</td><td>42.90</td></tr><tr><td>Char. OCR</td><td>43.72</td><td>43.70</td><td>39.31</td><td>41.02</td><td>43.73</td><td>45.91</td><td>46.02</td><td>45.11</td><td>42.95</td><td>43.02</td><td>43.83</td></tr><tr><td>Char. Delete</td><td>43.98</td><td>43.33</td><td>38.99</td><td>40.50</td><td>43.08</td><td>46.01</td><td>45.04</td><td>44.02</td><td>42.53</td><td>42.77</td><td>43.52</td></tr><tr><td>Char. Insert</td><td>45.03</td><td>43.14</td><td>39.22</td><td>40.61</td><td>42.77</td><td>46.69</td><td>44.61</td><td>42.44</td><td>43.29</td><td>42.86</td><td>42.37</td></tr><tr><td>Char. Subs.</td><td>43.87</td><td>43.07</td><td>39.39</td><td>40.22</td><td>42.46</td><td>46.11</td><td>45.22</td><td>43.25</td><td>43.41</td><td>42.65</td><td>43.08</td></tr><tr><td>Word Swap</td><td>43.83</td><td>44.56</td><td>39.15</td><td>40.84</td><td>43.51</td><td>46.25</td><td>46.21</td><td>44.93</td><td>43.24</td><td>43.54</td><td>44.30</td></tr><tr><td>Word Delete</td><td>44.24</td><td>42.38</td><td>39.54</td><td>40.34</td><td>43.12</td><td>45.70</td><td>44.49</td><td>45.56</td><td>43.41</td><td>42.02</td><td>44.28</td></tr><tr><td>Spell Error</td><td>44.82</td><td>42.66</td><td>38.68</td><td>41.03</td><td>42.85</td><td>46.34</td><td>44.85</td><td>44.97</td><td>42.71</td><td>43.08</td><td>43.71</td></tr><tr><td>GPT2 Insert</td><td>43.37</td><td>43.36</td><td>39.79</td><td>40.92</td><td>43.69</td><td>45.07</td><td>45.93</td><td>46.11</td><td>42.65</td><td>43.07</td><td>44.44</td></tr><tr><td>Word2vec Insert</td><td>46.47</td><td>41.80</td><td>39.67</td><td>40.80</td><td>42.66</td><td>48.36</td><td>45.13</td><td>43.78</td><td>44.18</td><td>42.95</td><td>42.86</td></tr><tr><td>BERT Subs.</td><td>44.88</td><td>44.01</td><td>39.57</td><td>40.23</td><td>42.86</td><td>46.87</td><td>45.32</td><td>46.35</td><td>43.17</td><td>42.92</td><td>43.79</td></tr><tr><td>Word2vec Insert</td><td>44.24</td><td>43.89</td><td>38.90</td><td>40.21</td><td>42.80</td><td>46.17</td><td>45.58</td><td>44.34</td><td>43.07</td><td>42.76</td><td>44.06</td></tr><tr><td>Back Translation</td><td>44.19</td><td>44.00</td><td>39.78</td><td>40.89</td><td>42.96</td><td>45.97</td><td>45.31</td><td>47.39</td><td>43.24</td><td>43.60</td><td>43.95</td></tr><tr><td>Average</td><td>44.42</td><td>43.23</td><td>39.34</td><td>40.62</td><td>43.01</td><td>46.34</td><td>45.26</td><td>44.79</td><td>43.21</td><td>42.89</td><td>43.62</td></tr><tr><td>Best</td><td>46.47</td><td>44.56</td><td>39.79</td><td>41.03</td><td>43.73</td><td>48.36</td><td>46.21</td><td>47.39</td><td>44.18</td><td>43.47</td><td>44.30</td></tr><tr><td rowspan="15">Manual-LLMDA</td><td>Char. Perturb</td><td>44.72</td><td>43.56</td><td>39.39</td><td>40.61</td><td>45.54</td><td>47.02</td><td>45.32</td><td>46.25</td><td>43.68</td><td>43.27</td><td>44.06</td></tr><tr><td>Word Insert/Delete</td><td>44.80</td><td>44.27</td><td>38.81</td><td>40.97</td><td>44.49</td><td>46.58</td><td>45.93</td><td>46.59</td><td>42.95</td><td>43.68</td><td>44.46</td></tr><tr><td>Word Swap</td><td>45.45</td><td>43.88</td><td>39.01</td><td>40.66</td><td>43.28</td><td>47.04</td><td>45.20</td><td>46.80</td><td>43.63</td><td>42.63</td><td>45.79</td></tr><tr><td>Word Replace</td><td>45.29</td><td>45.94</td><td>39.14</td><td>41.05</td><td>43.10</td><td>47.36</td><td>47.98</td><td>47.62</td><td>43.52</td><td>44.57</td><td>45.89</td></tr><tr><td>Grammar Transform</td><td>45.39</td><td>45.62</td><td>39.29</td><td>41.34</td><td>43.00</td><td>46.81</td><td>46.76</td><td>46.41</td><td>43.24</td><td>43.75</td><td>45.53</td></tr><tr><td>Data Mix</td><td>44.63</td><td>44.01</td><td>38.50</td><td>41.29</td><td>43.67</td><td>46.02</td><td>45.68</td><td>43.93</td><td>42.76</td><td>43.26</td><td>43.01</td></tr><tr><td>Paragraph Shuffle</td><td>45.78</td><td>42.03</td><td>39.29</td><td>40.70</td><td>43.28</td><td>47.70</td><td>45.02</td><td>48.02</td><td>44.43</td><td>43.25</td><td>45.41</td></tr><tr><td>Mask Predict</td><td>42.83</td><td>43.15</td><td>39.58</td><td>40.59</td><td>44.05</td><td>46.02</td><td>44.60</td><td>45.70</td><td>42.59</td><td>43.16</td><td>44.64</td></tr><tr><td>Sentence Reorder</td><td>44.18</td><td>43.84</td><td>39.37</td><td>40.76</td><td>43.20</td><td>46.30</td><td>46.35</td><td>48.46</td><td>43.35</td><td>43.70</td><td>45.64</td></tr><tr><td>Contextual Replace</td><td>43.52</td><td>46.26</td><td>38.73</td><td>40.99</td><td>43.49</td><td>45.38</td><td>47.62</td><td>43.92</td><td>42.58</td><td>45.26</td><td>44.76</td></tr><tr><td>Paraphrase</td><td>45.56</td><td>45.30</td><td>39.44</td><td>40.94</td><td>41.09</td><td>48.02</td><td>47.31</td><td>43.99</td><td>43.97</td><td>44.10</td><td>42.80</td></tr><tr><td>POS Augment</td><td>44.97</td><td>44.48</td><td>37.51</td><td>39.12</td><td>41.14</td><td>47.06</td><td>46.07</td><td>47.25</td><td>43.49</td><td>43.21</td><td>45.75</td></tr><tr><td>Back Translation</td><td>45.79</td><td>43.38</td><td>39.18</td><td>41.12</td><td>43.80</td><td>47.02</td><td>45.58</td><td>45.10</td><td>43.36</td><td>43.20</td><td>44.41</td></tr><tr><td>Average</td><td>44.83</td><td>44.28</td><td>39.14</td><td>40.93</td><td>43.50</td><td>46.79</td><td>46.10</td><td>46.15</td><td>43.35</td><td>43.61</td><td>44.78</td></tr><tr><td>Best</td><td>45.79</td><td>46.26</td><td>39.58</td><td>41.34</td><td>44.49</td><td>48.02</td><td>47.98</td><td>48.02</td><td>44.43</td><td>45.26</td><td>45.89</td></tr><tr><td>Self-LLMDA</td><td>51.72</td><td>54.98</td><td>48.80</td><td>40.02</td><td>42.80</td><td>43.80</td><td>50.00</td><td>52.75</td><td>49.48</td><td>46.64</td><td>48.83</td><td>48.95</td></tr></table>
+
+Table 1: The performance of different data augmentation methods. Char. and Subs. are the abbreviations of character and substitute, respectively. Underlined indicates best performance under each augmentation method group while Bold indicates the best result of the whole table. In each group, the last two rows represent the aggregated result of the whole group of augmentation methods (e.g. average and best).
+
+- Manual-LLMDA methods. This set comprises 13 manually designed augmentation instructions for LLM, including: Character-Level: Perturbations similar to those in Non-LLMDA. Word-Level: Swaps, replacements, and part-of-speech (POS) enhancements. Sentence-Level: Reordering and data mixing strategies. Contextual-Level: Predictive masking, contextual substitutions, and back-translation.
+
+We also report the average and best performance of Non-LLMDA and Manual-LLMDA for better comparison. An extensive ablation study of our task-informed selection model, presented in § 5.3.
+
+# 5.2 Main Results
+
+The analysis of experimental results presented in Tab. 1 reveals several findings: Firstly, there is performance inconsistency among the different instructions from Manual-LLMDA. The impact of augmentation instructions varies across different downstream tasks and models. This highlights the difficulty in creating universally effective data aug-
+
+mentation instruction. Secondly, Manual-LLMDA is not always better than Non-LLMDA. In controlled comparisons focusing on specific augmentation topics, Manual-LLMDA's advantages over Non-LLMDA were not clearly evident. For example, in the contexts of "Back Translation" and "Word Swap", Non-LLMDA outperformed Manual-LLMDA in 5 out of 12 and 7 out of 12 cases, respectively. Lastly, the experimental results show the superiority of Self-LLMDA. Our proposed model consistently outperformed these baseline methods, highlighting the effectiveness of integrating automatic instruction generation with targeted task-specific instruction selection. This approach not only optimizes performance but also reduces the manual efforts typically required to design effective augmentation strategies, showcasing the potential of our model in enhancing data augmentation practices.
+
+# 5.3 Ablation Study
+
+We add an ablation study to understand the impact of two key components in our framework: aug-
+
+<table><tr><td>Select. Method</td><td colspan="3">Class→Non-Class</td><td colspan="3">Non-Class→Class</td></tr><tr><td></td><td>125m</td><td>350m</td><td>1.3b</td><td>125m</td><td>350m</td><td>1.3b</td></tr><tr><td>Manual-LLMDA+</td><td>39.62</td><td>41.74</td><td>44.38</td><td>48.02</td><td>48.51</td><td>48.07</td></tr><tr><td>Random-Select</td><td>39.34</td><td>40.31</td><td>42.68</td><td>46.15</td><td>44.34</td><td>43.98</td></tr><tr><td>Empirical-Select</td><td>39.17</td><td>41.18</td><td>43.14</td><td>47.19</td><td>47.30</td><td>44.41</td></tr><tr><td>LLM-Select</td><td>38.77</td><td>41.06</td><td>43.07</td><td>46.81</td><td>48.02</td><td>46.42</td></tr><tr><td>Self-LLMDA</td><td>40.02</td><td>42.80</td><td>43.80</td><td>50.00</td><td>52.75</td><td>49.48</td></tr></table>
+
+Table 2: Ablation study of Self-LLMDA.
+
+mentation instruction self-generation and the task-informed instruction selection. Firstly, we train a task-informed instruction selection model  $S$  on manually-crafted instructions from Manual-LLMDA and named it Manual-LLMDA+ to understand the contribution of the contribution of LLM self-generated augmentation instructions. Secondly, we test the efficacy of our selection model by comparing three alternative selection strategies: (1) Random-select, which randomly select instruction from the pool of augmentation methods for each task; (2) Empirical-select, which selects the prompt that yielded the highest average performance across training tasks, under the assumption that successful prompts on training tasks will generalize well to test tasks; and (3) LLM-Select, which prompts the LLM to choose the most suitable instruction from candidates based on its internal decision-making processes. The Results in Tab. 2 show that Self-LLMDA consistently outperforms these alternative methods, indicating the benefits of instruction self-generation and task-informed selection in enhancing model performance.
+
+# 5.4 Hyperparameter Analysis
+
+Here, we closely examined the impact of two critical hyperparameters on the training of our task-informed instruction model:  $n$  and  $m$ . The hyperparameter  $n$  specifies the number of augmentation instructions to be sampled for optimizing Eq. 6. It should be noticed that, we only vary  $n$  at training time, while at inference, we will calculate the score for all the generated instructions and choose the one with the largest score. On the other hand,  $m$  determines the number of examples from the task dataset that are used to represent the task, influencing the model's performance during both the optimization and inference phases. Our analysis, depicted in Fig. 5, highlights several key findings: (1) Optimal Number of Instructions: We found that setting  $n = 2$  leads to the best performance, outperforming other configurations. This suggests that a pairwise comparison, as formulated in Eq. 6, is most effective for our model's learning process. (2)
+
+![](images/32fdf2baadc302724c9d94390b5c11ece373e25bcb8737a014ec79f6f9707d16.jpg)  
+(a) Class  $\rightarrow$  Non-Class
+
+![](images/2cda81379020719efdc9da2b164bee57a9f1f9ef7c14de1b25030970169bddcf.jpg)  
+(b) Non-Class  $\rightarrow$  Class
+
+![](images/d530e17aaa204f6639c90e2954602f17c8300fcc6885e3b371c402ca7b4ad825.jpg)  
+Figure 4: Hyperparameter analysis of  $n$ , which dictates the number of augmentation instructions sampled during the training of the selection model.  
+Figure 5: Analysis of the hyperparameter  $m$ , which determines the number of examples randomly sampled to represent a task.
+
+![](images/eafacb531676a010bbe6f3296b46f4202a5b0479a6a74109215cae574be62ba3.jpg)  
+(a) Class  $\rightarrow$  Non-Class  
+(b) Non-Class  $\rightarrow$  Class
+
+Representative Examples: Interestingly, a smaller number of examples  $(m)$  appear to better capture the essence of the tasks. This observation indicates that a larger set of examples could introduce noise, potentially detracting from the model's ability to accurately represent tasks for instruction selection.
+
+# 5.5 In-Depth Analysis of the Task-Informed Instruction Selection Model
+
+In this section, we provide a detailed analysis of the performance and generalization capabilities of our instruction selection model  $S$ , focusing on its generalizability to unknown augmentation instructions, unknown target models, and the specific case studies of the augmentation instructions it selects.
+
+Generalization to Unknown Augmentation Instructions. In this analysis, we delve into the selection model's adaptability to unknown augmentation instructions by simulating a dynamic environment where new instructions are generated asynchronously by the LLMs. This scenario mirrors practical applications where the augmentation instruction set can expand without necessitating retraining of the selection model. To test this, we constrained the training phase of the selection model to a limited subset of self-generated augmentation instructions (30% of all generated by the LLMs), utilizing the whole generated instructions for evaluation at inference time.
+
+As the results shown in Fig. 6, we can ob
+
+![](images/512953d8abb8e23169b4c70d2bb4c2344ca366d92ddfa7339664b24cc0e58cd3.jpg)  
+(a) Class  $\rightarrow$  Non-Class.
+
+![](images/a5fee75af7d9741756adbc5f772f0caaddd5729c7800471356e22a9e3193846c.jpg)  
+(b) Non-Class  $\rightarrow$  Class.  
+Figure 6: Result of generalization to unknown augmentation instruction selection.
+
+serve a performance improvement of our selection model over the best performance of Non-LLMDA and Manual-LLMDA. This indicates the robustness of our selection model in adapting to incremental augmentation instructions, effectively selecting suitable instructions even when faced with previously unknown instructions. These observations highlight the efficacy of our selection model in a dynamic augmentation scenario.
+
+Generalization to Unknown Target Models. Our study extended to evaluate the adaptability of our task-informed selection model across diverse target models. By applying the selection model, initially trained on the task performance of a specific target model, to different models. The results of these experiments are presented in Tab. 3. Our findings show that the augmentation instructions selected by our model remain effective even when applied to different target models. Notably, in most scenarios, our model Self-LLMDA, when transferred to alternate target models, outperformed the best results obtained using Non-LLMDA and Manual-LLMDA. This indicates that the underlying pattern determining instruction effectiveness via our instruction selection model is transferable.
+
+<table><tr><td rowspan="2">Train.</td><td colspan="3">Class → Non-Class</td><td colspan="3">Non-Class → Class</td></tr><tr><td>125m</td><td>350m</td><td>1.3b</td><td>125m</td><td>350m</td><td>1.3b</td></tr><tr><td>Best Non-LLMDA</td><td>39.79</td><td>41.03</td><td>43.73</td><td>48.36</td><td>46.21</td><td>47.39</td></tr><tr><td>Best Manual-LLMDA</td><td>39.58</td><td>41.34</td><td>44.49</td><td>48.02</td><td>47.98</td><td>48.02</td></tr><tr><td>125m</td><td>40.02</td><td>41.97</td><td>43.56</td><td>50.00</td><td>54.12</td><td>49.83</td></tr><tr><td>350m</td><td>39.96</td><td>42.80</td><td>43.66</td><td>49.96</td><td>52.75</td><td>48.82</td></tr><tr><td>1.3b</td><td>39.85</td><td>42.42</td><td>43.80</td><td>49.04</td><td>51.22</td><td>49.48</td></tr></table>
+
+Table 3: Transferability of the Task-Informed Selection Model. Our selection model, initially trained on a specific target model (indicated by each row in the second group), when applied to alternate target models (represented in each column).
+
+Analysis of Selected Instructions. We conducted a detailed analysis of the augmentation instructions chosen by our selection model, and the findings visualized in Fig. 7. The key insights from this analysis are as follows: (1) Diversity of Selected In
+
+![](images/bfa5a5f77a3f211872c1befb79bdb140a495a99806e4e6506363d08373cb3693.jpg)  
+(a) Class  $\rightarrow$  Non-Class.  
+Figure 7: Selected augmentation instructions from task-informed augmentation selection model.
+
+![](images/3b1fe28c6b513c270e6f5260da0d948c386cff2a629bfad3518deac48ea2d90c.jpg)  
+(b) Non-Class  $\rightarrow$  Class.
+
+**Structions:** The distribution of selected instructions showcases a wide variety in the types of augmentations chosen by the model, with 3, 2, and 6 unique data augmentation instructions identified for the  $125\mathrm{m}$ ,  $350\mathrm{m}$ , and  $1.3\mathrm{b}$  models under Class  $\rightarrow$  Non-Class, respectively. This demonstrates the model's ability to adapt and select from a broad spectrum of augmentation strategies to meet the specific requirements of different tasks. (2) Variability across Models: The selection patterns exhibit notable differences when the model is applied to various target models. This variability indicates preference differences across different target models. (3) Preference for Paraphrase-Based Instructions: A significant portion of the selected instructions fall into the category of paraphrase-based augmentations, such as "Text Paraphrase", "Paraphrase", "Contextual Paraphrase", and "Sentence Paraphrase". This preference not only highlights the effectiveness and general applicability of paraphrase-based augmentations but also illustrates our task-informed selection model's nuanced capability to discern and recommend the most suitable paraphrase variation for a given task.
+
+# 6 Conclusion
+
+In this work, we introduced Self-LLMDA, a novel framework that leverages the capabilities of LLMs for textual data augmentation. Our approach addresses the challenges associated with traditional data augmentation methods and the limitations of manual instruction generation in LLM-based augmentation. Self-LLMDA automates the generation and selection of augmentation instructions, thereby significantly enhancing the quality and applicability of augmented data across diverse downstream tasks. Tested across 26 diverse few-shot learning tasks, Self-LLMDA consistently outperforms both Non-LLMDA and Manual-LLMDA methods, showcasing its effectiveness and applicability.
+
+# 7 Limitations
+
+This study acknowledges several constraints that delineate the scope of our current work and outline directions for future research:
+
+# - Evaluation on a Limited Range of LLMs:
+
+Our experiments were conducted primarily with GPT 3.5 Turbo due to the high costs associated with using OpenAI models. While promising results in Tab. 1 suggest that our proposed Self-LLMDA method could potentially perform even better on more advanced models like GPT 4 Turbo, comprehensive testing was not feasible. Similarly, the computational demands of evaluating open-source LLMs such as LLAMA-70b-chat (Touvron et al., 2023), coupled with the extensive number of tasks in our study, exceeded our resources. Despite these limitations, we are optimistic that Self-LLMDA would exhibit enhanced performance across a broader spectrum of LLMs.  
+- Meta-Prompting Exploration: Within the Self-LLMDA framework, we employed one meta-prompt to guide the LLM in generating diverse and relevant augmentation instructions. However, our exploration of meta-prompting techniques was limited. We acknowledge that more sophisticated prompt engineering could further refine the quality and effectiveness of generated instructions. Investigating more advanced meta-prompting strategies remains an area for future exploration.  
+- Analysis of Ensemble Augmentation Methods: Our research did not investigate the potential benefits of combining multiple sets of augmented data (e.g.,  $\mathcal{D} \cup \mathcal{D}_1' \cup \mathcal{D}_2'$ ). Such ensemble approaches introduce additional complexities, such as determining the optimal number of augmentation instructions to include. While we hypothesize that ensemble augmentation could improve model performance, this aspect falls outside the current study's scope and is earmarked for subsequent investigation.
+
+# References
+
+Markus Bayer, Marc-Andre Kaufhold, and Christian Reuter. 2022. A survey on data augmentation for text classification. ACM Computing Surveys, 55(7):1-39.
+
+Yonatan Belinkov and Yonatan Bisk. 2018. Synthetic and natural noise both break neural machine translation.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners.  
+Qiao Cheng, Jin Huang, and Yitao Duan. 2022. Semantically consistent data augmentation for neural machine translation via conditional masked language model.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022. Scaling instruction-finetuned language models.  
+John Joon Young Chung, Ece Kamar, and Saleema Amershi. 2023. Increasing diversity while maintaining accuracy: Text data generation with large language models and human interventions. arXiv preprint arXiv:2306.04140.  
+Claude Coulombe. 2018. Text data augmentation made simple by leveraging nlp cloud apis.  
+Haixing Dai, Zheng Liu, Wenxiong Liao, Xiaoke Huang, Zihao Wu, Lin Zhao, Wei Liu, Ninghao Liu, Sheng Li, Dajiang Zhu, Hongmin Cai, Quanzheng Li, Dinggang Shen, Tianming Liu, and Xiang Li. 2023. Chataug: Leveraging chatgpt for text data augmentation. ArXiv, abs/2302.13007.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.  
+Aleksandra Edwards, Asahi Ushio, Jose Camacho-Collados, Hélène de Ribaupierre, and Alun Preece. 2023. Guiding generative language models for data augmentation in few-shot text classification.  
+Marzieh Fadaee, Arianna Bisazza, and Christof Monz. 2017. Data augmentation for low-resource neural machine translation. arXiv preprint arXiv:1705.00440.  
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2022. Simcse: Simple contrastive learning of sentence embeddings.
+
+Hongyu Guo, Yongyi Mao, and Richong Zhang. 2019. Augmenting data with mixup for sentence classification: An empirical study.  
+Yoichi Ishibashi, Danushka Bollegala, Katsuhito Sudoh, and Satoshi Nakamura. 2023. Evaluating the robustness of discrete prompts. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2373-2384, Dubrovnik, Croatia. Association for Computational Linguistics.  
+Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. Spanbert: Improving pre-training by representing and predicting spans.  
+Akbar Karimi, Leonardo Rossi, and Andrea Prati. 2021. AEDA: An easier data augmentation technique for text classification. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2748-2754, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UNIFIEDQA: Crossing format boundaries with a single QA system. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1896-1907, Online. Association for Computational Linguistics.  
+Varun Kumar, Ashutosh Choudhary, and Eunah Cho. 2021. Data augmentation using pre-trained transformer models.  
+Gakuto Kurata, Bing Xiang, and Bowen Zhou. 2016. Labeled Data Generation with Encoder-Decoder LSTM for Semantic Slot Filling. In Proc. Interspeech 2016, pages 725-729.  
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension.  
+Zhuoyan Li, Hangxiao Zhu, Zhuoran Lu, and Ming Yin. 2023. Synthetic data generation with large language models for text classification: Potential and limitations. arXiv preprint arXiv:2310.07849.  
+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.  
+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.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization.  
+Edward Ma. 2019. Nlp augmentation. https://github.com/makcedward/nlpaug.
+
+Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. 2022. Generating training data with language models: Towards zero-shot language understanding.  
+Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering.  
+Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2022. MetaICL: Learning to learn in context. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2791-2809, Seattle, United States. Association for Computational Linguistics.  
+John X. Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020. Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp.  
+Frédéric Piedboeuf and Philippe Langlais. 2023. Is ChatGPT the ultimate data augmentation algorithm? In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15606-15615, Singapore. Association for Computational Linguistics.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2023. Exploring the limits of transfer learning with a unified text-to-text transformer.  
+Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902-4912, Online. Association for Computational Linguistics.  
+Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. In International Conference on Machine Learning, pages 4596-4604. PMLR.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée 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.  
+Solomon Ubani, Suleyman Olcay Polat, and Rodney Nielsen. 2023. Zeroshotdataaug: Generating and augmenting training data with chatgpt.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. Glue: A multi-task benchmark and analysis platform for natural language understanding.
+
+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. arXiv preprint arXiv:2212.10560.  
+Jason Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks.  
+Xing Wu, Shangwen Lv, Liangjun Zang, Jizhong Han, and Songlin Hu. 2018. Conditional bert contextual augmentation.  
+Jiacheng Ye, Jiahui Gao, Qintong Li, Hang Xu, Jiangtao Feng, Zhiyong Wu, Tao Yu, and Lingpeng Kong. 2022. Zerogen: Efficient zero-shot learning via dataset generation.  
+Qinyuan Ye, Bill Yuchen Lin, and Xiang Ren. 2021. CrossFit: A few-shot learning challenge for cross-task generalization in NLP. pages 7163-7189.  
+Kang Min Yoo, Dongju Park, Jaewook Kang, Sang-Woo Lee, and Woomyeong Park. 2021. Gpt3mix: Leveraging large-scale language models for text augmentation.  
+Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander Ratner, Ranjay Krishna, Jiaming Shen, and Chao Zhang. 2023. Large language model as attributed training data generator: A tale of diversity and bias.  
+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 pre-trained transformer language models.  
+Xiang Zhang, Junbo Zhao, and Yann LeCun. 2016. Character-level convolutional networks for text classification.  
+Kaijie Zhu, Jindong Wang, Jiaheng Zhou, Zichen Wang, Hao Chen, Yidong Wang, Linyi Yang, Wei Ye, Yue Zhang, Neil Zhenqiang Gong, and Xing Xie. 2023. Promptbench: Towards evaluating the robustness of large language models on adversarial prompts.
+
+# A Detailed Experiment Settings
+
+Generation Configuration. We utilize gpt-3.5-turbo as our backbone LLM for augmentation instruction generation and data augmentation. We set the temperature for both of them as 0.7. For the instruction generation, we follow the generation hyper-parameter setting from Wang et al. (2022). For data augmentation, we utilize the default generation hyper-parameter from Chat Completion. The whole experiment including generating augmentation instructions and generating augmentation data costs us  $82 USD in total, according to OpenAI's pricing (Input$ 0.0005 / 1K tokens and output $0.0015 / 1K tokens). However, the total experiment cost around $200 USD for debugging and exploration.
+
+Meta Prompts for Data Augmentation As shown in step ③ in Fig. 2, we also need a meta prompt to encourage Self-LLMDA to augment high quality data. The main reason for this meta-prompt setting is because in some augmentation instructions they will discuss some external tools like word-embedding, other language models, if we did not provide the meta-prompt, the LLM will reject the generation of augmented data. The design of meta prompt is as follows:
+
+Please do the following data augmentation steps to the text delimited by triple backsticks. If you need any external resources or data, you can just simulate the environment by yourself and finish that step based on your own knowledge since you are the best language model in word. Augmentation Instructions:  $\mathbf{I}_j$ , Input Data:  $\mathbf{x}_i\{\mathbf{I}_{seed}\}$
+
+Task-informed Instruction Selection. The instruction ranking model is initialized with FLAN-T5-Large(Radford et al., 2019) and is trained using Adafactor (Shazeer and Stern, 2018) with learning rate 5e-5 and dropout 0.1. We train the selection model for 100 epochs and set the early stop with patience 20 epochs. We employ the validation set from training tasks to select the best checkpoint. The search space for different hyperparameter analysis are as follows:
+
+Target Model Finetuning. We use OPT (Zhang et al., 2022) from  $125\mathrm{m}$ ,  $350\mathrm{m}$ , 1.3b different sizes. For all of them we use AdamW(Loshchilov and Hutter, 2019) as our optimizer with learning rate 5e-5
+
+<table><tr><td>Symbol</td><td>Description</td><td>Search Space</td></tr><tr><td>n</td><td>Number of sampled augmentation instruction</td><td>{2,3,4,5}</td></tr><tr><td>m</td><td>Number of sampled examples from task dataset</td><td>{1,2,3,4}</td></tr><tr><td>-</td><td>batch size</td><td>{4,8,16}</td></tr><tr><td>-</td><td>epochs</td><td>100</td></tr></table>
+
+with 10 training epochs. Due to the constraint of GPU memory, for  $125\mathrm{m}$  and  $350\mathrm{m}$  we set the batch size as 8, while 1.3b we set the batch size as 2. All these experiments is tested on one NVIDIA A100 A100-40G GPU cards.
+
+# B Analysis of Self-Generated Instructions
+
+In our analysis, we delve into the characteristics and diversity of the self-generated augmentation prompts created by Manual-LLMDA.
+
+Statistical Information. To facilitate a structured examination, we categorize these prompts based on the textual data augmentation taxonomy outlined by Bayer et al. (2022). The distribution and basic statistics of these various augmentation methods are detailed in Tab. 5.
+
+Table 4: The search space of augmentation selection.  
+
+<table><tr><td>Augmentation Type</td><td>Count</td><td>AVG Length</td></tr><tr><td>Manual - LLMDA</td><td>13</td><td>28.15</td></tr><tr><td>- character level</td><td>1</td><td>32.00</td></tr><tr><td>- word level</td><td>3</td><td>20.67</td></tr><tr><td>- phrase level</td><td>4</td><td>31.75</td></tr><tr><td>- document level</td><td>5</td><td>29.00</td></tr><tr><td>Manual - LLMDA</td><td>51</td><td>25.58</td></tr><tr><td>- misc.</td><td>1</td><td>22.00</td></tr><tr><td>- character level</td><td>2</td><td>24.50</td></tr><tr><td>- word level</td><td>10</td><td>22.80</td></tr><tr><td>- phrase level</td><td>18</td><td>26.72</td></tr><tr><td>- document level</td><td>20</td><td>26.25</td></tr></table>
+
+Table 5: Statistics of augmentation prompts.
+
+Naming Conventions. A notable aspect of our analysis involves examining the naming conventions of the augmentation methods. Recognizing that the method names often provide a high-level summary of the augmentation approach (e.g., <method name>), we further explore the linguistic patterns within these names. Specifically, we conduct an analysis focusing on the first and last words of each method name. This approach allows us to gain insights into the thematic and functional aspects of the augmentation methods. The distribution of these first and last words in method names is visually represented in Fig. 8. This visual representation aids in understanding the range and focus of the augmentation techniques generated by Manua1-LLMDA.
+
+![](images/cf2944d5150c08f4f45328d1b9f7702cbe531bbc7407d57beb9c744454602c42.jpg)  
+(a) First word.
+
+![](images/992a26ce2875af3e62d82a87748f2c1ffe8a3496804afd541fc041207bab41d0.jpg)  
+(b) Last word.  
+Figure 8: The first words and last words from the Chat-Self. We filter these words by appearing more than once.
+
+By analyzing these key linguistic elements, we aim to shed light on the creative breadth and thematic focus of the self-generated augmentation instructions.
+
+![](images/df77fafe5ef586e7d9ed77c5dd1865c9b282833b7fff77424209f612da85f2cc.jpg)  
+Figure 9: Distribution of the ROUGE-L scores between generated instructions and their most similar human-designed instructions.
+
+# C Analysis of Generated Data Across Different Augmentation Methods
+
+In this analysis, we aim to discern the differences among the original dataset, non-LLM augmented data, data augmented via human-designed instructions, and data augmented using Manual-LLMDA generated instructions. Our focus is on the surface-level characteristics of the augmented content, and we consolidate data across all tasks for a comprehensive view. Key observations from the analysis, as detailed in Tab. 6, include the following:
+
+Length of Content. Data augmented by LLM-based methods, on average, exhibits longer content compared to both the original and traditionally augmented datasets. This increase in length could offer a broader spectrum of training examples, potentially aiding in better generalization of target models. However, it also introduces a challenge of dataset inconsistency and the risk of adding unwanted variations.
+
+Perplexity Scores. Interestingly, LLM-augmented content achieves lower perplexity scores (as measured on GPT2-small) compared to traditional augmentation methods. This suggests that the target model like GPT2-small has a better grasp of content augmented by LLMs. A possible explanation for the higher perplexity scores observed in non-LLM text augmentations is that the character and word-level changes might introduce new, irrelevant tokens into the text, thereby increasing complexity.
+
+Closeness to Original Examples. Compared to non-LLM augmentation methods, LLM-based augmentations tend to produce content that is more closely related or less diverse relative to the original examples. This observation points to a potential trade-off between relevance and diversity in the augmented content generated by LLMs.
+
+<table><tr><td>Metrics</td><td>Method</td><td>Mean</td><td>Std.</td><td>25%</td><td>50%</td><td>75%</td></tr><tr><td rowspan="4">Sentiment</td><td>Ori.</td><td>0.05</td><td>0.24</td><td>0.00</td><td>0.00</td><td>0.14</td></tr><tr><td>Traditional</td><td>0.05</td><td>0.23</td><td>0.00</td><td>0.00</td><td>0.12</td></tr><tr><td>Human</td><td>0.06</td><td>0.23</td><td>0.00</td><td>0.00</td><td>0.15</td></tr><tr><td>ChatGPT</td><td>0.06</td><td>0.23</td><td>0.00</td><td>0.00</td><td>0.16</td></tr><tr><td rowspan="2">Grammar</td><td>Ori.</td><td>1.50</td><td>2.34</td><td>0.00</td><td>1.00</td><td>2.00</td></tr><tr><td>Traditional</td><td>4.92</td><td>5.02</td><td>2.00</td><td>4.00</td><td>7.00</td></tr><tr><td rowspan="2">Error</td><td>Human</td><td>7.78</td><td>47.76</td><td>0.00</td><td>1.00</td><td>4.00</td></tr><tr><td>ChatGPT</td><td>4.19</td><td>27.51</td><td>0.00</td><td>1.00</td><td>3.00</td></tr><tr><td rowspan="4">Words</td><td>Ori.</td><td>29.06</td><td>38.84</td><td>11.00</td><td>20.00</td><td>34.00</td></tr><tr><td>Non-LLMDA</td><td>30.01</td><td>39.01</td><td>11.00</td><td>21.00</td><td>36.00</td></tr><tr><td>Manual-LLMDA</td><td>95.34</td><td>285.95</td><td>14.00</td><td>30.00</td><td>71.00</td></tr><tr><td>Self-LLMDA</td><td>66.04</td><td>182.91</td><td>14.00</td><td>27.00</td><td>57.00</td></tr><tr><td>Perplexity</td><td>Ori.</td><td>530.13</td><td>5867.93</td><td>46.42</td><td>93.78</td><td>253.85</td></tr><tr><td>On</td><td>Non-LLMDA</td><td>1354.91</td><td>5912.87</td><td>231.22</td><td>595.03</td><td>1305.86</td></tr><tr><td rowspan="2">GPT2-small</td><td>Manual-LLMDA</td><td>614.80</td><td>12284.76</td><td>17.72</td><td>58.94</td><td>184.42</td></tr><tr><td>Self-LLMDA</td><td>491.83</td><td>9715.57</td><td>21.28</td><td>59.85</td><td>173.28</td></tr><tr><td rowspan="3">Distance to Original</td><td>Non-LLMDA</td><td>0.26</td><td>0.15</td><td>0.12</td><td>0.25</td><td>0.39</td></tr><tr><td>Manual-LLMDA</td><td>0.17</td><td>0.04</td><td>0.15</td><td>0.17</td><td>0.18</td></tr><tr><td>Self-LLMDA</td><td>0.15</td><td>0.05</td><td>0.12</td><td>0.15</td><td>0.18</td></tr></table>
+
+Table 6: Characteristics of augmented data.  
+
+<table><tr><td>Aug. Type</td><td>small</td><td>medium</td><td>large</td></tr><tr><td>Trad.</td><td>42.37/44.18</td><td>46.92/49.72</td><td>44.81/47.33</td></tr><tr><td>Human.</td><td>42.43/44.52</td><td>47.25/49.52</td><td>48.44/53.32</td></tr><tr><td>LLM.</td><td>42.03/48.47</td><td>47.80/51.09</td><td>45.88/52.17</td></tr></table>
+
+Table 7: Main results, using target models from GPT2 family. Two numbers indicate the single best augmentation method across tasks and the task specific best augmentation method. Bold indicates the best average result except results.
+
+Augmentation Instruction Pitfalls Across Tasks The effectiveness of augmentation instructions can vary depending on the specific characteristics of the tasks at hand (Ribeiro et al., 2020; Wei and Zou, 2019). To illustrate this, we present a case study focusing on the augmentation instruction Pronoun replacement: replace pronouns in the text with their corresponding nouns or vice versa, maintaining the
+
+semantic meaning of the sentence. For the sake of brevity, we will use the abbreviation PR to refer to pronoun replacement. We consider two categories of tasks: text entailment (TE) and question answering (QA). As shown in Tab. 8, the results indicate that PR yields suboptimal performance on TE tasks, while it achieves good performance on QA tasks. This discrepancy can be attributed to the inherent characteristics of these tasks. TE tasks heavily rely on capturing the overall semantic meaning and logical relationships within the text, which may not always be preserved when applying pronoun replacement (Gao et al., 2022). In contrast, QA tasks aim to locate and provide specific information relevant to the given question (Joshi et al., 2020). By replacing pronouns with their corresponding nouns, the model can more easily identify the relevant entities and establish a clearer connection between the question and the answer, ultimately benefiting the QA task performance.
+
+<table><tr><td>Type</td><td>Task</td><td>Rank (/51)</td></tr><tr><td rowspan="4">TE</td><td>glue-mrpc</td><td>48</td></tr><tr><td>glue-rte</td><td>44</td></tr><tr><td>glue-wnli</td><td>33</td></tr><tr><td>medicalQuestions_pairs</td><td>48</td></tr><tr><td rowspan="5">QA</td><td>qasc</td><td>1</td></tr><tr><td>openbookqa</td><td>27</td></tr><tr><td>commonsense_qa</td><td>4</td></tr><tr><td>quartz-no_knowledge</td><td>34</td></tr><tr><td>quartz-with_knowledge</td><td>9</td></tr></table>
+
+Table 8: Performance comparison of the pronoun replacement (PR) augmentation instruction on text entailment (TE) and question answering (QA) tasks.
+
+# D Dataset Collection
+
+In Tab. 9, we list all 26 tasks and how we splitting them into training and testing for evaluating the model generalization to unknown downstream tasks. Each task will have 16 training and validation examples but with full test examples. We utilize the code from CrossFit(Ye et al., 2021) to extract and split the training, validation and testing for each task.
+
+# E Details of Baseline Methods
+
+# E.1 Augmentation Methods of Non-LLMDA
+
+All of the implementation of Non-LLMDA are from Ma (2019). Here is an elaboration on each of the mentioned Non-LLMDA augmentation methods:
+
+Character-Level Augmentations Random Swap (Belinkov and Bisk, 2018): This involves swapping adjacent characters within words to simulate typos that might occur during typing. For example, "example" might become "exmaple". OCR Replace: Simulating errors commonly introduced by Optical Character Recognition (OCR) software when digitizing text. Characters that look similar, like 'o' and '0' or '1' and '1', might be substituted for one another. Delete: Randomly removing characters from words to mimic typographical errors or omissions. Insert: Adding extra characters into words at random positions, simulating common typos or spelling errors. Substitute: Replacing characters in words with other characters, not necessarily similar in appearance, to create variations in the text.
+
+Word-Level Augmentations. Swap (Wei and Zou, 2019): Changing the positions of two adjacent words in a sentence to add syntactic variability while largely preserving the sentence's meaning. Delete: Removing words from sentences randomly to simulate information loss and encourage the model to learn from incomplete data. Spell Error (Coulombe, 2018): Introducing common spelling mistakes into words to mimic human error and increase the model's exposure to varied spellings. Word2Vector Insert (Morris et al., 2020): Identifying suitable locations in a sentence to insert synonyms or related words based on word embeddings (like word2vec representations), enhancing semantic diversity.
+
+Contextual-Level Augmentations Insert Word using GPT2 (Kumar et al., 2021): Leveraging a pretrained model like GPT2 to generate contextually relevant words to insert into sentences, increasing the complexity and variability of the sentence structures. Substitute Word using BERT (Kumar et al., 2021): Using a model like BERT to identify and replace words with contextually appropriate synonyms or related terms, maintaining the sentence's overall meaning while altering its surface form. Back-Translation (Fadaee et al., 2017): Translating a sentence into another language and then back into the original language. This process often introduces syntactic and lexical variations, providing a paraphrased version of the original sentence that retains its semantic content.
+
+<table><tr><td>Category</td><td>Train/Test</td><td colspan="4">Task Names</td></tr><tr><td rowspan="2">Class→Class</td><td>Train</td><td>financial Phrasebank</td><td>ethos-religion</td><td>glue-wnli</td><td>glue-mrpc</td></tr><tr><td>Test</td><td>tweet.eval-stance_feminist ethos-race superglue-cb</td><td>climate_fever tweet.eval-stance_atheism ethos-national_origin</td><td>poem_sentiment sick medicalquestions_pairs</td><td>tweet.eval-hate glue-rt hateSpeech18</td></tr><tr><td rowspan="2">Class→Non-Class</td><td>Train</td><td>tweet.eval-stance_atheism tweet.eval-stance_feminist glue-rt sick</td><td>superglue-cb climate_fever ethos-national.origin poem_sentiment</td><td>financial Phrasebank glue-mrpc glue-wnli hateSpeech18</td><td>ethos-religion tweet.eval-hate medical/questions_pairs ethos-race</td></tr><tr><td>Test</td><td>quarel superglue-copa quartz-with knowledge</td><td>codah qasc commonsense_qa</td><td>ai2.arc quartz-no_knowledge</td><td>openbookqa dream</td></tr><tr><td rowspan="2">Non-Class→Class</td><td>Train</td><td>quarel superglue-copa quartz-with knowledge</td><td>codah qasc commonsense_qa</td><td>ai2.arc quartz-no_knowledge</td><td>openbookqa dream</td></tr><tr><td>Test</td><td>tweet.eval-stance_atheism tweet.eval-stance_feminist glue-rtseick</td><td>superglue-cb climate_fever ethos-national.origin poem_sentiment</td><td>financial Phrasebank glue-mrpc glue-wnli hateSpeech18</td><td>ethos-religion tweet.eval-hate medical/questions_pairs ethos-race</td></tr><tr><td rowspan="2">Random→Random</td><td>Train</td><td>quartz-with knowledge financial Phrasebank</td><td>tweet.eval-stance_feminist ai2_arc</td><td>codah superglue-cb</td><td>ethos-race</td></tr><tr><td>Test</td><td>quartz-no_knowledge glue-rt gleamrpc climate_fever tweet.eval-stance_atheism</td><td>hateSpeech18 commonsense_qa poem_sentiment tweet.eval-hate openbookqa</td><td>medical这些问题_pairs superglue-copa quarel gasc sick</td><td>ethos-religion ethos-national_orig dream glue-wnli</td></tr></table>
+
+Table 9: All tasks used in this paper. We split them into training and testing sets under different experiment setting.
+
+# F Additional Experiment
+
+We also compare our method with other data augmentation techniques from Non-LLMDA and Manual-LLMDA. The Non-LLMDA includes EDA (Wei and Zou, 2019) and AEDA (Karimi et al., 2021), while Manual-LLMDA includes GPT3Mix (Yoo et al., 2021) and ZeroGen (Ye et al., 2022). As shown in Tab. 10, our proposed method Self-LLMDA significantly outperforms these baseline methods in the Class  $\rightarrow$  Class and Random  $\rightarrow$  Random settings. However, in the Non-Class  $\rightarrow$  Class setting, Self-LLMDA falls behind GPT3Mix. This may indicate suboptimal transferability of Self-LLMDA in this specific scenario. It is worth noting that GPT3Mix is designed specifically for classification tasks, whereas Self-LLMDA can be applied to a wide range of text-related tasks, demonstrating its versatility and broader applicability.
+
+# F.1 Augmentation Instructions of Manual-LLMDA
+
+In Tab. 11, we will represent the manually crafted augmentation instructions. The format of these augmentation instructions is "<method name>: <method instruction>".
+
+# G Self Instructions Generation
+
+The instructions automatically generated by LLM is shown in Tab. 12 and Tab. 13.
+
+<table><tr><td rowspan="2">TextualDA</td><td colspan="2">Class→Class</td><td colspan="2">Class→Non-Class</td><td colspan="2">Non-Class→Class</td><td colspan="2">Random→Random</td></tr><tr><td>125m</td><td>350m</td><td>125m</td><td>350m</td><td>125m</td><td>350m</td><td>125m</td><td>350m</td></tr><tr><td>EDA</td><td>45.28</td><td>44.55</td><td>39.34</td><td>41.27</td><td>47.60</td><td>46.97</td><td>43.89</td><td>44.16</td></tr><tr><td>AEDA</td><td>43.85</td><td>42.92</td><td>39.56</td><td>41.53</td><td>42.61</td><td>43.50</td><td>42.61</td><td>43.50</td></tr><tr><td>ZeroGen-LLM</td><td>45.66</td><td>45.81</td><td>40.43</td><td>40.62</td><td>47.69</td><td>48.58</td><td>44.14</td><td>44.74</td></tr><tr><td>GPT3Mix</td><td>51.62</td><td>51.36</td><td>-</td><td>-</td><td>55.81</td><td>54.04</td><td>-</td><td>-</td></tr><tr><td>Self-LLMDA</td><td>51.72</td><td>54.98</td><td>40.02</td><td>42.80</td><td>50.00</td><td>52.75</td><td>46.64</td><td>48.83</td></tr></table>
+
+Table 10: Performance comparison with other non-LLM-based and LLM-based textual data augmentations.  
+
+<table><tr><td>Name</td><td>Augmentation Instruction</td></tr><tr><td>Word Replace</td><td>Synonym Replacement: Replace certain words in the text with their synonyms while keeping the sentence structure intact. This can be done using pre-built synonym databases or word embeddings.</td></tr><tr><td>Back Translation</td><td>Back Translation: Translate the text into another language using machine translation and then translate it back to the original language. This process introduces variations in the sentence structure and wording.</td></tr><tr><td>Paraphrase</td><td>Paraphrase: Render the same text in different words without losing the meaning of the text itself. More often than not, a paraphrased text can convey its meaning better than the original words.</td></tr><tr><td>Word Insert/Delete</td><td>Random Insertion/Deletion: Randomly insert or delete words in the text to create new variations of the original sentences.</td></tr><tr><td>Word Swap</td><td>Random Swapping: Randomly swap the positions of words in the text to create new sentence arrangements.</td></tr><tr><td>Mask Predict</td><td>Masking/Prediction: Mask certain words in the text and train the model to predict those masked words. This is similar to the concept of masked language modeling used in models like BERT.</td></tr><tr><td>Char. Perturb</td><td>Character-level Perturbation: Instead of operating at the word level, perform data augmentation at the character level. This can involve randomly replacing, inserting, or deleting characters within the text, leading to novel variations.</td></tr><tr><td>Sentence Reorder</td><td>Sentence Reordering: Randomly reorder the sentences in the text while maintaining the coherence of the overall passage. This can help the model become more robust in understanding different sentence arrangements.</td></tr><tr><td>Contextual Replace</td><td>Contextual Synonym Replacement: Instead of blindly replacing words with synonyms, consider the context of the sentence to choose appropriate synonyms. This can be achieved by using contextual word embeddings or language models like ELMo or BERT.</td></tr><tr><td>POS Augment</td><td>Part-of-Speech Augmentation: Identify the part-of-speech tags of words in the text and replace words with synonyms that have the same part-of-speech. This ensures that the grammatical structure of the sentence remains intact.</td></tr><tr><td>Grammar Transform</td><td>Grammar Transformation: Apply various grammar rules to the text, such as changing active voice to passive voice, transforming affirmative sentences to negative, or converting declarative sentences into questions.</td></tr><tr><td>Data Mix</td><td>Data Mixing: Combine two or more texts from different sources to create a new mixed-text data point. This can introduce diversity in the content and writing style.</td></tr><tr><td>Paragraph Shuffle</td><td>Paragraph Shuffling: Shuffle the order of paragraphs in longer texts to create new document structures. This can be particularly useful for tasks that involve document-level understanding.</td></tr></table>
+
+Table 11: Manually crafted augmentation instructions.
+
+<table><tr><td>Augmentation Instruction</td></tr><tr><td>Contextual word replacement: replace certain words in the text with contextually similar words. this can be done by using word embeddings to find words that have similar meanings or are used in similar contexts.</td></tr><tr><td>Sentiment flipping: change the sentiment of the text by flipping positive sentiments to negative and vice versa. this can help generate diverse data for sentiment analysis tasks.</td></tr><tr><td>Style transfer: transform the writing style of the text while maintaining its original content. this can involve converting formal language to informal, changing the tone, or adapting the text to a specific genre.</td></tr><tr><td>Contextual paraphrasing: paraphrase sentences in the text while considering the surrounding context. this ensures that the paraphrased sentences maintain the same meaning within the given context.</td></tr><tr><td>Domain adaptation: modify the text to make it more suitable for a different domain or topic. this can involve replacing domain-specific terms, adjusting terminology, or incorporating relevant keywords.</td></tr><tr><td>Text summarization: generate a summary of the text by condensing its main points into a shorter version. this can help create diverse data for summarization tasks and provide alternative perspectives on the original text.</td></tr><tr><td>Contextual word insertion: insert new words into the text that are contextually relevant and maintain the semantic meaning of the sentence. this can be done by using word embeddings to find words that are commonly used in similar contexts.</td></tr><tr><td>Text paraphrasing: paraphrase the text by rephrasing sentences while preserving the original meaning. this can be done using techniques such as sentence splitting, word substitution, and sentence rearrangement.</td></tr><tr><td>Contextual sentence deletion: delete certain sentences from the text while maintaining the coherence and semantic meaning of the remaining sentences. this can help create diverse data for document summarization tasks.</td></tr><tr><td>Text expansion: expand the text by adding additional details, examples, or explanations to enhance its content. this can be done by incorporating information from external sources or generating new sentences based on the existing text.</td></tr><tr><td>Emotion augmentation: add emotional expressions or sentiments to the text to convey different emotions. this can help create diverse data for emotion classification or sentiment analysis tasks.</td></tr><tr><td>Named entity substitution: replace named entities (such as names of people, organizations, or locations) in the text with other similar entities. this can introduce variations while maintaining the context of the sentence.</td></tr><tr><td>Sentence reordering: rearrange the order of sentences within the text while keeping the semantic coherence intact. this can create new narrative structures and perspectives.</td></tr><tr><td>Grammatical error injection: introduce grammatical errors into the text by randomly modifying verb tenses, subject-verb agreement, or punctuation. this can simulate noisy real-world data and improve model robustness.</td></tr><tr><td>Word embedding replacement: replace certain words in the text with their word embeddings. this can introduce variations while preserving the semantic meaning of the sentence.</td></tr><tr><td>Synonym replacement: replace words in the text with their synonyms. this can create alternative phrasing while maintaining the overall meaning.</td></tr><tr><td>Contextual word substitution: substitute words in the text with other words that have similar contextual meanings. this ensures that the replacement maintains the intended semantics.</td></tr><tr><td>Sentence splitting: split longer sentences into shorter ones, or vice versa, to create new sentence structures. this can help the model understand different sentence lengths and improve its generalization ability.</td></tr><tr><td>Sentence combination: combine multiple shorter sentences into a single longer sentence or vice versa. this can create diverse sentence structures and test the models comprehension of complex sentences.</td></tr><tr><td>Negation insertion: introduce negations into the text by adding words like &quot;not&quot; or &quot;no&quot; to change the polarity of certain statements. this can help the model understand negative contexts better.</td></tr><tr><td>Word masking: randomly mask certain words in the text by replacing them with a special token. this forces the model to rely on the surrounding context to understand the meaning of the masked word.</td></tr><tr><td>Character-level augmentation: modify individual characters within words, such as changing vowels or consonants, to generate diverse textual variations.</td></tr><tr><td>Sentence shuffling: shuffle the order of sentences within a paragraph or document to create new arrangements and test the model&#x27;s ability to understand different contexts.</td></tr><tr><td>Paraphrasing: rewrite the text using different phrasing or sentence structures while maintaining the original meaning.</td></tr><tr><td>Numerical value perturbation: add or subtract small random values to numerical values in the text to create slight variations.</td></tr><tr><td>Pos tagging variation: randomly change the part-of-speech tags of words in the text while ensuring grammatical correctness. this can introduce different syntactic patterns and word usages.</td></tr><tr><td>Sentence deletion: remove one or more sentences from the text to create a shorter version while still maintaining coherence and semantic meaning.</td></tr><tr><td>Sentiment modification: change the sentiment or emotion expressed in the text while keeping the content intact. this can involve altering positive statements to negative ones or vice versa.</td></tr><tr><td>Negation addition: add negation words (e.g., &quot;not,&quot; &quot;never&quot;) to the text to create negative versions of the original sentences.</td></tr><tr><td>Word order variation: randomly change the order of words within phrases or clauses in the text to generate new sentence arrangements.</td></tr><tr><td>Word sense disambiguation: replace ambiguous words in the text with their different senses to create diverse interpretations of the sentence.</td></tr></table>
+
+Table 12: Automatic generated augmentation instructions.
+
+<table><tr><td>Augmentation Instruction</td></tr><tr><td>Negation transformation: transform positive statements into negative ones by adding negation words or phrases, or vice versa.</td></tr><tr><td>Contradiction generation: introduce contradictions within the text by modifying certain statements to be opposite to what they originally conveyed.</td></tr><tr><td>Named entity replacement: identify named entities in the text (e.g., names of people, organizations) and replace them with similar entities to generate new variations.</td></tr><tr><td>Word order shuffling: randomly shuffle the order of words within a sentence to create new sentence structures.</td></tr><tr><td>Grammar error injection: introduce grammatical errors such as incorrect verb conjugation, subject-verb agreement, or punctuation mistakes to simulate natural language variation.</td></tr><tr><td>Text shuffling: shuffle the order of sentences within a paragraph or paragraphs within a document to create new arrangements. this helps diversify the structure and flow of the text.</td></tr><tr><td>Contextual deletion: remove certain words or phrases from the text while ensuring that the remaining content still conveys the same overall meaning. this tests the model&#x27;s ability to understand and fill in missing information.</td></tr><tr><td>Sentence paraphrasing: generate paraphrases of the original sentences while preserving their meaning. this can be achieved through techniques like back-translation or paraphrase generation models.</td></tr><tr><td>Named entity variation: replace named entities (such as names, locations, organizations) in the text with different variations to create diverse instances of the same sentence.</td></tr><tr><td>Numerical variation: modify numerical values in the text by adding/subtracting a small random value or replacing them with synonyms/alternative representations.</td></tr><tr><td>Negation augmentation: introduce negations in the text by adding &quot;not&quot; or other negation words to certain phrases or sentences. this helps the model understand negative contexts better.</td></tr><tr><td>Data sampling: randomly sample subsets of the data to create smaller training sets for faster experimentation and exploration of different data distributions.</td></tr><tr><td>Backtranslation: translate the text into another language and then translate it back to the original language. this can introduce variations in sentence structure and word choice while preserving the overall meaning.</td></tr><tr><td>Sentence splitting/merging: split long sentences into shorter ones or merge short sentences into longer ones to create new sentence structures.</td></tr><tr><td>Pronoun replacement: replace pronouns in the text with their corresponding nouns or vice versa, maintaining the semantic meaning of the sentence.</td></tr><tr><td>Word deletion: randomly delete certain words from the text to create shorter or more concise sentences. this can simulate scenarios where some information is missing or incomplete.</td></tr><tr><td>Character-level perturbation: introduce noise at the character level by randomly changing, deleting, or inserting characters in the text. this can help models become more robust to noisy inputs and improve generalization.</td></tr><tr><td>Domain-specific term replacement: identify domain-specific terms in the text and replace them with synonyms or related terms specific to another domain. this can help models generalize better across different domains.</td></tr><tr><td>Negation/positive conversion: convert negative statements to positive ones or vice versa to generate different perspectives or sentiments.</td></tr><tr><td>Part-of-speech tagging: modify the part-of-speech tags of words in the text to create new grammatical arrangements and sentence structures.</td></tr></table>
+
+Table 13: Automatic generated augmentation instructions.

2024/Empowering Large Language Models for Textual Data Augmentation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64c34c3352cad11116c80a9b15504558f1dd1dbd6ce2711f986ca6cb32199d47
+size 1823916

2024/Empowering cross-lingual abilities of instruction-tuned large language models by translation-following demonstrations/f7644dfc-cd06-4320-a14e-02a1e95a33d9_content_list.json

@@ -0,0 +1,1779 @@
+[
+    {
+        "type": "text",
+        "text": "Empowering Cross-lingual Abilities of Instruction-tuned Large Language Models by Translation-following Demonstrations",
+        "text_level": 1,
+        "bbox": [
+            117,
+            87,
+            878,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Leonardo Ranaldi  $(\\dagger)$ , Giulia Pucci  $(\\star)$ , André Freitas  $(\\dagger, *)$",
+        "bbox": [
+            257,
+            140,
+            739,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$(\\dagger)$  Idiap Research Institute, Martigny, Switzerland",
+        "bbox": [
+            294,
+            159,
+            707,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$(\\star)$  Department of Computing Science, University of Aberdeen, UK",
+        "bbox": [
+            225,
+            175,
+            774,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$(\\ast)$ Department of Computer Science, University of Manchester, UK",
+        "bbox": [
+            225,
+            192,
+            774,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "[first_name].[last_name]@idiap.ch",
+        "bbox": [
+            331,
+            210,
+            663,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The language ability of Large Language Models (LLMs) is often unbalanced towards English because of the imbalance in the distribution of the pre-training data. This disparity is demanded in further fine-tuning and affecting the cross-lingual abilities of LLMs. In this paper, we propose to empower Instruction-tuned LLMs (It-LLMs) in languages other than English by building semantic alignment between them. Hence, we propose CrossAlpaca, an It-LLM with cross-lingual Instruction-following and Translation-following demonstrations to improve semantic alignment between languages. We validate our approach on the multilingual Question Answering (QA) benchmarks XQUAD and MLQA and adapted versions of MMLU and BBH. Our models, tested over six different languages, outperform the It-LLMs tuned on monolingual data. The final results show that instruction tuning on non-English data is not enough and that semantic alignment can be further improved by Translation-following demonstrations.",
+        "bbox": [
+            141,
+            279,
+            460,
+            606
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            619,
+            258,
+            634
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large Language Models (LLMs) achieve comprehensive language abilities through pre-training on large corpora (Brown et al., 2020; Touvron et al., 2023; Ranaldi and Pucci, 2023b). Hence, the acquired language abilities follow the corpora features, mostly available in English (Lin et al., 2021; Pucci and Ranaldi, 2024). This phenomenon produces an imbalance in both pre-training (Blevins and Zettlemoyer, 2022) and fine-tuning (Le et al., 2021). Consequently, LLMs underperform in languages different from English (Huang et al., 2023).",
+        "bbox": [
+            112,
+            645,
+            489,
+            821
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Efforts to increase multilingual abilities propose continuing pre-training with large-scale monolingual data (Imani et al., 2023; Cui et al., 2023; Yang et al., 2023). However, learning a language from monolingual data requires considerable data and computational resources.",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose CrossAlpaca, an Instruction-tuned LLM (It-LLM) empowered with a semantic alignment between English and other languages. We analyze how to elicit the non-English abilities of It-LLMs by focusing on the crucial phase: instruction-tuning over Instruction-following demonstrations. Hence, we study the effect of cross-lingual alignment by proposing Translation-following demonstrations to improve the instruction-tuning phase.",
+        "bbox": [
+            507,
+            253,
+            884,
+            413
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In our experiments, we use  $LLaMA-7B$  (Touvron et al., 2023) as the LLM backbone and consider six target languages selected from among the available data (shown in Table 4); where data are absent, we perform a translation task. As Instruction-following demonstrations, we use the Stanford Alpaca dataset (Taori et al., 2023) and translated versions in the corresponding languages, while for the Translation-following, we use a publicly available translation resource news Commentary (Tiedemann, 2012), the most accessible and extendable to multiple languages (i.e., CrossAlpaca demonstrations on Figure 1).",
+        "bbox": [
+            507,
+            416,
+            884,
+            625
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Behind an instruction-tuning phase, we evaluate the performance of our six language-specific CrossAlpacas using four different benchmarks: two native-multilingual, i.e., XQUAD (Artetxe et al., 2019) and MLQA (Lewis et al., 2020), and two native-monolinguals, MMLU (Hendrycks et al., 2021) and BBH (Suzgun et al., 2022). The results show that CrossAlpacas instructed with language-specific instruction and translation data far outperform Alpacas instructed only with non-English demonstrations. However, even though CrossAlpaca reduces the gap, the Translation-following models of the original Alpaca still perform best. This result shows that LLaMA's learning abilities on English data are superior to those on non-English data. The semantic alignment task also improves the cross-lingual abilities of It-LLMs.",
+        "bbox": [
+            507,
+            627,
+            884,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Our findings can be summarized as follows:",
+        "bbox": [
+            527,
+            903,
+            853,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "7961",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2024, pages 7961-7973",
+        "bbox": [
+            228,
+            945,
+            766,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "August 11-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            284,
+            958,
+            712,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f77b86829fd878d10ff49daafa9ab637bf2b7f9343234db1ad3e8a9a29f08489.jpg",
+        "image_caption": [
+            "Figure 1: Our  $x$ -CrossAlpacas are fine-tuned on Instruction-following and Translation-following demonstrations. This example shows the it-CrossAlpaca, fine-tuned on it-Alpaca and Translation-following demonstrations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            243,
+            137,
+            909,
+            281
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The learning abilities of LLMs on non-English Instruction-tuning tasks are limited;",
+            "- The multi-lingual abilities of Instruction-tuned LLMs could be empowered through cross-lingual alignment;",
+            "- Thus, we propose CrossAlpaca, an Instruction-tuning approach for non-English models that are based on Instruction-following and Translation-following demonstration for target language;",
+            "- We show that It-LLMs are able to semantically align through cross-lingual Translation-following demonstrations via an extensive evaluation on four multi-lingual QA benchmarks."
+        ],
+        "bbox": [
+            136,
+            346,
+            489,
+            618
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            631,
+            270,
+            646
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Pre-trained Language Model on Cross-Linguual Corpora",
+        "text_level": 1,
+        "bbox": [
+            112,
+            657,
+            411,
+            689
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The next token prediction based on the prefix sequence, as well-known as language modeling, is the everlasting task of modern NLP (Tenney et al., 2019). The extensive knowledge of today's Large Language Models (LLMs) depends on the billions of neurons trained on large-scale corpora with derivatives of the language modeling task. Consequently, the pre-training corpora are predominantly in English, e.g., BooksCorpus (Zhu et al., 2015), MEGATRON-LM (Shoeybi et al., 2019), Gutenberg Dataset (Lahiri, 2014) PILE (Gao et al., 2020), C4 (Dodge et al., 2021), RefinedWeb (Penedo et al., 2023); therefore, LLMs usually have much better knowledge of English than other languages.",
+        "bbox": [
+            112,
+            694,
+            489,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Aulamo and Tiedemann (2019); Abadji et al. (2022), in order to solve this problem, propose forward corpora translated into several languages. However, these corpora are not as huge as their competitors, and the lack of massively parallel data in the pre-training corpora also prevents LLMs from aligning the different languages well (Li et al., 2023).",
+        "bbox": [
+            507,
+            346,
+            884,
+            473
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 The Instruction-tuning Paradigm",
+        "text_level": 1,
+        "bbox": [
+            507,
+            485,
+            816,
+            500
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Ouyang et al. (2022); Wei et al. (2022) fine-tuned LLMs using the Instruction-tuning method based on Instruction-tuning data, which are instruction-response corpora, to make LLMs more scalable and improve zero-shot performance. In this method, the LLM backbone is fed with data from the instruction  $(I,X,Y)$ , where  $I$  is an instruction describing the task's requirements,  $X$  is the input, which can be optional, and  $Y$  is the output for the given task. The goal of this method is to minimize the function  $f(Y)$ :",
+        "bbox": [
+            507,
+            506,
+            882,
+            682
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (Y) = \\arg \\min  _ {\\theta} \\log p _ {\\theta} (Y \\mid I, X) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            694,
+            882,
+            715
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta$  are model learnable parameters.",
+        "bbox": [
+            507,
+            726,
+            810,
+            741
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Earlier studies show that the instruction-tuning method of LLMs with both human (Wang et al., 2023) and synthetic-generated instructions (Taori et al., 2023; Xu et al., 2023) empowers the ability of LLMs to solve considerable tasks in zero-shot scenarios.",
+        "bbox": [
+            507,
+            741,
+            882,
+            837
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, we state that the generally used instruction-tuning datasets, alpaca (Taori et al., 2023), Self-Instruct (Wang et al., 2023), Self-Chat (Xu et al., 2023) or a teacher-student alignment approaches (Ranaldi and Freitas, 2024), conceived",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "header",
+        "text": "Instruction-following (original)",
+        "bbox": [
+            171,
+            83,
+            383,
+            95
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "header",
+        "text": "Instruction: Classify the sentence as true or false.  \nInput: Sentence: the Earth is flat.  \nOutput: False.",
+        "bbox": [
+            144,
+            99,
+            413,
+            124
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "header",
+        "text": "Stanford Alpaca",
+        "bbox": [
+            433,
+            87,
+            492,
+            110
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "header",
+        "text": "#",
+        "bbox": [
+            495,
+            86,
+            532,
+            112
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "7962",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/2b168cb7e40c1e66c87ac707be95f1fbb15447888cbfc812e0b1a71b5f5cbef0.jpg",
+        "image_caption": [
+            "Figure 2: Examples of Translation-following demonstrations. In this particular example, there are two demonstrations with two different directions from English to German (en-x) and vice versa."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            245,
+            83,
+            786,
+            259
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "in English, which limits the prospect of LLMs to follow non-English instructions and therefore solve related tasks.",
+        "bbox": [
+            112,
+            322,
+            487,
+            369
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Instruction-tuning is at hand",
+        "text_level": 1,
+        "bbox": [
+            112,
+            382,
+            386,
+            397
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Large Language Models have considerable success with many techniques in vogue, such as Instruction-tuning. However, their prohibitive size only allows part of the scientific community to experiment with these models.",
+        "bbox": [
+            112,
+            403,
+            487,
+            482
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The latest advances that make these models and techniques accessible involve efficient parameter tuning. Parameter-Efficient Tuning (PEFT) is an efficient technique to adjust a small part of the model parameters and freeze the others. The main goal is to significantly reduce computational and storage costs while maintaining the performance of the original models. The standard techniques for PEFT are LoRA (Hu et al., 2021a), Prefix Tuning (Li and Liang, 2021), P-Tuning (Liu et al., 2022). The basic idea is to keep the pre-trained model weights and incorporate low-rank matrices in each architecture layer. This approach significantly reduces the number of parameters that require training for subsequent tasks, thereby increasing overall efficiency. This building block has been and continues to be very important because it skills the scientific community to fair research and enables the development of multiple open-source works.",
+        "bbox": [
+            112,
+            483,
+            489,
+            789
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.4 Cross-lingual Instruction-tuning Challenge",
+        "text_level": 1,
+        "bbox": [
+            112,
+            801,
+            415,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recent work has shown the remarkable abilities of LLMs in learning instruction in different languages. Santilli and Rodola (2023); Chen et al. (2023) proposed monolingual fine-tuning LLaMA adaptations on language-specific translated instructions. The",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/073682b8cd52959b51552dadcd521b1fa629e07d4c1d8ba49c28403fa222cb84.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Language</td><td>Name</td></tr><tr><td>Alpaca (Taori et al., 2023)</td><td>English</td><td>en-Alpaca</td></tr><tr><td>Alpaca-Chinese (Chen et al., 2023)</td><td>Chinese</td><td>zh-Alpaca</td></tr><tr><td>Arabic Alpaca (Yasbok)</td><td>Arabic</td><td>ar-Alpaca</td></tr><tr><td>Camascio (Santilli and Rodola, 2023)</td><td>Italian</td><td>it-Alpaca</td></tr><tr><td>Guanaco (Kohaku-Blueleaf, 2023)</td><td>Spanish</td><td>es-Alpaca</td></tr><tr><td>German Alpaca (Thissen, 2023)</td><td>German</td><td>de-Alpaca</td></tr></table>",
+        "bbox": [
+            546,
+            318,
+            847,
+            381
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Details and names of mono-lingual Instruction-tuned Large Language Models that use a language-specific version of Alpaca as instruction-tuning data.",
+        "bbox": [
+            507,
+            390,
+            884,
+            432
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "use of optimization techniques introduces in Section 2.3, to propose engaging custom adapters for various tasks. Zhang et al. (2023); Ranaldi and Pucci (2023a) investigated the cross-lingual abilities of It-LLMs by promoting the effects of augmenting instruction demonstrations. In particular, Zhang et al. (2023) use fine-grained custom translations that are not easily accessible, while Ranaldi and Pucci (2023a) propose a multilingual approach by considerably increasing instruction and translation tasks. However, these studies have opened an interesting avenue in analyzing the cross-lingual abilities of It-LLMs.",
+        "bbox": [
+            507,
+            458,
+            882,
+            668
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we take the next step by proposing our CrossAlpaca. Our methods show of the cross-lingual learning abilities of It-LLMs on a large scale. Using four well-known benchmarks, we show that the weaknesses of It-LLMs trained on non-English data can be strengthened with cross-lingual alignment approaches. Hence, our analyses aim to understand the role of Instruction-following and Translation-following demonstrations in closing the gap in learning and adapting LLMs' abilities to non-English languages.",
+        "bbox": [
+            507,
+            670,
+            882,
+            845
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Data & Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            860,
+            687,
+            873
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Pre-training from scratch a Large Language Model (LLM) in multiple languages is cost-prohibitive for",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "7963",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "data collection and parameter learning. This is the reason why the trend is to do further fine-tuning to strengthen the models' abilities in a specific language (Tanti et al., 2021; Moslem et al., 2023). In this paper, we aim to exploit the abilities of pretrained LLMs for non-English languages by further improving the alignment between English and the target language. Hence, in Section 3.1, we introduce the challenge of fine-tuning an LLM on a mono-lingual custom scenario. Furthermore, in Section 3.2, we propose our approach to fine-tune cross-lingual LLMs.",
+        "bbox": [
+            112,
+            84,
+            492,
+            277
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 Mono-lingual Instruction-tuning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            291,
+            416,
+            307
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The limited accessibility and transparency of paid API services of state-of-the-art LLMs have pushed research toward developing open-source models. Using the instruction-tuning paradigm (Section 2.2) and Stanford Alpaca (Taori et al., 2023), a corpus consisting of  $52\\mathrm{k}$  of English instruction-output pairs generated by text-davinci-003, several Instruction-tuned versions of LLaMA were released.",
+        "bbox": [
+            112,
+            313,
+            489,
+            456
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Following this approach, multiple monolingual Instruction-tuned versions of LLaMA were proposed by translating the Stanford Alpaca Instruction-following data into the specific language. Table 4 (Appendix B) shows a set of alpaca versions available as open source. Following a systematic analysis of the translated versions of Alpaca in official repositories<sup>1</sup>, the languages of the benchmark datasets, and the translation pairs present in news/commentary, which will be introduced later, we selected the languages that share the most already available data. Table 1 shows the custom versions used in this work, which for simplicity will be renamed  $x$ -Alpaca, where  $x$  indicates the specific language.",
+        "bbox": [
+            112,
+            458,
+            489,
+            701
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Cross-lingual Instruction-tuning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            714,
+            415,
+            730
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Mono-lingual Instruction-tuning approaches (Section 3.1) reward LLMs' multi-lingual abilities. However, the use of translated Alpaca alone for the specific language is not sufficient to elicit the non-English ability of LLMs. For this reason, we propose CrossAlpaca (Figure 1), which enriches the cross-lingual Instruction-tuning with the Translation-following demonstrations. We aim to highlight LLMs' English and non-English abilities",
+        "bbox": [
+            112,
+            736,
+            489,
+            881
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "by proposing a semantic alignment task to enrich versions of Alpaca.",
+        "bbox": [
+            507,
+            84,
+            882,
+            116
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Instruction-following demonstrations The original version of Alpaca is in English. However, as described in Section 3.1, different open-source translations are produced with a translation engine. In our experiments, we propose the Instruction-tuning phase with the original English version (en-Alpaca) and the translated language-specific versions. Moreover, we propose the CrossAlpaca versions built with the different Alpacas translated into the specific language and the cross-lingual translations (introduced later). In this way, we investigate the abilities of the LLM backbone in understanding multilingual instructions and cross-lingual alignment.",
+        "bbox": [
+            507,
+            127,
+            885,
+            351
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Translation-following demonstrations Learning general instruction data is a sensible strategy for building models to solve multi-tasks directed by instructions (Wang et al., 2023; Zeng et al., 2023). However, translation data could contribute to learning semantic alignment. Zhang et al. (2023) showed that the translation performance of LLMs could be improved by using expertly annotated translation data, and Zhu et al. (2023) showed that it could be beneficial for instruction-tuning.",
+        "bbox": [
+            507,
+            363,
+            884,
+            523
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inspired by Zhu et al. (2023), we use publicly available sentence-level translation datasets, such as news Commentary (Tiedemann, 2012), to construct the translation task instruction demonstrations. We also propose extending this to additional languages, which we release as an open-source dataset. In particular, for each specific language, we constructed 20k demonstrations. Hence, following the Alpaca style (Instruction, Input, and Output) (see Figure 1), we randomly selected 10k English to non-English translations and 10k random non-English to English translations (all translations come from news commentary). We have constructed these correspondences for the languages mentioned above in Section 3.1. Figure 2 shows a case of Translation-following for English-German direction (en-de) and vice versa (de-en).",
+        "bbox": [
+            507,
+            524,
+            885,
+            797
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            810,
+            655,
+            828
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to observe the English and non-English abilities of Large Language Models (LLMs) and the impact of the Instruction-tuning approach in cross-lingual scenarios, we propose CrossAlpaca. Our approach is based on instruction-tuning on",
+        "bbox": [
+            507,
+            839,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "1official versions on https://github.com/tloen/ alpaca-lora and https://huggingface.co/models",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "7964",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/14c6bd50518b7dad6a053ed9076ed641c959d0781773a676e56ce2d278bfd5ee.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            86,
+            492,
+            239
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/2307526836f0fcee0a0a3e4889ec62229154f0fed6fa8c2f2202bb61cfd60840.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            505,
+            86,
+            838,
+            239
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7779c4f54976419118b8e6f202bfe60808a9c74554d0c25a2d2682e01cdd6559.jpg",
+        "image_caption": [
+            "Figure 3: Evaluation results on proposed benchmarks. The dotted line represents the performance of the original Alpaca (Taori et al., 2023) on English tasks."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            157,
+            249,
+            489,
+            401
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/b7b4db4804e084e1cbac624dc9cc564401c2a46611542b195046a0b9f7d6b947.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            505,
+            250,
+            836,
+            401
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "language-specific data augmented with a crosslingual semantic alignment. Hence, we set several baseline models (Section 4.1), which we augmented with our CrossAlpaca approach (Section 4.2). Finally, we performed a series of systematic evaluations (Section 4.3.1) to observe the impact of the proposed intervention.",
+        "bbox": [
+            112,
+            473,
+            489,
+            585
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Baseline Instruction-tuned LLMs",
+        "text_level": 1,
+        "bbox": [
+            112,
+            599,
+            421,
+            614
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The common denominator among the It-LLMs shown in Table 1 is the LLM backbone, LLaMA7B (Touvron et al., 2023). Starting from Instruction-following data from the original Alpaca (Taori et al., 2023) and its open-source non-English versions $^{2}$ , we reproduced  $x$ -Alpaca for  $x$  specific languages: Chinese (zh), Italian (it), Arabic (ar), Spanish (es), German (de) and the original English version (en).",
+        "bbox": [
+            112,
+            621,
+            489,
+            765
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Cross-lingual Instruction-tuned LLMs",
+        "text_level": 1,
+        "bbox": [
+            112,
+            780,
+            463,
+            796
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The CrossAlpacas are instruction-tuned on Instruction-following and Translation-following demonstrations (CrossAlpacas demonstrations). The first ones stem from the resources introduced in Section 4.1. The second comes from",
+        "bbox": [
+            112,
+            803,
+            489,
+            883
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "news COMMENTary (Tiedemann, 2012).",
+        "bbox": [
+            507,
+            473,
+            800,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our approach generates a series of instruction-tuned versions of the data shown in Figure 1. We have named the versions  $x$ -CrossAlpaca where  $x$  denotes Chinese (zh), Arabic (ar), Italian (it), Spanish (es), and German (de).",
+        "bbox": [
+            507,
+            489,
+            884,
+            569
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            580,
+            714,
+            596
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In order to assess the performance of the CrossAlpaca, we defined several benchmarks (Section 4.3.1) on which we applied systematic tuning (Section 4.3.2) and evaluation (Section 4.3.3) pipelines.",
+        "bbox": [
+            505,
+            601,
+            884,
+            665
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3.1 Benchmarks",
+        "text_level": 1,
+        "bbox": [
+            507,
+            674,
+            670,
+            688
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the performance of the It-LLMs and the impact of the translation-based semantic alignment approach, we used two cross-lingual (XQUAD (Artetxe et al., 2019), MLQA (Lewis et al., 2020)) and two multi-task (MMLU (Hendrycks et al., 2021) and BBH (Suzgun et al., 2022)) benchmarks. XQUAD and MLQA focus on understanding questions and answers through translation into different languages. MMLU and BBH, being multi-task benchmarks, include subtasks related to Boolean expressions and QA on basic-level subjects (e.g., chemistry, physics). However, we decided to introduce them to observe whether our approach degrades performance in these tasks. The first two",
+        "bbox": [
+            505,
+            694,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "2open-source code is available on https://github.com/tloen/alpaca-lora",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "7965",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "datasets selected are appropriately constructed for multi-language testing, while the second two are available only in English. So we do a preliminary translation step as outlined below.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "MultiLingual Question Answering (MLQA)",
+        "text_level": 1,
+        "bbox": [
+            112,
+            162,
+            489,
+            178
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "(Lewis et al., 2020) evaluates cross-lingual question answering performance. The benchmark comprises over 5K extractive QA instances in the SQuAD (Rajpurkar et al., 2016) format in several languages. MLQA is highly parallel, with QA instances aligned across four languages on average. Although comprising different languages, some languages are not represented, such as Italian. To conduct the experiments uniformly, we have translated the examples as also done in the forthcoming MMLU and BBH.",
+        "bbox": [
+            112,
+            179,
+            489,
+            355
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Cross-lingual Question Answering Dataset (XQUAD) (Artetxe et al., 2019) consists of a subset of 240 paragraphs and 1190 question-answer pairs from the development set of SQuAD v1.1 (Rajpurkar et al., 2016) with their manual translations into several languages. Consequently, the dataset is entirely parallel across 11 languages.",
+        "bbox": [
+            112,
+            370,
+            489,
+            483
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Massive Multitask Language Understanding (MMLU) (Hendrycks et al., 2021) measures knowledge of the world and problem-solving problems in multiple subjects with 57 subjects across STEM, humanities, social sciences, and other areas. The benchmark is native in English; however, we translated it into five additional languages<sup>3</sup>.",
+        "bbox": [
+            112,
+            498,
+            489,
+            609
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "BIG-Bench Hard (BBH) (Suzgun et al., 2022) is a subset of challenging tasks related to navigation, logical deduction, and fallacy detection. Here again, the benchmark is native English, and we have translated it into five languages<sup>3</sup>.",
+        "bbox": [
+            112,
+            625,
+            489,
+            705
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3.2 Models Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            720,
+            282,
+            736
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In order to align the results with the state-of-the-art models, we used the alpaca_LoRA (Hu et al., 2021b) code $^2$ , adopting the same hyperparameters.",
+        "bbox": [
+            112,
+            743,
+            489,
+            790
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We performed the fine-tuning with a single epoch and a batch-size of 128 examples, running our experiments on a workstation equipped with two Nvidia RTX A6000 with 48 GB of VRAM.",
+        "bbox": [
+            112,
+            793,
+            487,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3.3 Evaluation",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            655,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As an evaluation metric, we use accuracy. Hence, we estimate accuracy by measuring exact match values in the zero-shot setting. For each model, the parts of benchmarks related to the specific language are used (e.g., for zh-Alpaca and zh-CrossAlpaca, data from MLQA, XQUAD, MMLU, and BBH in Chinese are used).",
+        "bbox": [
+            507,
+            107,
+            884,
+            218
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            236,
+            608,
+            250
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Improving non-English abilities in Instruction-tuned Large Language Models (It-LLMs) remains challenging. However, the  $x$ -CrossAlpacas revealed improved results in cross-lingual Question Answering (QA) benchmarks and maintained logical-mathematical skills. From the results of Figure 3 (further detailed in Table 3), it is possible to observe the weaknesses emerging from the fine-tuning of the translated versions of Alpaca (Section 5.1), the improvement obtained from the alignment phase is encouraging (Section 5.2) but it is not enough to outperform the English one. Therefore, we investigated the impact of demonstrations on downstream performance (Section 5.3).",
+        "bbox": [
+            507,
+            266,
+            884,
+            489
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The fine-grained analysis highlighted the importance of cross-lingual alignment data and the critical issues with non-English data. This opens the way for new hypotheses regarding the imbalance of pre-training languages and learning abilities via instruction-tuning.",
+        "bbox": [
+            507,
+            493,
+            882,
+            589
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Translating the Alpaca is not the right way",
+        "text_level": 1,
+        "bbox": [
+            507,
+            606,
+            880,
+            621
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The instruction-tuning on LLaMA, predominantly pre-trained in English, affects x-Alpaca. Figure 3 and in terms of numbers, Table 3 show that in both MLQA and XQUAD, there is a gap of 55 and 53 average points between en-Alpaca and the x-Alpacas. This phenomenon is mitigated for MMLU and BBH, where we observed an average gap of 18 and 14 points. Instructing an LLM on Alpaca-style demonstrations translated into different languages is not always a good strategy. However, some x-Alpacas, such as zh-Alpaca and de-Alpaca, have performed better. We hypothesize that this phenomenon is related to the scale of the pre-training data in the respective languages and, thus, the abilities of the LLaMA. In future developments, we plan to extend the study on other LLMs beyond LLaMA to observe whether the phenomenon is similar, milder, or more significant.",
+        "bbox": [
+            507,
+            629,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>We performed translations using the Google translator API from English to Chinese (zh), Italian (it), Arabic (ar), Spanish (es), German (de). Resources will be made available along with the publication",
+        "bbox": [
+            112,
+            868,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "7966",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/25ef11c65468ba47dd11e717a806e1e69369c5fc506224068e386e0e360baf74.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>QA\nTask</td><td>en-Alpaca</td><td>avg-Alpaca</td><td>avg-CrossAlpaca</td><td>δ</td></tr><tr><td>MLQA</td><td>0.89</td><td>0.34</td><td>0.64</td><td>+0.30</td></tr><tr><td>XQUAD</td><td>0.97</td><td>0.31</td><td>0.65</td><td>+0.30</td></tr><tr><td>MMLU</td><td>0.42</td><td>0.24</td><td>0.32</td><td>+0.08</td></tr><tr><td>BBH</td><td>0.30</td><td>0.24</td><td>0.28</td><td>+0.04</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            487,
+            168
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Averages of the results on proposed benchmarks. The column  $\\delta$  indicates the difference between avg-Alpacas and our avg-CrossAlpacas.",
+        "bbox": [
+            112,
+            177,
+            489,
+            219
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 CrossAlpaca: A cross-lingual solution",
+        "text_level": 1,
+        "bbox": [
+            112,
+            247,
+            452,
+            263
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Semantic alignment through Translation-following demonstrations during fine-tuning could have a valuable impact on the cross-lingual abilities of It-LLMs. The  $x$ -CrossAlpacas outperformed the x-Alpacas by 30 average points on MLQA, 34 average points on XQUAD, 8 on MMLU, and 4 on BBH (see Table 2 and more detailed in Table 3). They also brought their performances closer to the sota obtained from the original Alpaca by 25 points on MLQA and 22 average points on XQUAD. In MMLU and BBH, the gap became very close, with averages of 10 and 2 points (see Table 2 or the line 'en-Alpaca vs avg-CrossAlpaca' in Table 3).",
+        "bbox": [
+            112,
+            269,
+            487,
+            478
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Enriching Translation-following demonstrations has outstanding influences on the cross-lingual abilities of the It-LLMs. However, even in this case, Chinese and German models (zh- and de-CrossAlpaca) outperformed Arabic by many points and, in some specific cases, Spanish and Italian as well. This phenomenon, we hypothesize, is related to the diversity in corpus representation within the pre-training data, as shown in (Yang et al., 2023). Therefore, cross-lingual approaches do not have an incisive impact as in languages less present in the pre-training phases of the language model.",
+        "bbox": [
+            112,
+            480,
+            489,
+            674
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            687,
+            280,
+            703
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Our CrossAlpacas, distinguished by the construction of the demonstrations pairs (Section 4.2), achieves significant performance improvements and contributes to closing the gap between the original Alpaca (en-Alpaca) and a series of x-Alpacas in different languages. In order to show the impact of enrichment with cross-lingual demonstrations, we propose two different analyses. In the first analysis, we incrementally decrease the training data, in particular on the Translation-following demonstrations side (Section 5.4). In the second, still working on the Translation-following part (defined by half en-x and half x-en demonstrations), we an",
+        "bbox": [
+            112,
+            709,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a08f94c795058f3708498e8634dde81f022dcbb5670e723cfaf3da442e1a6342.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            556,
+            85,
+            836,
+            218
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/161681fe38c2d95b24c8e37313135ddf3d4616b082402bf0274a128c9684f67d.jpg",
+        "image_caption": [
+            "Figure 4: Evaluation of proposed benchmarks of the demonstrations used for instruction-tuning our CrossAlpacas."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            554,
+            225,
+            836,
+            357
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "alyze the impact of the demonstrations by splitting the experiments into en-x and x-en (Section 5.4.1).",
+        "bbox": [
+            507,
+            442,
+            882,
+            473
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 Translation-following demonstrations empower non-English abilities.",
+        "text_level": 1,
+        "bbox": [
+            507,
+            488,
+            850,
+            520
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 (for a more extensive analysis please refer to Figure 6 in the Appendix) shows the non-English performance of the x-CrossAlpaca on the benchmarks under different scales of Translation-following demonstrations<sup>4</sup>. The x-CrossAlpacas achieve higher accuracy when more demonstrations are used, revealing the benefits of cross-lingual alignment to improve non-English performance. However, Translation-following demonstrations are built in two directions: x-en and en-x (foreign-English and English-foreign). Although equally distributed, we are still determining whether there is an asymmetry between the two types of contributions.",
+        "bbox": [
+            505,
+            526,
+            884,
+            750
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Finally, this result is less pronounced for the multi-task benchmarks (see Figure 6 MMLU and BBH). This could be related to the fact that in the benchmarks, there are logical-mathematical subtasks which are less language dependent (comparatively simpler syntax, smaller vocabularies), and hence, our cross-lingual approach does not influence the results for MLQA and XQUAD as",
+        "bbox": [
+            507,
+            752,
+            882,
+            881
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "4The Traslation-following demonstrations are equally selected in a random way from the 10k en-x and x-en.",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "7967",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "much. Furthermore, it is possible to observe a trend in language groups, particularly Chinese-German, Italian-Spanish, and Arabic. This trend, which recurs frequently in the results of our experiments, will be further investigated in future work.",
+        "bbox": [
+            112,
+            84,
+            489,
+            165
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/bfe2c307a8eced7305a024eef5b448dfe50ac3468bd0a2b7f826a6c8f7f8fa08.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            193,
+            442,
+            326
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ee2382f505e4bd9f77f0855bbcee68ef9431fdc6bf8ef4c70df5db557eea9517.jpg",
+        "image_caption": [
+            "Figure 5: Evaluation of proposed benchmarks using one-direction Translation-following demonstrations. For en-x for English-foreigner and x-en for foreign English."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            332,
+            442,
+            464
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4.1 Demonstration direction matters on non-English abilities",
+        "text_level": 1,
+        "bbox": [
+            112,
+            580,
+            455,
+            612
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 5 shows the evaluation conducted in varying directions of the translation-following demonstrations. In particular, demonstrations with English-foreign direction (en-x) better impact downstream models. Conversely, foreign-English (x-en) demonstrations perform better than baselines but underperform demonstrations in the opposite direction. However, as shown in Figure 5 (and complete Figure 7), the x-CrossAlpacas continue to outperform. Nevertheless, the trend of translation-following demonstrations with one direction is interesting. Again, as was the case in the previous ablation study, the multi-task benchmarks (see MMLU and BBH in Figure 7) do not seem to obtain significant influences, which reinforces the hypothesis that the models are greatly affected by cross-lingual capabilities in tasks where there is a strong presence of natural language.",
+        "bbox": [
+            112,
+            629,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Future Works",
+        "text_level": 1,
+        "bbox": [
+            507,
+            83,
+            665,
+            98
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The cross-lingual abilities of Intruction-tuned Large Language Models (It-LLMs) seem to be supported by LLMs, such as in the case of Alpaca, the LLaMA backbone. However, it seems that low-impact demonstrations at the data level can enrich these abilities. We obtained valuable results from our experiments by proposing strategic demonstrations, i.e., Translation-following demonstrations. These results were proposed by performing fine-tuning on LLaMA-7B as done by Taori et al. (2023).",
+        "bbox": [
+            505,
+            108,
+            884,
+            284
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In future work, we would like to continue to investigate by increasing the number of parameters in LLaMA and including additional backbone models. In addition, it might be interesting to evaluate the impact on low-resource languages as done for in-context settings in (Ranaldi et al., 2024). Hence, we would like to get to the underside of performances obtained in some experiments (see Section 5.4) by extending previous epistemic approaches (Ranaldi et al., 2023a,c,b) to It-LLMs. In parallel, plans include analyzing the translation capabilities of general It-LLMs and those enhanced with translation tasks, including some specialized translation tasks among our evaluation benchmarks. Finally, we would like to investigate the learning abilities of the original Alpaca as the translation data change, proposing different probing experiments on (original) English data enhanced with translations.",
+        "bbox": [
+            507,
+            286,
+            884,
+            575
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            587,
+            640,
+            602
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we proposed CrossAlpaca, an approach to empowering the instruction-tuning of LLMs on non-English data. Specifically, we coupled Instruction-following (Alpaca-style) demonstrations with Translation-following demonstrations. Our method seeks to instruct the LLM to semantic alignment between English and non-English overperforms models instructed on non-English texts. In particular, thanks to our CrossAlpaca demonstrations, the instructed models achieved significant performance improvements on four Question Answering benchmarks XQUAD, MLQA, MMLU, and BBH. In addition, we observe that semantic alignment strengthens with increasing Translation-following data; this demonstrates the de-facto abilities of It-LLMs to learn from instructions. Our approach and results contribute to improved research on the potential for producing more powerful LLMs for non-English languages.",
+        "bbox": [
+            505,
+            612,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "7968",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Julien Abadji, Pedro Ortiz Suarez, Laurent Romary, and Benoit Sagot. 2022. Towards a cleaner document-oriented multilingual crawled corpus. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4344-4355, Marseille, France. European Language Resources Association.",
+            "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2019. On the cross-lingual transferability of monolingual representations. In Annual Meeting of the Association for Computational Linguistics.",
+            "Mikko Aulamo and Jörg Tiedemann. 2019. The OPUS resource repository: An open package for creating parallel corpora and machine translation services. In Proceedings of the 22nd Nordic Conference on Computational Linguistics, pages 389-394, Turku, Finland. Linköping University Electronic Press.",
+            "Terra Blevins and Luke Zettlemoyer. 2022. Language contamination helps explains the cross-lingual capabilities of english pretrained models. In Conference on Empirical Methods in Natural Language Processing.",
+            "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners.",
+            "Wei-Lin Chen, Cheng-Kuang Wu, and Hsin-Hsi Chen. 2023. Traditional-chinese alpaca: Models and datasets. https://github.com/ntunlplab/traditional-chinese-alpaca.",
+            "Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and effective text encoding for chinese llama and alpaca.",
+            "Jesse Dodge, Maarten Sap, Ana Marasovic, William Agnew, Gabriel Ilharco, Dirk Groeneveld, Margaret Mitchell, and Matt Gardner. 2021. Documenting large webtext corpora: A case study on the colossal clean crawled corpus.",
+            "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The pile: An 800gb dataset of diverse text for language modeling.",
+            "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding."
+        ],
+        "bbox": [
+            115,
+            108,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021a. Lora: Low-rank adaptation of large language models.",
+            "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021b. Lora: Low-rank adaptation of large language models.",
+            "Haoyang Huang, Tianyi Tang, Dongdong Zhang, Wayne Xin Zhao, Ting Song, Yan Xia, and Furu Wei. 2023. Not all languages are created equal in llms: Improving multilingual capability by cross-lingual-thought prompting.",
+            "Ayyoob Imani, Peiqin Lin, Amir Hossein Kargaran, Silvia Severini, Masoud Jalili Sabet, Nora Kassner, Chunlan Ma, Helmut Schmid, André F. T. Martins, François Yvon, and Hinrich Schütze. 2023. Glot500: Scaling multilingual corpora and language models to 500 languages.",
+            "Kohaku-Blueleaf. 2023. Guanaco-lora: LoRA for trainin Multilingual Instruction-following LM based on LLaMA. https://huggingface.co/plncmm/guanaco-lora-7b.",
+            "Shibamouli Lahiri. 2014. Complexity of Word Collocation Networks: A Preliminary Structural Analysis. In Proceedings of the Student Research Workshop at the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 96-105, Gothenburg, Sweden. Association for Computational Linguistics.",
+            "Hang Le, Juan Miguel Pino, Changhan Wang, Jiatao Gu, Didier Schwab, and Laurent Besacier. 2021. Lightweight adapter tuning for multilingual speech translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 817-824. Association for Computational Linguistics.",
+            "Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7315-7330, Online. Association for Computational Linguistics.",
+            "Jiahuan Li, Hao Zhou, Shujian Huang, Shanbo Cheng, and Jiajun Chen. 2023. Eliciting the translation ability of large language models via multilingual finetuning with translation instructions.",
+            "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. 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 4582-4597, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "7969",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zehui Lin, Xiao Pan, Mingxuan Wang, Xipeng Qiu, Jiangtao Feng, Hao Zhou, and Lei Li. 2021. Pretraining multilingual neural machine translation by leveraging alignment information.",
+            "Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, Dublin, Ireland. Association for Computational Linguistics.",
+            "Yasmin Moslem, Rejwanul Haque, John D. Kelleher, and Andy Way. 2023. Adaptive machine translation with large language models.",
+            "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.",
+            "Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. 2023. The refined web dataset for falcon llm: Outperforming curated corpora with web data, and web data only.",
+            "Giulia Pucci and Leonardo Ranaldi. 2024. Does the language matter? curriculum learning over neo-Latin languages. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 5212-5220, Torino, Italia. ELRA and ICCL.",
+            "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text.",
+            "Federico Ranaldi, Elena Sofia Ruzzetti, Felicia Logozzo, Michele Mastromattei, Leonardo Ranaldi, and Fabio Massimo Zanzotto. 2023a. Exploring linguistic properties of monolingual berts with typological classification among languages.",
+            "Leonardo Ranaldi and Andre Freitas. 2024. Aligning large and small language models via chain-of-thought reasoning. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1812-1827, St. Julian's, Malta. Association for Computational Linguistics.",
+            "Leonardo Ranaldi, Aria Nourbakhsh, Elena Sofia Ruzzetti, Arianna Patrizi, Dario Onorati, Michele Mastromattei, Francesca Fallucchi, and Fabio Massimo Zanzotto. 2023b. The dark side of the language: Pre-trained transformers in the DarkNet. In Proceedings of the 14th International Conference on Recent"
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Advances in Natural Language Processing, pages 949-960, Varna, Bulgaria. INCOMA Ltd., Shoumen, Bulgaria.",
+            "Leonardo Ranaldi and Giulia Pucci. 2023a. Does the English matter? elicit cross-lingual abilities of large language models. In Proceedings of the 3rd Workshop on Multi-lingual Representation Learning (MRL), pages 173–183, Singapore. Association for Computational Linguistics.",
+            "Leonardo Ranaldi and Giulia Pucci. 2023b. Knowing knowledge: Epistemological study of knowledge in transformers. Applied Sciences, 13(2).",
+            "Leonardo Ranaldi, Giulia Pucci, Federico Ranaldi, Elena Sofia Ruzzetti, and Fabio Massimo Zanzotto. 2024. Empowering multi-step reasoning across languages via tree-of-thoughts.",
+            "Leonardo Ranaldi, Elena Sofia Ruzzetti, Davide Venditti, Dario Onorati, and Fabio Massimo Zanzotto. 2023c. A trip towards fairness: Bias and de-biasing in large language models.",
+            "Andrea Santilli and Emanuele Rodola. 2023. Camoscio: an italian instruction-tuned llama.",
+            "Mohammad Shoeybi, Mostofa Ali Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. 2019. Megatron-LM: Training multi-billion parameter language models using model parallelism. ArXiv, abs/1909.08053.",
+            "Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, , and Jason Wei. 2022. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261.",
+            "Marc Tanti, Lonneke van der Plas, Claudia Borg, and Albert Gatt. 2021. On the language-specificity of multilingual bert and the impact of fine-tuning.",
+            "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.",
+            "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics.",
+            "Martin Thissen. 2023. Fine-tune alpaca for any language. https://github.com/thisserand/alpaca-lora-finetune-language.",
+            "Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12), pages 2214-2218, Istanbul,"
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "7970",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Turkey. European Language Resources Association (ELRA).",
+            "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, 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. 2023. Self-instruct: Aligning language models 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. 2022. Finetuned language models are zero-shot learners.",
+            "Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. 2023. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. arXiv preprint arXiv:2304.01196.",
+            "Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. 2023. Bigtranslate: Augmenting large language models with multilingual translation capability over 100 languages.",
+            "Yasbok. Alpaca Instruction Fine-Tuning for Arabic. https://huggingface.co/Yasbok.",
+            "Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2023. Tim: Teaching large language models to translate with comparison.",
+            "Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang, Zhengrui Ma, Yan Zhou, Langlin Huang, Mengyu Bu, Shangtong Gui, Yunji Chen, Xilin Chen, and Yang Feng. 2023. Bayling: Bridging cross-lingual alignment and instruction following through interactive translation for large language models.",
+            "Wenhao Zhu, Yunzhe Lv, Qingxiu Dong, Fei Yuan, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. 2023. Extrapolating large language models to non-english by aligning languages.",
+            "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV)."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            778
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "7971",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/179c8bcc1f61fc743d5aa8eb19dbf3c437cf9c92d5b2edc8b06e7d60c778572e.jpg",
+        "table_caption": [
+            "A Appendix"
+        ],
+        "table_footnote": [
+            "Table 3: Evaluation results on proposed benchmarks. The avg- lines are the averages of x-Alpacas and x-CorssAlpacas. The last three lines indicate the comparisons between en-Alpaca. The last line indicates the comparisons between avg-Alpacas and avg-CrossAlpacas."
+        ],
+        "table_body": "<table><tr><td></td><td>Model</td><td>MLQA</td><td>XQUAD</td><td>MMLU</td><td>BBH</td></tr><tr><td></td><td>en-Alpaca</td><td>0.89</td><td>0.87</td><td>0.42</td><td>0.30</td></tr><tr><td rowspan=\"5\">Alpaca</td><td>zh-Alpaca</td><td>0.48</td><td>0.38</td><td>0.26</td><td>0.25</td></tr><tr><td>ar-Alpaca</td><td>0.17</td><td>0.16</td><td>0.18</td><td>0.20</td></tr><tr><td>it-Alpaca</td><td>0.35</td><td>0.32</td><td>0.21</td><td>0.25</td></tr><tr><td>es-Alpaca</td><td>0.32</td><td>0.33</td><td>0.19</td><td>0.24</td></tr><tr><td>de-Alpaca</td><td>0.36</td><td>0.39</td><td>0.24</td><td>0.25</td></tr><tr><td></td><td>avg-Alpaca</td><td>0.34</td><td>0.31</td><td>0.24</td><td>0.24</td></tr><tr><td rowspan=\"5\">CrossAlpaca</td><td>zh-CrossAlpaca</td><td>0.70</td><td>0.69</td><td>0.36</td><td>0.28</td></tr><tr><td>ar-CrossAlpaca</td><td>0.56</td><td>0.60</td><td>0.25</td><td>0.25</td></tr><tr><td>it-CrossAlpaca</td><td>0.64</td><td>0.65</td><td>0.28</td><td>0.27</td></tr><tr><td>es-CrossAlpaca</td><td>0.65</td><td>0.64</td><td>0.28</td><td>0.28</td></tr><tr><td>de-CrossAlpaca</td><td>0.64</td><td>0.67</td><td>0.32</td><td>0.29</td></tr><tr><td></td><td>avg-CrossAlpaca</td><td>0.64</td><td>0.65</td><td>0.32</td><td>0.28</td></tr><tr><td></td><td>en-Alpaca vs avg-Alpaca</td><td>0.34(-0.55)</td><td>0.31(-0.56)</td><td>0.24(-0.18)</td><td>0.24(-0.06)</td></tr><tr><td></td><td>en-Alpaca vs avg-CrossAlpaca</td><td>0.64(-0.25)</td><td>0.65(-0.22)</td><td>0.32(-0.10)</td><td>0.28(-0.20)</td></tr><tr><td></td><td>avg-CrossAlpaca vs avg-Alpaca</td><td>(+0.30)</td><td>(+0.34)</td><td>(+0.08)</td><td>(+0.04)</td></tr></table>",
+        "bbox": [
+            136,
+            154,
+            862,
+            456
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/7ddce01a20951b832eda20c507de34d282222a7d8c3bb76835b71a586faecb81.jpg",
+        "table_caption": [
+            "B Appendix"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Language</td><td>Alpaca</td><td>MLQA</td><td>XQUAD</td><td>MMLU</td><td>BBH</td><td>News/commentary</td></tr><tr><td>Arabic</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Chinese</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>English</td><td>x</td><td>x</td><td>x</td><td>x</td><td>x</td><td>x</td></tr><tr><td>German</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Greek</td><td>-</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Hindi</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Italian</td><td>x</td><td>-</td><td>-</td><td>x</td><td>x</td><td>x</td></tr><tr><td>Russian</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Spanish</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Turkish</td><td>x</td><td>-</td><td>x</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Vietnamese</td><td>-</td><td>x</td><td>x</td><td>-</td><td>-</td><td>-</td></tr></table>",
+        "bbox": [
+            169,
+            642,
+            826,
+            840
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 4: List of available state-of-the-art resources.",
+        "bbox": [
+            322,
+            848,
+            670,
+            863
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "7972",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/95b2576b4f00518064c21a463161694c253da331496aa0f224304cff16647117.jpg",
+        "image_caption": [
+            "C Appendix"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            120,
+            489,
+            273
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/97e188ad7570aa8e4c8d091a81fc470daae5c534cea96487512ff9d81aad08bf.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            505,
+            118,
+            836,
+            273
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/43ed6c5e9c2d6d92547db2ea9bfa863bdb6642ab0aacbf2a9931aca7a2573578.jpg",
+        "image_caption": [
+            "Figure 6: Evaluation of all proposed benchmarks of the demonstrations used for instruction-tuning our CrossAlpacas."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            285,
+            489,
+            436
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/a275604db23ecc34bfac52ad45e2279e361a3693d98de3a571a3ee324bcb7d1e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            507,
+            285,
+            836,
+            436
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/c5ccce64f4266f9b341fdd9b793ea081c0f51cff55dced769c3f2f4cce303d12.jpg",
+        "image_caption": [
+            "D Appendix"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            527,
+            489,
+            678
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/3f986813402f6745ae4a46f0d55f1f2eb512387b0bf7e536193d91f199e31a3f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            507,
+            527,
+            836,
+            678
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/7cd59f68ce3891a7d078cecc93fa5f0d4f1244504d22ebed5f2b327b8db5e8c7.jpg",
+        "image_caption": [
+            "Figure 7: Evaluation of all proposed benchmarks using one-direction Translation-following demonstrations. For en-x for English-foreigner and x-en for foreign English. With the red cross, we indicate the results of the x-Alpaca standards, and with the green diamond, the results of our x-CrossAlpaca."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            690,
+            489,
+            841
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/2d2056d9b83128c5ae6440a8bd75de06f61e2bb4b874c5122cfe93559b72a19a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            507,
+            690,
+            836,
+            841
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "7973",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2024/Empowering cross-lingual abilities of instruction-tuned large language models by translation-following demonstrations/f7644dfc-cd06-4320-a14e-02a1e95a33d9_model.json

@@ -0,0 +1,2404 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.119,
+                0.089,
+                0.88,
+                0.128
+            ],
+            "angle": 0,
+            "content": "Empowering Cross-lingual Abilities of Instruction-tuned Large Language Models by Translation-following Demonstrations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.258,
+                0.141,
+                0.741,
+                0.159
+            ],
+            "angle": 0,
+            "content": "Leonardo Ranaldi \\((\\dagger)\\), Giulia Pucci \\((\\star)\\), André Freitas \\((\\dagger, *)\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.295,
+                0.16,
+                0.709,
+                0.176
+            ],
+            "angle": 0,
+            "content": "\\((\\dagger)\\) Idiap Research Institute, Martigny, Switzerland"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.226,
+                0.177,
+                0.776,
+                0.193
+            ],
+            "angle": 0,
+            "content": "\\((\\star)\\) Department of Computing Science, University of Aberdeen, UK"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.226,
+                0.193,
+                0.776,
+                0.209
+            ],
+            "angle": 0,
+            "content": "\\((\\ast)\\)Department of Computer Science, University of Manchester, UK"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.332,
+                0.211,
+                0.665,
+                0.226
+            ],
+            "angle": 0,
+            "content": "[first_name].[last_name]@idiap.ch"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.28,
+                0.461,
+                0.607
+            ],
+            "angle": 0,
+            "content": "The language ability of Large Language Models (LLMs) is often unbalanced towards English because of the imbalance in the distribution of the pre-training data. This disparity is demanded in further fine-tuning and affecting the cross-lingual abilities of LLMs. In this paper, we propose to empower Instruction-tuned LLMs (It-LLMs) in languages other than English by building semantic alignment between them. Hence, we propose CrossAlpaca, an It-LLM with cross-lingual Instruction-following and Translation-following demonstrations to improve semantic alignment between languages. We validate our approach on the multilingual Question Answering (QA) benchmarks XQUAD and MLQA and adapted versions of MMLU and BBH. Our models, tested over six different languages, outperform the It-LLMs tuned on monolingual data. The final results show that instruction tuning on non-English data is not enough and that semantic alignment can be further improved by Translation-following demonstrations."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.62,
+                0.26,
+                0.635
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.646,
+                0.49,
+                0.822
+            ],
+            "angle": 0,
+            "content": "Large Language Models (LLMs) achieve comprehensive language abilities through pre-training on large corpora (Brown et al., 2020; Touvron et al., 2023; Ranaldi and Pucci, 2023b). Hence, the acquired language abilities follow the corpora features, mostly available in English (Lin et al., 2021; Pucci and Ranaldi, 2024). This phenomenon produces an imbalance in both pre-training (Blevins and Zettlemoyer, 2022) and fine-tuning (Le et al., 2021). Consequently, LLMs underperform in languages different from English (Huang et al., 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.823,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Efforts to increase multilingual abilities propose continuing pre-training with large-scale monolingual data (Imani et al., 2023; Cui et al., 2023; Yang et al., 2023). However, learning a language from monolingual data requires considerable data and computational resources."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.254,
+                0.885,
+                0.414
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose CrossAlpaca, an Instruction-tuned LLM (It-LLM) empowered with a semantic alignment between English and other languages. We analyze how to elicit the non-English abilities of It-LLMs by focusing on the crucial phase: instruction-tuning over Instruction-following demonstrations. Hence, we study the effect of cross-lingual alignment by proposing Translation-following demonstrations to improve the instruction-tuning phase."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.417,
+                0.885,
+                0.626
+            ],
+            "angle": 0,
+            "content": "In our experiments, we use \\(LLaMA-7B\\) (Touvron et al., 2023) as the LLM backbone and consider six target languages selected from among the available data (shown in Table 4); where data are absent, we perform a translation task. As Instruction-following demonstrations, we use the Stanford Alpaca dataset (Taori et al., 2023) and translated versions in the corresponding languages, while for the Translation-following, we use a publicly available translation resource news Commentary (Tiedemann, 2012), the most accessible and extendable to multiple languages (i.e., CrossAlpaca demonstrations on Figure 1)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.628,
+                0.885,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Behind an instruction-tuning phase, we evaluate the performance of our six language-specific CrossAlpacas using four different benchmarks: two native-multilingual, i.e., XQUAD (Artetxe et al., 2019) and MLQA (Lewis et al., 2020), and two native-monolinguals, MMLU (Hendrycks et al., 2021) and BBH (Suzgun et al., 2022). The results show that CrossAlpacas instructed with language-specific instruction and translation data far outperform Alpacas instructed only with non-English demonstrations. However, even though CrossAlpaca reduces the gap, the Translation-following models of the original Alpaca still perform best. This result shows that LLaMA's learning abilities on English data are superior to those on non-English data. The semantic alignment task also improves the cross-lingual abilities of It-LLMs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.528,
+                0.904,
+                0.855,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Our findings can be summarized as follows:"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7961"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.229,
+                0.946,
+                0.768,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2024, pages 7961-7973"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.285,
+                0.959,
+                0.714,
+                0.972
+            ],
+            "angle": 0,
+            "content": "August 11-16, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.172,
+                0.084,
+                0.384,
+                0.096
+            ],
+            "angle": 0,
+            "content": "Instruction-following (original)"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.145,
+                0.101,
+                0.415,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Instruction: Classify the sentence as true or false.  \nInput: Sentence: the Earth is flat.  \nOutput: False."
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.435,
+                0.088,
+                0.494,
+                0.111
+            ],
+            "angle": 0,
+            "content": "Stanford Alpaca"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.496,
+                0.087,
+                0.534,
+                0.113
+            ],
+            "angle": 0,
+            "content": "#"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.244,
+                0.138,
+                0.91,
+                0.282
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.292,
+                0.885,
+                0.322
+            ],
+            "angle": 0,
+            "content": "Figure 1: Our \\(x\\)-CrossAlpacas are fine-tuned on Instruction-following and Translation-following demonstrations. This example shows the it-CrossAlpaca, fine-tuned on it-Alpaca and Translation-following demonstrations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.347,
+                0.486,
+                0.378
+            ],
+            "angle": 0,
+            "content": "- The learning abilities of LLMs on non-English Instruction-tuning tasks are limited;"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.389,
+                0.488,
+                0.438
+            ],
+            "angle": 0,
+            "content": "- The multi-lingual abilities of Instruction-tuned LLMs could be empowered through cross-lingual alignment;"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.449,
+                0.488,
+                0.529
+            ],
+            "angle": 0,
+            "content": "- Thus, we propose CrossAlpaca, an Instruction-tuning approach for non-English models that are based on Instruction-following and Translation-following demonstration for target language;"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.54,
+                0.49,
+                0.619
+            ],
+            "angle": 0,
+            "content": "- We show that It-LLMs are able to semantically align through cross-lingual Translation-following demonstrations via an extensive evaluation on four multi-lingual QA benchmarks."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.347,
+                0.49,
+                0.619
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.632,
+                0.271,
+                0.647
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.658,
+                0.412,
+                0.69
+            ],
+            "angle": 0,
+            "content": "2.1 Pre-trained Language Model on Cross-Linguual Corpora"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.695,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "The next token prediction based on the prefix sequence, as well-known as language modeling, is the everlasting task of modern NLP (Tenney et al., 2019). The extensive knowledge of today's Large Language Models (LLMs) depends on the billions of neurons trained on large-scale corpora with derivatives of the language modeling task. Consequently, the pre-training corpora are predominantly in English, e.g., BooksCorpus (Zhu et al., 2015), MEGATRON-LM (Shoeybi et al., 2019), Gutenberg Dataset (Lahiri, 2014) PILE (Gao et al., 2020), C4 (Dodge et al., 2021), RefinedWeb (Penedo et al., 2023); therefore, LLMs usually have much better knowledge of English than other languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.347,
+                0.885,
+                0.474
+            ],
+            "angle": 0,
+            "content": "Aulamo and Tiedemann (2019); Abadji et al. (2022), in order to solve this problem, propose forward corpora translated into several languages. However, these corpora are not as huge as their competitors, and the lack of massively parallel data in the pre-training corpora also prevents LLMs from aligning the different languages well (Li et al., 2023)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.486,
+                0.818,
+                0.501
+            ],
+            "angle": 0,
+            "content": "2.2 The Instruction-tuning Paradigm"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.507,
+                0.884,
+                0.683
+            ],
+            "angle": 0,
+            "content": "Ouyang et al. (2022); Wei et al. (2022) fine-tuned LLMs using the Instruction-tuning method based on Instruction-tuning data, which are instruction-response corpora, to make LLMs more scalable and improve zero-shot performance. In this method, the LLM backbone is fed with data from the instruction \\((I,X,Y)\\), where \\(I\\) is an instruction describing the task's requirements, \\(X\\) is the input, which can be optional, and \\(Y\\) is the output for the given task. The goal of this method is to minimize the function \\(f(Y)\\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.568,
+                0.695,
+                0.883,
+                0.717
+            ],
+            "angle": 0,
+            "content": "\\[\nf (Y) = \\arg \\min  _ {\\theta} \\log p _ {\\theta} (Y \\mid I, X) \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.727,
+                0.811,
+                0.742
+            ],
+            "angle": 0,
+            "content": "where \\(\\theta\\) are model learnable parameters."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.743,
+                0.884,
+                0.838
+            ],
+            "angle": 0,
+            "content": "Earlier studies show that the instruction-tuning method of LLMs with both human (Wang et al., 2023) and synthetic-generated instructions (Taori et al., 2023; Xu et al., 2023) empowers the ability of LLMs to solve considerable tasks in zero-shot scenarios."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "However, we state that the generally used instruction-tuning datasets, alpaca (Taori et al., 2023), Self-Instruct (Wang et al., 2023), Self-Chat (Xu et al., 2023) or a teacher-student alignment approaches (Ranaldi and Freitas, 2024), conceived"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7962"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.246,
+                0.084,
+                0.788,
+                0.26
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.268,
+                0.885,
+                0.298
+            ],
+            "angle": 0,
+            "content": "Figure 2: Examples of Translation-following demonstrations. In this particular example, there are two demonstrations with two different directions from English to German (en-x) and vice versa."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.323,
+                0.489,
+                0.37
+            ],
+            "angle": 0,
+            "content": "in English, which limits the prospect of LLMs to follow non-English instructions and therefore solve related tasks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.383,
+                0.388,
+                0.398
+            ],
+            "angle": 0,
+            "content": "2.3 Instruction-tuning is at hand"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.404,
+                0.489,
+                0.483
+            ],
+            "angle": 0,
+            "content": "Large Language Models have considerable success with many techniques in vogue, such as Instruction-tuning. However, their prohibitive size only allows part of the scientific community to experiment with these models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.485,
+                0.49,
+                0.79
+            ],
+            "angle": 0,
+            "content": "The latest advances that make these models and techniques accessible involve efficient parameter tuning. Parameter-Efficient Tuning (PEFT) is an efficient technique to adjust a small part of the model parameters and freeze the others. The main goal is to significantly reduce computational and storage costs while maintaining the performance of the original models. The standard techniques for PEFT are LoRA (Hu et al., 2021a), Prefix Tuning (Li and Liang, 2021), P-Tuning (Liu et al., 2022). The basic idea is to keep the pre-trained model weights and incorporate low-rank matrices in each architecture layer. This approach significantly reduces the number of parameters that require training for subsequent tasks, thereby increasing overall efficiency. This building block has been and continues to be very important because it skills the scientific community to fair research and enables the development of multiple open-source works."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.802,
+                0.416,
+                0.833
+            ],
+            "angle": 0,
+            "content": "2.4 Cross-lingual Instruction-tuning Challenge"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Recent work has shown the remarkable abilities of LLMs in learning instruction in different languages. Santilli and Rodola (2023); Chen et al. (2023) proposed monolingual fine-tuning LLaMA adaptations on language-specific translated instructions. The"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.547,
+                0.319,
+                0.848,
+                0.382
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>Language</td><td>Name</td></tr><tr><td>Alpaca (Taori et al., 2023)</td><td>English</td><td>en-Alpaca</td></tr><tr><td>Alpaca-Chinese (Chen et al., 2023)</td><td>Chinese</td><td>zh-Alpaca</td></tr><tr><td>Arabic Alpaca (Yasbok)</td><td>Arabic</td><td>ar-Alpaca</td></tr><tr><td>Camascio (Santilli and Rodola, 2023)</td><td>Italian</td><td>it-Alpaca</td></tr><tr><td>Guanaco (Kohaku-Blueleaf, 2023)</td><td>Spanish</td><td>es-Alpaca</td></tr><tr><td>German Alpaca (Thissen, 2023)</td><td>German</td><td>de-Alpaca</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.391,
+                0.885,
+                0.434
+            ],
+            "angle": 0,
+            "content": "Table 1: Details and names of mono-lingual Instruction-tuned Large Language Models that use a language-specific version of Alpaca as instruction-tuning data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.46,
+                0.884,
+                0.669
+            ],
+            "angle": 0,
+            "content": "use of optimization techniques introduces in Section 2.3, to propose engaging custom adapters for various tasks. Zhang et al. (2023); Ranaldi and Pucci (2023a) investigated the cross-lingual abilities of It-LLMs by promoting the effects of augmenting instruction demonstrations. In particular, Zhang et al. (2023) use fine-grained custom translations that are not easily accessible, while Ranaldi and Pucci (2023a) propose a multilingual approach by considerably increasing instruction and translation tasks. However, these studies have opened an interesting avenue in analyzing the cross-lingual abilities of It-LLMs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.671,
+                0.884,
+                0.846
+            ],
+            "angle": 0,
+            "content": "In this paper, we take the next step by proposing our CrossAlpaca. Our methods show of the cross-lingual learning abilities of It-LLMs on a large scale. Using four well-known benchmarks, we show that the weaknesses of It-LLMs trained on non-English data can be strengthened with cross-lingual alignment approaches. Hence, our analyses aim to understand the role of Instruction-following and Translation-following demonstrations in closing the gap in learning and adapting LLMs' abilities to non-English languages."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.861,
+                0.689,
+                0.875
+            ],
+            "angle": 0,
+            "content": "3 Data & Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.888,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Pre-training from scratch a Large Language Model (LLM) in multiple languages is cost-prohibitive for"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7963"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.278
+            ],
+            "angle": 0,
+            "content": "data collection and parameter learning. This is the reason why the trend is to do further fine-tuning to strengthen the models' abilities in a specific language (Tanti et al., 2021; Moslem et al., 2023). In this paper, we aim to exploit the abilities of pretrained LLMs for non-English languages by further improving the alignment between English and the target language. Hence, in Section 3.1, we introduce the challenge of fine-tuning an LLM on a mono-lingual custom scenario. Furthermore, in Section 3.2, we propose our approach to fine-tune cross-lingual LLMs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.292,
+                0.418,
+                0.308
+            ],
+            "angle": 0,
+            "content": "3.1 Mono-lingual Instruction-tuning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.314,
+                0.49,
+                0.457
+            ],
+            "angle": 0,
+            "content": "The limited accessibility and transparency of paid API services of state-of-the-art LLMs have pushed research toward developing open-source models. Using the instruction-tuning paradigm (Section 2.2) and Stanford Alpaca (Taori et al., 2023), a corpus consisting of \\(52\\mathrm{k}\\) of English instruction-output pairs generated by text-davinci-003, several Instruction-tuned versions of LLaMA were released."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.46,
+                0.49,
+                0.702
+            ],
+            "angle": 0,
+            "content": "Following this approach, multiple monolingual Instruction-tuned versions of LLaMA were proposed by translating the Stanford Alpaca Instruction-following data into the specific language. Table 4 (Appendix B) shows a set of alpaca versions available as open source. Following a systematic analysis of the translated versions of Alpaca in official repositories<sup>1</sup>, the languages of the benchmark datasets, and the translation pairs present in news/commentary, which will be introduced later, we selected the languages that share the most already available data. Table 1 shows the custom versions used in this work, which for simplicity will be renamed \\( x \\)-Alpaca, where \\( x \\) indicates the specific language."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.715,
+                0.416,
+                0.731
+            ],
+            "angle": 0,
+            "content": "3.2 Cross-lingual Instruction-tuning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.737,
+                0.49,
+                0.882
+            ],
+            "angle": 0,
+            "content": "Mono-lingual Instruction-tuning approaches (Section 3.1) reward LLMs' multi-lingual abilities. However, the use of translated Alpaca alone for the specific language is not sufficient to elicit the non-English ability of LLMs. For this reason, we propose CrossAlpaca (Figure 1), which enriches the cross-lingual Instruction-tuning with the Translation-following demonstrations. We aim to highlight LLMs' English and non-English abilities"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.085,
+                0.883,
+                0.117
+            ],
+            "angle": 0,
+            "content": "by proposing a semantic alignment task to enrich versions of Alpaca."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.128,
+                0.886,
+                0.352
+            ],
+            "angle": 0,
+            "content": "Instruction-following demonstrations The original version of Alpaca is in English. However, as described in Section 3.1, different open-source translations are produced with a translation engine. In our experiments, we propose the Instruction-tuning phase with the original English version (en-Alpaca) and the translated language-specific versions. Moreover, we propose the CrossAlpaca versions built with the different Alpacas translated into the specific language and the cross-lingual translations (introduced later). In this way, we investigate the abilities of the LLM backbone in understanding multilingual instructions and cross-lingual alignment."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.364,
+                0.885,
+                0.524
+            ],
+            "angle": 0,
+            "content": "Translation-following demonstrations Learning general instruction data is a sensible strategy for building models to solve multi-tasks directed by instructions (Wang et al., 2023; Zeng et al., 2023). However, translation data could contribute to learning semantic alignment. Zhang et al. (2023) showed that the translation performance of LLMs could be improved by using expertly annotated translation data, and Zhu et al. (2023) showed that it could be beneficial for instruction-tuning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.525,
+                0.887,
+                0.799
+            ],
+            "angle": 0,
+            "content": "Inspired by Zhu et al. (2023), we use publicly available sentence-level translation datasets, such as news Commentary (Tiedemann, 2012), to construct the translation task instruction demonstrations. We also propose extending this to additional languages, which we release as an open-source dataset. In particular, for each specific language, we constructed 20k demonstrations. Hence, following the Alpaca style (Instruction, Input, and Output) (see Figure 1), we randomly selected 10k English to non-English translations and 10k random non-English to English translations (all translations come from news commentary). We have constructed these correspondences for the languages mentioned above in Section 3.1. Figure 2 shows a case of Translation-following for English-German direction (en-de) and vice versa (de-en)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.812,
+                0.656,
+                0.829
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In order to observe the English and non-English abilities of Large Language Models (LLMs) and the impact of the Instruction-tuning approach in cross-lingual scenarios, we propose CrossAlpaca. Our approach is based on instruction-tuning on"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.892,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "1official versions on https://github.com/tloen/ alpaca-lora and https://huggingface.co/models"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7964"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.087,
+                0.493,
+                0.24
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.506,
+                0.087,
+                0.839,
+                0.24
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.158,
+                0.25,
+                0.49,
+                0.402
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.506,
+                0.251,
+                0.838,
+                0.402
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.419,
+                0.882,
+                0.448
+            ],
+            "angle": 0,
+            "content": "Figure 3: Evaluation results on proposed benchmarks. The dotted line represents the performance of the original Alpaca (Taori et al., 2023) on English tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.474,
+                0.49,
+                0.586
+            ],
+            "angle": 0,
+            "content": "language-specific data augmented with a crosslingual semantic alignment. Hence, we set several baseline models (Section 4.1), which we augmented with our CrossAlpaca approach (Section 4.2). Finally, we performed a series of systematic evaluations (Section 4.3.1) to observe the impact of the proposed intervention."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.6,
+                0.423,
+                0.615
+            ],
+            "angle": 0,
+            "content": "4.1 Baseline Instruction-tuned LLMs"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.622,
+                0.49,
+                0.766
+            ],
+            "angle": 0,
+            "content": "The common denominator among the It-LLMs shown in Table 1 is the LLM backbone, LLaMA7B (Touvron et al., 2023). Starting from Instruction-following data from the original Alpaca (Taori et al., 2023) and its open-source non-English versions\\(^{2}\\), we reproduced \\(x\\)-Alpaca for \\(x\\) specific languages: Chinese (zh), Italian (it), Arabic (ar), Spanish (es), German (de) and the original English version (en)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.781,
+                0.465,
+                0.797
+            ],
+            "angle": 0,
+            "content": "4.2 Cross-lingual Instruction-tuned LLMs"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.804,
+                0.49,
+                0.884
+            ],
+            "angle": 0,
+            "content": "The CrossAlpacas are instruction-tuned on Instruction-following and Translation-following demonstrations (CrossAlpacas demonstrations). The first ones stem from the resources introduced in Section 4.1. The second comes from"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.474,
+                0.8,
+                0.489
+            ],
+            "angle": 0,
+            "content": "news COMMENTary (Tiedemann, 2012)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.49,
+                0.885,
+                0.57
+            ],
+            "angle": 0,
+            "content": "Our approach generates a series of instruction-tuned versions of the data shown in Figure 1. We have named the versions \\( x \\)-CrossAlpaca where \\( x \\) denotes Chinese (zh), Arabic (ar), Italian (it), Spanish (es), and German (de)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.581,
+                0.715,
+                0.597
+            ],
+            "angle": 0,
+            "content": "4.3 Experimental Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.602,
+                0.885,
+                0.667
+            ],
+            "angle": 0,
+            "content": "In order to assess the performance of the CrossAlpaca, we defined several benchmarks (Section 4.3.1) on which we applied systematic tuning (Section 4.3.2) and evaluation (Section 4.3.3) pipelines."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.675,
+                0.671,
+                0.689
+            ],
+            "angle": 0,
+            "content": "4.3.1 Benchmarks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.695,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "To evaluate the performance of the It-LLMs and the impact of the translation-based semantic alignment approach, we used two cross-lingual (XQUAD (Artetxe et al., 2019), MLQA (Lewis et al., 2020)) and two multi-task (MMLU (Hendrycks et al., 2021) and BBH (Suzgun et al., 2022)) benchmarks. XQUAD and MLQA focus on understanding questions and answers through translation into different languages. MMLU and BBH, being multi-task benchmarks, include subtasks related to Boolean expressions and QA on basic-level subjects (e.g., chemistry, physics). However, we decided to introduce them to observe whether our approach degrades performance in these tasks. The first two"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.892,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "2open-source code is available on https://github.com/tloen/alpaca-lora"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7965"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.149
+            ],
+            "angle": 0,
+            "content": "datasets selected are appropriately constructed for multi-language testing, while the second two are available only in English. So we do a preliminary translation step as outlined below."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.113,
+                0.164,
+                0.49,
+                0.179
+            ],
+            "angle": 0,
+            "content": "MultiLingual Question Answering (MLQA)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.18,
+                0.49,
+                0.356
+            ],
+            "angle": 0,
+            "content": "(Lewis et al., 2020) evaluates cross-lingual question answering performance. The benchmark comprises over 5K extractive QA instances in the SQuAD (Rajpurkar et al., 2016) format in several languages. MLQA is highly parallel, with QA instances aligned across four languages on average. Although comprising different languages, some languages are not represented, such as Italian. To conduct the experiments uniformly, we have translated the examples as also done in the forthcoming MMLU and BBH."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.372,
+                0.49,
+                0.485
+            ],
+            "angle": 0,
+            "content": "Cross-lingual Question Answering Dataset (XQUAD) (Artetxe et al., 2019) consists of a subset of 240 paragraphs and 1190 question-answer pairs from the development set of SQuAD v1.1 (Rajpurkar et al., 2016) with their manual translations into several languages. Consequently, the dataset is entirely parallel across 11 languages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.499,
+                0.49,
+                0.611
+            ],
+            "angle": 0,
+            "content": "Massive Multitask Language Understanding (MMLU) (Hendrycks et al., 2021) measures knowledge of the world and problem-solving problems in multiple subjects with 57 subjects across STEM, humanities, social sciences, and other areas. The benchmark is native in English; however, we translated it into five additional languages<sup>3</sup>."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.626,
+                0.49,
+                0.706
+            ],
+            "angle": 0,
+            "content": "BIG-Bench Hard (BBH) (Suzgun et al., 2022) is a subset of challenging tasks related to navigation, logical deduction, and fallacy detection. Here again, the benchmark is native English, and we have translated it into five languages<sup>3</sup>."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.721,
+                0.283,
+                0.737
+            ],
+            "angle": 0,
+            "content": "4.3.2 Models Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.744,
+                0.49,
+                0.791
+            ],
+            "angle": 0,
+            "content": "In order to align the results with the state-of-the-art models, we used the alpaca_LoRA (Hu et al., 2021b) code\\(^2\\), adopting the same hyperparameters."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.794,
+                0.489,
+                0.856
+            ],
+            "angle": 0,
+            "content": "We performed the fine-tuning with a single epoch and a batch-size of 128 examples, running our experiments on a workstation equipped with two Nvidia RTX A6000 with 48 GB of VRAM."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.085,
+                0.657,
+                0.099
+            ],
+            "angle": 0,
+            "content": "4.3.3 Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.108,
+                0.885,
+                0.219
+            ],
+            "angle": 0,
+            "content": "As an evaluation metric, we use accuracy. Hence, we estimate accuracy by measuring exact match values in the zero-shot setting. For each model, the parts of benchmarks related to the specific language are used (e.g., for zh-Alpaca and zh-CrossAlpaca, data from MLQA, XQUAD, MMLU, and BBH in Chinese are used)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.237,
+                0.61,
+                0.252
+            ],
+            "angle": 0,
+            "content": "5 Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.267,
+                0.885,
+                0.491
+            ],
+            "angle": 0,
+            "content": "Improving non-English abilities in Instruction-tuned Large Language Models (It-LLMs) remains challenging. However, the \\(x\\)-CrossAlpacas revealed improved results in cross-lingual Question Answering (QA) benchmarks and maintained logical-mathematical skills. From the results of Figure 3 (further detailed in Table 3), it is possible to observe the weaknesses emerging from the fine-tuning of the translated versions of Alpaca (Section 5.1), the improvement obtained from the alignment phase is encouraging (Section 5.2) but it is not enough to outperform the English one. Therefore, we investigated the impact of demonstrations on downstream performance (Section 5.3)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.494,
+                0.884,
+                0.59
+            ],
+            "angle": 0,
+            "content": "The fine-grained analysis highlighted the importance of cross-lingual alignment data and the critical issues with non-English data. This opens the way for new hypotheses regarding the imbalance of pre-training languages and learning abilities via instruction-tuning."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.607,
+                0.882,
+                0.622
+            ],
+            "angle": 0,
+            "content": "5.1 Translating the Alpaca is not the right way"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.63,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The instruction-tuning on LLaMA, predominantly pre-trained in English, affects x-Alpaca. Figure 3 and in terms of numbers, Table 3 show that in both MLQA and XQUAD, there is a gap of 55 and 53 average points between en-Alpaca and the x-Alpacas. This phenomenon is mitigated for MMLU and BBH, where we observed an average gap of 18 and 14 points. Instructing an LLM on Alpaca-style demonstrations translated into different languages is not always a good strategy. However, some x-Alpacas, such as zh-Alpaca and de-Alpaca, have performed better. We hypothesize that this phenomenon is related to the scale of the pre-training data in the respective languages and, thus, the abilities of the LLaMA. In future developments, we plan to extend the study on other LLMs beyond LLaMA to observe whether the phenomenon is similar, milder, or more significant."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.869,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>3</sup>We performed translations using the Google translator API from English to Chinese (zh), Italian (it), Arabic (ar), Spanish (es), German (de). Resources will be made available along with the publication"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7966"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.081,
+                0.488,
+                0.169
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>QA\nTask</td><td>en-Alpaca</td><td>avg-Alpaca</td><td>avg-CrossAlpaca</td><td>δ</td></tr><tr><td>MLQA</td><td>0.89</td><td>0.34</td><td>0.64</td><td>+0.30</td></tr><tr><td>XQUAD</td><td>0.97</td><td>0.31</td><td>0.65</td><td>+0.30</td></tr><tr><td>MMLU</td><td>0.42</td><td>0.24</td><td>0.32</td><td>+0.08</td></tr><tr><td>BBH</td><td>0.30</td><td>0.24</td><td>0.28</td><td>+0.04</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.178,
+                0.49,
+                0.221
+            ],
+            "angle": 0,
+            "content": "Table 2: Averages of the results on proposed benchmarks. The column \\(\\delta\\) indicates the difference between avg-Alpacas and our avg-CrossAlpacas."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.248,
+                0.453,
+                0.264
+            ],
+            "angle": 0,
+            "content": "5.2 CrossAlpaca: A cross-lingual solution"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.271,
+                0.489,
+                0.479
+            ],
+            "angle": 0,
+            "content": "Semantic alignment through Translation-following demonstrations during fine-tuning could have a valuable impact on the cross-lingual abilities of It-LLMs. The \\(x\\)-CrossAlpacas outperformed the x-Alpacas by 30 average points on MLQA, 34 average points on XQUAD, 8 on MMLU, and 4 on BBH (see Table 2 and more detailed in Table 3). They also brought their performances closer to the sota obtained from the original Alpaca by 25 points on MLQA and 22 average points on XQUAD. In MMLU and BBH, the gap became very close, with averages of 10 and 2 points (see Table 2 or the line 'en-Alpaca vs avg-CrossAlpaca' in Table 3)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.481,
+                0.49,
+                0.675
+            ],
+            "angle": 0,
+            "content": "Enriching Translation-following demonstrations has outstanding influences on the cross-lingual abilities of the It-LLMs. However, even in this case, Chinese and German models (zh- and de-CrossAlpaca) outperformed Arabic by many points and, in some specific cases, Spanish and Italian as well. This phenomenon, we hypothesize, is related to the diversity in corpus representation within the pre-training data, as shown in (Yang et al., 2023). Therefore, cross-lingual approaches do not have an incisive impact as in languages less present in the pre-training phases of the language model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.688,
+                0.281,
+                0.704
+            ],
+            "angle": 0,
+            "content": "5.3 Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.71,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Our CrossAlpacas, distinguished by the construction of the demonstrations pairs (Section 4.2), achieves significant performance improvements and contributes to closing the gap between the original Alpaca (en-Alpaca) and a series of x-Alpacas in different languages. In order to show the impact of enrichment with cross-lingual demonstrations, we propose two different analyses. In the first analysis, we incrementally decrease the training data, in particular on the Translation-following demonstrations side (Section 5.4). In the second, still working on the Translation-following part (defined by half en-x and half x-en demonstrations), we an"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.557,
+                0.086,
+                0.837,
+                0.219
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.556,
+                0.226,
+                0.837,
+                0.358
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.372,
+                0.885,
+                0.416
+            ],
+            "angle": 0,
+            "content": "Figure 4: Evaluation of proposed benchmarks of the demonstrations used for instruction-tuning our CrossAlpacas."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.443,
+                0.884,
+                0.474
+            ],
+            "angle": 0,
+            "content": "alyze the impact of the demonstrations by splitting the experiments into en-x and x-en (Section 5.4.1)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.489,
+                0.851,
+                0.521
+            ],
+            "angle": 0,
+            "content": "5.4 Translation-following demonstrations empower non-English abilities."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.527,
+                0.885,
+                0.751
+            ],
+            "angle": 0,
+            "content": "Figure 4 (for a more extensive analysis please refer to Figure 6 in the Appendix) shows the non-English performance of the x-CrossAlpaca on the benchmarks under different scales of Translation-following demonstrations<sup>4</sup>. The x-CrossAlpacas achieve higher accuracy when more demonstrations are used, revealing the benefits of cross-lingual alignment to improve non-English performance. However, Translation-following demonstrations are built in two directions: x-en and en-x (foreign-English and English-foreign). Although equally distributed, we are still determining whether there is an asymmetry between the two types of contributions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.753,
+                0.884,
+                0.882
+            ],
+            "angle": 0,
+            "content": "Finally, this result is less pronounced for the multi-task benchmarks (see Figure 6 MMLU and BBH). This could be related to the fact that in the benchmarks, there are logical-mathematical subtasks which are less language dependent (comparatively simpler syntax, smaller vocabularies), and hence, our cross-lingual approach does not influence the results for MLQA and XQUAD as"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.893,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "4The Traslation-following demonstrations are equally selected in a random way from the 10k en-x and x-en."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7967"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.166
+            ],
+            "angle": 0,
+            "content": "much. Furthermore, it is possible to observe a trend in language groups, particularly Chinese-German, Italian-Spanish, and Arabic. This trend, which recurs frequently in the results of our experiments, will be further investigated in future work."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.195,
+                0.443,
+                0.327
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.334,
+                0.443,
+                0.466
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.481,
+                0.49,
+                0.525
+            ],
+            "angle": 0,
+            "content": "Figure 5: Evaluation of proposed benchmarks using one-direction Translation-following demonstrations. For en-x for English-foreigner and x-en for foreign English."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.581,
+                0.456,
+                0.613
+            ],
+            "angle": 0,
+            "content": "5.4.1 Demonstration direction matters on non-English abilities"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.63,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Figure 5 shows the evaluation conducted in varying directions of the translation-following demonstrations. In particular, demonstrations with English-foreign direction (en-x) better impact downstream models. Conversely, foreign-English (x-en) demonstrations perform better than baselines but underperform demonstrations in the opposite direction. However, as shown in Figure 5 (and complete Figure 7), the x-CrossAlpacas continue to outperform. Nevertheless, the trend of translation-following demonstrations with one direction is interesting. Again, as was the case in the previous ablation study, the multi-task benchmarks (see MMLU and BBH in Figure 7) do not seem to obtain significant influences, which reinforces the hypothesis that the models are greatly affected by cross-lingual capabilities in tasks where there is a strong presence of natural language."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.084,
+                0.667,
+                0.099
+            ],
+            "angle": 0,
+            "content": "6 Future Works"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.109,
+                0.885,
+                0.285
+            ],
+            "angle": 0,
+            "content": "The cross-lingual abilities of Intruction-tuned Large Language Models (It-LLMs) seem to be supported by LLMs, such as in the case of Alpaca, the LLaMA backbone. However, it seems that low-impact demonstrations at the data level can enrich these abilities. We obtained valuable results from our experiments by proposing strategic demonstrations, i.e., Translation-following demonstrations. These results were proposed by performing fine-tuning on LLaMA-7B as done by Taori et al. (2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.287,
+                0.885,
+                0.576
+            ],
+            "angle": 0,
+            "content": "In future work, we would like to continue to investigate by increasing the number of parameters in LLaMA and including additional backbone models. In addition, it might be interesting to evaluate the impact on low-resource languages as done for in-context settings in (Ranaldi et al., 2024). Hence, we would like to get to the underside of performances obtained in some experiments (see Section 5.4) by extending previous epistemic approaches (Ranaldi et al., 2023a,c,b) to It-LLMs. In parallel, plans include analyzing the translation capabilities of general It-LLMs and those enhanced with translation tasks, including some specialized translation tasks among our evaluation benchmarks. Finally, we would like to investigate the learning abilities of the original Alpaca as the translation data change, proposing different probing experiments on (original) English data enhanced with translations."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.588,
+                0.642,
+                0.603
+            ],
+            "angle": 0,
+            "content": "7 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.613,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this paper, we proposed CrossAlpaca, an approach to empowering the instruction-tuning of LLMs on non-English data. Specifically, we coupled Instruction-following (Alpaca-style) demonstrations with Translation-following demonstrations. Our method seeks to instruct the LLM to semantic alignment between English and non-English overperforms models instructed on non-English texts. In particular, thanks to our CrossAlpaca demonstrations, the instructed models achieved significant performance improvements on four Question Answering benchmarks XQUAD, MLQA, MMLU, and BBH. In addition, we observe that semantic alignment strengthens with increasing Translation-following data; this demonstrates the de-facto abilities of It-LLMs to learn from instructions. Our approach and results contribute to improved research on the potential for producing more powerful LLMs for non-English languages."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7968"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.085,
+                0.214,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.109,
+                0.49,
+                0.188
+            ],
+            "angle": 0,
+            "content": "Julien Abadji, Pedro Ortiz Suarez, Laurent Romary, and Benoit Sagot. 2022. Towards a cleaner document-oriented multilingual crawled corpus. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4344-4355, Marseille, France. European Language Resources Association."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.2,
+                0.49,
+                0.254
+            ],
+            "angle": 0,
+            "content": "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2019. On the cross-lingual transferability of monolingual representations. In Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.265,
+                0.489,
+                0.344
+            ],
+            "angle": 0,
+            "content": "Mikko Aulamo and Jörg Tiedemann. 2019. The OPUS resource repository: An open package for creating parallel corpora and machine translation services. In Proceedings of the 22nd Nordic Conference on Computational Linguistics, pages 389-394, Turku, Finland. Linköping University Electronic Press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.357,
+                0.489,
+                0.423
+            ],
+            "angle": 0,
+            "content": "Terra Blevins and Luke Zettlemoyer. 2022. Language contamination helps explains the cross-lingual capabilities of english pretrained models. In Conference on Empirical Methods in Natural Language Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.435,
+                0.489,
+                0.58
+            ],
+            "angle": 0,
+            "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.591,
+                0.489,
+                0.644
+            ],
+            "angle": 0,
+            "content": "Wei-Lin Chen, Cheng-Kuang Wu, and Hsin-Hsi Chen. 2023. Traditional-chinese alpaca: Models and datasets. https://github.com/ntunlplab/traditional-chinese-alpaca."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.657,
+                0.489,
+                0.697
+            ],
+            "angle": 0,
+            "content": "Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and effective text encoding for chinese llama and alpaca."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.709,
+                0.489,
+                0.775
+            ],
+            "angle": 0,
+            "content": "Jesse Dodge, Maarten Sap, Ana Marasovic, William Agnew, Gabriel Ilharco, Dirk Groeneveld, Margaret Mitchell, and Matt Gardner. 2021. Documenting large webtext corpora: A case study on the colossal clean crawled corpus."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.787,
+                0.489,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The pile: An 800gb dataset of diverse text for language modeling."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.866,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.109,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.14
+            ],
+            "angle": 0,
+            "content": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021a. Lora: Low-rank adaptation of large language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.149,
+                0.883,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021b. Lora: Low-rank adaptation of large language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.213,
+                0.883,
+                0.28
+            ],
+            "angle": 0,
+            "content": "Haoyang Huang, Tianyi Tang, Dongdong Zhang, Wayne Xin Zhao, Ting Song, Yan Xia, and Furu Wei. 2023. Not all languages are created equal in llms: Improving multilingual capability by cross-lingual-thought prompting."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.29,
+                0.883,
+                0.369
+            ],
+            "angle": 0,
+            "content": "Ayyoob Imani, Peiqin Lin, Amir Hossein Kargaran, Silvia Severini, Masoud Jalili Sabet, Nora Kassner, Chunlan Ma, Helmut Schmid, André F. T. Martins, François Yvon, and Hinrich Schütze. 2023. Glot500: Scaling multilingual corpora and language models to 500 languages."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.379,
+                0.883,
+                0.433
+            ],
+            "angle": 0,
+            "content": "Kohaku-Blueleaf. 2023. Guanaco-lora: LoRA for trainin Multilingual Instruction-following LM based on LLaMA. https://huggingface.co/plncmm/guanaco-lora-7b."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.443,
+                0.883,
+                0.535
+            ],
+            "angle": 0,
+            "content": "Shibamouli Lahiri. 2014. Complexity of Word Collocation Networks: A Preliminary Structural Analysis. In Proceedings of the Student Research Workshop at the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 96-105, Gothenburg, Sweden. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.545,
+                0.883,
+                0.637
+            ],
+            "angle": 0,
+            "content": "Hang Le, Juan Miguel Pino, Changhan Wang, Jiatao Gu, Didier Schwab, and Laurent Besacier. 2021. Lightweight adapter tuning for multilingual speech translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 817-824. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.648,
+                0.883,
+                0.74
+            ],
+            "angle": 0,
+            "content": "Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7315-7330, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.75,
+                0.883,
+                0.803
+            ],
+            "angle": 0,
+            "content": "Jiahuan Li, Hao Zhou, Shujian Huang, Shanbo Cheng, and Jiajun Chen. 2023. Eliciting the translation ability of large language models via multilingual finetuning with translation instructions."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.813,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. 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 4582-4597, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7969"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.139
+            ],
+            "angle": 0,
+            "content": "Zehui Lin, Xiao Pan, Mingxuan Wang, Xipeng Qiu, Jiangtao Feng, Hao Zhou, and Lei Li. 2021. Pretraining multilingual neural machine translation by leveraging alignment information."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.148,
+                0.487,
+                0.254
+            ],
+            "angle": 0,
+            "content": "Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.263,
+                0.487,
+                0.303
+            ],
+            "angle": 0,
+            "content": "Yasmin Moslem, Rejwanul Haque, John D. Kelleher, and Andy Way. 2023. Adaptive machine translation with large language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.312,
+                0.487,
+                0.415
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.426,
+                0.487,
+                0.504
+            ],
+            "angle": 0,
+            "content": "Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. 2023. The refined web dataset for falcon llm: Outperforming curated corpora with web data, and web data only."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.514,
+                0.487,
+                0.605
+            ],
+            "angle": 0,
+            "content": "Giulia Pucci and Leonardo Ranaldi. 2024. Does the language matter? curriculum learning over neo-Latin languages. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 5212-5220, Torino, Italia. ELRA and ICCL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.615,
+                0.486,
+                0.655
+            ],
+            "angle": 0,
+            "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.664,
+                0.487,
+                0.73
+            ],
+            "angle": 0,
+            "content": "Federico Ranaldi, Elena Sofia Ruzzetti, Felicia Logozzo, Michele Mastromattei, Leonardo Ranaldi, and Fabio Massimo Zanzotto. 2023a. Exploring linguistic properties of monolingual berts with typological classification among languages."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.739,
+                0.487,
+                0.831
+            ],
+            "angle": 0,
+            "content": "Leonardo Ranaldi and Andre Freitas. 2024. Aligning large and small language models via chain-of-thought reasoning. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1812-1827, St. Julian's, Malta. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.84,
+                0.487,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Leonardo Ranaldi, Aria Nourbakhsh, Elena Sofia Ruzzetti, Arianna Patrizi, Dario Onorati, Michele Mastromattei, Francesca Fallucchi, and Fabio Massimo Zanzotto. 2023b. The dark side of the language: Pre-trained transformers in the DarkNet. In Proceedings of the 14th International Conference on Recent"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Advances in Natural Language Processing, pages 949-960, Varna, Bulgaria. INCOMA Ltd., Shoumen, Bulgaria."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.136,
+                0.882,
+                0.215
+            ],
+            "angle": 0,
+            "content": "Leonardo Ranaldi and Giulia Pucci. 2023a. Does the English matter? elicit cross-lingual abilities of large language models. In Proceedings of the 3rd Workshop on Multi-lingual Representation Learning (MRL), pages 173–183, Singapore. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.225,
+                0.881,
+                0.265
+            ],
+            "angle": 0,
+            "content": "Leonardo Ranaldi and Giulia Pucci. 2023b. Knowing knowledge: Epistemological study of knowledge in transformers. Applied Sciences, 13(2)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.275,
+                0.882,
+                0.327
+            ],
+            "angle": 0,
+            "content": "Leonardo Ranaldi, Giulia Pucci, Federico Ranaldi, Elena Sofia Ruzzetti, and Fabio Massimo Zanzotto. 2024. Empowering multi-step reasoning across languages via tree-of-thoughts."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.337,
+                0.882,
+                0.39
+            ],
+            "angle": 0,
+            "content": "Leonardo Ranaldi, Elena Sofia Ruzzetti, Davide Venditti, Dario Onorati, and Fabio Massimo Zanzotto. 2023c. A trip towards fairness: Bias and de-biasing in large language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.4,
+                0.882,
+                0.426
+            ],
+            "angle": 0,
+            "content": "Andrea Santilli and Emanuele Rodola. 2023. Camoscio: an italian instruction-tuned llama."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.437,
+                0.882,
+                0.501
+            ],
+            "angle": 0,
+            "content": "Mohammad Shoeybi, Mostofa Ali Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. 2019. Megatron-LM: Training multi-billion parameter language models using model parallelism. ArXiv, abs/1909.08053."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.513,
+                0.882,
+                0.591
+            ],
+            "angle": 0,
+            "content": "Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, , and Jason Wei. 2022. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.602,
+                0.882,
+                0.642
+            ],
+            "angle": 0,
+            "content": "Marc Tanti, Lonneke van der Plas, Claudia Borg, and Albert Gatt. 2021. On the language-specificity of multilingual bert and the impact of fine-tuning."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.651,
+                0.882,
+                0.717
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.727,
+                0.882,
+                0.806
+            ],
+            "angle": 0,
+            "content": "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.816,
+                0.882,
+                0.856
+            ],
+            "angle": 0,
+            "content": "Martin Thissen. 2023. Fine-tune alpaca for any language. https://github.com/thisserand/alpaca-lora-finetune-language."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.866,
+                0.882,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12), pages 2214-2218, Istanbul,"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7970"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.135,
+                0.086,
+                0.486,
+                0.113
+            ],
+            "angle": 0,
+            "content": "Turkey. European Language Resources Association (ELRA)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.123,
+                0.487,
+                0.202
+            ],
+            "angle": 0,
+            "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.212,
+                0.486,
+                0.264
+            ],
+            "angle": 0,
+            "content": "Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.274,
+                0.487,
+                0.327
+            ],
+            "angle": 0,
+            "content": "Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.337,
+                0.487,
+                0.39
+            ],
+            "angle": 0,
+            "content": "Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. 2023. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. arXiv preprint arXiv:2304.01196."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.4,
+                0.487,
+                0.453
+            ],
+            "angle": 0,
+            "content": "Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. 2023. Bigtranslate: Augmenting large language models with multilingual translation capability over 100 languages."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.463,
+                0.487,
+                0.489
+            ],
+            "angle": 0,
+            "content": "Yasbok. Alpaca Instruction Fine-Tuning for Arabic. https://huggingface.co/Yasbok."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.499,
+                0.487,
+                0.539
+            ],
+            "angle": 0,
+            "content": "Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2023. Tim: Teaching large language models to translate with comparison."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.549,
+                0.487,
+                0.628
+            ],
+            "angle": 0,
+            "content": "Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang, Zhengrui Ma, Yan Zhou, Langlin Huang, Mengyu Bu, Shangtong Gui, Yunji Chen, Xilin Chen, and Yang Feng. 2023. Bayling: Bridging cross-lingual alignment and instruction following through interactive translation for large language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.638,
+                0.487,
+                0.691
+            ],
+            "angle": 0,
+            "content": "Wenhao Zhu, Yunzhe Lv, Qingxiu Dong, Fei Yuan, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. 2023. Extrapolating large language models to non-english by aligning languages."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.7,
+                0.487,
+                0.78
+            ],
+            "angle": 0,
+            "content": "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV)."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.78
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.518,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7971"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.115,
+                0.133,
+                0.239,
+                0.149
+            ],
+            "angle": 0,
+            "content": "A Appendix"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.137,
+                0.155,
+                0.863,
+                0.457
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Model</td><td>MLQA</td><td>XQUAD</td><td>MMLU</td><td>BBH</td></tr><tr><td></td><td>en-Alpaca</td><td>0.89</td><td>0.87</td><td>0.42</td><td>0.30</td></tr><tr><td rowspan=\"5\">Alpaca</td><td>zh-Alpaca</td><td>0.48</td><td>0.38</td><td>0.26</td><td>0.25</td></tr><tr><td>ar-Alpaca</td><td>0.17</td><td>0.16</td><td>0.18</td><td>0.20</td></tr><tr><td>it-Alpaca</td><td>0.35</td><td>0.32</td><td>0.21</td><td>0.25</td></tr><tr><td>es-Alpaca</td><td>0.32</td><td>0.33</td><td>0.19</td><td>0.24</td></tr><tr><td>de-Alpaca</td><td>0.36</td><td>0.39</td><td>0.24</td><td>0.25</td></tr><tr><td></td><td>avg-Alpaca</td><td>0.34</td><td>0.31</td><td>0.24</td><td>0.24</td></tr><tr><td rowspan=\"5\">CrossAlpaca</td><td>zh-CrossAlpaca</td><td>0.70</td><td>0.69</td><td>0.36</td><td>0.28</td></tr><tr><td>ar-CrossAlpaca</td><td>0.56</td><td>0.60</td><td>0.25</td><td>0.25</td></tr><tr><td>it-CrossAlpaca</td><td>0.64</td><td>0.65</td><td>0.28</td><td>0.27</td></tr><tr><td>es-CrossAlpaca</td><td>0.65</td><td>0.64</td><td>0.28</td><td>0.28</td></tr><tr><td>de-CrossAlpaca</td><td>0.64</td><td>0.67</td><td>0.32</td><td>0.29</td></tr><tr><td></td><td>avg-CrossAlpaca</td><td>0.64</td><td>0.65</td><td>0.32</td><td>0.28</td></tr><tr><td></td><td>en-Alpaca vs avg-Alpaca</td><td>0.34(-0.55)</td><td>0.31(-0.56)</td><td>0.24(-0.18)</td><td>0.24(-0.06)</td></tr><tr><td></td><td>en-Alpaca vs avg-CrossAlpaca</td><td>0.64(-0.25)</td><td>0.65(-0.22)</td><td>0.32(-0.10)</td><td>0.28(-0.20)</td></tr><tr><td></td><td>avg-CrossAlpaca vs avg-Alpaca</td><td>(+0.30)</td><td>(+0.34)</td><td>(+0.08)</td><td>(+0.04)</td></tr></table>"
+        },
+        {
+            "type": "table_footnote",
+            "bbox": [
+                0.113,
+                0.466,
+                0.884,
+                0.511
+            ],
+            "angle": 0,
+            "content": "Table 3: Evaluation results on proposed benchmarks. The avg- lines are the averages of x-Alpacas and x-CorssAlpacas. The last three lines indicate the comparisons between en-Alpaca. The last line indicates the comparisons between avg-Alpacas and avg-CrossAlpacas."
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.619,
+                0.238,
+                0.636
+            ],
+            "angle": 0,
+            "content": "B Appendix"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.171,
+                0.643,
+                0.828,
+                0.841
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Language</td><td>Alpaca</td><td>MLQA</td><td>XQUAD</td><td>MMLU</td><td>BBH</td><td>News/commentary</td></tr><tr><td>Arabic</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Chinese</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>English</td><td>x</td><td>x</td><td>x</td><td>x</td><td>x</td><td>x</td></tr><tr><td>German</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Greek</td><td>-</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Hindi</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Italian</td><td>x</td><td>-</td><td>-</td><td>x</td><td>x</td><td>x</td></tr><tr><td>Russian</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Spanish</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Turkish</td><td>x</td><td>-</td><td>x</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Vietnamese</td><td>-</td><td>x</td><td>x</td><td>-</td><td>-</td><td>-</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.324,
+                0.849,
+                0.672,
+                0.864
+            ],
+            "angle": 0,
+            "content": "Table 4: List of available state-of-the-art resources."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7972"
+        }
+    ],
+    [
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.115,
+                0.093,
+                0.24,
+                0.11
+            ],
+            "angle": 0,
+            "content": "C Appendix"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.121,
+                0.49,
+                0.274
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.507,
+                0.12,
+                0.837,
+                0.274
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.286,
+                0.49,
+                0.437
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.508,
+                0.286,
+                0.837,
+                0.437
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.453,
+                0.882,
+                0.47
+            ],
+            "angle": 0,
+            "content": "Figure 6: Evaluation of all proposed benchmarks of the demonstrations used for instruction-tuning our CrossAlpacas."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.115,
+                0.499,
+                0.24,
+                0.516
+            ],
+            "angle": 0,
+            "content": "D Appendix"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.528,
+                0.49,
+                0.68
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.508,
+                0.528,
+                0.837,
+                0.68
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.691,
+                0.49,
+                0.842
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.508,
+                0.691,
+                0.837,
+                0.842
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.86,
+                0.885,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Figure 7: Evaluation of all proposed benchmarks using one-direction Translation-following demonstrations. For en-x for English-foreigner and x-en for foreign English. With the red cross, we indicate the results of the x-Alpaca standards, and with the green diamond, the results of our x-CrossAlpaca."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7973"
+        }
+    ]
+]

2024/Empowering cross-lingual abilities of instruction-tuned large language models by translation-following demonstrations/f7644dfc-cd06-4320-a14e-02a1e95a33d9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d18f30a7ae0bb4c6a339a2e9d132c9bbcebe10e3ec7ee66f9d33c2053392361a
+size 669251

2024/Empowering cross-lingual abilities of instruction-tuned large language models by translation-following demonstrations/full.md

@@ -0,0 +1,315 @@
+# Empowering Cross-lingual Abilities of Instruction-tuned Large Language Models by Translation-following Demonstrations
+
+Leonardo Ranaldi  $(\dagger)$ , Giulia Pucci  $(\star)$ , André Freitas  $(\dagger, *)$
+
+$(\dagger)$  Idiap Research Institute, Martigny, Switzerland
+
+$(\star)$  Department of Computing Science, University of Aberdeen, UK
+
+$(\ast)$ Department of Computer Science, University of Manchester, UK
+
+[first_name].[last_name]@idiap.ch
+
+# Abstract
+
+The language ability of Large Language Models (LLMs) is often unbalanced towards English because of the imbalance in the distribution of the pre-training data. This disparity is demanded in further fine-tuning and affecting the cross-lingual abilities of LLMs. In this paper, we propose to empower Instruction-tuned LLMs (It-LLMs) in languages other than English by building semantic alignment between them. Hence, we propose CrossAlpaca, an It-LLM with cross-lingual Instruction-following and Translation-following demonstrations to improve semantic alignment between languages. We validate our approach on the multilingual Question Answering (QA) benchmarks XQUAD and MLQA and adapted versions of MMLU and BBH. Our models, tested over six different languages, outperform the It-LLMs tuned on monolingual data. The final results show that instruction tuning on non-English data is not enough and that semantic alignment can be further improved by Translation-following demonstrations.
+
+# 1 Introduction
+
+Large Language Models (LLMs) achieve comprehensive language abilities through pre-training on large corpora (Brown et al., 2020; Touvron et al., 2023; Ranaldi and Pucci, 2023b). Hence, the acquired language abilities follow the corpora features, mostly available in English (Lin et al., 2021; Pucci and Ranaldi, 2024). This phenomenon produces an imbalance in both pre-training (Blevins and Zettlemoyer, 2022) and fine-tuning (Le et al., 2021). Consequently, LLMs underperform in languages different from English (Huang et al., 2023).
+
+Efforts to increase multilingual abilities propose continuing pre-training with large-scale monolingual data (Imani et al., 2023; Cui et al., 2023; Yang et al., 2023). However, learning a language from monolingual data requires considerable data and computational resources.
+
+In this paper, we propose CrossAlpaca, an Instruction-tuned LLM (It-LLM) empowered with a semantic alignment between English and other languages. We analyze how to elicit the non-English abilities of It-LLMs by focusing on the crucial phase: instruction-tuning over Instruction-following demonstrations. Hence, we study the effect of cross-lingual alignment by proposing Translation-following demonstrations to improve the instruction-tuning phase.
+
+In our experiments, we use  $LLaMA-7B$  (Touvron et al., 2023) as the LLM backbone and consider six target languages selected from among the available data (shown in Table 4); where data are absent, we perform a translation task. As Instruction-following demonstrations, we use the Stanford Alpaca dataset (Taori et al., 2023) and translated versions in the corresponding languages, while for the Translation-following, we use a publicly available translation resource news Commentary (Tiedemann, 2012), the most accessible and extendable to multiple languages (i.e., CrossAlpaca demonstrations on Figure 1).
+
+Behind an instruction-tuning phase, we evaluate the performance of our six language-specific CrossAlpacas using four different benchmarks: two native-multilingual, i.e., XQUAD (Artetxe et al., 2019) and MLQA (Lewis et al., 2020), and two native-monolinguals, MMLU (Hendrycks et al., 2021) and BBH (Suzgun et al., 2022). The results show that CrossAlpacas instructed with language-specific instruction and translation data far outperform Alpacas instructed only with non-English demonstrations. However, even though CrossAlpaca reduces the gap, the Translation-following models of the original Alpaca still perform best. This result shows that LLaMA's learning abilities on English data are superior to those on non-English data. The semantic alignment task also improves the cross-lingual abilities of It-LLMs.
+
+Our findings can be summarized as follows:
+
+![](images/f77b86829fd878d10ff49daafa9ab637bf2b7f9343234db1ad3e8a9a29f08489.jpg)  
+Figure 1: Our  $x$ -CrossAlpacas are fine-tuned on Instruction-following and Translation-following demonstrations. This example shows the it-CrossAlpaca, fine-tuned on it-Alpaca and Translation-following demonstrations.
+
+- The learning abilities of LLMs on non-English Instruction-tuning tasks are limited;  
+- The multi-lingual abilities of Instruction-tuned LLMs could be empowered through cross-lingual alignment;  
+- Thus, we propose CrossAlpaca, an Instruction-tuning approach for non-English models that are based on Instruction-following and Translation-following demonstration for target language;  
+- We show that It-LLMs are able to semantically align through cross-lingual Translation-following demonstrations via an extensive evaluation on four multi-lingual QA benchmarks.
+
+# 2 Related Work
+
+# 2.1 Pre-trained Language Model on Cross-Linguual Corpora
+
+The next token prediction based on the prefix sequence, as well-known as language modeling, is the everlasting task of modern NLP (Tenney et al., 2019). The extensive knowledge of today's Large Language Models (LLMs) depends on the billions of neurons trained on large-scale corpora with derivatives of the language modeling task. Consequently, the pre-training corpora are predominantly in English, e.g., BooksCorpus (Zhu et al., 2015), MEGATRON-LM (Shoeybi et al., 2019), Gutenberg Dataset (Lahiri, 2014) PILE (Gao et al., 2020), C4 (Dodge et al., 2021), RefinedWeb (Penedo et al., 2023); therefore, LLMs usually have much better knowledge of English than other languages.
+
+Aulamo and Tiedemann (2019); Abadji et al. (2022), in order to solve this problem, propose forward corpora translated into several languages. However, these corpora are not as huge as their competitors, and the lack of massively parallel data in the pre-training corpora also prevents LLMs from aligning the different languages well (Li et al., 2023).
+
+# 2.2 The Instruction-tuning Paradigm
+
+Ouyang et al. (2022); Wei et al. (2022) fine-tuned LLMs using the Instruction-tuning method based on Instruction-tuning data, which are instruction-response corpora, to make LLMs more scalable and improve zero-shot performance. In this method, the LLM backbone is fed with data from the instruction  $(I,X,Y)$ , where  $I$  is an instruction describing the task's requirements,  $X$  is the input, which can be optional, and  $Y$  is the output for the given task. The goal of this method is to minimize the function  $f(Y)$ :
+
+$$
+f (Y) = \arg \min  _ {\theta} \log p _ {\theta} (Y \mid I, X) \tag {1}
+$$
+
+where  $\theta$  are model learnable parameters.
+
+Earlier studies show that the instruction-tuning method of LLMs with both human (Wang et al., 2023) and synthetic-generated instructions (Taori et al., 2023; Xu et al., 2023) empowers the ability of LLMs to solve considerable tasks in zero-shot scenarios.
+
+However, we state that the generally used instruction-tuning datasets, alpaca (Taori et al., 2023), Self-Instruct (Wang et al., 2023), Self-Chat (Xu et al., 2023) or a teacher-student alignment approaches (Ranaldi and Freitas, 2024), conceived
+
+![](images/2b168cb7e40c1e66c87ac707be95f1fbb15447888cbfc812e0b1a71b5f5cbef0.jpg)  
+Figure 2: Examples of Translation-following demonstrations. In this particular example, there are two demonstrations with two different directions from English to German (en-x) and vice versa.
+
+in English, which limits the prospect of LLMs to follow non-English instructions and therefore solve related tasks.
+
+# 2.3 Instruction-tuning is at hand
+
+Large Language Models have considerable success with many techniques in vogue, such as Instruction-tuning. However, their prohibitive size only allows part of the scientific community to experiment with these models.
+
+The latest advances that make these models and techniques accessible involve efficient parameter tuning. Parameter-Efficient Tuning (PEFT) is an efficient technique to adjust a small part of the model parameters and freeze the others. The main goal is to significantly reduce computational and storage costs while maintaining the performance of the original models. The standard techniques for PEFT are LoRA (Hu et al., 2021a), Prefix Tuning (Li and Liang, 2021), P-Tuning (Liu et al., 2022). The basic idea is to keep the pre-trained model weights and incorporate low-rank matrices in each architecture layer. This approach significantly reduces the number of parameters that require training for subsequent tasks, thereby increasing overall efficiency. This building block has been and continues to be very important because it skills the scientific community to fair research and enables the development of multiple open-source works.
+
+# 2.4 Cross-lingual Instruction-tuning Challenge
+
+Recent work has shown the remarkable abilities of LLMs in learning instruction in different languages. Santilli and Rodola (2023); Chen et al. (2023) proposed monolingual fine-tuning LLaMA adaptations on language-specific translated instructions. The
+
+<table><tr><td>Model</td><td>Language</td><td>Name</td></tr><tr><td>Alpaca (Taori et al., 2023)</td><td>English</td><td>en-Alpaca</td></tr><tr><td>Alpaca-Chinese (Chen et al., 2023)</td><td>Chinese</td><td>zh-Alpaca</td></tr><tr><td>Arabic Alpaca (Yasbok)</td><td>Arabic</td><td>ar-Alpaca</td></tr><tr><td>Camascio (Santilli and Rodola, 2023)</td><td>Italian</td><td>it-Alpaca</td></tr><tr><td>Guanaco (Kohaku-Blueleaf, 2023)</td><td>Spanish</td><td>es-Alpaca</td></tr><tr><td>German Alpaca (Thissen, 2023)</td><td>German</td><td>de-Alpaca</td></tr></table>
+
+Table 1: Details and names of mono-lingual Instruction-tuned Large Language Models that use a language-specific version of Alpaca as instruction-tuning data.
+
+use of optimization techniques introduces in Section 2.3, to propose engaging custom adapters for various tasks. Zhang et al. (2023); Ranaldi and Pucci (2023a) investigated the cross-lingual abilities of It-LLMs by promoting the effects of augmenting instruction demonstrations. In particular, Zhang et al. (2023) use fine-grained custom translations that are not easily accessible, while Ranaldi and Pucci (2023a) propose a multilingual approach by considerably increasing instruction and translation tasks. However, these studies have opened an interesting avenue in analyzing the cross-lingual abilities of It-LLMs.
+
+In this paper, we take the next step by proposing our CrossAlpaca. Our methods show of the cross-lingual learning abilities of It-LLMs on a large scale. Using four well-known benchmarks, we show that the weaknesses of It-LLMs trained on non-English data can be strengthened with cross-lingual alignment approaches. Hence, our analyses aim to understand the role of Instruction-following and Translation-following demonstrations in closing the gap in learning and adapting LLMs' abilities to non-English languages.
+
+# 3 Data & Methods
+
+Pre-training from scratch a Large Language Model (LLM) in multiple languages is cost-prohibitive for
+
+data collection and parameter learning. This is the reason why the trend is to do further fine-tuning to strengthen the models' abilities in a specific language (Tanti et al., 2021; Moslem et al., 2023). In this paper, we aim to exploit the abilities of pretrained LLMs for non-English languages by further improving the alignment between English and the target language. Hence, in Section 3.1, we introduce the challenge of fine-tuning an LLM on a mono-lingual custom scenario. Furthermore, in Section 3.2, we propose our approach to fine-tune cross-lingual LLMs.
+
+# 3.1 Mono-lingual Instruction-tuning
+
+The limited accessibility and transparency of paid API services of state-of-the-art LLMs have pushed research toward developing open-source models. Using the instruction-tuning paradigm (Section 2.2) and Stanford Alpaca (Taori et al., 2023), a corpus consisting of  $52\mathrm{k}$  of English instruction-output pairs generated by text-davinci-003, several Instruction-tuned versions of LLaMA were released.
+
+Following this approach, multiple monolingual Instruction-tuned versions of LLaMA were proposed by translating the Stanford Alpaca Instruction-following data into the specific language. Table 4 (Appendix B) shows a set of alpaca versions available as open source. Following a systematic analysis of the translated versions of Alpaca in official repositories<sup>1</sup>, the languages of the benchmark datasets, and the translation pairs present in news/commentary, which will be introduced later, we selected the languages that share the most already available data. Table 1 shows the custom versions used in this work, which for simplicity will be renamed  $x$ -Alpaca, where  $x$  indicates the specific language.
+
+# 3.2 Cross-lingual Instruction-tuning
+
+Mono-lingual Instruction-tuning approaches (Section 3.1) reward LLMs' multi-lingual abilities. However, the use of translated Alpaca alone for the specific language is not sufficient to elicit the non-English ability of LLMs. For this reason, we propose CrossAlpaca (Figure 1), which enriches the cross-lingual Instruction-tuning with the Translation-following demonstrations. We aim to highlight LLMs' English and non-English abilities
+
+by proposing a semantic alignment task to enrich versions of Alpaca.
+
+Instruction-following demonstrations The original version of Alpaca is in English. However, as described in Section 3.1, different open-source translations are produced with a translation engine. In our experiments, we propose the Instruction-tuning phase with the original English version (en-Alpaca) and the translated language-specific versions. Moreover, we propose the CrossAlpaca versions built with the different Alpacas translated into the specific language and the cross-lingual translations (introduced later). In this way, we investigate the abilities of the LLM backbone in understanding multilingual instructions and cross-lingual alignment.
+
+Translation-following demonstrations Learning general instruction data is a sensible strategy for building models to solve multi-tasks directed by instructions (Wang et al., 2023; Zeng et al., 2023). However, translation data could contribute to learning semantic alignment. Zhang et al. (2023) showed that the translation performance of LLMs could be improved by using expertly annotated translation data, and Zhu et al. (2023) showed that it could be beneficial for instruction-tuning.
+
+Inspired by Zhu et al. (2023), we use publicly available sentence-level translation datasets, such as news Commentary (Tiedemann, 2012), to construct the translation task instruction demonstrations. We also propose extending this to additional languages, which we release as an open-source dataset. In particular, for each specific language, we constructed 20k demonstrations. Hence, following the Alpaca style (Instruction, Input, and Output) (see Figure 1), we randomly selected 10k English to non-English translations and 10k random non-English to English translations (all translations come from news commentary). We have constructed these correspondences for the languages mentioned above in Section 3.1. Figure 2 shows a case of Translation-following for English-German direction (en-de) and vice versa (de-en).
+
+# 4 Experiments
+
+In order to observe the English and non-English abilities of Large Language Models (LLMs) and the impact of the Instruction-tuning approach in cross-lingual scenarios, we propose CrossAlpaca. Our approach is based on instruction-tuning on
+
+![](images/14c6bd50518b7dad6a053ed9076ed641c959d0781773a676e56ce2d278bfd5ee.jpg)
+
+![](images/2307526836f0fcee0a0a3e4889ec62229154f0fed6fa8c2f2202bb61cfd60840.jpg)
+
+![](images/7779c4f54976419118b8e6f202bfe60808a9c74554d0c25a2d2682e01cdd6559.jpg)  
+Figure 3: Evaluation results on proposed benchmarks. The dotted line represents the performance of the original Alpaca (Taori et al., 2023) on English tasks.
+
+![](images/b7b4db4804e084e1cbac624dc9cc564401c2a46611542b195046a0b9f7d6b947.jpg)
+
+language-specific data augmented with a crosslingual semantic alignment. Hence, we set several baseline models (Section 4.1), which we augmented with our CrossAlpaca approach (Section 4.2). Finally, we performed a series of systematic evaluations (Section 4.3.1) to observe the impact of the proposed intervention.
+
+# 4.1 Baseline Instruction-tuned LLMs
+
+The common denominator among the It-LLMs shown in Table 1 is the LLM backbone, LLaMA7B (Touvron et al., 2023). Starting from Instruction-following data from the original Alpaca (Taori et al., 2023) and its open-source non-English versions $^{2}$ , we reproduced  $x$ -Alpaca for  $x$  specific languages: Chinese (zh), Italian (it), Arabic (ar), Spanish (es), German (de) and the original English version (en).
+
+# 4.2 Cross-lingual Instruction-tuned LLMs
+
+The CrossAlpacas are instruction-tuned on Instruction-following and Translation-following demonstrations (CrossAlpacas demonstrations). The first ones stem from the resources introduced in Section 4.1. The second comes from
+
+news COMMENTary (Tiedemann, 2012).
+
+Our approach generates a series of instruction-tuned versions of the data shown in Figure 1. We have named the versions  $x$ -CrossAlpaca where  $x$  denotes Chinese (zh), Arabic (ar), Italian (it), Spanish (es), and German (de).
+
+# 4.3 Experimental Setup
+
+In order to assess the performance of the CrossAlpaca, we defined several benchmarks (Section 4.3.1) on which we applied systematic tuning (Section 4.3.2) and evaluation (Section 4.3.3) pipelines.
+
+# 4.3.1 Benchmarks
+
+To evaluate the performance of the It-LLMs and the impact of the translation-based semantic alignment approach, we used two cross-lingual (XQUAD (Artetxe et al., 2019), MLQA (Lewis et al., 2020)) and two multi-task (MMLU (Hendrycks et al., 2021) and BBH (Suzgun et al., 2022)) benchmarks. XQUAD and MLQA focus on understanding questions and answers through translation into different languages. MMLU and BBH, being multi-task benchmarks, include subtasks related to Boolean expressions and QA on basic-level subjects (e.g., chemistry, physics). However, we decided to introduce them to observe whether our approach degrades performance in these tasks. The first two
+
+datasets selected are appropriately constructed for multi-language testing, while the second two are available only in English. So we do a preliminary translation step as outlined below.
+
+# MultiLingual Question Answering (MLQA)
+
+(Lewis et al., 2020) evaluates cross-lingual question answering performance. The benchmark comprises over 5K extractive QA instances in the SQuAD (Rajpurkar et al., 2016) format in several languages. MLQA is highly parallel, with QA instances aligned across four languages on average. Although comprising different languages, some languages are not represented, such as Italian. To conduct the experiments uniformly, we have translated the examples as also done in the forthcoming MMLU and BBH.
+
+Cross-lingual Question Answering Dataset (XQUAD) (Artetxe et al., 2019) consists of a subset of 240 paragraphs and 1190 question-answer pairs from the development set of SQuAD v1.1 (Rajpurkar et al., 2016) with their manual translations into several languages. Consequently, the dataset is entirely parallel across 11 languages.
+
+Massive Multitask Language Understanding (MMLU) (Hendrycks et al., 2021) measures knowledge of the world and problem-solving problems in multiple subjects with 57 subjects across STEM, humanities, social sciences, and other areas. The benchmark is native in English; however, we translated it into five additional languages<sup>3</sup>.
+
+BIG-Bench Hard (BBH) (Suzgun et al., 2022) is a subset of challenging tasks related to navigation, logical deduction, and fallacy detection. Here again, the benchmark is native English, and we have translated it into five languages<sup>3</sup>.
+
+# 4.3.2 Models Setup
+
+In order to align the results with the state-of-the-art models, we used the alpaca_LoRA (Hu et al., 2021b) code $^2$ , adopting the same hyperparameters.
+
+We performed the fine-tuning with a single epoch and a batch-size of 128 examples, running our experiments on a workstation equipped with two Nvidia RTX A6000 with 48 GB of VRAM.
+
+# 4.3.3 Evaluation
+
+As an evaluation metric, we use accuracy. Hence, we estimate accuracy by measuring exact match values in the zero-shot setting. For each model, the parts of benchmarks related to the specific language are used (e.g., for zh-Alpaca and zh-CrossAlpaca, data from MLQA, XQUAD, MMLU, and BBH in Chinese are used).
+
+# 5 Results
+
+Improving non-English abilities in Instruction-tuned Large Language Models (It-LLMs) remains challenging. However, the  $x$ -CrossAlpacas revealed improved results in cross-lingual Question Answering (QA) benchmarks and maintained logical-mathematical skills. From the results of Figure 3 (further detailed in Table 3), it is possible to observe the weaknesses emerging from the fine-tuning of the translated versions of Alpaca (Section 5.1), the improvement obtained from the alignment phase is encouraging (Section 5.2) but it is not enough to outperform the English one. Therefore, we investigated the impact of demonstrations on downstream performance (Section 5.3).
+
+The fine-grained analysis highlighted the importance of cross-lingual alignment data and the critical issues with non-English data. This opens the way for new hypotheses regarding the imbalance of pre-training languages and learning abilities via instruction-tuning.
+
+# 5.1 Translating the Alpaca is not the right way
+
+The instruction-tuning on LLaMA, predominantly pre-trained in English, affects x-Alpaca. Figure 3 and in terms of numbers, Table 3 show that in both MLQA and XQUAD, there is a gap of 55 and 53 average points between en-Alpaca and the x-Alpacas. This phenomenon is mitigated for MMLU and BBH, where we observed an average gap of 18 and 14 points. Instructing an LLM on Alpaca-style demonstrations translated into different languages is not always a good strategy. However, some x-Alpacas, such as zh-Alpaca and de-Alpaca, have performed better. We hypothesize that this phenomenon is related to the scale of the pre-training data in the respective languages and, thus, the abilities of the LLaMA. In future developments, we plan to extend the study on other LLMs beyond LLaMA to observe whether the phenomenon is similar, milder, or more significant.
+
+<table><tr><td>QA
+Task</td><td>en-Alpaca</td><td>avg-Alpaca</td><td>avg-CrossAlpaca</td><td>δ</td></tr><tr><td>MLQA</td><td>0.89</td><td>0.34</td><td>0.64</td><td>+0.30</td></tr><tr><td>XQUAD</td><td>0.97</td><td>0.31</td><td>0.65</td><td>+0.30</td></tr><tr><td>MMLU</td><td>0.42</td><td>0.24</td><td>0.32</td><td>+0.08</td></tr><tr><td>BBH</td><td>0.30</td><td>0.24</td><td>0.28</td><td>+0.04</td></tr></table>
+
+Table 2: Averages of the results on proposed benchmarks. The column  $\delta$  indicates the difference between avg-Alpacas and our avg-CrossAlpacas.
+
+# 5.2 CrossAlpaca: A cross-lingual solution
+
+Semantic alignment through Translation-following demonstrations during fine-tuning could have a valuable impact on the cross-lingual abilities of It-LLMs. The  $x$ -CrossAlpacas outperformed the x-Alpacas by 30 average points on MLQA, 34 average points on XQUAD, 8 on MMLU, and 4 on BBH (see Table 2 and more detailed in Table 3). They also brought their performances closer to the sota obtained from the original Alpaca by 25 points on MLQA and 22 average points on XQUAD. In MMLU and BBH, the gap became very close, with averages of 10 and 2 points (see Table 2 or the line 'en-Alpaca vs avg-CrossAlpaca' in Table 3).
+
+Enriching Translation-following demonstrations has outstanding influences on the cross-lingual abilities of the It-LLMs. However, even in this case, Chinese and German models (zh- and de-CrossAlpaca) outperformed Arabic by many points and, in some specific cases, Spanish and Italian as well. This phenomenon, we hypothesize, is related to the diversity in corpus representation within the pre-training data, as shown in (Yang et al., 2023). Therefore, cross-lingual approaches do not have an incisive impact as in languages less present in the pre-training phases of the language model.
+
+# 5.3 Ablation Study
+
+Our CrossAlpacas, distinguished by the construction of the demonstrations pairs (Section 4.2), achieves significant performance improvements and contributes to closing the gap between the original Alpaca (en-Alpaca) and a series of x-Alpacas in different languages. In order to show the impact of enrichment with cross-lingual demonstrations, we propose two different analyses. In the first analysis, we incrementally decrease the training data, in particular on the Translation-following demonstrations side (Section 5.4). In the second, still working on the Translation-following part (defined by half en-x and half x-en demonstrations), we an
+
+![](images/a08f94c795058f3708498e8634dde81f022dcbb5670e723cfaf3da442e1a6342.jpg)
+
+![](images/161681fe38c2d95b24c8e37313135ddf3d4616b082402bf0274a128c9684f67d.jpg)  
+Figure 4: Evaluation of proposed benchmarks of the demonstrations used for instruction-tuning our CrossAlpacas.
+
+alyze the impact of the demonstrations by splitting the experiments into en-x and x-en (Section 5.4.1).
+
+# 5.4 Translation-following demonstrations empower non-English abilities.
+
+Figure 4 (for a more extensive analysis please refer to Figure 6 in the Appendix) shows the non-English performance of the x-CrossAlpaca on the benchmarks under different scales of Translation-following demonstrations<sup>4</sup>. The x-CrossAlpacas achieve higher accuracy when more demonstrations are used, revealing the benefits of cross-lingual alignment to improve non-English performance. However, Translation-following demonstrations are built in two directions: x-en and en-x (foreign-English and English-foreign). Although equally distributed, we are still determining whether there is an asymmetry between the two types of contributions.
+
+Finally, this result is less pronounced for the multi-task benchmarks (see Figure 6 MMLU and BBH). This could be related to the fact that in the benchmarks, there are logical-mathematical subtasks which are less language dependent (comparatively simpler syntax, smaller vocabularies), and hence, our cross-lingual approach does not influence the results for MLQA and XQUAD as
+
+much. Furthermore, it is possible to observe a trend in language groups, particularly Chinese-German, Italian-Spanish, and Arabic. This trend, which recurs frequently in the results of our experiments, will be further investigated in future work.
+
+![](images/bfe2c307a8eced7305a024eef5b448dfe50ac3468bd0a2b7f826a6c8f7f8fa08.jpg)
+
+![](images/ee2382f505e4bd9f77f0855bbcee68ef9431fdc6bf8ef4c70df5db557eea9517.jpg)  
+Figure 5: Evaluation of proposed benchmarks using one-direction Translation-following demonstrations. For en-x for English-foreigner and x-en for foreign English.
+
+# 5.4.1 Demonstration direction matters on non-English abilities
+
+Figure 5 shows the evaluation conducted in varying directions of the translation-following demonstrations. In particular, demonstrations with English-foreign direction (en-x) better impact downstream models. Conversely, foreign-English (x-en) demonstrations perform better than baselines but underperform demonstrations in the opposite direction. However, as shown in Figure 5 (and complete Figure 7), the x-CrossAlpacas continue to outperform. Nevertheless, the trend of translation-following demonstrations with one direction is interesting. Again, as was the case in the previous ablation study, the multi-task benchmarks (see MMLU and BBH in Figure 7) do not seem to obtain significant influences, which reinforces the hypothesis that the models are greatly affected by cross-lingual capabilities in tasks where there is a strong presence of natural language.
+
+# 6 Future Works
+
+The cross-lingual abilities of Intruction-tuned Large Language Models (It-LLMs) seem to be supported by LLMs, such as in the case of Alpaca, the LLaMA backbone. However, it seems that low-impact demonstrations at the data level can enrich these abilities. We obtained valuable results from our experiments by proposing strategic demonstrations, i.e., Translation-following demonstrations. These results were proposed by performing fine-tuning on LLaMA-7B as done by Taori et al. (2023).
+
+In future work, we would like to continue to investigate by increasing the number of parameters in LLaMA and including additional backbone models. In addition, it might be interesting to evaluate the impact on low-resource languages as done for in-context settings in (Ranaldi et al., 2024). Hence, we would like to get to the underside of performances obtained in some experiments (see Section 5.4) by extending previous epistemic approaches (Ranaldi et al., 2023a,c,b) to It-LLMs. In parallel, plans include analyzing the translation capabilities of general It-LLMs and those enhanced with translation tasks, including some specialized translation tasks among our evaluation benchmarks. Finally, we would like to investigate the learning abilities of the original Alpaca as the translation data change, proposing different probing experiments on (original) English data enhanced with translations.
+
+# 7 Conclusion
+
+In this paper, we proposed CrossAlpaca, an approach to empowering the instruction-tuning of LLMs on non-English data. Specifically, we coupled Instruction-following (Alpaca-style) demonstrations with Translation-following demonstrations. Our method seeks to instruct the LLM to semantic alignment between English and non-English overperforms models instructed on non-English texts. In particular, thanks to our CrossAlpaca demonstrations, the instructed models achieved significant performance improvements on four Question Answering benchmarks XQUAD, MLQA, MMLU, and BBH. In addition, we observe that semantic alignment strengthens with increasing Translation-following data; this demonstrates the de-facto abilities of It-LLMs to learn from instructions. Our approach and results contribute to improved research on the potential for producing more powerful LLMs for non-English languages.
+
+# References
+
+Julien Abadji, Pedro Ortiz Suarez, Laurent Romary, and Benoit Sagot. 2022. Towards a cleaner document-oriented multilingual crawled corpus. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4344-4355, Marseille, France. European Language Resources Association.  
+Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2019. On the cross-lingual transferability of monolingual representations. In Annual Meeting of the Association for Computational Linguistics.  
+Mikko Aulamo and Jörg Tiedemann. 2019. The OPUS resource repository: An open package for creating parallel corpora and machine translation services. In Proceedings of the 22nd Nordic Conference on Computational Linguistics, pages 389-394, Turku, Finland. Linköping University Electronic Press.  
+Terra Blevins and Luke Zettlemoyer. 2022. Language contamination helps explains the cross-lingual capabilities of english pretrained models. In Conference on Empirical Methods in Natural Language Processing.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners.  
+Wei-Lin Chen, Cheng-Kuang Wu, and Hsin-Hsi Chen. 2023. Traditional-chinese alpaca: Models and datasets. https://github.com/ntunlplab/traditional-chinese-alpaca.  
+Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and effective text encoding for chinese llama and alpaca.  
+Jesse Dodge, Maarten Sap, Ana Marasovic, William Agnew, Gabriel Ilharco, Dirk Groeneveld, Margaret Mitchell, and Matt Gardner. 2021. Documenting large webtext corpora: A case study on the colossal clean crawled corpus.  
+Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The pile: An 800gb dataset of diverse text for language modeling.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding.
+
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021a. Lora: Low-rank adaptation of large language models.  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021b. Lora: Low-rank adaptation of large language models.  
+Haoyang Huang, Tianyi Tang, Dongdong Zhang, Wayne Xin Zhao, Ting Song, Yan Xia, and Furu Wei. 2023. Not all languages are created equal in llms: Improving multilingual capability by cross-lingual-thought prompting.  
+Ayyoob Imani, Peiqin Lin, Amir Hossein Kargaran, Silvia Severini, Masoud Jalili Sabet, Nora Kassner, Chunlan Ma, Helmut Schmid, André F. T. Martins, François Yvon, and Hinrich Schütze. 2023. Glot500: Scaling multilingual corpora and language models to 500 languages.  
+Kohaku-Blueleaf. 2023. Guanaco-lora: LoRA for trainin Multilingual Instruction-following LM based on LLaMA. https://huggingface.co/plncmm/guanaco-lora-7b.  
+Shibamouli Lahiri. 2014. Complexity of Word Collocation Networks: A Preliminary Structural Analysis. In Proceedings of the Student Research Workshop at the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 96-105, Gothenburg, Sweden. Association for Computational Linguistics.  
+Hang Le, Juan Miguel Pino, Changhan Wang, Jiatao Gu, Didier Schwab, and Laurent Besacier. 2021. Lightweight adapter tuning for multilingual speech translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 817-824. Association for Computational Linguistics.  
+Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7315-7330, Online. Association for Computational Linguistics.  
+Jiahuan Li, Hao Zhou, Shujian Huang, Shanbo Cheng, and Jiajun Chen. 2023. Eliciting the translation ability of large language models via multilingual finetuning with translation instructions.  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. 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 4582-4597, Online. Association for Computational Linguistics.
+
+Zehui Lin, Xiao Pan, Mingxuan Wang, Xipeng Qiu, Jiangtao Feng, Hao Zhou, and Lei Li. 2021. Pretraining multilingual neural machine translation by leveraging alignment information.  
+Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, Dublin, Ireland. Association for Computational Linguistics.  
+Yasmin Moslem, Rejwanul Haque, John D. Kelleher, and Andy Way. 2023. Adaptive machine translation with large language models.  
+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.  
+Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. 2023. The refined web dataset for falcon llm: Outperforming curated corpora with web data, and web data only.  
+Giulia Pucci and Leonardo Ranaldi. 2024. Does the language matter? curriculum learning over neo-Latin languages. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 5212-5220, Torino, Italia. ELRA and ICCL.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text.  
+Federico Ranaldi, Elena Sofia Ruzzetti, Felicia Logozzo, Michele Mastromattei, Leonardo Ranaldi, and Fabio Massimo Zanzotto. 2023a. Exploring linguistic properties of monolingual berts with typological classification among languages.  
+Leonardo Ranaldi and Andre Freitas. 2024. Aligning large and small language models via chain-of-thought reasoning. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1812-1827, St. Julian's, Malta. Association for Computational Linguistics.  
+Leonardo Ranaldi, Aria Nourbakhsh, Elena Sofia Ruzzetti, Arianna Patrizi, Dario Onorati, Michele Mastromattei, Francesca Fallucchi, and Fabio Massimo Zanzotto. 2023b. The dark side of the language: Pre-trained transformers in the DarkNet. In Proceedings of the 14th International Conference on Recent
+
+Advances in Natural Language Processing, pages 949-960, Varna, Bulgaria. INCOMA Ltd., Shoumen, Bulgaria.  
+Leonardo Ranaldi and Giulia Pucci. 2023a. Does the English matter? elicit cross-lingual abilities of large language models. In Proceedings of the 3rd Workshop on Multi-lingual Representation Learning (MRL), pages 173–183, Singapore. Association for Computational Linguistics.  
+Leonardo Ranaldi and Giulia Pucci. 2023b. Knowing knowledge: Epistemological study of knowledge in transformers. Applied Sciences, 13(2).  
+Leonardo Ranaldi, Giulia Pucci, Federico Ranaldi, Elena Sofia Ruzzetti, and Fabio Massimo Zanzotto. 2024. Empowering multi-step reasoning across languages via tree-of-thoughts.  
+Leonardo Ranaldi, Elena Sofia Ruzzetti, Davide Venditti, Dario Onorati, and Fabio Massimo Zanzotto. 2023c. A trip towards fairness: Bias and de-biasing in large language models.  
+Andrea Santilli and Emanuele Rodola. 2023. Camoscio: an italian instruction-tuned llama.  
+Mohammad Shoeybi, Mostofa Ali Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. 2019. Megatron-LM: Training multi-billion parameter language models using model parallelism. ArXiv, abs/1909.08053.  
+Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, , and Jason Wei. 2022. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261.  
+Marc Tanti, Lonneke van der Plas, Claudia Borg, and Albert Gatt. 2021. On the language-specificity of multilingual bert and the impact of fine-tuning.  
+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.  
+Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics.  
+Martin Thissen. 2023. Fine-tune alpaca for any language. https://github.com/thisserand/alpaca-lora-finetune-language.  
+Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12), pages 2214-2218, Istanbul,
+
+Turkey. European Language Resources Association (ELRA).  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, 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. 2023. Self-instruct: Aligning language models 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. 2022. Finetuned language models are zero-shot learners.  
+Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. 2023. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. arXiv preprint arXiv:2304.01196.  
+Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. 2023. Bigtranslate: Augmenting large language models with multilingual translation capability over 100 languages.  
+Yasbok. Alpaca Instruction Fine-Tuning for Arabic. https://huggingface.co/Yasbok.  
+Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2023. Tim: Teaching large language models to translate with comparison.  
+Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang, Zhengrui Ma, Yan Zhou, Langlin Huang, Mengyu Bu, Shangtong Gui, Yunji Chen, Xilin Chen, and Yang Feng. 2023. Bayling: Bridging cross-lingual alignment and instruction following through interactive translation for large language models.  
+Wenhao Zhu, Yunzhe Lv, Qingxiu Dong, Fei Yuan, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. 2023. Extrapolating large language models to non-english by aligning languages.  
+Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV).
+
+A Appendix  
+
+<table><tr><td></td><td>Model</td><td>MLQA</td><td>XQUAD</td><td>MMLU</td><td>BBH</td></tr><tr><td></td><td>en-Alpaca</td><td>0.89</td><td>0.87</td><td>0.42</td><td>0.30</td></tr><tr><td rowspan="5">Alpaca</td><td>zh-Alpaca</td><td>0.48</td><td>0.38</td><td>0.26</td><td>0.25</td></tr><tr><td>ar-Alpaca</td><td>0.17</td><td>0.16</td><td>0.18</td><td>0.20</td></tr><tr><td>it-Alpaca</td><td>0.35</td><td>0.32</td><td>0.21</td><td>0.25</td></tr><tr><td>es-Alpaca</td><td>0.32</td><td>0.33</td><td>0.19</td><td>0.24</td></tr><tr><td>de-Alpaca</td><td>0.36</td><td>0.39</td><td>0.24</td><td>0.25</td></tr><tr><td></td><td>avg-Alpaca</td><td>0.34</td><td>0.31</td><td>0.24</td><td>0.24</td></tr><tr><td rowspan="5">CrossAlpaca</td><td>zh-CrossAlpaca</td><td>0.70</td><td>0.69</td><td>0.36</td><td>0.28</td></tr><tr><td>ar-CrossAlpaca</td><td>0.56</td><td>0.60</td><td>0.25</td><td>0.25</td></tr><tr><td>it-CrossAlpaca</td><td>0.64</td><td>0.65</td><td>0.28</td><td>0.27</td></tr><tr><td>es-CrossAlpaca</td><td>0.65</td><td>0.64</td><td>0.28</td><td>0.28</td></tr><tr><td>de-CrossAlpaca</td><td>0.64</td><td>0.67</td><td>0.32</td><td>0.29</td></tr><tr><td></td><td>avg-CrossAlpaca</td><td>0.64</td><td>0.65</td><td>0.32</td><td>0.28</td></tr><tr><td></td><td>en-Alpaca vs avg-Alpaca</td><td>0.34(-0.55)</td><td>0.31(-0.56)</td><td>0.24(-0.18)</td><td>0.24(-0.06)</td></tr><tr><td></td><td>en-Alpaca vs avg-CrossAlpaca</td><td>0.64(-0.25)</td><td>0.65(-0.22)</td><td>0.32(-0.10)</td><td>0.28(-0.20)</td></tr><tr><td></td><td>avg-CrossAlpaca vs avg-Alpaca</td><td>(+0.30)</td><td>(+0.34)</td><td>(+0.08)</td><td>(+0.04)</td></tr></table>
+
+Table 3: Evaluation results on proposed benchmarks. The avg- lines are the averages of x-Alpacas and x-CorssAlpacas. The last three lines indicate the comparisons between en-Alpaca. The last line indicates the comparisons between avg-Alpacas and avg-CrossAlpacas.
+
+B Appendix  
+
+<table><tr><td>Language</td><td>Alpaca</td><td>MLQA</td><td>XQUAD</td><td>MMLU</td><td>BBH</td><td>News/commentary</td></tr><tr><td>Arabic</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Chinese</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>English</td><td>x</td><td>x</td><td>x</td><td>x</td><td>x</td><td>x</td></tr><tr><td>German</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Greek</td><td>-</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Hindi</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Italian</td><td>x</td><td>-</td><td>-</td><td>x</td><td>x</td><td>x</td></tr><tr><td>Russian</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Spanish</td><td>x</td><td>x</td><td>x</td><td>-</td><td>-</td><td>x</td></tr><tr><td>Turkish</td><td>x</td><td>-</td><td>x</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Vietnamese</td><td>-</td><td>x</td><td>x</td><td>-</td><td>-</td><td>-</td></tr></table>
+
+Table 4: List of available state-of-the-art resources.
+
+![](images/95b2576b4f00518064c21a463161694c253da331496aa0f224304cff16647117.jpg)  
+C Appendix
+
+![](images/97e188ad7570aa8e4c8d091a81fc470daae5c534cea96487512ff9d81aad08bf.jpg)
+
+![](images/43ed6c5e9c2d6d92547db2ea9bfa863bdb6642ab0aacbf2a9931aca7a2573578.jpg)  
+Figure 6: Evaluation of all proposed benchmarks of the demonstrations used for instruction-tuning our CrossAlpacas.
+
+![](images/a275604db23ecc34bfac52ad45e2279e361a3693d98de3a571a3ee324bcb7d1e.jpg)
+
+![](images/c5ccce64f4266f9b341fdd9b793ea081c0f51cff55dced769c3f2f4cce303d12.jpg)  
+D Appendix
+
+![](images/3f986813402f6745ae4a46f0d55f1f2eb512387b0bf7e536193d91f199e31a3f.jpg)
+
+![](images/7cd59f68ce3891a7d078cecc93fa5f0d4f1244504d22ebed5f2b327b8db5e8c7.jpg)  
+Figure 7: Evaluation of all proposed benchmarks using one-direction Translation-following demonstrations. For en-x for English-foreigner and x-en for foreign English. With the red cross, we indicate the results of the x-Alpaca standards, and with the green diamond, the results of our x-CrossAlpaca.
+
+![](images/2d2056d9b83128c5ae6440a8bd75de06f61e2bb4b874c5122cfe93559b72a19a.jpg)

2024/Empowering cross-lingual abilities of instruction-tuned large language models by translation-following demonstrations/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:016d1e7026e117b82824701bc49792889ebebb7b981e1b3bc153a41e37f09654
+size 500993

2024/Encoding Hierarchical Schema via Concept Flow for Multifaceted Ideology Detection/5e9641cb-740e-48d6-bdd3-e4eca45bddbc_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ce8f3ed3a8d0f5da751186c4693463f38011fbb8d457492a998bcb578800a0f
+size 744785

2024/Encoding Hierarchical Schema via Concept Flow for Multifaceted Ideology Detection/full.md

@@ -0,0 +1,479 @@
+# Encoding Hierarchical Schema via Concept Flow for Multifaceted Ideology Detection
+
+Songtao Liu $^{1}$ , Bang Wang $^{1,*}$ , Wei Xiang $^{2}$ , Han Xu $^{3}$  and Minghua Xu $^{3,*}$
+
+$\{^1$  School of Electronic Information and Communications,
+
+$^{2}$ School of Software Engineering,
+
+$^{3}$ School of Journalism and Information Communication\},
+
+Huazhong University of Science and Technology, Wuhan, China
+
+{liusongtao, wangbang, xiangwei, xuh, xuminghua}@hust.edu.cn
+
+# Abstract
+
+Multifaceted ideology detection (MID) aims to detect the ideological leanings of texts towards multiple facets. Previous studies on ideology detection mainly focus on one generic facet and ignore label semantics and explanatory descriptions of ideologies, which are a kind of instructive information and reveal the specific concepts of ideologies. In this paper, we develop a novel concept semantics-enhanced framework for the MID task. Specifically, we propose a bidirectional iterative concept flow (BICo) method to encode multifaceted ideologies. BICo enables the concepts to flow across levels of the schema tree and enriches concept representations with multi-granularity semantics. Furthermore, we explore concept attentive matching and concept-guided contrastive learning strategies to guide the model to capture ideology features with the learned concept semantics. Extensive experiments on the benchmark dataset show that our approach achieves state-of-the-art performance in MID, including in the cross-topic scenario. $^{1}$
+
+# 1 Introduction
+
+Multifaceted ideology detection (MID) aims to identify the ideological leanings (e.g., Left, Center, Right, etc.) expressed in texts towards multiple facets, as shown in Figure 1. It is crucial for understanding public opinion and detecting potential extremism (Kannangara, 2018; Grover and Mark, 2019; Demszky et al., 2019), which is helpful for governments and cybersecurity organizations (Stefanov et al., 2020; Aldera et al., 2021). It can also facilitate downstream research and applications in social sciences (Kabir and Madria, 2022).
+
+In most of related work, researches generally focus on modeling the text content with diversified cues, such as sentiment polarities (Bhatia and P,
+
+![](images/5cc4f86365a128c7f365501802b034057382b1e82fb713af6f09758f0b377107.jpg)
+
+![](images/2188c931700d950b1799325610cba1ae822215f3cd0f566d392c98f2a7879f4b.jpg)
+
+![](images/b896a87c6c5574930fafe8139a7d8aea2b9626d83c1882cfec5b8e5bf9813ff6.jpg)  
+Figure 1: Upper: the multifaceted ideology schema and concepts of facets and ideologies (Liu et al., 2023). Lower left: the tree-like hierarchical structure of the schema. Lower right: an example of MID. "L" denotes Left, "R" denotes Right.
+
+![](images/f3b38d09630387dd1be8f01748d7329c83db9ab93277664c37565264ec859300.jpg)
+
+2018; Kabir and Madria, 2022), named entities (Liu et al., 2022), and discourse structure (Devatine et al., 2023; Hong et al., 2023), or jointly learning with other related tasks (Baly et al., 2019). There are also approaches that incorporate information sources beyond text to facilitate ideology mining. Hyperlink structure (Kulkarni et al., 2018), social networks (Stefanov et al., 2020; Li and Goldwasser, 2021), external knowledge from knowledge graphs (Zhang et al., 2022) as well as information from other modalities (Qiu et al., 2022), are introduced in the task of ideology detection.
+
+Although achieving promising performance, those methods limit the ideology prediction to a generic facet. In other words, they only label a text as ideologically left- or right-leaning as a whole, regardless whether the text containing one or more different facets. Furthermore, they ignore a crucial clue, label semantics, that is, what exactly does an ideology mean? In this case, ideological categories are represented as one-hot vectors without any semantic information, and models can only rely on
+
+the training data distribution to analyze latent ideology features, which could be unfavorable for the generalization ability of models (Wang et al., 2021; Wen and Hauptmann, 2023).
+
+So how can we effectively detect multifaceted ideologies? And what exactly does an ideology mean? Liu et al. (2023) propose the multifaceted ideology detection task for the first time and design a multifaceted ideology schema which contains 12 facets covering 5 domains in a tree-like hierarchical structure (see Figure 1, details in Appendix A). Each facet, as well as the ideological attributes under each facet, are defined using natural texts, which can be regarded as concepts. These concepts describe the meaning of facets and ideologies, thus making it natural to represent the label semantics. In addition, in the hierarchical schema, higher-level concepts (like Domain and Facet) have general semantics shared by their child concepts, while lower-level concepts (like Ideology and Facet) describe their parents from various views, which can be seen as the semantic divisions of higher-level concepts. This meaningful hierarchical structure can be utilized to enrich the concept semantics.
+
+Based on the motivation above, to incorporate the concept semantics and leverage the hierarchical structure of the schema in MID, we propose a novel Bidirectional Iterative Concept Flow (BICo) method to encode the hierarchical schema. Specifically, BICo allows concepts to flow in two directions on the schema tree, enabling them to perceive both high-level general semantics and low-level specific perspectives. On the one hand, inspired by the relation rotation in complex space (Sun et al., 2018), we design Concept Metapath Diffusion to perform message passing from root to leaf. On the other hand, in the direction of leaf to root, we propose Concept Hierarchy Aggregation to aggregate concept semantics in lower levels to the ones in higher levels based on the parent-child relation. Concept flow in the two directions is iterated multiple times and the final concept representations are enriched by multi-granularity semantics. For example, the Facet representations capture the meanings of different ideologies in the corresponding facet, while the Ideology representations also perceive information about the Facet and Domain they belong to. We match the text and Facet representations based on the attention mechanism to recognize text-related facets. Furthermore, we explore a Concept-Guided Contrastive Learning strategy to learn more distinguishable text representations
+
+under the guidance of Ideology concepts.
+
+The main contributions of our work are summarized as follows:
+
+(1) We propose a concept semantics-enhanced MID framework. To our best knowledge, this is the first work that incorporates label semantics and explanatory descriptions in the MID task.  
+(2) We propose a Bidirectional Iterative Concept Flow (BICo) method to encode the hierarchical schema. Concepts flow on the schema tree in two directions iteratively to capture multi-granularity concept semantics.  
+(3) We design Concept Attentive Matching and Concept-Guided Contrastive Learning strategies to enable the model to extract ideology features with the help of concept semantics.  
+(4) Extensive experiments on the MITweet benchmark demonstrate the effectiveness of our approach, including in the cross-topic scenario.
+
+# 2 Task Description
+
+Given an input text and a set of facets, Multi-faceted Ideology Detection (MID) is divided into two sub-tasks: (1) Relevance Recognition aims to recognize the facets that the text is related to; (2) Ideology Analysis predicts which ideology the text holds towards the related facets. Formally, a sample instance can be considered as a triple  $\left(x,\{y_R^i\}_{i = 1}^n,\{y_I^i\}_{i = 1}^m\right)$ , where  $x$  is the input text,  $y_{R}^{i}\in \{\text{Related, Unrelated}\}$  represents the relevance label of  $i$ -th facet,  $n$  is the number of given facets. For each facet that the text is related to, we have an ideology label  $y_{I}^{i}\in \{\text{Left, Center, Right}\}$ ,  $m$  is the number of related facets.
+
+# 3 Approach
+
+In this section, we first introduce the proposed Bidirectional Iterative Concept Flow (BICo) for encoding the hierarchical schema, and then discuss how we augment multifaceted ideology detection based on the learned concept encodings. Figure 2 illustrates the overall structure of our model.
+
+# 3.1 Bidirectional Iterative Concept Flow
+
+# 3.1.1 Concept Hierarchy Tree
+
+Liu et al. (2023) define the first hierarchical schema of multifaceted ideology, which contains 12 facets covering 5 domains. We construct a concept hierarchy tree  $T = (N,E)$  based on the schema, as shown in Figure 2. The node set  $N$  contains four types of nodes, i.e., Root, Domain, Facet and
+
+![](images/6baa03f36b7cfd44a15f25c43091fb2ce9b22afd1f8ac46c75ca150458416a79.jpg)  
+Figure 2: Overview of our concept-enhanced multifaceted ideology detection framework. The blue box in the middle shows the proposed bidirectional iterative concept flow (BICo), which includes root-to-leaf concept metapath diffusion and leaf-to-root concept hierarchy aggregation. The concept representations are enriched gradually by bidirectional iteration, and are then used to enhance the two subtasks of MID through concept attentive matching and concept-guided contrastive learning.
+
+Ideology. The edge set  $E$  indicates the subordination relation between nodes. The Ideology, or leaf, nodes represent the three ideologies (Left, Center, Right) of each facet.
+
+To initialize node embeddings in the concept hierarchy tree, we leverage the concepts of facets and ideologies in the schema. Specifically, we adopt a pre-trained language model as the concept encoder and feed the concepts of facets and ideologies in the schema into the encoder. We then extract the hidden state of [CLS] token as initial representations of Facet and Ideology nodes, i.e.,  $\mathbf{h}_F$  and  $\mathbf{h}_I$ . For Root and Domain nodes, we obtain their initial embeddings  $(\mathbf{h}_R$  and  $\mathbf{h}_D)$  by average-pooling their child node embeddings.
+
+# 3.1.2 Concept Metapath Diffusion
+
+In the concept hierarchy tree, higher-level nodes (like Root and Domain) have general and abstract concepts, which are shared by their child nodes and could be beneficial for enriching the representations of lower-level nodes (like Ideology and Facet). In order to allow lower-level nodes to perceive higher-level abstract semantics, we adopt the relation rotation in complex space (Sun et al., 2018), which is effective for information transfer along edges in a sequential structure.
+
+Specifically, we define concept metapath as a path from root to leaf (Root - Domain - Facet - Ideology). Given node representations in a metapath  $(\mathbf{h}_R,\mathbf{h}_D,\mathbf{h}_F,\mathbf{h}_I) = (\mathbf{h}_0,\mathbf{h}_1,\mathbf{h}_2,\mathbf{h}_3)$ , let  $\mathbf{r}_i$  be the representation of edge between node  $h_i$  and  $h_{i + 1}$ , the concept metapath diffusion from root to leaf through relation rotation is formulated as:
+
+$$
+\mathbf {o} _ {0} = \mathbf {h} _ {0} ^ {\prime} = \mathbf {h} _ {0} \tag {1}
+$$
+
+$$
+\mathbf {h} _ {i} ^ {\prime} = \mathbf {h} _ {i} + \mathbf {h} _ {i - 1} ^ {\prime} \odot \mathbf {r} _ {i - 1} \tag {2}
+$$
+
+$$
+\mathbf {o} _ {i} = \frac {\mathbf {h} _ {i} ^ {\prime}}{i + 1} \tag {3}
+$$
+
+where  $\mathbf{h}_i$ ,  $\mathbf{r}_i$  and  $\mathbf{o}_i$  are all complex vectors,  $\mathbf{o}_i$  is the updated embedding,  $\odot$  is the element-wise complex product and performs vector rotation in complex space. Here we can easily interpret a real vector of dimension  $d$  as a complex vector of dimension  $d/2$  by treating the first half of the vector as the real part and the second half as the imaginary part. We perform concept diffusion on all metapaths in the tree and  $\mathbf{r}_i$  is shared between each two consecutive levels of nodes. Note that in relation rotation, edges represent the rotation angles of vectors in complex space. Therefore,  $\mathbf{r}_i$  is first randomly initialized in the range of  $(- \pi, \pi)$ , and then its real and imaginary parts are obtained
+
+by the Euler's formula.
+
+# 3.1.3 Concept Hierarchy Aggregation
+
+In contrast to metapath diffusion, concept hierarchy aggregation enables concept flow from leaf to root. In the concept hierarchy tree, child nodes describe their parent node from different views, and thus can be regarded as more fine-grained concepts. Through concept hierarchy aggregation, we aggregate concept semantics of child nodes to their parent node, so as to enrich the representations of higher-level nodes.
+
+We utilize the graph attention network (GAT), which aggregates features through attention mechanism in a graph. Considering the characteristics of concept hierarchy tree, we modify it to explicitly model the hierarchical structure and quantitatively measure the compatibility between hierarchies in the tree. Specifically, we only establish aggregation between the parent node and its own child nodes, which is different from aggregating over all one-hop neighbors in GAT. Additionally, we use different attention parameters at different levels to distinguish the aggregation features of each hierarchy.
+
+Formally, for a parent node  $p$  with embedding  $\mathbf{h}_p$ , we compute an aggregation weight for each child node  $i$  and then weighted sum all child nodes' embeddings:
+
+$$
+e _ {p i} = \text {L e a k y R e L U} \left(\mathbf {A} _ {l} \left(\mathbf {h} _ {p} \| \mathbf {h} _ {i}\right)\right) \tag {4}
+$$
+
+$$
+\alpha_ {p i} = \frac {\exp \left(e _ {p i}\right)}{\sum_ {j \in \mathcal {C} _ {p} \cup \{p \}} \exp \left(e _ {p j}\right)} \tag {5}
+$$
+
+$$
+\mathbf {h} _ {p} ^ {\prime} = \sigma \left(\sum_ {i \in \mathcal {C} _ {p} \cup \{p \}} \alpha_ {p i} \mathbf {h} _ {i}\right) \tag {6}
+$$
+
+where  $\mathbf{A}_l$  is the learnable parameter for aggregation of nodes in level  $l$ ,  $\mathcal{C}_p$  is the child node set of  $p$ ,  $\mathbf{h}_p^{\prime}$  is the updated representation for  $p$ .
+
+# 3.1.4 Bidirectional Iteration
+
+The root-to-leaf metapath diffusion and leaf-to-root hierarchy aggregation are iterated multiple times to update node encodings. Finally, The new generated concept representations can be fully aware of higher-level general semantics and constructed with concepts from different aspects. Next we will enhance the MID task with the enriched Facet and Ideology representations,  $\mathbf{c}_F$  and  $\mathbf{c}_I$ .
+
+# 3.2 Concept-Enhanced MID
+
+# 3.2.1 Text Encoder
+
+We select a pre-trained language model as the text encoder. In the subtask of Relevance Recognition, the encoder processes input sequence and outputs a hidden representation for each token:  $\mathbf{X} = \{\mathbf{x}_i\}_{i=1}^L$ , where  $L$  is the length of text. For Ideology Analysis, we concatenate the text and its related facet concept, and then feed the sequence into text encoder to acquire the hidden state of [CLS] as text representation  $\mathbf{t}$ .
+
+# 3.2.2 Concept Attentive Matching
+
+In Relevance Recognition subtask, to enable the text to be aware of label semantics (i.e., Facet concepts) and measure the importance of each token in relevance feature extraction, we adopt the cross-attention mechanism (Vaswani et al., 2017) to match the Facet and input token representations:
+
+$$
+\mathbf {t} ^ {i} = \operatorname {s o f t m a x} \left(\frac {\mathbf {c} _ {F} ^ {i} \mathbf {X} ^ {T}}{\sqrt {d}}\right) \mathbf {X} \tag {7}
+$$
+
+where  $d$  is the dimension of vectors in the equation, the superscript  $i$  represents  $i$ -th facet,  $\mathbf{c}_F^i$  is  $i$ -th facet representation and  $\mathbf{t}^i$  is  $i$ -th facet-aware text representation.
+
+# 3.2.3 Concept-Guided Contrastive Learning
+
+To inject label semantics (i.e., Ideology concepts) into Ideology Analysis subtask, we further explore a Concept-Guided Contrastive Learning strategy (CGCL), which tries to make intra-ideology representations more compact in the feature space and inter-ideology ones more distinguishable with the ideology concepts as anchors. The motivation is that ideology concepts describe the general meaning of ideologies. In the embedding space, this property can be interpreted as clustering, where an ideology concept anchor is the semantic center of samples with that ideological category.
+
+Specifically, given text representations  $\mathcal{B} = \{\mathbf{t}_i\}_{i=1}^B$  in a batch ( $B$  is the batch size), and three Ideology representations  $\mathcal{A} = \{\mathbf{c}_{I,L}, \mathbf{c}_{I,C}, \mathbf{c}_{I,R}\}$  (corresponding to Left, Center and Right respectively) which will be used as concept anchors in the vector space, the concept-guided contrastive
+
+loss is formulated as:
+
+$$
+\begin{array}{l} \mathcal {L} _ {C G C L} = \frac {1}{3} \sum_ {i \in \{L, C, R \}} \mathcal {L} _ {i} (8) \\ \mathcal {L} _ {i} = \log \frac {\sum_ {j \in \{j \mid y _ {I , j} = i \}} \exp \left(f \left(\mathbf {c} _ {I , i} , \mathbf {t} _ {j}\right) / \tau\right)}{\sum_ {\mathbf {v} \in \mathcal {B} \cup \mathcal {A} \backslash \{c _ {I , i} \}} \exp \left(f \left(\mathbf {c} _ {I , i} , \mathbf {v}\right) / \tau\right)} (9) \\ \end{array}
+$$
+
+where  $y_{I,j}$  is the ideology label of  $\mathbf{t}_j$ ,  $f$  is the cosine similarity function,  $\tau$  is temperature parameter. Note that  $\mathcal{L}_{CGCL}$  is computed for each facet, and we omit the facet superscript for clarity.
+
+# 3.2.4 Classification and Training
+
+Considering the varying ideology features among different facets, we set up a classification head with a softmax function for each facet in both subtasks:
+
+$$
+y ^ {i} = \operatorname {s o f t m a x} \left(\mathbf {W} ^ {i} \mathbf {t} ^ {i} + \mathbf {b} ^ {i}\right) \tag {10}
+$$
+
+where the superscript  $i$  represents  $i$ -th facet,  $\mathbf{W}^i$  and  $\mathbf{b}^i$  are trainable parameters.
+
+Note that in Relevance Recognition, we also incorporate contrastive learning (CL), which is similar to the concept-guided CL in Sec. 3.2.3, but the anchors here are text representations themselves:
+
+$$
+\mathcal {L} _ {C L} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\sum_ {j \in \mathcal {B} _ {i}} \exp \left(f \left(\mathbf {t} _ {i} , \mathbf {t} _ {j}\right) / \tau\right)}{\sum_ {k \in \{k \mid i \neq k \}} \exp \left(f \left(\mathbf {t} _ {i} , \mathbf {t} _ {k}\right) / \tau\right)} \tag {11}
+$$
+
+where  $\mathbf{t}_i$  is the facet-aware text representation,  $\mathcal{B}_i = \{j|i\neq j,y_{R,i} = y_{R,j}\}$ $y_{R,i}$  is the relevance label of  $\mathbf{t}_i$ $B$  is the batch size,  $\tau$  is temperature parameter. Here  $\mathcal{L}_{CL}$  is also computed for each facet, and we omit the facet superscript for clarity.
+
+Finally, the training loss of both subtasks is the weighted sum of cross-entropy classification loss and contrastive learning loss across all facets:
+
+$$
+\mathcal {L} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\mathcal {L} _ {C E} ^ {i} + \lambda \mathcal {L} _ {C L} ^ {i}\right) \tag {12}
+$$
+
+where  $\mathcal{L}_{CE}^{i}$  is the cross-entropy loss of  $i$ -th facet,  $\lambda$  is a hyper-parameter controlling the weight of contrastive loss,  $n$  is the total number of facets.
+
+# 4 Experiments
+
+# 4.1 Dataset and Evaluation Metrics
+
+We conduct experiments on the MITweet (Liu et al., 2023) dataset, which contains 12,594 English tweets and covers 14 highly controversial topics in
+
+recent years. Each instance in MITweet is annotated with a relevance label and an ideology label (if the relevance label is “Related”) for each of the 12 facets in the multifaceted ideology schema. The statistics of MITweet is shown in Table 6.
+
+We follow the original training/validation/test split and use the same evaluation metrics as Liu et al. (2023). First we calculate the Accuracy (Acc) and F1 score for each facet. Then we utilize both Macro and Micro methods to integrate metrics from all facets to obtain overall results of model performance. Macro-F1 and Macro-Acc are calculated by averaging F1 and Acc across all facets. Micro-F1 and Micro-Acc are the aggregated F1 and Acc scores obtained by concatenating the predictions of all facets. Note that, following existing work, we only report F1-related metrics for Relevance Recognition due to the highly imbalanced data distribution in this subtask.
+
+# 4.2 Implement Details
+
+The pre-trained BERTweet-base (Nguyen et al., 2020) is used as the concept and text encoder, and the two encoders share weights as this gave better results in preliminary experiments. We train the Relevance Recognition model and the Ideology Analysis model independently. Each model includes the BICo module and is trained end-to-end. We use AdamW (Loshchilov and Hutter, 2018) as the optimizer. The learning rate is set to 2e-5. The batch size  $B$  is set to 64. The iteration number of BICo is set to 4 for relevance recognition and 2 for ideology analysis. For contrastive loss, we set the temperature parameter  $\tau$  to 0.5 for relevance recognition and 0.1 for ideology analysis. The contrastive loss weight  $\lambda$  is set to 0.3 for both subtasks. The classification head is a two-layer fully connected network, in which the hidden size is 512. The above parameters are selected based on the validation set. We report the average results of 5 runs with different random seeds.
+
+# 4.3 Comparison Models
+
+We compare our approach with the latest benchmark in the MID task, BERTweetInd (Liu et al., 2023), which uses BERTweet as the backbone and detects indicator words from training set as the textual descriptions of facets. In addition, we test the zero/few-shot performance of advanced large language models (LLMs) in this task. Specifically, we select two popular LLMs, LLaMA2 (Touvron
+
+<table><tr><td>Model</td><td>Macro-F1</td><td>Micro-F1</td><td>Macro-Acc</td><td>Micro-Acc</td></tr><tr><td colspan="5">Subtask 1: Relevance Recognition</td></tr><tr><td>BERTweetInd</td><td>57.48</td><td>70.32</td><td>-</td><td>-</td></tr><tr><td>LLaMA2-13B°</td><td>27.45</td><td>32.28</td><td>-</td><td>-</td></tr><tr><td>ChatGPT°</td><td>33.11</td><td>40.07</td><td>-</td><td>-</td></tr><tr><td>LLaMA2-13B△</td><td>29.35</td><td>38.17</td><td>-</td><td>-</td></tr><tr><td>ChatGPT△</td><td>38.83</td><td>44.78</td><td>-</td><td>-</td></tr><tr><td>Our approach</td><td>59.22†</td><td>72.14†</td><td>-</td><td>-</td></tr><tr><td>w/o CL</td><td>58.56</td><td>71.42</td><td>-</td><td>-</td></tr><tr><td>w/o BICo</td><td>58.14</td><td>70.41</td><td>-</td><td>-</td></tr><tr><td>w/o CL&amp;BICo</td><td>57.85</td><td>70.42</td><td>-</td><td>-</td></tr><tr><td colspan="5">Subtask 2: Ideology Analysis</td></tr><tr><td>BERTweetInd</td><td>42.68</td><td>69.28</td><td>65.88</td><td>76.38</td></tr><tr><td>LLaMA2-13B°</td><td>35.60</td><td>47.33</td><td>45.98</td><td>49.69</td></tr><tr><td>ChatGPT°</td><td>37.11</td><td>53.41</td><td>48.57</td><td>57.95</td></tr><tr><td>LLaMA2-13B△</td><td>38.51</td><td>47.22</td><td>46.13</td><td>48.90</td></tr><tr><td>ChatGPT△</td><td>42.64</td><td>60.54</td><td>58.44</td><td>68.25</td></tr><tr><td>Our approach</td><td>47.32†</td><td>70.90†</td><td>66.79</td><td>78.60†</td></tr><tr><td>w/o BICo</td><td>46.02</td><td>68.58</td><td>66.18</td><td>76.63</td></tr><tr><td>w/o Concept anchors</td><td>45.08</td><td>68.38</td><td>66.04</td><td>77.30</td></tr><tr><td>w/o CGCL</td><td>44.21</td><td>67.54</td><td>65.15</td><td>76.79</td></tr></table>
+
+Table 1: Overall results of different models and ablation study.  $\circ$  and  $\bigtriangleup$  denote 0-shot and 3-shot,respectively. † denotes the significance test over BERTweetInd at p-value<0.05. Bold values are the best results in the corresponding subtask.
+
+et al., 2023) and ChatGPT $^2$ , which exhibit superior capacities in communicating with humans, including solving a wide range of complex tasks without further training. We use the Llama-2-13b-chat and gpt-3.5-turbo-1106 versions. The prompts designed for LLMs can be found in Appendix B.
+
+We also provide variants of our proposed approach in the ablation study:
+
+- Relevance Recognition
+
+(1) "w/o CL" denotes without contrastive learning.  
+(2) "w/o BICo" denotes without bidirectional iterative concept flow, in which case the facet representations in Concept Attentive Matching are directly from the concept encoder.  
+(3) "w/o CL&BICo" denotes the combination of the above two cases.
+
+- Ideology Analysis
+
+(1) "w/o BICo" denotes without bidirectional iterative concept flow. In this case, the concept anchors (i.e., ideology representations) are directly
+
+from the concept encoder.
+
+(2) "w/o concept anchors" denotes performing the contrastive learning without the guidance of concept anchors, i.e., the anchors are text representations themselves, which is the case of Eq. (11).  
+(3) "w/o CGCL" denotes discarding the concept-guided contrastive learning.
+
+# 4.4 Main Results
+
+We present the overall results of our approach and other models in Table 1. First, we can observe that our concept-enhanced method performs consistently better than other baseline models, including the advanced large language models, indicating the superiority of our approach for the MID task. Second, compared with BERTweetInd, which is also a BERTweet-based model, our approach achieves significant improvements in both subtasks. This suggests that the application of concept semantics in the hierarchical schema helps the model to capture the correlation between text and labels, thus improving the performance. Third, for the LLMs, although ChatGPT performs better than LLaMA2-13B, and the few in-context demonstrations improve the results, there is still a large gap between LLMs and other task-specific models. This indicates that the MID task remains challenging for current LLMs. One possible reason is that, this task requires not only strong text understanding and semantic reasoning abilities, but also the integration of specialized sociological knowledge and background information on relevant topics, which is difficult for general-purpose LLMs.
+
+In more detail, F1 scores of different models on each facet are shown in Table 2. In the subtask of Relevance Recognition, our approach achieves the best results on 10 out of 12 facets, surpassing the second-place by over 4 points on 4 facets (PoR, EE, DS, SD). This again demonstrates the effectiveness of our concept-enhanced framework in the MID task. However, on the facets of CSR and CV, our approach is inferior to BERTweetInd, especially on CSR. We think this is likely because there are too few related samples in CSR (as shown in Table 6), and our method uses a separate classification head for each facet, resulting in even more insufficient training for CSR. Although this issue affects the results, it is only an edge case. The two LLMs still perform poorly, especially on PoR, SS and CV. By analyzing the responses generated by LLMs, we find that LLMs are more likely to ignore or generalize the definitions in prompts on these facets.
+
+<table><tr><td>Model</td><td>PoR</td><td>SS</td><td>EO</td><td>EE</td><td>EP</td><td>CSR</td><td>CV</td><td>DS</td><td>MF</td><td>SD</td><td>JO</td><td>PeR</td></tr><tr><td colspan="13">Subtask 1: Relevance Recognition</td></tr><tr><td>BERTweetInd</td><td>46.92</td><td>32.71</td><td>71.05</td><td>63.29</td><td>82.26</td><td>35.04</td><td>19.52</td><td>62.73</td><td>85.99</td><td>44.07</td><td>75.55</td><td>70.71</td></tr><tr><td>LLaMA2-13B△</td><td>3.33</td><td>9.41</td><td>31.30</td><td>20.48</td><td>47.23</td><td>5.19</td><td>4.33</td><td>30.82</td><td>56.42</td><td>26.32</td><td>56.06</td><td>61.34</td></tr><tr><td>ChatGPT△</td><td>6.48</td><td>10.45</td><td>54.27</td><td>37.33</td><td>53.04</td><td>10.27</td><td>7.96</td><td>47.02</td><td>78.87</td><td>33.52</td><td>63.92</td><td>62.79</td></tr><tr><td>Our approach</td><td>52.63†</td><td>33.33†</td><td>71.65†</td><td>67.98†</td><td>83.00</td><td>31.58</td><td>19.35</td><td>66.80†</td><td>86.59†</td><td>50.55†</td><td>76.16†</td><td>70.98†</td></tr><tr><td colspan="13">Subtask 2: Ideology Analysis</td></tr><tr><td>BERTweetInd</td><td>24.40</td><td>27.59</td><td>52.26</td><td>41.25</td><td>52.43</td><td>49.93</td><td>43.37</td><td>57.00</td><td>48.39</td><td>43.92</td><td>36.55</td><td>35.04</td></tr><tr><td>LLaMA2-13B△</td><td>37.23</td><td>44.47</td><td>52.28</td><td>36.97</td><td>45.56</td><td>44.81</td><td>29.60</td><td>41.60</td><td>29.87</td><td>39.90</td><td>33.47</td><td>26.40</td></tr><tr><td>ChatGPT△</td><td>24.44</td><td>43.91</td><td>50.79</td><td>43.73</td><td>54.06</td><td>33.33</td><td>51.59</td><td>52.92</td><td>33.03</td><td>36.03</td><td>41.95</td><td>45.87</td></tr><tr><td>Our approach</td><td>33.16†</td><td>45.18†</td><td>50.25</td><td>61.70†</td><td>59.65†</td><td>35.56</td><td>49.31†</td><td>62.35†</td><td>47.25</td><td>45.10†</td><td>37.10</td><td>41.24†</td></tr></table>
+
+For the Ideology Analysis subtask, the baseline models achieve the best or second-best result on some facets. Nevertheless, our approach ranks in the top two on 10 out of 12 facets and shows overall superior performance.
+
+# 4.5 Ablation Study
+
+We conduct ablation studies to inspect the importance of major components in our model and the results are reported in Table 1. It is clear that the removal of either one of the modules causes a drop in performance. The Micro-F1 decreases by 1.73 and 2.32 points on the two subtasks, respectively, when BICo is removed, which validates that it is important to further model the schema hierarchy and concept interactions on top of the concept encoder. BICo iteratively performs concept diffusion and aggregation on the hierarchy tree, and the updated concept representations are enriched by higher-level general semantics and lower-level concrete perspectives, which are helpful for the model to understand the deep meaning of facet and ideology labels.
+
+In Ideology Analysis, the removal of concept anchors leads to noticeable performance degradation. This suggests that relying solely on text content to identify ideology is insufficient, and injecting label semantics can guide the model to capture ideology features and distinguish among different ideologies more accurately, so as to improve the performance of MID. Moreover, the results of "w/o CL" in Relevance Recognition and "w/o CGCL" in Ideology Analysis verify the effectiveness of contrastive learning strategies in two subtasks.
+
+We also conduct ablation study for the modules of Concept Metapath Diffusion and Concept Hierar-
+
+Table 2: F1 scores of different models on each facet. † denotes the significance test over BERTweetInd at p-value<0.05. △ denotes 3-shot. Bold and underlined values are the best and second-best results in the subtask, respectively. Full names of 12 facets in the first row can be found in Appendix A.  
+
+<table><tr><td rowspan="2">Test Topics</td><td rowspan="2">Model</td><td>Relevance Recognition</td><td colspan="2">Ideology Analysis</td></tr><tr><td>Micro-F1</td><td>Micro-Acc</td><td>Micro-F1</td></tr><tr><td rowspan="4">CHR&amp;GF</td><td>BERTweetInd</td><td>59.60</td><td>70.20</td><td>52.41</td></tr><tr><td>LLaMA2-13B</td><td>28.29</td><td>56.79</td><td>44.22</td></tr><tr><td>ChatGPT</td><td>36.20</td><td>69.70</td><td>51.87</td></tr><tr><td>Our approach</td><td>61.00†</td><td>72.43†</td><td>54.76†</td></tr><tr><td rowspan="4">BLM&amp;Dm</td><td>BERTweetInd</td><td>54.69</td><td>80.64</td><td>58.89</td></tr><tr><td>LLaMA2-13B</td><td>31.90</td><td>58.93</td><td>46.27</td></tr><tr><td>ChatGPT</td><td>39.54</td><td>73.45</td><td>54.09</td></tr><tr><td>Our approach</td><td>62.88†</td><td>83.31†</td><td>61.04†</td></tr></table>
+
+Table 3: Cross-topic results of different models. CHR means Capitol Hill Riot, GF means George Floyd, BLM means Black Lives Matter, Dm means Democracy. † denotes the significance test over BERTweetInd at p-value<0.05. Bold values are the best results in the corresponding test topics.
+
+chy Aggregation in BICo. The results are presented in Appendix C.
+
+# 4.6 Cross-Topic Generalization
+
+In our approach, label concepts are incorporated to enhance the model and they are enriched by multi-granularity concepts from different levels in the hierarchical schema through BICo. Intuitively, concepts provide a general description of a label. Therefore, our model should have better generalization to new topics with the help of concept semantics. To validate this viewpoint, we test and compare the cross-topic generalization ability of different models.
+
+In the cross-topic scenario, the models are trained on some topics and then tested on the rest topics. To reduce randomness, we conduct experiments on two sets of test topics and the results
+
+![](images/efddb35871281d95e499e2dfe9aada2879beb68712ebbdfde3454bb2c72847f1.jpg)  
+(a) w/o CGCL
+
+![](images/941af308eb7ef15d66cb845b82a7e28afad70b4a72f01d4e8cf0fcf324871f97.jpg)  
+(b) w/ CL
+
+![](images/83418b8cc9b3337eb751e95fb0c38e0626bb05212f25558472ff9b081fe5dd89.jpg)  
+(c) w/ CGCL
+
+![](images/d28a65b1f785487c729fde7057b249ae2efd3296a60498cef291bb7183e1578f.jpg)  
+Figure 3: T-SNE visualization of text representations learned by different model variants in the Ideology Analysis subtask. CGCL denotes our Concept-Guided Contrastive Learning. CL denotes the Contrastive Learning without concept anchors. Red, green and blue dots represent Left, Center and Right samples, respectively.
+
+![](images/0a8c9b3fd21651b93c6b91a79b973099134fd513179e58cf7028347f602e540d.jpg)  
+Figure 4: Results of different numbers of iterations.
+
+are shown in Table 3. It can be observed that our approach consistently outperforms other models in both subtasks, which verifies that our approach can better generalize the learning ability to deal with cross-topic scenarios. LLMs lag behind other models by a significant margin. This shows that task-specific models still have advantages even in cross-topic scenarios. However, for the test topics of CHR&GF, ChatGPT performs closely to task-specific models in the Ideology Analysis subtask, indicating that ChatGPT may have practical value in specific cross-topic scenarios.
+
+# 4.7 Effect of Number of Iterations
+
+To analyze the effect of using different numbers of iterations in BICo, we conduct experiments on both subtasks and present the results in Figure 4. We can observe a clear upward and then downward trend in model performance as the number of iter
+
+tions increases. The optimal number of iterations for Relevance Recognition is 4 and for Ideology Analysis is 2. One possible reason for this trend is that, when the number of iterations is too small, the concept diffusion and aggregation are insufficient, and the concept representations do not fully perceive the semantics of different granularities in the hierarchical structure. In contrast, when the number of iterations is too large, there will be redundancy in information transfer, and the semantic features of the concept itself will be lost.
+
+# 4.8 Visualization
+
+To qualitatively examine the role of label semantics (concept anchors) in the concept-guided contrastive learning, we randomly select a facet (Diplomatic Strategy) and show the t-SNE projections of text representations from test set in Figure 3. As observed, for the case of "w/o CGCL", all samples are almost scattered without separations. There is a similar but better distribution for the model trained with CL. While for our CGCL (i.e., the full model), instances are well clustered by labels with only a slight overlap and the concept anchors are approximately cluster centers. This confirms that concept representations learned from BICo guide the model to better distinguish among different ideologies in the embedding space, which is helpful for subsequent classification.
+
+# 5 Related Work
+
+Ideology Detection This task detects the ideology of texts in a generic facet. Many studies rely on text analysis techniques and try to leverage various textual cues (Bhatia and P, 2018; Baly et al., 2019, 2020; Chen et al., 2020; Kabir and Madria,
+
+2022; Liu et al., 2022; Kim and Johnson, 2022; Devatine et al., 2023; Hong et al., 2023; Chen et al., 2023). In addition to text content, social networks (Li and Goldwasser, 2019; Stefanov et al., 2020; Xiao et al., 2020; Li and Goldwasser, 2021), external knowledge (Kulkarni et al., 2018; Zhang et al., 2022) and multimodal information (Dinkov et al., 2019; Qiu et al., 2022) are utilized to identify the ideology of online texts.
+
+Multifaceted Ideology Detection Considering that some texts may contain descriptions of different issues and reflect the author's ideology from various aspects, some recent work study ideology detection on multiple facets. Sinno et al. (2022) investigate the political ideology of news articles from three facets, social, economic and foreign. Liu et al. (2023) first propose the MID task and design the first multifaceted ideology schema which defines 5 domains and 12 facets in a hierarchical structure. They also manually annotate a high-quality MITweet dataset and build baselines for MID. We follow Liu et al. (2023) and introduce label semantics into models through encoding the hierarchical schema.
+
+# 6 Conclusion
+
+In this paper, we have proposed a concept semantics-enhanced framework for the MID task. We have also designed a novel bidirectional iterative concept flow method to capture multi-granularity concept semantics. Moreover, we have explored concept attentive matching and concept-guided contrastive learning strategies to enable the model to extract ideology features with the help of concept semantics. Experiment results have validated the superiority of our approach.
+
+# Acknowledgement
+
+This work is supported in part by Major Project of National Social Science Foundation of China: "AI and Precise International Communication" (Grant No. 22&ZD317) and National Natural Science Foundation of China (Grant No. 62172167). The computation is supported by the HPC Platform of Huazhong University of Science and Technology.
+
+# Limitations
+
+- Following Liu et al. (2023), we divide multifaceted ideology detection into two subtasks in a pipeline manner. However, this modeling approach increases the computational cost in
+
+both training and inference stages. In addition, error propagation in this pipeline mode is also a problem that cannot be ignored. We will investigate how to solve this task in an end-to-end manner in future work.
+
+- While we attempt to tune the concepts defined in the schema to better fit our approach, we are constrained by computational resources and time, so we directly adopt the concepts in the schema. Although these concepts are representative, there may be better ones that could lead to better performance.
+
+# Ethical Considerations
+
+We carry out this work and conduct the experiments in accordance with the general ethics in social science research. The proposed concept-enhanced framework could automatically detect the multifaceted ideology of given texts, which is helpful for policy-makers and social statisticians. However, the algorithm is not perfect and may make incorrect predictions. Therefore, researches should realize the potential harm from the misuse of the ideology detection system, and cannot rely solely on the system to make judgments.
+
+# References
+
+Saja Aldera, Ahmad Emam, Muhammad Al-Qurishi, Majed Alrubaian, and Abdulrahman Alothaim. 2021. Online extremism detection in textual content: a systematic literature review. IEEE Access, 9:42384-42396.  
+Ramy Baly, Giovanni Da San Martino, James Glass, and Preslav Nakov. 2020. We can detect your bias: Predicting the political ideology of news articles. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4982-4991, Online. Association for Computational Linguistics.  
+Ramy Baly, Georgi Karadzhov, Abdelrhman Saleh, James Glass, and Preslav Nakov. 2019. Multi-task ordinal regression for jointly predicting the trustworthiness and the leading political ideology of news media. 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 2109-2116, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Sumit Bhatia and Deepak P. 2018. Topic-specific sentiment analysis can help identify political ideology.
+
+In Proceedings of the 9th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 79-84, Brussels, Belgium. Association for Computational Linguistics.  
+Chen Chen, Dylan Walker, and Venkatesh Saligrama. 2023. Ideology prediction from scarce and biased supervision: Learn to disregard the "what" and focus on the "how!" In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9529-9549, Toronto, Canada. Association for Computational Linguistics.  
+Wei-Fan Chen, Khalid Al Khatib, Henning Wachsmuth, and Benno Stein. 2020. Analyzing political bias and unfairness in news articles at different levels of granularity. In Proceedings of the Fourth Workshop on Natural Language Processing and Computational Social Science, pages 149-154, Online. Association for Computational Linguistics.  
+Dorottya Demszky, Nikhil Garg, Rob Voigt, James Zou, Jesse Shapiro, Matthew Gentzkow, and Dan Jurafsky. 2019. Analyzing polarization in social media: Method and application to tweets on 21 mass shootings. 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 2970-3005, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Nicolas Devatine, Philippe Muller, and Chloé Braud. 2023. An integrated approach for political bias prediction and explanation based on discursive structure. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 11196-11211, Toronto, Canada. Association for Computational Linguistics.  
+Yoan Dinkov, Ahmed Ali, Ivan Koychev, and Preslav Nakov. 2019. Predicting the leading political ideology of youtube channels using acoustic, textual, and metadata information.  
+Ted Grover and Gloria Mark. 2019. Detecting potential warning behaviors of ideological radicalization in an alt-right subreddit. In Proceedings of the International AAAI Conference on Web and Social Media, volume 13, pages 193-204.  
+Jiwoo Hong, Yejin Cho, Jiyoung Han, Jaemin Jung, and James Thorne. 2023. Disentangling structure and style: Political bias detection in news by inducing document hierarchy. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5664-5686, Singapore. Association for Computational Linguistics.  
+Md Yasin Kabir and Sanjay Madria. 2022. A deep learning approach for ideology detection and polarization analysis using Covid-19 tweets. In International Conference on Conceptual Modeling, pages 209-223. Springer.
+
+Sandeepa Kannangara. 2018. Mining twitter for fine-grained political opinion polarity classification, ideology detection and sarcasm detection. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, pages 751-752.  
+Michelle YoungJin Kim and Kristen Marie Johnson. 2022. CLoSE: Contrastive learning of subframe embeddings for political bias classification of news media. In Proceedings of the 29th International Conference on Computational Linguistics, pages 2780-2793, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Vivek Kulkarni, Junting Ye, Steve Skiena, and William Yang Wang. 2018. Multi-view models for political ideology detection of news articles. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3518-3527, Brussels, Belgium. Association for Computational Linguistics.  
+Chang Li and Dan Goldwasser. 2019. Encoding social information with graph convolutional networks for Political perspective detection in news media. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2594-2604, Florence, Italy. Association for Computational Linguistics.  
+Chang Li and Dan Goldwasser. 2021. Using social and linguistic information to adapt pretrained representations for political perspective identification. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4569-4579, Online. Association for Computational Linguistics.  
+Songtao Liu, Ziling Luo, Minghua Xu, Lixiao Wei, Ziyao Wei, Han Yu, Wei Xiang, and Bang Wang. 2023. Ideology takes multiple looks: A high-quality dataset for multifaceted ideology detection. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4200-4213, Singapore. Association for Computational Linguistics.  
+Yujian Liu, Xinliang Frederick Zhang, David Wegsman, Nicholas Beauchamp, and Lu Wang. 2022. POLITICS: Pretraining with same-story article comparison for ideology prediction and stance detection. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 1354-1374, Seattle, United States. Association for Computational Linguistics.  
+Ilya Loshchilov and Frank Hutter. 2018. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Dat Quoc Nguyen, Thanh Vu, and Anh Tuan Nguyen. 2020. BERTweet: A pre-trained language model for English tweets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 9-14, Online. Association for Computational Linguistics.
+
+Changyuan Qiu, Winston Wu, Xinliang Frederick Zhang, and Lu Wang. 2022. Late fusion with triplet margin objective for multimodal ideology prediction and analysis. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9720-9736, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+Barea Sinno, Bernardo Oviedo, Katherine Atwell, Malihe Alikhani, and Junyi Jessy Li. 2022. Political ideology and polarization: A multi-dimensional approach. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 231-243, Seattle, United States. Association for Computational Linguistics.
+
+Peter Stefanov, Kareem Darwish, Atanas Atanasov, and Preslav Nakov. 2020. Predicting the topical stance and political leaning of media using tweets. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 527-537, Online. Association for Computational Linguistics.
+
+Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. 2018. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations.
+
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.
+
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.
+
+Xuepeng Wang, Li Zhao, Bing Liu, Tao Chen, Feng Zhang, and Di Wang. 2021. Concept-based label embedding via dynamic routing for hierarchical text classification. 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 5010-5019, Online. Association for Computational Linguistics.
+
+Haoyang Wen and Alexander Hauptmann. 2023. Zero-shot and few-shot stance detection on varied topics via conditional generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1491-1499, Toronto, Canada. Association for Computational Linguistics.
+
+Zhiping Xiao, Weiping Song, Haoyan Xu, Zhicheng Ren, and Yizhou Sun. 2020. Timme: Twitter ideology-detection via multi-task multi-relational embedding. In Proceedings of the 26th ACM SIGKDD
+
+International Conference on Knowledge Discovery & Data Mining, KDD '20, page 2258-2268, New York, NY, USA. Association for Computing Machinery.
+
+Wenqian Zhang, Shangbin Feng, Zilong Chen, Zhenyu Lei, Jundong Li, and Minnan Luo. 2022. KCD: Knowledge walks and textual cues enhanced political perspective detection in news media. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4129-4140, Seattle, United States. Association for Computational Linguistics.
+
+# A Multifaceted Ideology Schema
+
+We present the multifaceted ideology schema in Table 4. Concepts of facets and ideologies defined in the schema can be found in Liu et al. (2023). Note that the original schema does not give the concepts of "Center", so we define them based on the concepts of "Left" and "Right", as follows:
+
+# A.1 Domain 1: Politics
+
+# - Political Regime (PoR)
+
+Center: A moderate stance advocating for a mix of public and private ownership, seeking a balanced approach to property control and means of production.
+
+# - State Structure (SS)
+
+Center: A moderate stance advocating for a balanced power structure, combining elements of central authority and power distribution.
+
+# A.2 Domain 2: Economy
+
+# - Economic Orientation (EO)
+
+Center: A moderate stance advocating for combining government intervention in important economic decisions with the role of individuals, organizations, and market interactions.
+
+# - Economic Equality (EE)
+
+Center: A moderate position advocating for an economic system that balances equal treatment and access to resources with considerations for distribution outcomes among different groups.
+
+# A.3 Domain 3: Culture
+
+# - Ethical Pursuit (EP)
+
+Center: The mainstream culture should consider individual freedoms and cultural norms while promoting inclusivity dialogue on controversial issues.
+
+<table><tr><td>Domain</td><td>Facet</td><td>Left</td><td>Right</td></tr><tr><td rowspan="2">Politics</td><td>Political Regime (PoR)</td><td>Socialism</td><td>Capitalism</td></tr><tr><td>State Structure (SS)</td><td>Centralism</td><td>Federalism</td></tr><tr><td rowspan="2">Economy</td><td>Economic Orientation (EO)</td><td>Command Economy</td><td>Market Economy</td></tr><tr><td>Economic Equality (EE)</td><td>Outcome Equality</td><td>Opportunity Equality</td></tr><tr><td rowspan="3">Culture</td><td>Ethical Pursuit (EP)</td><td>Ethical Liberalism</td><td>Ethical Conservatism</td></tr><tr><td>Church-State Relations (CSR)</td><td>Secularism</td><td>Caesaropapism</td></tr><tr><td>Cultural Value (CV)</td><td>Collectivism</td><td>Individualism</td></tr><tr><td rowspan="2">Diplomacy</td><td>Diplomatic Strategy (DS)</td><td>Globalism</td><td>Isolationism</td></tr><tr><td>Military Force (MF)</td><td>Militarism</td><td>Pacifism</td></tr><tr><td rowspan="3">Society</td><td>Social Development (SD)</td><td>Revolutionism</td><td>Reformism</td></tr><tr><td>Justice Orientation (JO)</td><td>Result Justice</td><td>Procedural Justice</td></tr><tr><td>Personal Right (PeR)</td><td>Social Responsibility</td><td>Individual Right</td></tr></table>
+
+Table 4: Multifaceted ideology schema (Liu et al., 2023).
+
+# - Church-State Relations (CSR)
+
+Center: A moderate position advocating for a balanced and cooperative relationship between the church and state, respecting both religious autonomy and the principles of secular governance.
+
+# - Cultural Value (CV)
+
+Center: A moderate stance that recognizes the importance of both social collectives and individual autonomy in shaping and preserving a diverse and inclusive society.
+
+# A.4 Domain 4: Diplomacy
+
+# - Diplomatic Strategy (DS)
+
+Center: A moderate position that balances international cooperation and national interests, recognizing the value of engagement while cautiously managing political and economic entanglements with other countries.
+
+# - Military Force (MF)
+
+Center: A moderate stance that recognizes the need for armed defense and security while prioritizing non-violent resolution for conflicts.
+
+# A.5 Domain 5: Society
+
+# - Social Development (SD)
+
+Center: A moderate position that advocates combining direct action when necessary with a recognition of the value of gradual and sustainable change to achieve social goals.
+
+# - Justice Orientation (JO)
+
+Center: A moderate stance that seeks a balance between fair distribution and fair decision-making, considering both the outcomes and procedure of justice.
+
+# - Personal Right (PeR)
+
+Center: A moderate position that recognizes the importance of both fulfilling individual responsibilities and protecting individual rights in an equitable manner.
+
+# B Prompts for LLMs
+
+The prompt templates designed for LLMs in two subtasks are as follows. We fill the templates with the facet names and definitions in the multifaceted ideology schema. In few-shot experiments, we provide LLMs with a few in-context demonstrations, which are manually selected for each facet to ensure diversity. We also provide a brief analysis as chain-of-thought for each demonstration. In zero-shot experiments, the demonstrations in the prompts will be removed.
+
+# B.1 Relevance Recognition
+
+# - System prompt
+
+You will be provided with a piece of text. Determine if the text is related to {"facet)}.
+
+{facet} is defined as: {facet_def}
+
+First give your analysis briefly and then select your answer from ["Related", "Unrelated"].
+
+Here are some demonstrations: {demonstrations}
+
+- User prompt
+
+Text:"""{text}"
+
+# B.2 Ideology Analysis
+
+- System prompt
+
+You will be provided with a piece of text. Determine the orientation of the text towards {"facet"}
+
+The orientation towards {"facet"} can be divided into ["Left", "Right", "Center"]. The definitions are as follows:
+
+-Left: {left_def}  
+-Right:{right_def}  
+-Center: {center_def}
+
+First give your analysis briefly and then select your answer from ["Left", "Right", "Center"].
+
+Here are some demonstrations: {demonstrations}
+
+- User prompt
+
+Text:"""{text}"
+
+# C Additional Ablation Study
+
+As shown in Table 5, the removal of Concept Metapath Diffusion or Concept Hierarchy Aggregation causes a drop in performance. And removing both of them (w/o BICo) leads to a more significant performance degradation. The concept diffusion from root to leaf enables the high-level general semantics to propagate to lower-level nodes, while the concept aggregation from leaf to root allows the high-level nodes to perceive multifaceted concepts from lower levels. Both contribute to enriching label representations. The results further validate the effectiveness of both modules.
+
+<table><tr><td>Model</td><td>Macro-F1</td><td>Micro-F1</td><td>Macro-Acc</td><td>Micro-Acc</td></tr></table>
+
+Subtask 1: Relevance Recognition  
+
+<table><tr><td>Our Approach</td><td>59.22</td><td>72.14</td><td>-</td><td>-</td></tr><tr><td>w/o CMD</td><td>57.92</td><td>70.81</td><td>-</td><td>-</td></tr><tr><td>w/o CHA</td><td>58.73</td><td>71.03</td><td>-</td><td>-</td></tr><tr><td>w/o BICo</td><td>58.14</td><td>70.41</td><td>-</td><td>-</td></tr></table>
+
+Subtask 2: Ideology Analysis  
+
+<table><tr><td>Our Approach</td><td>47.32</td><td>70.90</td><td>66.79</td><td>78.60</td></tr><tr><td>w/o CMD</td><td>46.38</td><td>68.93</td><td>65.90</td><td>77.71</td></tr><tr><td>w/o CHA</td><td>46.13</td><td>69.48</td><td>66.75</td><td>77.85</td></tr><tr><td>w/o BICo</td><td>46.02</td><td>68.58</td><td>66.18</td><td>76.63</td></tr></table>
+
+Table 5: Results of ablation study for the modules of Concept Metapath Diffusion (CMD) and Concept Hierarchy Aggregation (CHA) in BICo. Note that "w/o BICo" is equivalent to "w/o CMD&CHA"
+
+<table><tr><td rowspan="2">Domain</td><td rowspan="2">Facet</td><td colspan="2">Relevance</td><td colspan="2">Ideology</td></tr><tr><td>#Related</td><td>#Left</td><td>#Center</td><td>#Right</td></tr><tr><td rowspan="2">Politics</td><td>PoR</td><td>112</td><td>39</td><td>14</td><td>59</td></tr><tr><td>SS</td><td>291</td><td>67</td><td>88</td><td>136</td></tr><tr><td rowspan="2">Economy</td><td>EO</td><td>759</td><td>294</td><td>297</td><td>168</td></tr><tr><td>EE</td><td>672</td><td>520</td><td>119</td><td>33</td></tr><tr><td rowspan="3">Culture</td><td>EP</td><td>2935</td><td>1976</td><td>465</td><td>494</td></tr><tr><td>CSR</td><td>68</td><td>33</td><td>17</td><td>18</td></tr><tr><td>CV</td><td>154</td><td>95</td><td>11</td><td>48</td></tr><tr><td rowspan="2">Diplomacy</td><td>DS</td><td>1572</td><td>711</td><td>421</td><td>440</td></tr><tr><td>MF</td><td>1837</td><td>132</td><td>575</td><td>1130</td></tr><tr><td rowspan="3">Society</td><td>SD</td><td>1737</td><td>1236</td><td>287</td><td>214</td></tr><tr><td>JO</td><td>3452</td><td>3058</td><td>281</td><td>113</td></tr><tr><td>PeR</td><td>3516</td><td>171</td><td>241</td><td>3104</td></tr></table>
+
+Table 6: Statistics of the MITweet dataset.

2024/Encoding Hierarchical Schema via Concept Flow for Multifaceted Ideology Detection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad4f3dcdc55b884be9fa88ea96e0786a1cae6b1c37098c97658478df4ba423b6
+size 668915

2024/End-to-End Emotion Semantic Parsing/4ffbaf63-28e0-4983-95b1-695face241d4_content_list.json

@@ -0,0 +1,1648 @@
+[
+    {
+        "type": "text",
+        "text": "End-to-End Emotion Semantic Parsing",
+        "text_level": 1,
+        "bbox": [
+            294,
+            90,
+            702,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xiaotong Jiang, Zhongqing Wang\\*, and Guodong Zhou",
+        "bbox": [
+            263,
+            145,
+            737,
+            162
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Natural Language Processing Lab, Soochow University, Suzhou, China xtjiang21@stu.suda.edu.cn, {wangzq, gdzhou}@suda.edu.cn",
+        "bbox": [
+            208,
+            162,
+            794,
+            195
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Emotion detection is the task of automatically associating one or more emotions with a text. The emotions are experienced, targeted, and caused by different semantic constituents. Therefore, it is necessary to incorporate these semantic constituents into the process of emotion detection. In this study, we propose a new task called emotion semantic parsing which aims to parse the emotion and semantic constituents into an abstract semantic tree structure. In particular, we design an end-to-end generation model to capture the relations between emotion and all the semantic constituents, and to generate them jointly. Furthermore, we employ a task decomposition strategy to capture the semantic relation among these constituents in a more cognitive and structural way. Experimental results demonstrate the importance of the proposed task, and indicate the proposed model gives superior performance compared to other models.",
+        "bbox": [
+            141,
+            288,
+            460,
+            587
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            600,
+            258,
+            615
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The goal of emotion detection is to automatically detect or categorize the emotional states of human according to some inputs. Nowadays, emotion detection can be found in a broad range of applications, including but not limited to emotional support (Tu et al., 2022; Pavarini et al., 2023), human-computer interaction (Chowdary et al., 2021; Alrowais et al., 2023) and healthcare surveillance (Dhuheir et al., 2021; Fang et al., 2023). Henceforth, emotion detection has attracted increasing attention from both research community and industry in recent years (Hu et al., 2021; Zanwar et al., 2023).",
+        "bbox": [
+            112,
+            626,
+            489,
+            833
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In literature, most studies focus on detecting emotion from the given text. For example, feature-based works (Balahur et al., 2011) and deep learning approaches, including RNNs (Majumder et al.,",
+        "bbox": [
+            112,
+            835,
+            489,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/70c77cfa0599318b2e305e2e78b8b7ecbe29fb3eb098f449c76f1f6e4ef3ac57.jpg",
+        "image_caption": [
+            "He clenched his fists as he watched the bully taunt his little brother.",
+            "Figure 1: Example of abstract semantic tree."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            546,
+            287,
+            848,
+            448
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2019), Transformers (Yu et al., 2018), and pretrained language models (Zheng et al., 2022; Zanwar et al., 2023) are used in emotion detection. Nevertheless, over the past few years, numerous studies (Xia and Ding, 2019; Kim and Klinger, 2018; Campagnano et al., 2022) have indicated that the emotions may be experienced, targeted, and/or caused by different semantic constituents, and can be linked to form abstract semantic structures.",
+        "bbox": [
+            507,
+            511,
+            884,
+            655
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, most existing works only take \"cause\" into consideration, neglecting the intricate relationships between emotion constituents, such as the strong relationship between trigger and emotion, as well as the similarity of sentiment and emotion labels. This limitation makes acquiring a nuanced understanding of the interplay between emotion and semantic constituents challenging, thereby hindering the model's capability to thoroughly comprehend emotional expression. By addressing these complexities, we believe it can serve as a foundational component for improving results in downstream real-world tasks by providing a more comprehensive understanding of emotional dynamics and their influence.",
+        "bbox": [
+            507,
+            659,
+            884,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To tackle the above problem, we propose a new",
+        "bbox": [
+            526,
+            904,
+            882,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding author",
+        "bbox": [
+            141,
+            906,
+            294,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "37",
+        "bbox": [
+            489,
+            927,
+            510,
+            939
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2024, pages 37-47 August 11-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            243,
+            945,
+            752,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "task called emotion semantic parsing which aims to parse the emotion and semantic constituents into an abstract semantic tree structure. In this work, we consider five emotion-related semantic constituents into this task, including emotion, sentiment, trigger, holder, and target. As shown in Figure 1, emotion and sentiment refer to the specific emotion and sentiment labels of the sentence; trigger refers to a trigger word or expression that describes an emotion; and the constituents of holder and target correspond to the person or entities towards whom or from whom the emotion is directed, respectively.",
+        "bbox": [
+            112,
+            84,
+            492,
+            277
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Moreover, we use the form of the abstract semantic tree structure to model the relationship between these constituents, we use the Trigger node to represent the whole abstract semantic tree, and we use the virtual Feelings and Entities nodes to compass other constituent nodes. The tree structure can be considered as a semantic representation in order to better represent the structure of the emotion-related constituents. As shown in Figure 1, the tree structure models a sentence using a rooted directed acyclic graph, highlighting its main elements (e.g. emotion, triggers) and semantic relations. It can thus potentially reveal a more comprehensive and complete semantic structure for extracting these constituents. Furthermore, since the emotion and semantic constituents are correlated in the form of the abstract semantic tree structure, we design an end-to-end generative model to predict them jointly. Specifically, our proposed model begins by decomposing the original task into several sub-tasks, as shown in Figure 2. Subsequently, the proposed model utilizes template-based prompt learning to extract each semantic constituent and semantic constituents pair separately for each sub-tasks. Finally, the proposed model leverages the acquired knowledge from sub-tasks to generate all semantic constituents in the main task.",
+        "bbox": [
+            112,
+            278,
+            490,
+            712
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conducted extensive experiments on the unified benchmark. Experimental results demonstrate the importance of the proposed task, and indicate the proposed model gives superior performance compared to other models. Further analysis shows the effect of the proposed abstract semantic tree structure and decomposition strategies.",
+        "bbox": [
+            112,
+            714,
+            490,
+            829
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Works",
+        "text_level": 1,
+        "bbox": [
+            112,
+            844,
+            280,
+            859
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In the domain of emotion analysis, previous studies have mainly focused on text-based emotion detection. Early feature-based work often relied on",
+        "bbox": [
+            112,
+            873,
+            490,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "traditional machine learning algorithms and manually crafted word-level emotion lexicons (Balahure et al., 2011). As deep neural networks progressed, emotion detection methods began utilizing various networks, such as RNNs (Majumder et al., 2019) and Transformers (Vaswani et al., 2017; Yu et al., 2018), to learn the emotion-related representation for detection. Recently, pre-trained language models (Devlin et al., 2019; Raffel et al., 2020) have dominated this area with its flexibility and rich internal knowledge (Alvarez-Gonzalez et al., 2021; Zheng et al., 2022).",
+        "bbox": [
+            507,
+            84,
+            885,
+            275
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The emotions are not expressed individually, and can be experienced, targeted, and caused by different semantic constituents (Xia and Ding, 2019; Xiao et al., 2023; Campagnano et al., 2022). Most recent studies focus on emotion cause extraction task (Xia and Ding, 2019; Wang et al., 2023; Gu et al., 2023), with the goal of extracting the potential pairs of emotions and corresponding causes in the document and conversation scenario. Many end-to-end approaches have been proposed. These works can be broadly categorized into two main groups: some focus on multi-task learning (Ding et al., 2020a,b; Chen et al., 2022), while others transform the original task into a unified sequence labeling problem (Yuan et al., 2020; Chen et al., 2020). Recently, Xiao et al. (2023) developed a rule-based system to extract emotion causes on social media from constituency parsing trees. Wang et al. (2023) proposed emotion cause triplet extraction to extract emotion-cause-category triplets for fine-grained emotion cause analysis.",
+        "bbox": [
+            507,
+            278,
+            885,
+            617
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Different from prior studies, we introduce a novel task, termed emotion semantic parsing. This task seeks to comprehensively examine the impact of various semantic constituents on emotional expressions. To achieve this, we develop an end-to-end generative model specifically designed to capture the intricate relationships between emotion and a wide range of semantic constituents.",
+        "bbox": [
+            507,
+            618,
+            885,
+            747
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            507,
+            763,
+            662,
+            778
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 2, we introduce a novel task termed emotion semantic parsing. This task focuses on the generation of an abstract semantic tree structure that encapsulates both emotion and semantic constituents. To achieve this, we develop an end-to-end generative model specifically tailored for the generation of the abstract semantic tree structure. Our proposed model adopts a",
+        "bbox": [
+            507,
+            791,
+            885,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "38",
+        "bbox": [
+            490,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "decomposition-based strategy, breaking down the complex task into several more manageable subtasks. Specifically, our proposed model begins by decomposing the original task into several subtasks. Finally, the proposed model leverages the acquired knowledge from sub-tasks to generate all semantic constituents in the main task.",
+        "bbox": [
+            112,
+            84,
+            487,
+            195
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce two important issues of this task: abstract semantic tree construction and the abstract semantic tree parsing model. The decomposition-based strategy for abstract semantic tree parsing will be discussed in the next section.",
+        "bbox": [
+            112,
+            197,
+            489,
+            292
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Abstract Semantic Tree Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            304,
+            452,
+            318
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Inspired by the widespread impact of tree structures in parsing tasks (Li et al., 2023; Groschwitz et al., 2023), we represent the relationships among multiple constituents in a tree structure. This approach allows us to capture the complex dynamics inherent in emotional expression. We first introduce the constituents in the abstract semantic tree, then we describe the construction process of the tree.",
+        "bbox": [
+            112,
+            324,
+            489,
+            467
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Definitions of Semantic Constituents",
+        "text_level": 1,
+        "bbox": [
+            112,
+            476,
+            405,
+            491
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "There are five types of semantic constituents, each serving as a node in the abstract semantic tree. Their definitions are provided below:",
+        "bbox": [
+            112,
+            495,
+            489,
+            544
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Emotion is the specific emotion label represented from the original text. In this study, we use Plutchik (2001)'s taxonomy of emotion.",
+            "- Sentiment is the specific sentiment label represented from the original text, including \"positive\", \"negative\", and \"neutral\".",
+            "- Trigger refers to a trigger word or expression that describes an emotion, even if it is implicit.",
+            "- Holder refers to a person or entity that feels or experiences the emotion identified by the trigger.",
+            "- Target refers to a person or entity towards whom/which the emotion identified by the trigger is directed."
+        ],
+        "bbox": [
+            136,
+            552,
+            489,
+            813
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Process of Tree Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            822,
+            344,
+            835
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 1, the abstract semantic tree represents a sentence through a rooted directed acyclic graph utilizing the aforementioned semantic constituents. The construction process is outlined below:",
+        "bbox": [
+            112,
+            841,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We take the \"Trigger\" node as the root node, since it is the most important constituent in the emotion parsing.",
+            "- The virtual nodes \"Feelings\" and \"Entities\" are linked to Trigger node.",
+            "- The \"Sentiment\" and \"Emotion\" are linked to \"Feelings\" as terminal nodes, while the \"Holder\" and \"Target\" are linked to \"Entities\" node."
+        ],
+        "bbox": [
+            531,
+            84,
+            884,
+            252
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As Figure 1 shows, the trigger \"clenched his fists\" implies the author's emotion. The holder and target are \"he\" and \"the bully taunt his little brother\", respectively. By leveraging these emotion-related semantic constituents, the model can accurately predict the \"anger\" emotion and \"negative\" polarity of the sentence.",
+        "bbox": [
+            507,
+            267,
+            884,
+            378
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Basic Semantic Tree Parsing",
+        "text_level": 1,
+        "bbox": [
+            507,
+            392,
+            779,
+            407
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "After we construct the abstract semantic tree, we then employ a generative model to parse the emotion and all the semantic constituents into a linearized tree structure. As shown in Figure 2, given an input text, the output can be represented as a linearized semantic tree.",
+        "bbox": [
+            507,
+            414,
+            882,
+            508
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We adopt the pre-trained language model T5 (Chung et al., 2022) as our transformer-based encoder-decoder architecture. Therefore, by formulating parsing tasks as a text generation problem, we can tackle them in a unified sequence-to-sequence framework without a task-specific model design.",
+        "bbox": [
+            507,
+            511,
+            882,
+            621
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The main challenge of emotion semantic parsing is the complex structure inherently in the emotional expression. Specifically, the complex structure will bring out two challenges: First, the large number of semantic constituents raises the complexity to the parsing task. Consequently, the precise extraction of every constituent becomes challenging yet indispensable. Second, effectively modeling these internal relationships also presents a significant challenge in the task.",
+        "bbox": [
+            507,
+            624,
+            882,
+            784
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Decomposition-based Semantic Tree Parsing Model",
+        "text_level": 1,
+        "bbox": [
+            507,
+            797,
+            855,
+            831
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this study, we propose a decomposed-based generative model to tackle the above challenges. The motivation is to alleviate the complexity of the original task by decomposing it into several sub-tasks. As shown in Figure 2, our proposed model can be",
+        "bbox": [
+            507,
+            841,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "39",
+        "bbox": [
+            490,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/c8db31deb594d79748377026b359f71ac241c28c0a504b872213340c37690e9a.jpg",
+        "image_caption": [
+            "Figure 2: Overview of the proposed model."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            176,
+            80,
+            823,
+            329
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/23ffac2dbeaf8a802478e94064ac3283fe2a48403710f281026c73d1148deb2e.jpg",
+        "image_caption": [
+            "Figure 3: Examples of task decomposition strategies. Due to the space limit, prompt texts in this figure are shortened. Original prompt texts are presented in Section 4.1."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            136,
+            378,
+            468,
+            629
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "separated into two stages: decomposition and recomposition. Our model firstly breaks down the original abstract semantic parsing task into several sub-tasks, and then designs individual instruction-based templates for each task. All sub-tasks are fine-tuned by a generative backbone at the same time. Based on the model parameters from the last re-composition stage, our model further finetunes its parameters again with the original task, which involves jointly parsing all semantic constituents.",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Task Decomposition Strategy",
+        "text_level": 1,
+        "bbox": [
+            507,
+            376,
+            784,
+            393
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To implement a generalized framework that is applicable to all tasks, we design templates for each individual task. The input template for each task comprises the task-specific instruction, and the original text to be parsed. The output template for each task is the sub-tree containing the constituents to be parsed. The following templates use the example in Figure 2 as the Original Text.",
+        "bbox": [
+            507,
+            398,
+            884,
+            526
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Mono Parsing includes trigger, emotion, sentiment, holder, and target parsing, individually. As shown in Figure 3(a), these sub-tasks aim to capture each specific constituent:",
+        "bbox": [
+            507,
+            527,
+            882,
+            590
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "//taking emotion parsing task as an example",
+        "bbox": [
+            544,
+            600,
+            847,
+            630
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "INPUT Detect the emotion label in the emotional expression: + Original Text",
+        "bbox": [
+            544,
+            634,
+            843,
+            667
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Output [Emotion (Anger)]",
+        "bbox": [
+            544,
+            671,
+            749,
+            688
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Trigger-related Pair Parsing includes trigger-emotion, trigger-sentiment, trigger-holder, trigger-target, trigger-feelings, and trigger-entities. As shown in Figure 3(b), these sub-tasks aim to capture the relationship between the trigger and each specific constituent:",
+        "bbox": [
+            507,
+            695,
+            884,
+            791
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "//taking trigger-holder parsing task as an example",
+        "bbox": [
+            544,
+            799,
+            843,
+            831
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "INPUT Extract both the trigger and holder in the emotional expression: + Original Text",
+        "bbox": [
+            544,
+            835,
+            845,
+            882
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Output [Trigger (clench his fists), [Entities, [Holder (he)]]]",
+        "bbox": [
+            544,
+            889,
+            847,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "40",
+        "bbox": [
+            489,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Entities Pair Parsing aims to capture the relationship between holder and target, and can gain insights into the interpersonal dynamics and relational aspects that influence emotional experiences. The example is shown in Figure 3(c), with the template of prompt provided below:",
+        "bbox": [
+            112,
+            84,
+            489,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "INPUT Extract both holder and target in the emotional expression: + Original Text",
+        "bbox": [
+            149,
+            193,
+            450,
+            241
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Output [Entities, [Holder (he), Target (the bully taunt his little brother) ]]",
+        "bbox": [
+            149,
+            249,
+            452,
+            281
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Feelings Pair Parsing aims to capture the relationship between sentiment and emotion. The example is shown in Figure 3(d), with the template of prompt provided below:",
+        "bbox": [
+            112,
+            294,
+            489,
+            359
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "INPUT Extract both the emotion and sentiment labels in the emotional expression: + Original Text",
+        "bbox": [
+            147,
+            373,
+            453,
+            420
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Output [Feelings, [Emotion (Anger), Sentiment (Negative) ]]",
+        "bbox": [
+            149,
+            428,
+            453,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Task Re-composition Strategy",
+        "text_level": 1,
+        "bbox": [
+            112,
+            473,
+            396,
+            489
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After gaining knowledge from sub-tasks, the model has learned the pattern of the template and the relationships from different emotion-related semantic constituents. Thus it can boost the final end-to-end emotion semantic parsing task, which is to predict the emotion label and outputs all emotion-related semantic constituents in the form of an abstract semantic tree. The example is shown in Figure 2, and the template of prompt is the below:",
+        "bbox": [
+            112,
+            495,
+            489,
+            640
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "INPUT Extract emotion trigger, holder, target, and emotion and sentiment labels in the emotional expression: + Original Text",
+        "bbox": [
+            149,
+            653,
+            453,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Output [Trigger (clenched his fists), Feelings, [Emotion (Anger), Sentiment (Negative)]], [Entities, [Holder (he), Target (the bully taunt his little brother) ]]",
+        "bbox": [
+            149,
+            725,
+            453,
+            789
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Objective Functions and Training",
+        "text_level": 1,
+        "bbox": [
+            112,
+            802,
+            426,
+            819
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we show the objective functions and training process of the proposed model.",
+        "bbox": [
+            112,
+            824,
+            485,
+            856
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The goal is to maximize the output linearized tree  $X_{T}$  probability given the sentence  $X$ . Therefore, we optimize the negative log-likelihood loss function:",
+        "bbox": [
+            112,
+            857,
+            489,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/f16ba66ccbc268e24499e77421b9733a2dbbdeb514daa2977d09f345196476a6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Amount</td></tr><tr><td>#Sample</td><td>7,234</td></tr><tr><td>Avg. Sent Len.</td><td>26.3</td></tr><tr><td>#Emotion Labels</td><td>9</td></tr><tr><td>#Sentiment Label</td><td>3</td></tr><tr><td>#Trigger</td><td>7,111</td></tr><tr><td>#Holder</td><td>5,633</td></tr><tr><td>#Target</td><td>6,127</td></tr></table>",
+        "bbox": [
+            576,
+            80,
+            818,
+            218
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Distribution of dataset.",
+        "bbox": [
+            584,
+            227,
+            803,
+            240
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = - \\frac {1}{| \\tau |} \\sum_ {(X, X _ {T}) \\in \\tau} \\log p (X _ {T} | X; \\theta) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            561,
+            288,
+            882,
+            329
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta$  is the model parameters, and  $(X,X_{T})$  is a (sentence, tree) pair in training set  $\\tau$ , then",
+        "bbox": [
+            507,
+            338,
+            882,
+            370
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\log p (X _ {T} | X; \\theta) \\\\ = \\sum_ {i = 1} ^ {n} \\log p \\left(x _ {T} ^ {i} \\mid x _ {T} ^ {1}, x _ {T} ^ {2}, \\dots x _ {T} ^ {i - 1}, X; \\theta\\right) \\tag {2} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            552,
+            394,
+            880,
+            454
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $p(x_{T}^{i}|x_{T}^{1},x_{T}^{2},\\dots x_{T}^{i - 1},X;\\theta)$  is calculated by the decoder.",
+        "bbox": [
+            507,
+            466,
+            882,
+            498
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Experimentation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            511,
+            690,
+            527
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the datasets used for evaluation and the baseline methods employed for comparison. We then report the experimental results conducted from different perspectives, and analyze the effectiveness of the proposed model with different factors.",
+        "bbox": [
+            507,
+            536,
+            882,
+            631
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            643,
+            618,
+            658
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the effectiveness of our proposed method on a benchmark proposed in Campagnano et al. (2022), which combines several datasets. We specifically choose three datasets (Mohammad et al., 2015; Kim and Klinger, 2018; Bostan et al., 2020) from this benchmark because they are the only ones that encompass annotations for all constituents. We enhance the quality of the original dataset by deleting duplicate or irrelevant examples (e.g., those with all trigger, holder, and target are \"null\"), while ensuring the trigger is not \"null\". The distribution of dataset is shown in Table 1. We choose  $90\\%$  of the data as the training set, and the remaining  $10\\%$  of the data as the testing set. Additionally,  $10\\%$  of the training data is randomly set aside for validation during the training process.",
+        "bbox": [
+            505,
+            664,
+            884,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "41",
+        "bbox": [
+            489,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/11dc1887948cc8884e59fb7da2825f758fe8bcde3c77d265780a80295a951870.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>Emotion</td><td>Sentiment</td><td>Trigger</td><td>Holder</td><td>Target</td><td>Micro F1.</td></tr><tr><td>BERT-CRF</td><td>40.3</td><td>58.6</td><td>19.0</td><td>23.3</td><td>24.0</td><td>33.1</td></tr><tr><td>BERT-MRC</td><td>42.6</td><td>60.3</td><td>15.2</td><td>22.4</td><td>19.3</td><td>30.8</td></tr><tr><td>TaskFusion</td><td>45.7</td><td>62.3</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>USSA</td><td>41.3</td><td>-</td><td>17.3</td><td>32.6</td><td>14.3</td><td>-</td></tr><tr><td>T5</td><td>40.2</td><td>59.4</td><td>21.0</td><td>47.2</td><td>32.1</td><td>38.2</td></tr><tr><td>ChatGPT</td><td>22.3</td><td>30.7</td><td>10.4</td><td>21.3</td><td>9.6</td><td>18.9</td></tr><tr><td>LLaMA</td><td>42.4</td><td>57.6</td><td>22.4</td><td>43.5</td><td>31.3</td><td>39.4</td></tr><tr><td>Ours</td><td>46.3</td><td>63.6</td><td>30.8</td><td>55.0</td><td>37.4</td><td>46.7</td></tr></table>",
+        "bbox": [
+            183,
+            80,
+            815,
+            241
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Comparison with baselines.",
+        "bbox": [
+            371,
+            249,
+            623,
+            265
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our proposed method utilizes the generative backbone Flan-T5 (Chung et al., 2022) from HuggingFace's hub. All experiments are conducted on a single NVIDIA Tesla V100 GPU with 32GB memory. The model is trained for 10 epochs in each of the two stages, using the AdamW (Loshchilov and Hutter) optimizer, with a maximum sequence length of 256, a batch size of 16, and a learning rate of 3e-4. And the average results over 5 runs with different random seeds are reported. We adopt F1-score as the metric for all experiments, the micro F1 is employed to evaluate the overall performance.",
+        "bbox": [
+            112,
+            288,
+            489,
+            483
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            497,
+            265,
+            511
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To better evaluate the effectiveness of our proposed model, we compare its performance against several strong baseline models, which are detailed below,",
+        "bbox": [
+            112,
+            520,
+            489,
+            569
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- BERT-CRF (Campagnano et al., 2022) employs bidirectional LSTM layers over BERT representation, then uses CRF to generate token-level labels to determine the span boundary of each semantic constituent.",
+            "- BERT-MRC (Devlin et al., 2019) is used to predict the start and end indexes of all required semantic constituents.",
+            "- TaskFusion (Zanwar et al., 2023) is a multi-task framework with a pre-trained language model for emotion-related tasks. We adopt it for the proposed emotion semantic parsing task.",
+            "- USSA (Zhai et al., 2023) is a joint learning framework, which transfers aspect-based sentiment analysis into a 2D table-filling scheme. We adopt it for the proposed emotion semantic parsing task."
+        ],
+        "bbox": [
+            136,
+            585,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- T5 (Chung et al., 2022) is a pre-trained language model which naturally suits the sequence-to-sequence task. It also can be considered as a joint learning framework, which generates all the semantic constituents jointly.",
+            "- ChatGPT (Ouyang et al., 2022) uses dialog-based instructional data in the supervised and RL-based meta-training stages. We use ChatGPT API to generate all required semantic constituents.",
+            "- LLaMA (Touvron et al., 2023) is a series of large language models trained on trillions of tokens. We use Alpaca-LoRA² to fine-tune LLaMA-7B to generate required semantic constituents."
+        ],
+        "bbox": [
+            531,
+            288,
+            885,
+            590
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As demonstrated in Table 2, several key observations can be made: 1) Aligning with the current trend favoring generative frameworks, models that adopt a generative paradigm (i.e., T5, LLaMA, and our proposed model) generally outpace models based on a classification paradigm (i.e., BERT-CRF, BERT-MRC, Task Fusion, USSA) in terms of performance. 2) A standout observation is the substantial performance disparity between ChatGPT and all other methodologies. This discrepancy might be attributed to our limited usage of ChatGPT's API, without fine-tuning its internal parameters using insights from our training dataset.",
+        "bbox": [
+            507,
+            626,
+            885,
+            835
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In contrast, our proposed model consistently outperforms prior studies  $(p < 0.05)$  in all settings. This underscores the effectiveness of our model, employing task decomposition strategies and the abstract semantic tree structure.",
+        "bbox": [
+            507,
+            841,
+            885,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "42",
+        "bbox": [
+            489,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/f1305e83c2f21d4ba51501bf115fedd5ad29d5eb119ba222d6355bbaf6b65239.jpg",
+        "image_caption": [
+            "Figure 4: Effect of tree structures."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            247,
+            82,
+            742,
+            200
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/53c852f7c76c5b9b63dbd6a9d1e8152b75edf4f27c61b8a9c42ea2c813e45a35.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Tuples</td><td>e</td><td>s</td><td>t</td><td>h</td><td>a</td></tr><tr><td>(e)</td><td>42.5</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>(t)</td><td>-</td><td>-</td><td>28.4</td><td>-</td><td>-</td></tr><tr><td>(e,s)</td><td>43.5</td><td>65.1</td><td>-</td><td>-</td><td>-</td></tr><tr><td>(e,t)</td><td>45.2</td><td>-</td><td>29.1</td><td>-</td><td>-</td></tr><tr><td>(e,s,t)</td><td>45.7</td><td>67.5</td><td>30.7</td><td>-</td><td>-</td></tr><tr><td>(t,h)</td><td>-</td><td>-</td><td>29.3</td><td>43.1</td><td>-</td></tr><tr><td>(t,a)</td><td>-</td><td>-</td><td>29.1</td><td>-</td><td>32.4</td></tr><tr><td>(t,h,a)</td><td>-</td><td>-</td><td>29.3</td><td>50.0</td><td>35.5</td></tr><tr><td>(e,t,h,a)</td><td>45.9</td><td>-</td><td>29.4</td><td>52.4</td><td>35.3</td></tr><tr><td>Ours</td><td>46.3</td><td>63.6</td><td>30.8</td><td>55.0</td><td>37.4</td></tr></table>",
+        "bbox": [
+            139,
+            253,
+            460,
+            394
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/7ff76400426a81adfe836d30f18aed045fac2e7be217037dbbc6d0a05ade2740.jpg",
+        "table_caption": [
+            "Table 3: Correlations among semantic constituents."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>Micro F1.</td></tr><tr><td>Ours</td><td>46.7</td></tr><tr><td>- Mono</td><td>39.8</td></tr><tr><td>- TriggerPair</td><td>40.3</td></tr><tr><td>- EntitiesPair</td><td>41.8</td></tr><tr><td>- FeelingsPair</td><td>44.3</td></tr></table>",
+        "bbox": [
+            179,
+            432,
+            423,
+            539
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Impact of decomposition strategies.",
+        "bbox": [
+            147,
+            548,
+            450,
+            562
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Influence of Semantic Constituents",
+        "text_level": 1,
+        "bbox": [
+            112,
+            589,
+            433,
+            602
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We then analyze the influence of the semantic constituents and the correlations among them. As shown in Table 3,  $e$  denotes the emotion label,  $s$  is the sentiment polarity,  $t, h$  and  $a$  are trigger, holder, and target, separately. In addition, the rows in the table mean different combinations of emotion-related semantic constituents. For example,  $(e, s)$  means that we used the proposed model to generate the emotion label and the sentiment polarity jointly.",
+        "bbox": [
+            112,
+            609,
+            489,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "From the table, it is evident that there is a clear correlation between emotion and sentiment polarity, as the combined performance of  $(e,s)$  surpasses that of generating emotion  $e$  alone. This underscores the importance of considering sentiment polarity when analyzing and interpreting emotions. Additionally, our findings reveal that the majority of semantic constituents contribute significantly to",
+        "bbox": [
+            112,
+            756,
+            489,
+            883
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "emotion detection, especially trigger  $t$ . For example, the combined performance of  $(e, t)$  surpasses that of generating emotion  $e$  alone by a large margin. This underscores the importance of these semantic elements in understanding and contextualizing emotions within a given text.",
+        "bbox": [
+            507,
+            256,
+            884,
+            351
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, our results highlight a correlation between the holder  $h$  and target  $a$  with the trigger  $t$ , as they often refer to the person or entity experiencing or feeling the emotion triggered by a specific event or situation. This emphasizes the interconnected nature of these semantic constituents and their collective role in shaping emotional expressions.",
+        "bbox": [
+            507,
+            353,
+            884,
+            480
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Notably, we observe that incorporating all semantic constituents leads to the highest performance, highlighting the effectiveness of our proposed model with semantic tree parsing. This underscores the utility and relevance of these semantic constituents in enhancing emotion detection accuracy.",
+        "bbox": [
+            507,
+            483,
+            885,
+            595
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 Effect of Tree Structures",
+        "text_level": 1,
+        "bbox": [
+            507,
+            608,
+            749,
+            621
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To assess the impact of the abstract semantic tree structure, we evaluated our model's performance using four distinct tree structures in Figure 4.",
+        "bbox": [
+            507,
+            630,
+            880,
+            678
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In our initial assessment, we removed the entire tree structure, treating the semantic constituents as a linear sequence (a). This modification resulted in a substantial decrease in performance, from  $46.7\\%$  to  $40.7\\%$ , emphasizing the importance of the tree structure in capturing the intricate relationships among semantic constituents.",
+        "bbox": [
+            507,
+            678,
+            880,
+            790
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Subsequently, we constructed two simplified tree structures, with either Trigger (b) or Emotion (c) serving as the root node. Interestingly, both of these structures yielded comparable performance results, suggesting that both of them are pivotal components in the emotion parsing task, and their roles are equally significant. Furthermore, we introduced two virtual nodes, labeled as Feelings and",
+        "bbox": [
+            507,
+            793,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "1https://api.openai.com/v1/chat/completions",
+        "bbox": [
+            134,
+            892,
+            462,
+            906
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://github.com/tloen/alpaca-lora",
+        "bbox": [
+            136,
+            906,
+            418,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "43",
+        "bbox": [
+            489,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/85aab6252d1d830e27f052a1140b0f66e8a7a77c2b570ab01bff61f9a7d2390f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Input</td><td>Model</td><td>Output (Linearized Abstract Semantic Tree)</td></tr><tr><td rowspan=\"2\">Jean Valjean was sincere. This sincerity from the very grief that it caused him, rendered inquiries useless, and conferred authority on all that that man had said. Here, for Marius, there was a strange reversal of situations.</td><td>T5</td><td>[Trigger, sincerity, [Feelings, [Emotion, trust], [Sentiment, positive]], [Entity, [Holder, him], [Target, null]]]</td></tr><tr><td>Ours</td><td>[Trigger, grief, [Feelings, [Emotion, sadness], [Sentiment, negative]], [Entity, [Holder, him], [Target, null]]]</td></tr><tr><td rowspan=\"2\">&quot;Do you think me so very proud?&quot; he suddenly asked. &quot;I think you very strange.&quot;</td><td>T5</td><td>[Trigger, so very proud, [Feelings, [Emotion, joy], [Sentiment, positive]], [Entity, [Holder, you], [Target, me]]]</td></tr><tr><td>Ours</td><td>[Trigger, suddenly, [Feelings, [Emotion, surprise], [Sentiment, other]], [Entity, [Holder, he], [Target, null]]]</td></tr><tr><td rowspan=\"2\">Last year I was wild to marry you, and you wouldn&#x27;t look at me.</td><td>T5</td><td>[Trigger, wouldn&#x27;t look at me, [Feelings, [Emotion, anger], [Sentiment, negative]], [Entity, [Holder, I], [Target, you]]]</td></tr><tr><td>Ours</td><td>[Trigger, wouldn&#x27;t look at me, [Feelings, [Emotion, disgust], [Sentiment, negative]], [Entity, [Holder, you], [Target, I]]]</td></tr></table>",
+        "bbox": [
+            161,
+            82,
+            836,
+            299
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 5: Examples of case study. The text in green means an exact match, while text in red indicates inaccuracies. \"Null\" means the extracted constituent is missing.",
+        "bbox": [
+            110,
+            309,
+            882,
+            338
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Entities into the tree structure (d). The inclusion of these nodes led to a noticeable improvement in performance, highlighting the efficacy of our proposed tree structures and reinforcing the importance of considering semantic constituents within a structured framework for accurate emotion parsing.",
+        "bbox": [
+            112,
+            363,
+            487,
+            461
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.5 Impact of Decomposition Strategies",
+        "text_level": 1,
+        "bbox": [
+            112,
+            474,
+            440,
+            491
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We further investigated the influence of various decomposition strategies, and the results are presented in Table 4, where Mono denotes Mono Parsing, TriggerPair denotes Trigger-related Pair Parsing, EntitiesPair denotes Entities Pair Parsing, and FeelingsPair denotes Feelings Pair Parsing. All of these can be found in Section 4.1.",
+        "bbox": [
+            112,
+            497,
+            489,
+            607
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our analysis reveals that among all the strategies, mono parsing exerts the most substantial impact on model performance, aligning with our initial expectations. Mono parsing can be regarded as the foundational and indivisible task that serves as the basis for other higher-level tasks. Moreover, our findings indicate that all the decomposition strategies employed have a positive influence on the task, emphasizing the importance of breaking down the original task into smaller sub-tasks and the efficacy of reintegrating these sub-tasks to address the main task.",
+        "bbox": [
+            112,
+            611,
+            489,
+            802
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.6 Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            816,
+            248,
+            833
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To highlight our model's significance, we've chosen several challenging and illustrative examples shown in Figure 5.",
+        "bbox": [
+            112,
+            839,
+            487,
+            887
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "These challenging sentences share the same feature: they express more than one emotion expres",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "sion, each associated with its own trigger, holder, and target. However, in each case, there is only one overall emotion. The presence of secondary emotions makes it difficult to fully grasp the overall emotional tone of the sentence.",
+        "bbox": [
+            507,
+            363,
+            882,
+            443
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In the first two examples, the mis-detection of emotion expressions by T5 underscores the intricate and nuanced nature of the task, as well as the significance of each element in constructing a precise semantic representation. However, even when the correct emotion expression is identified, modeling the complex relationship between emotion and other semantic components remains challenging. As illustrated in the final example, although T5 correctly identifies the expression and its trigger, its inability to accurately detect the holder and target leads to an incorrect prediction of emotion. This highlights the interconnected nature of these elements and underscores their collective importance in achieving accurate emotion detection.",
+        "bbox": [
+            507,
+            445,
+            884,
+            684
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In contrast, our proposed model, by leveraging task decomposition, demonstrates enhanced efficacy in navigating complexities present in difficult scenarios such as negotiation, emotion shifts, and multiple possible emotion expressions. This not only facilitates accurate detection of individual constituents but also enables a more nuanced understanding of their interrelationships, leading to improved overall performance.",
+        "bbox": [
+            507,
+            687,
+            882,
+            832
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            846,
+            640,
+            860
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this study, we introduce a novel task called emotion semantic parsing, which focuses on analyzing emotions and semantic elements to construct an ab",
+        "bbox": [
+            507,
+            873,
+            882,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "44",
+        "bbox": [
+            489,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tract semantic tree structure. To achieve this, we develop an end-to-end generation model capable of capturing intricate relationships between emotions and various semantic constituents, generating them collaboratively. Furthermore, our proposed model employs template-based prompt learning, allowing for the separate extraction of each semantic constituent and pairs of semantic constituents for individual sub-tasks. Experimental results not only underscore the significance of the introduced task but also showcase the superior performance of our proposed model.",
+        "bbox": [
+            112,
+            84,
+            492,
+            280
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            290,
+            220,
+            306
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our work focuses on employing a unified framework for emotion semantic parsing. However, one of the limitations is the computational complexity of the proposed model, whose training speed is slower than the classification-based models. Secondly, more tasks could be further explored, including nested emotion parsing, cross-domain and cross-lingual emotion semantic parsing. Finally, a large language model (LLM) can achieve superior performance in nearly all downstream tasks including our task, attributed to their ability to introduce rich knowledge learned during the pretraining stage for our proposed task. For open-source large language models like LLaMA (Touvron et al., 2023), limitations include high computational costs, high-quality data costs, as well as the need to address LLM-related issues such as safety, hallucinations, and ethical concerns, which remain to be resolved in future work.",
+        "bbox": [
+            112,
+            316,
+            492,
+            623
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            634,
+            285,
+            651
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We would like to thank Shufan Fu for her invaluable advice and assistance with the experiments during this work. We also gratefully thank the anonymous reviewers for their insightful and valuable comments. This work is supported by the National Natural Science Foundation of China (No. 62076175, No.61976146), and Jiangsu Innovation Doctor Plan.",
+        "bbox": [
+            112,
+            661,
+            489,
+            789
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            818,
+            213,
+            832
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "ref_text",
+        "text": "Fadwa Alrowais, Noha Negm, Majdi Khalid, Nabil Almalki, Radwa Marzouk, Abdullah Mohamed, Mesfer Al Duhayyim, and Amani a Alneil. 2023. Modified earthworm optimization with deep learning assisted emotion recognition for human computer interface. IEEE Access, 11:35089-35096.",
+        "bbox": [
+            115,
+            841,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Nurudin Alvarez-Gonzalez, Andreas Kaltenbrunner, and Vicenç Gómez. 2021. Uncovering the limits of text-based emotion detection. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2560-2583, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Alexandra Balahur, Jesús M. Hermida, Andres Montoyo, and Rafael Muñoz. 2011. Emotinet: A knowledge base for emotion detection in text built on the appraisal theories. In *Natural Language Processing and Information Systems*, pages 27-39, Berlin, Heidelberg. Springer Berlin Heidelberg.",
+            "Laura Ana Maria Bostan, Evgeny Kim, and Roman Klinger. 2020. GoodNewsEveryone: A corpus of news headlines annotated with emotions, semantic roles, and reader perception. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 1554-1566, Marseille, France. European Language Resources Association.",
+            "Cesare Campagnano, Simone Conia, and Roberto Navigli. 2022. SRL4E – Semantic Role Labeling for Emotions: A Unified Evaluation Framework. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4586–4601, Dublin, Ireland. Association for Computational Linguistics.",
+            "Hao Chen, Zepeng Zhai, Fangxiang Feng, Ruifan Li, and Xiaojie Wang. 2022. Enhanced Multi-Channel Graph Convolutional Network for Aspect Sentiment Triplet Extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2974–2985, Dublin, Ireland. Association for Computational Linguistics.",
+            "Xinhong Chen, Qing Li, and Jianping Wang. 2020. A unified sequence labeling model for emotion cause pair extraction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 208-218, Barcelona, Spain (Online). International Committee on Computational Linguistics.",
+            "Kalpana Chowdary, Tu Nguyen, and D. Hemanth. 2021. Deep learning-based facial emotion recognition for human-computer interaction applications. *Neural Computing and Applications*, pages 1-18.",
+            "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, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "45",
+        "bbox": [
+            490,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Marwan Dhuheir, Abdullahif Albaseer, Emna Baccour, Aiman Erbad, Mohamed M. Abdallah, and Mounir Hamdi. 2021. Emotion recognition for healthcare surveillance systems using neural networks: A survey. CoRR, abs/2107.05989.",
+            "Zixiang Ding, Rui Xia, and Jianfei Yu. 2020a. ECPE-2D: Emotion-cause pair extraction based on joint two-dimensional representation, interaction and prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3161-3170, Online. Association for Computational Linguistics.",
+            "Zixiang Ding, Rui Xia, and Jianfei Yu. 2020b. End-to-end emotion-cause pair extraction based on sliding window multi-label learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3574-3583, Online. Association for Computational Linguistics.",
+            "Fang Fang, Yin Zhou, Shi Ying, and Zhijuan Li. 2023. A study of the ping an health app based on user reviews with sentiment analysis. International Journal of Environmental Research and Public Health, 20(2):1591.",
+            "Jonas Groschwitz, Shay Cohen, Lucia Donatelli, and Meaghan Fowlie. 2023. AMR parsing is far from solved: GrAPES, the granular AMR parsing evaluation suite. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10728-10752, Singapore. Association for Computational Linguistics.",
+            "Xiaojie Gu, Renze Lou, Lin Sun, and Shangxin Li. 2023. Page: A position-aware graph-based model for emotion cause entailment in conversation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5.",
+            "Dou Hu, Lingwei Wei, and Xiaoyong Huai. 2021. DialogueCRN: Contextual reasoning networks for emotion recognition in conversations. 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 7042-7052, Online. Association for Computational Linguistics.",
+            "Evgeny Kim and Roman Klinger. 2018. Who feels what and why? annotation of a literature corpus with semantic roles of emotions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1345-1359, Santa Fe, New Mexico, USA. Association for Computational Linguistics.",
+            "Pan Li, Ping Li, and Kai Zhang. 2023. Dual-channel span for aspect sentiment triplet extraction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 248-261, Singapore. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations.",
+            "Navonil Majumder, Soujanya Poria, Devamanyu Hazarika, Rada Mihalcea, Alexander Gelbukh, and Erik Cambria. 2019. Dialoguernn: An attentive rnn for emotion detection in conversations. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):6818-6825.",
+            "Saif M. Mohammad, Xiaodan Zhu, Svetlana Kiritchenko, and Joel Martin. 2015. Sentiment, emotion, purpose, and style in electoral tweets. Information Processing Management, 51(4):480-499.",
+            "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744.",
+            "Gabriela Pavarini, Tessa Reardon, Anja Hollowell, Vanessa Bennett, Emma Lawrence, Peer Support Young People's Advisory Group Brooks-Hall Ellie Foster-Estwick Ashley Juma Damian Omari Lewis Peter Power Lucy Rogers Maia, Vanessa Pifold, and Ilina Singh. 2023. Online peer support training to promote adolescents' emotional support skills, mental health and agency during covid-19: Randomised controlled trial and qualitative evaluation. European Child & Adolescent Psychiatry, 32(6):1119-1130.",
+            "Robert Plutchik. 2001. The nature of emotions: Human emotions have deep evolutionary roots, a fact that may explain their complexity and provide tools for clinical practice. American Scientist, 89(4):344-350.",
+            "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. The Journal of Machine Learning Research, 21(1):5485-5551.",
+            "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. arXiv preprint arXiv:2302.13971.",
+            "Quan Tu, Yanran Li, Jianwei Cui, Bin Wang, Ji-Rong Wen, and Rui Yan. 2022. MISC: A mixed strategy-aware model integrating COMET for emotional support conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 308-319, Dublin, Ireland. Association for Computational Linguistics.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "46",
+        "bbox": [
+            489,
+            928,
+            510,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.",
+            "Fanfan Wang, Jianfei Yu, and Rui Xia. 2023. Generative emotion cause triplet extraction in conversations with commonsense knowledge. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 3952-3963, Singapore. Association for Computational Linguistics.",
+            "Rui Xia and Zixiang Ding. 2019. Emotion-cause pair extraction: A new task to emotion analysis in texts. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1003–1012, Florence, Italy. Association for Computational Linguistics.",
+            "Debin Xiao, Rui Xia, and Jianfei Yu. 2023. Emotion cause extraction on social media without human annotation. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 1455-1468, Toronto, Canada. Association for Computational Linguistics.",
+            "Jianfei Yu, Luís Marujo, Jing Jiang, Pradeep Karuturi, and William Brendel. 2018. Improving multi-label emotion classification via sentiment classification with dual attention transfer network. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1097-1102, Brussels, Belgium. Association for Computational Linguistics.",
+            "Chaofa Yuan, Chuang Fan, Jianzhu Bao, and Ruifeng Xu. 2020. Emotion-cause pair extraction as sequence labeling based on a novel tagging scheme. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3568-3573, Online. Association for Computational Linguistics.",
+            "Sourabh Zanwar, Xiaofei Li, Daniel Wiechmann, Yu Qiao, and Elma Kerz. 2023. What to fuse and how to fuse: Exploring emotion and personality fusion strategies for explainable mental disorder detection. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 8926-8940, Toronto, Canada. Association for Computational Linguistics.",
+            "Zepeng Zhai, Hao Chen, Ruifan Li, and Xiaojie Wang. 2023. USSA: A unified table filling scheme for structured sentiment analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14340-14353, Toronto, Canada. Association for Computational Linguistics.",
+            "Xiaopeng Zheng, Zhiyue Liu, Zizhen Zhang, Zhaoyang Wang, and Jiahai Wang. 2022. UECA-prompt: Universal prompt for emotion cause analysis. In Proceedings of the 29th International Conference on Computational Linguistics, pages 7031-7041, Gyeongju, Republic of Korea. International Committee on Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            901
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "47",
+        "bbox": [
+            489,
+            928,
+            510,
+            939
+        ],
+        "page_idx": 10
+    }
+]

2024/End-to-End Emotion Semantic Parsing/4ffbaf63-28e0-4983-95b1-695face241d4_model.json

@@ -0,0 +1,2191 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.295,
+                0.091,
+                0.703,
+                0.112
+            ],
+            "angle": 0,
+            "content": "End-to-End Emotion Semantic Parsing"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.265,
+                0.146,
+                0.738,
+                0.163
+            ],
+            "angle": 0,
+            "content": "Xiaotong Jiang, Zhongqing Wang\\*, and Guodong Zhou"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.21,
+                0.163,
+                0.796,
+                0.196
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)Natural Language Processing Lab, Soochow University, Suzhou, China xtjiang21@stu.suda.edu.cn, {wangzq, gdzhou}@suda.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.262,
+                0.341,
+                0.276
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.289,
+                0.461,
+                0.588
+            ],
+            "angle": 0,
+            "content": "Emotion detection is the task of automatically associating one or more emotions with a text. The emotions are experienced, targeted, and caused by different semantic constituents. Therefore, it is necessary to incorporate these semantic constituents into the process of emotion detection. In this study, we propose a new task called emotion semantic parsing which aims to parse the emotion and semantic constituents into an abstract semantic tree structure. In particular, we design an end-to-end generation model to capture the relations between emotion and all the semantic constituents, and to generate them jointly. Furthermore, we employ a task decomposition strategy to capture the semantic relation among these constituents in a more cognitive and structural way. Experimental results demonstrate the importance of the proposed task, and indicate the proposed model gives superior performance compared to other models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.601,
+                0.26,
+                0.616
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.627,
+                0.49,
+                0.834
+            ],
+            "angle": 0,
+            "content": "The goal of emotion detection is to automatically detect or categorize the emotional states of human according to some inputs. Nowadays, emotion detection can be found in a broad range of applications, including but not limited to emotional support (Tu et al., 2022; Pavarini et al., 2023), human-computer interaction (Chowdary et al., 2021; Alrowais et al., 2023) and healthcare surveillance (Dhuheir et al., 2021; Fang et al., 2023). Henceforth, emotion detection has attracted increasing attention from both research community and industry in recent years (Hu et al., 2021; Zanwar et al., 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.836,
+                0.49,
+                0.9
+            ],
+            "angle": 0,
+            "content": "In literature, most studies focus on detecting emotion from the given text. For example, feature-based works (Balahur et al., 2011) and deep learning approaches, including RNNs (Majumder et al.,"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.569,
+                0.262,
+                0.81,
+                0.284
+            ],
+            "angle": 0,
+            "content": "He clenched his fists as he watched the bully taunt his little brother."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.547,
+                0.288,
+                0.85,
+                0.449
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.545,
+                0.465,
+                0.848,
+                0.48
+            ],
+            "angle": 0,
+            "content": "Figure 1: Example of abstract semantic tree."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.512,
+                0.885,
+                0.656
+            ],
+            "angle": 0,
+            "content": "2019), Transformers (Yu et al., 2018), and pretrained language models (Zheng et al., 2022; Zanwar et al., 2023) are used in emotion detection. Nevertheless, over the past few years, numerous studies (Xia and Ding, 2019; Kim and Klinger, 2018; Campagnano et al., 2022) have indicated that the emotions may be experienced, targeted, and/or caused by different semantic constituents, and can be linked to form abstract semantic structures."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.661,
+                0.885,
+                0.901
+            ],
+            "angle": 0,
+            "content": "However, most existing works only take \"cause\" into consideration, neglecting the intricate relationships between emotion constituents, such as the strong relationship between trigger and emotion, as well as the similarity of sentiment and emotion labels. This limitation makes acquiring a nuanced understanding of the interplay between emotion and semantic constituents challenging, thereby hindering the model's capability to thoroughly comprehend emotional expression. By addressing these complexities, we believe it can serve as a foundational component for improving results in downstream real-world tasks by providing a more comprehensive understanding of emotional dynamics and their influence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.527,
+                0.906,
+                0.883,
+                0.922
+            ],
+            "angle": 0,
+            "content": "To tackle the above problem, we propose a new"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.142,
+                0.907,
+                0.295,
+                0.921
+            ],
+            "angle": 0,
+            "content": "* Corresponding author"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.928,
+                0.511,
+                0.94
+            ],
+            "angle": 0,
+            "content": "37"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.245,
+                0.946,
+                0.754,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2024, pages 37-47 August 11-16, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.278
+            ],
+            "angle": 0,
+            "content": "task called emotion semantic parsing which aims to parse the emotion and semantic constituents into an abstract semantic tree structure. In this work, we consider five emotion-related semantic constituents into this task, including emotion, sentiment, trigger, holder, and target. As shown in Figure 1, emotion and sentiment refer to the specific emotion and sentiment labels of the sentence; trigger refers to a trigger word or expression that describes an emotion; and the constituents of holder and target correspond to the person or entities towards whom or from whom the emotion is directed, respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.279,
+                0.492,
+                0.713
+            ],
+            "angle": 0,
+            "content": "Moreover, we use the form of the abstract semantic tree structure to model the relationship between these constituents, we use the Trigger node to represent the whole abstract semantic tree, and we use the virtual Feelings and Entities nodes to compass other constituent nodes. The tree structure can be considered as a semantic representation in order to better represent the structure of the emotion-related constituents. As shown in Figure 1, the tree structure models a sentence using a rooted directed acyclic graph, highlighting its main elements (e.g. emotion, triggers) and semantic relations. It can thus potentially reveal a more comprehensive and complete semantic structure for extracting these constituents. Furthermore, since the emotion and semantic constituents are correlated in the form of the abstract semantic tree structure, we design an end-to-end generative model to predict them jointly. Specifically, our proposed model begins by decomposing the original task into several sub-tasks, as shown in Figure 2. Subsequently, the proposed model utilizes template-based prompt learning to extract each semantic constituent and semantic constituents pair separately for each sub-tasks. Finally, the proposed model leverages the acquired knowledge from sub-tasks to generate all semantic constituents in the main task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.715,
+                0.491,
+                0.83
+            ],
+            "angle": 0,
+            "content": "We conducted extensive experiments on the unified benchmark. Experimental results demonstrate the importance of the proposed task, and indicate the proposed model gives superior performance compared to other models. Further analysis shows the effect of the proposed abstract semantic tree structure and decomposition strategies."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.845,
+                0.281,
+                0.86
+            ],
+            "angle": 0,
+            "content": "2 Related Works"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.874,
+                0.491,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In the domain of emotion analysis, previous studies have mainly focused on text-based emotion detection. Early feature-based work often relied on"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.887,
+                0.277
+            ],
+            "angle": 0,
+            "content": "traditional machine learning algorithms and manually crafted word-level emotion lexicons (Balahure et al., 2011). As deep neural networks progressed, emotion detection methods began utilizing various networks, such as RNNs (Majumder et al., 2019) and Transformers (Vaswani et al., 2017; Yu et al., 2018), to learn the emotion-related representation for detection. Recently, pre-trained language models (Devlin et al., 2019; Raffel et al., 2020) have dominated this area with its flexibility and rich internal knowledge (Alvarez-Gonzalez et al., 2021; Zheng et al., 2022)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.279,
+                0.886,
+                0.618
+            ],
+            "angle": 0,
+            "content": "The emotions are not expressed individually, and can be experienced, targeted, and caused by different semantic constituents (Xia and Ding, 2019; Xiao et al., 2023; Campagnano et al., 2022). Most recent studies focus on emotion cause extraction task (Xia and Ding, 2019; Wang et al., 2023; Gu et al., 2023), with the goal of extracting the potential pairs of emotions and corresponding causes in the document and conversation scenario. Many end-to-end approaches have been proposed. These works can be broadly categorized into two main groups: some focus on multi-task learning (Ding et al., 2020a,b; Chen et al., 2022), while others transform the original task into a unified sequence labeling problem (Yuan et al., 2020; Chen et al., 2020). Recently, Xiao et al. (2023) developed a rule-based system to extract emotion causes on social media from constituency parsing trees. Wang et al. (2023) proposed emotion cause triplet extraction to extract emotion-cause-category triplets for fine-grained emotion cause analysis."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.619,
+                0.886,
+                0.749
+            ],
+            "angle": 0,
+            "content": "Different from prior studies, we introduce a novel task, termed emotion semantic parsing. This task seeks to comprehensively examine the impact of various semantic constituents on emotional expressions. To achieve this, we develop an end-to-end generative model specifically designed to capture the intricate relationships between emotion and a wide range of semantic constituents."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.764,
+                0.663,
+                0.78
+            ],
+            "angle": 0,
+            "content": "3 Preliminaries"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.793,
+                0.886,
+                0.922
+            ],
+            "angle": 0,
+            "content": "As shown in Figure 2, we introduce a novel task termed emotion semantic parsing. This task focuses on the generation of an abstract semantic tree structure that encapsulates both emotion and semantic constituents. To achieve this, we develop an end-to-end generative model specifically tailored for the generation of the abstract semantic tree structure. Our proposed model adopts a"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.491,
+                0.929,
+                0.511,
+                0.941
+            ],
+            "angle": 0,
+            "content": "38"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.196
+            ],
+            "angle": 0,
+            "content": "decomposition-based strategy, breaking down the complex task into several more manageable subtasks. Specifically, our proposed model begins by decomposing the original task into several subtasks. Finally, the proposed model leverages the acquired knowledge from sub-tasks to generate all semantic constituents in the main task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.198,
+                0.49,
+                0.293
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce two important issues of this task: abstract semantic tree construction and the abstract semantic tree parsing model. The decomposition-based strategy for abstract semantic tree parsing will be discussed in the next section."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.305,
+                0.453,
+                0.319
+            ],
+            "angle": 0,
+            "content": "3.1 Abstract Semantic Tree Construction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.325,
+                0.49,
+                0.468
+            ],
+            "angle": 0,
+            "content": "Inspired by the widespread impact of tree structures in parsing tasks (Li et al., 2023; Groschwitz et al., 2023), we represent the relationships among multiple constituents in a tree structure. This approach allows us to capture the complex dynamics inherent in emotional expression. We first introduce the constituents in the abstract semantic tree, then we describe the construction process of the tree."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.478,
+                0.406,
+                0.492
+            ],
+            "angle": 0,
+            "content": "Definitions of Semantic Constituents"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.497,
+                0.49,
+                0.545
+            ],
+            "angle": 0,
+            "content": "There are five types of semantic constituents, each serving as a node in the abstract semantic tree. Their definitions are provided below:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.553,
+                0.488,
+                0.601
+            ],
+            "angle": 0,
+            "content": "- Emotion is the specific emotion label represented from the original text. In this study, we use Plutchik (2001)'s taxonomy of emotion."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.611,
+                0.49,
+                0.658
+            ],
+            "angle": 0,
+            "content": "- Sentiment is the specific sentiment label represented from the original text, including \"positive\", \"negative\", and \"neutral\"."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.668,
+                0.488,
+                0.701
+            ],
+            "angle": 0,
+            "content": "- Trigger refers to a trigger word or expression that describes an emotion, even if it is implicit."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.71,
+                0.488,
+                0.757
+            ],
+            "angle": 0,
+            "content": "- Holder refers to a person or entity that feels or experiences the emotion identified by the trigger."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.767,
+                0.488,
+                0.814
+            ],
+            "angle": 0,
+            "content": "- Target refers to a person or entity towards whom/which the emotion identified by the trigger is directed."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.553,
+                0.49,
+                0.814
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.823,
+                0.345,
+                0.837
+            ],
+            "angle": 0,
+            "content": "Process of Tree Construction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.842,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "As shown in Figure 1, the abstract semantic tree represents a sentence through a rooted directed acyclic graph utilizing the aforementioned semantic constituents. The construction process is outlined below:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.085,
+                0.883,
+                0.134
+            ],
+            "angle": 0,
+            "content": "- We take the \"Trigger\" node as the root node, since it is the most important constituent in the emotion parsing."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.146,
+                0.885,
+                0.178
+            ],
+            "angle": 0,
+            "content": "- The virtual nodes \"Feelings\" and \"Entities\" are linked to Trigger node."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.19,
+                0.885,
+                0.253
+            ],
+            "angle": 0,
+            "content": "- The \"Sentiment\" and \"Emotion\" are linked to \"Feelings\" as terminal nodes, while the \"Holder\" and \"Target\" are linked to \"Entities\" node."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.532,
+                0.085,
+                0.885,
+                0.253
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.268,
+                0.885,
+                0.379
+            ],
+            "angle": 0,
+            "content": "As Figure 1 shows, the trigger \"clenched his fists\" implies the author's emotion. The holder and target are \"he\" and \"the bully taunt his little brother\", respectively. By leveraging these emotion-related semantic constituents, the model can accurately predict the \"anger\" emotion and \"negative\" polarity of the sentence."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.393,
+                0.78,
+                0.408
+            ],
+            "angle": 0,
+            "content": "3.2 Basic Semantic Tree Parsing"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.415,
+                0.884,
+                0.509
+            ],
+            "angle": 0,
+            "content": "After we construct the abstract semantic tree, we then employ a generative model to parse the emotion and all the semantic constituents into a linearized tree structure. As shown in Figure 2, given an input text, the output can be represented as a linearized semantic tree."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.512,
+                0.884,
+                0.623
+            ],
+            "angle": 0,
+            "content": "We adopt the pre-trained language model T5 (Chung et al., 2022) as our transformer-based encoder-decoder architecture. Therefore, by formulating parsing tasks as a text generation problem, we can tackle them in a unified sequence-to-sequence framework without a task-specific model design."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.625,
+                0.883,
+                0.785
+            ],
+            "angle": 0,
+            "content": "The main challenge of emotion semantic parsing is the complex structure inherently in the emotional expression. Specifically, the complex structure will bring out two challenges: First, the large number of semantic constituents raises the complexity to the parsing task. Consequently, the precise extraction of every constituent becomes challenging yet indispensable. Second, effectively modeling these internal relationships also presents a significant challenge in the task."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.799,
+                0.857,
+                0.832
+            ],
+            "angle": 0,
+            "content": "4 Decomposition-based Semantic Tree Parsing Model"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.842,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In this study, we propose a decomposed-based generative model to tackle the above challenges. The motivation is to alleviate the complexity of the original task by decomposing it into several sub-tasks. As shown in Figure 2, our proposed model can be"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.491,
+                0.929,
+                0.511,
+                0.941
+            ],
+            "angle": 0,
+            "content": "39"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.178,
+                0.082,
+                0.825,
+                0.33
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.35,
+                0.338,
+                0.645,
+                0.353
+            ],
+            "angle": 0,
+            "content": "Figure 2: Overview of the proposed model."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.137,
+                0.379,
+                0.47,
+                0.63
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.645,
+                0.49,
+                0.701
+            ],
+            "angle": 0,
+            "content": "Figure 3: Examples of task decomposition strategies. Due to the space limit, prompt texts in this figure are shortened. Original prompt texts are presented in Section 4.1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.761,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "separated into two stages: decomposition and recomposition. Our model firstly breaks down the original abstract semantic parsing task into several sub-tasks, and then designs individual instruction-based templates for each task. All sub-tasks are fine-tuned by a generative backbone at the same time. Based on the model parameters from the last re-composition stage, our model further finetunes its parameters again with the original task, which involves jointly parsing all semantic constituents."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.378,
+                0.785,
+                0.394
+            ],
+            "angle": 0,
+            "content": "4.1 Task Decomposition Strategy"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.399,
+                0.885,
+                0.527
+            ],
+            "angle": 0,
+            "content": "To implement a generalized framework that is applicable to all tasks, we design templates for each individual task. The input template for each task comprises the task-specific instruction, and the original text to be parsed. The output template for each task is the sub-tree containing the constituents to be parsed. The following templates use the example in Figure 2 as the Original Text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.528,
+                0.884,
+                0.592
+            ],
+            "angle": 0,
+            "content": "Mono Parsing includes trigger, emotion, sentiment, holder, and target parsing, individually. As shown in Figure 3(a), these sub-tasks aim to capture each specific constituent:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.545,
+                0.601,
+                0.848,
+                0.631
+            ],
+            "angle": 0,
+            "content": "//taking emotion parsing task as an example"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.545,
+                0.636,
+                0.845,
+                0.668
+            ],
+            "angle": 0,
+            "content": "INPUT Detect the emotion label in the emotional expression: + Original Text"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.546,
+                0.673,
+                0.75,
+                0.689
+            ],
+            "angle": 0,
+            "content": "Output [Emotion (Anger)]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.696,
+                0.885,
+                0.792
+            ],
+            "angle": 0,
+            "content": "Trigger-related Pair Parsing includes trigger-emotion, trigger-sentiment, trigger-holder, trigger-target, trigger-feelings, and trigger-entities. As shown in Figure 3(b), these sub-tasks aim to capture the relationship between the trigger and each specific constituent:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.545,
+                0.8,
+                0.845,
+                0.832
+            ],
+            "angle": 0,
+            "content": "//taking trigger-holder parsing task as an example"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.545,
+                0.837,
+                0.846,
+                0.883
+            ],
+            "angle": 0,
+            "content": "INPUT Extract both the trigger and holder in the emotional expression: + Original Text"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.545,
+                0.89,
+                0.848,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Output [Trigger (clench his fists), [Entities, [Holder (he)]]]"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.929,
+                0.512,
+                0.941
+            ],
+            "angle": 0,
+            "content": "40"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.182
+            ],
+            "angle": 0,
+            "content": "Entities Pair Parsing aims to capture the relationship between holder and target, and can gain insights into the interpersonal dynamics and relational aspects that influence emotional experiences. The example is shown in Figure 3(c), with the template of prompt provided below:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.195,
+                0.452,
+                0.242
+            ],
+            "angle": 0,
+            "content": "INPUT Extract both holder and target in the emotional expression: + Original Text"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.25,
+                0.453,
+                0.282
+            ],
+            "angle": 0,
+            "content": "Output [Entities, [Holder (he), Target (the bully taunt his little brother) ]]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.296,
+                0.49,
+                0.36
+            ],
+            "angle": 0,
+            "content": "Feelings Pair Parsing aims to capture the relationship between sentiment and emotion. The example is shown in Figure 3(d), with the template of prompt provided below:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.149,
+                0.374,
+                0.454,
+                0.422
+            ],
+            "angle": 0,
+            "content": "INPUT Extract both the emotion and sentiment labels in the emotional expression: + Original Text"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.429,
+                0.454,
+                0.461
+            ],
+            "angle": 0,
+            "content": "Output [Feelings, [Emotion (Anger), Sentiment (Negative) ]]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.474,
+                0.398,
+                0.491
+            ],
+            "angle": 0,
+            "content": "4.2 Task Re-composition Strategy"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.496,
+                0.49,
+                0.641
+            ],
+            "angle": 0,
+            "content": "After gaining knowledge from sub-tasks, the model has learned the pattern of the template and the relationships from different emotion-related semantic constituents. Thus it can boost the final end-to-end emotion semantic parsing task, which is to predict the emotion label and outputs all emotion-related semantic constituents in the form of an abstract semantic tree. The example is shown in Figure 2, and the template of prompt is the below:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.655,
+                0.454,
+                0.717
+            ],
+            "angle": 0,
+            "content": "INPUT Extract emotion trigger, holder, target, and emotion and sentiment labels in the emotional expression: + Original Text"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.726,
+                0.454,
+                0.79
+            ],
+            "angle": 0,
+            "content": "Output [Trigger (clenched his fists), Feelings, [Emotion (Anger), Sentiment (Negative)]], [Entities, [Holder (he), Target (the bully taunt his little brother) ]]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.803,
+                0.427,
+                0.82
+            ],
+            "angle": 0,
+            "content": "4.3 Objective Functions and Training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.825,
+                0.487,
+                0.857
+            ],
+            "angle": 0,
+            "content": "In this subsection, we show the objective functions and training process of the proposed model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.858,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "The goal is to maximize the output linearized tree \\( X_{T} \\) probability given the sentence \\( X \\). Therefore, we optimize the negative log-likelihood loss function:"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.577,
+                0.082,
+                0.819,
+                0.219
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Amount</td></tr><tr><td>#Sample</td><td>7,234</td></tr><tr><td>Avg. Sent Len.</td><td>26.3</td></tr><tr><td>#Emotion Labels</td><td>9</td></tr><tr><td>#Sentiment Label</td><td>3</td></tr><tr><td>#Trigger</td><td>7,111</td></tr><tr><td>#Holder</td><td>5,633</td></tr><tr><td>#Target</td><td>6,127</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.586,
+                0.228,
+                0.805,
+                0.241
+            ],
+            "angle": 0,
+            "content": "Table 1: Distribution of dataset."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.562,
+                0.289,
+                0.884,
+                0.33
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} = - \\frac {1}{| \\tau |} \\sum_ {(X, X _ {T}) \\in \\tau} \\log p (X _ {T} | X; \\theta) \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.339,
+                0.883,
+                0.372
+            ],
+            "angle": 0,
+            "content": "where \\(\\theta\\) is the model parameters, and \\((X,X_{T})\\) is a (sentence, tree) pair in training set \\(\\tau\\), then"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.554,
+                0.395,
+                0.882,
+                0.455
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\log p (X _ {T} | X; \\theta) \\\\ = \\sum_ {i = 1} ^ {n} \\log p \\left(x _ {T} ^ {i} \\mid x _ {T} ^ {1}, x _ {T} ^ {2}, \\dots x _ {T} ^ {i - 1}, X; \\theta\\right) \\tag {2} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.467,
+                0.883,
+                0.499
+            ],
+            "angle": 0,
+            "content": "where \\(p(x_{T}^{i}|x_{T}^{1},x_{T}^{2},\\dots x_{T}^{i - 1},X;\\theta)\\) is calculated by the decoder."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.512,
+                0.692,
+                0.528
+            ],
+            "angle": 0,
+            "content": "5 Experimentation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.537,
+                0.884,
+                0.632
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce the datasets used for evaluation and the baseline methods employed for comparison. We then report the experimental results conducted from different perspectives, and analyze the effectiveness of the proposed model with different factors."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.644,
+                0.62,
+                0.659
+            ],
+            "angle": 0,
+            "content": "5.1 Settings"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.665,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "We evaluate the effectiveness of our proposed method on a benchmark proposed in Campagnano et al. (2022), which combines several datasets. We specifically choose three datasets (Mohammad et al., 2015; Kim and Klinger, 2018; Bostan et al., 2020) from this benchmark because they are the only ones that encompass annotations for all constituents. We enhance the quality of the original dataset by deleting duplicate or irrelevant examples (e.g., those with all trigger, holder, and target are \"null\"), while ensuring the trigger is not \"null\". The distribution of dataset is shown in Table 1. We choose \\(90\\%\\) of the data as the training set, and the remaining \\(10\\%\\) of the data as the testing set. Additionally, \\(10\\%\\) of the training data is randomly set aside for validation during the training process."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.929,
+                0.511,
+                0.941
+            ],
+            "angle": 0,
+            "content": "41"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.184,
+                0.081,
+                0.816,
+                0.242
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Methods</td><td>Emotion</td><td>Sentiment</td><td>Trigger</td><td>Holder</td><td>Target</td><td>Micro F1.</td></tr><tr><td>BERT-CRF</td><td>40.3</td><td>58.6</td><td>19.0</td><td>23.3</td><td>24.0</td><td>33.1</td></tr><tr><td>BERT-MRC</td><td>42.6</td><td>60.3</td><td>15.2</td><td>22.4</td><td>19.3</td><td>30.8</td></tr><tr><td>TaskFusion</td><td>45.7</td><td>62.3</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>USSA</td><td>41.3</td><td>-</td><td>17.3</td><td>32.6</td><td>14.3</td><td>-</td></tr><tr><td>T5</td><td>40.2</td><td>59.4</td><td>21.0</td><td>47.2</td><td>32.1</td><td>38.2</td></tr><tr><td>ChatGPT</td><td>22.3</td><td>30.7</td><td>10.4</td><td>21.3</td><td>9.6</td><td>18.9</td></tr><tr><td>LLaMA</td><td>42.4</td><td>57.6</td><td>22.4</td><td>43.5</td><td>31.3</td><td>39.4</td></tr><tr><td>Ours</td><td>46.3</td><td>63.6</td><td>30.8</td><td>55.0</td><td>37.4</td><td>46.7</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.373,
+                0.25,
+                0.625,
+                0.266
+            ],
+            "angle": 0,
+            "content": "Table 2: Comparison with baselines."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.29,
+                0.49,
+                0.484
+            ],
+            "angle": 0,
+            "content": "Our proposed method utilizes the generative backbone Flan-T5 (Chung et al., 2022) from HuggingFace's hub. All experiments are conducted on a single NVIDIA Tesla V100 GPU with 32GB memory. The model is trained for 10 epochs in each of the two stages, using the AdamW (Loshchilov and Hutter) optimizer, with a maximum sequence length of 256, a batch size of 16, and a learning rate of 3e-4. And the average results over 5 runs with different random seeds are reported. We adopt F1-score as the metric for all experiments, the micro F1 is employed to evaluate the overall performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.498,
+                0.267,
+                0.512
+            ],
+            "angle": 0,
+            "content": "5.2 Main Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.521,
+                0.49,
+                0.57
+            ],
+            "angle": 0,
+            "content": "To better evaluate the effectiveness of our proposed model, we compare its performance against several strong baseline models, which are detailed below,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.586,
+                0.49,
+                0.667
+            ],
+            "angle": 0,
+            "content": "- BERT-CRF (Campagnano et al., 2022) employs bidirectional LSTM layers over BERT representation, then uses CRF to generate token-level labels to determine the span boundary of each semantic constituent."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.682,
+                0.489,
+                0.73
+            ],
+            "angle": 0,
+            "content": "- BERT-MRC (Devlin et al., 2019) is used to predict the start and end indexes of all required semantic constituents."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.745,
+                0.49,
+                0.826
+            ],
+            "angle": 0,
+            "content": "- TaskFusion (Zanwar et al., 2023) is a multi-task framework with a pre-trained language model for emotion-related tasks. We adopt it for the proposed emotion semantic parsing task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.841,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "- USSA (Zhai et al., 2023) is a joint learning framework, which transfers aspect-based sentiment analysis into a 2D table-filling scheme. We adopt it for the proposed emotion semantic parsing task."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.586,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.29,
+                0.886,
+                0.371
+            ],
+            "angle": 0,
+            "content": "- T5 (Chung et al., 2022) is a pre-trained language model which naturally suits the sequence-to-sequence task. It also can be considered as a joint learning framework, which generates all the semantic constituents jointly."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.401,
+                0.886,
+                0.481
+            ],
+            "angle": 0,
+            "content": "- ChatGPT (Ouyang et al., 2022) uses dialog-based instructional data in the supervised and RL-based meta-training stages. We use ChatGPT API to generate all required semantic constituents."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.513,
+                0.886,
+                0.592
+            ],
+            "angle": 0,
+            "content": "- LLaMA (Touvron et al., 2023) is a series of large language models trained on trillions of tokens. We use Alpaca-LoRA² to fine-tune LLaMA-7B to generate required semantic constituents."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.532,
+                0.29,
+                0.886,
+                0.592
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.627,
+                0.886,
+                0.837
+            ],
+            "angle": 0,
+            "content": "As demonstrated in Table 2, several key observations can be made: 1) Aligning with the current trend favoring generative frameworks, models that adopt a generative paradigm (i.e., T5, LLaMA, and our proposed model) generally outpace models based on a classification paradigm (i.e., BERT-CRF, BERT-MRC, Task Fusion, USSA) in terms of performance. 2) A standout observation is the substantial performance disparity between ChatGPT and all other methodologies. This discrepancy might be attributed to our limited usage of ChatGPT's API, without fine-tuning its internal parameters using insights from our training dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.842,
+                0.886,
+                0.921
+            ],
+            "angle": 0,
+            "content": "In contrast, our proposed model consistently outperforms prior studies \\((p < 0.05)\\) in all settings. This underscores the effectiveness of our model, employing task decomposition strategies and the abstract semantic tree structure."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.929,
+                0.512,
+                0.941
+            ],
+            "angle": 0,
+            "content": "42"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.248,
+                0.083,
+                0.744,
+                0.202
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.381,
+                0.217,
+                0.615,
+                0.232
+            ],
+            "angle": 0,
+            "content": "Figure 4: Effect of tree structures."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.14,
+                0.254,
+                0.462,
+                0.395
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Tuples</td><td>e</td><td>s</td><td>t</td><td>h</td><td>a</td></tr><tr><td>(e)</td><td>42.5</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>(t)</td><td>-</td><td>-</td><td>28.4</td><td>-</td><td>-</td></tr><tr><td>(e,s)</td><td>43.5</td><td>65.1</td><td>-</td><td>-</td><td>-</td></tr><tr><td>(e,t)</td><td>45.2</td><td>-</td><td>29.1</td><td>-</td><td>-</td></tr><tr><td>(e,s,t)</td><td>45.7</td><td>67.5</td><td>30.7</td><td>-</td><td>-</td></tr><tr><td>(t,h)</td><td>-</td><td>-</td><td>29.3</td><td>43.1</td><td>-</td></tr><tr><td>(t,a)</td><td>-</td><td>-</td><td>29.1</td><td>-</td><td>32.4</td></tr><tr><td>(t,h,a)</td><td>-</td><td>-</td><td>29.3</td><td>50.0</td><td>35.5</td></tr><tr><td>(e,t,h,a)</td><td>45.9</td><td>-</td><td>29.4</td><td>52.4</td><td>35.3</td></tr><tr><td>Ours</td><td>46.3</td><td>63.6</td><td>30.8</td><td>55.0</td><td>37.4</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.127,
+                0.405,
+                0.474,
+                0.419
+            ],
+            "angle": 0,
+            "content": "Table 3: Correlations among semantic constituents."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.18,
+                0.433,
+                0.424,
+                0.54
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Methods</td><td>Micro F1.</td></tr><tr><td>Ours</td><td>46.7</td></tr><tr><td>- Mono</td><td>39.8</td></tr><tr><td>- TriggerPair</td><td>40.3</td></tr><tr><td>- EntitiesPair</td><td>41.8</td></tr><tr><td>- FeelingsPair</td><td>44.3</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.149,
+                0.549,
+                0.452,
+                0.563
+            ],
+            "angle": 0,
+            "content": "Table 4: Impact of decomposition strategies."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.59,
+                0.434,
+                0.604
+            ],
+            "angle": 0,
+            "content": "5.3 Influence of Semantic Constituents"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.611,
+                0.49,
+                0.755
+            ],
+            "angle": 0,
+            "content": "We then analyze the influence of the semantic constituents and the correlations among them. As shown in Table 3, \\( e \\) denotes the emotion label, \\( s \\) is the sentiment polarity, \\( t, h \\) and \\( a \\) are trigger, holder, and target, separately. In addition, the rows in the table mean different combinations of emotion-related semantic constituents. For example, \\( (e, s) \\) means that we used the proposed model to generate the emotion label and the sentiment polarity jointly."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.757,
+                0.49,
+                0.884
+            ],
+            "angle": 0,
+            "content": "From the table, it is evident that there is a clear correlation between emotion and sentiment polarity, as the combined performance of \\((e,s)\\) surpasses that of generating emotion \\(e\\) alone. This underscores the importance of considering sentiment polarity when analyzing and interpreting emotions. Additionally, our findings reveal that the majority of semantic constituents contribute significantly to"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.257,
+                0.885,
+                0.353
+            ],
+            "angle": 0,
+            "content": "emotion detection, especially trigger \\( t \\). For example, the combined performance of \\( (e, t) \\) surpasses that of generating emotion \\( e \\) alone by a large margin. This underscores the importance of these semantic elements in understanding and contextualizing emotions within a given text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.354,
+                0.885,
+                0.481
+            ],
+            "angle": 0,
+            "content": "Furthermore, our results highlight a correlation between the holder \\( h \\) and target \\( a \\) with the trigger \\( t \\), as they often refer to the person or entity experiencing or feeling the emotion triggered by a specific event or situation. This emphasizes the interconnected nature of these semantic constituents and their collective role in shaping emotional expressions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.484,
+                0.886,
+                0.596
+            ],
+            "angle": 0,
+            "content": "Notably, we observe that incorporating all semantic constituents leads to the highest performance, highlighting the effectiveness of our proposed model with semantic tree parsing. This underscores the utility and relevance of these semantic constituents in enhancing emotion detection accuracy."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.609,
+                0.75,
+                0.623
+            ],
+            "angle": 0,
+            "content": "5.4 Effect of Tree Structures"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.631,
+                0.882,
+                0.679
+            ],
+            "angle": 0,
+            "content": "To assess the impact of the abstract semantic tree structure, we evaluated our model's performance using four distinct tree structures in Figure 4."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.68,
+                0.882,
+                0.791
+            ],
+            "angle": 0,
+            "content": "In our initial assessment, we removed the entire tree structure, treating the semantic constituents as a linear sequence (a). This modification resulted in a substantial decrease in performance, from \\(46.7\\%\\) to \\(40.7\\%\\), emphasizing the importance of the tree structure in capturing the intricate relationships among semantic constituents."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.794,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Subsequently, we constructed two simplified tree structures, with either Trigger (b) or Emotion (c) serving as the root node. Interestingly, both of these structures yielded comparable performance results, suggesting that both of them are pivotal components in the emotion parsing task, and their roles are equally significant. Furthermore, we introduced two virtual nodes, labeled as Feelings and"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.136,
+                0.894,
+                0.463,
+                0.907
+            ],
+            "angle": 0,
+            "content": "1https://api.openai.com/v1/chat/completions"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.137,
+                0.907,
+                0.42,
+                0.921
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)https://github.com/tloen/alpaca-lora"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.136,
+                0.894,
+                0.463,
+                0.921
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.929,
+                0.511,
+                0.941
+            ],
+            "angle": 0,
+            "content": "43"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.162,
+                0.083,
+                0.837,
+                0.3
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Input</td><td>Model</td><td>Output (Linearized Abstract Semantic Tree)</td></tr><tr><td rowspan=\"2\">Jean Valjean was sincere. This sincerity from the very grief that it caused him, rendered inquiries useless, and conferred authority on all that that man had said. Here, for Marius, there was a strange reversal of situations.</td><td>T5</td><td>[Trigger, sincerity, [Feelings, [Emotion, trust], [Sentiment, positive]], [Entity, [Holder, him], [Target, null]]]</td></tr><tr><td>Ours</td><td>[Trigger, grief, [Feelings, [Emotion, sadness], [Sentiment, negative]], [Entity, [Holder, him], [Target, null]]]</td></tr><tr><td rowspan=\"2\">&quot;Do you think me so very proud?&quot; he suddenly asked. &quot;I think you very strange.&quot;</td><td>T5</td><td>[Trigger, so very proud, [Feelings, [Emotion, joy], [Sentiment, positive]], [Entity, [Holder, you], [Target, me]]]</td></tr><tr><td>Ours</td><td>[Trigger, suddenly, [Feelings, [Emotion, surprise], [Sentiment, other]], [Entity, [Holder, he], [Target, null]]]</td></tr><tr><td rowspan=\"2\">Last year I was wild to marry you, and you wouldn&#x27;t look at me.</td><td>T5</td><td>[Trigger, wouldn&#x27;t look at me, [Feelings, [Emotion, anger], [Sentiment, negative]], [Entity, [Holder, I], [Target, you]]]</td></tr><tr><td>Ours</td><td>[Trigger, wouldn&#x27;t look at me, [Feelings, [Emotion, disgust], [Sentiment, negative]], [Entity, [Holder, you], [Target, I]]]</td></tr></table>"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.31,
+                0.884,
+                0.34
+            ],
+            "angle": 0,
+            "content": "Figure 5: Examples of case study. The text in green means an exact match, while text in red indicates inaccuracies. \"Null\" means the extracted constituent is missing."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.365,
+                0.489,
+                0.462
+            ],
+            "angle": 0,
+            "content": "Entities into the tree structure (d). The inclusion of these nodes led to a noticeable improvement in performance, highlighting the efficacy of our proposed tree structures and reinforcing the importance of considering semantic constituents within a structured framework for accurate emotion parsing."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.475,
+                0.441,
+                0.492
+            ],
+            "angle": 0,
+            "content": "5.5 Impact of Decomposition Strategies"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.498,
+                0.49,
+                0.608
+            ],
+            "angle": 0,
+            "content": "We further investigated the influence of various decomposition strategies, and the results are presented in Table 4, where Mono denotes Mono Parsing, TriggerPair denotes Trigger-related Pair Parsing, EntitiesPair denotes Entities Pair Parsing, and FeelingsPair denotes Feelings Pair Parsing. All of these can be found in Section 4.1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.612,
+                0.49,
+                0.803
+            ],
+            "angle": 0,
+            "content": "Our analysis reveals that among all the strategies, mono parsing exerts the most substantial impact on model performance, aligning with our initial expectations. Mono parsing can be regarded as the foundational and indivisible task that serves as the basis for other higher-level tasks. Moreover, our findings indicate that all the decomposition strategies employed have a positive influence on the task, emphasizing the importance of breaking down the original task into smaller sub-tasks and the efficacy of reintegrating these sub-tasks to address the main task."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.818,
+                0.25,
+                0.834
+            ],
+            "angle": 0,
+            "content": "5.6 Case Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.489,
+                0.888
+            ],
+            "angle": 0,
+            "content": "To highlight our model's significance, we've chosen several challenging and illustrative examples shown in Figure 5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.89,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "These challenging sentences share the same feature: they express more than one emotion expres"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.365,
+                0.884,
+                0.444
+            ],
+            "angle": 0,
+            "content": "sion, each associated with its own trigger, holder, and target. However, in each case, there is only one overall emotion. The presence of secondary emotions makes it difficult to fully grasp the overall emotional tone of the sentence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.446,
+                0.885,
+                0.686
+            ],
+            "angle": 0,
+            "content": "In the first two examples, the mis-detection of emotion expressions by T5 underscores the intricate and nuanced nature of the task, as well as the significance of each element in constructing a precise semantic representation. However, even when the correct emotion expression is identified, modeling the complex relationship between emotion and other semantic components remains challenging. As illustrated in the final example, although T5 correctly identifies the expression and its trigger, its inability to accurately detect the holder and target leads to an incorrect prediction of emotion. This highlights the interconnected nature of these elements and underscores their collective importance in achieving accurate emotion detection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.688,
+                0.884,
+                0.833
+            ],
+            "angle": 0,
+            "content": "In contrast, our proposed model, by leveraging task decomposition, demonstrates enhanced efficacy in navigating complexities present in difficult scenarios such as negotiation, emotion shifts, and multiple possible emotion expressions. This not only facilitates accurate detection of individual constituents but also enables a more nuanced understanding of their interrelationships, leading to improved overall performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.847,
+                0.642,
+                0.862
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.874,
+                0.884,
+                0.921
+            ],
+            "angle": 0,
+            "content": "In this study, we introduce a novel task called emotion semantic parsing, which focuses on analyzing emotions and semantic elements to construct an ab"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.929,
+                0.512,
+                0.941
+            ],
+            "angle": 0,
+            "content": "44"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.281
+            ],
+            "angle": 0,
+            "content": "tract semantic tree structure. To achieve this, we develop an end-to-end generation model capable of capturing intricate relationships between emotions and various semantic constituents, generating them collaboratively. Furthermore, our proposed model employs template-based prompt learning, allowing for the separate extraction of each semantic constituent and pairs of semantic constituents for individual sub-tasks. Experimental results not only underscore the significance of the introduced task but also showcase the superior performance of our proposed model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.291,
+                0.221,
+                0.307
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.317,
+                0.493,
+                0.624
+            ],
+            "angle": 0,
+            "content": "Our work focuses on employing a unified framework for emotion semantic parsing. However, one of the limitations is the computational complexity of the proposed model, whose training speed is slower than the classification-based models. Secondly, more tasks could be further explored, including nested emotion parsing, cross-domain and cross-lingual emotion semantic parsing. Finally, a large language model (LLM) can achieve superior performance in nearly all downstream tasks including our task, attributed to their ability to introduce rich knowledge learned during the pretraining stage for our proposed task. For open-source large language models like LLaMA (Touvron et al., 2023), limitations include high computational costs, high-quality data costs, as well as the need to address LLM-related issues such as safety, hallucinations, and ethical concerns, which remain to be resolved in future work."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.636,
+                0.287,
+                0.652
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.662,
+                0.49,
+                0.79
+            ],
+            "angle": 0,
+            "content": "We would like to thank Shufan Fu for her invaluable advice and assistance with the experiments during this work. We also gratefully thank the anonymous reviewers for their insightful and valuable comments. This work is supported by the National Natural Science Foundation of China (No. 62076175, No.61976146), and Jiangsu Innovation Doctor Plan."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.819,
+                0.214,
+                0.833
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.842,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Fadwa Alrowais, Noha Negm, Majdi Khalid, Nabil Almalki, Radwa Marzouk, Abdullah Mohamed, Mesfer Al Duhayyim, and Amani a Alneil. 2023. Modified earthworm optimization with deep learning assisted emotion recognition for human computer interface. IEEE Access, 11:35089-35096."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.168
+            ],
+            "angle": 0,
+            "content": "Nurudin Alvarez-Gonzalez, Andreas Kaltenbrunner, and Vicenç Gómez. 2021. Uncovering the limits of text-based emotion detection. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2560-2583, Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.174,
+                0.885,
+                0.254
+            ],
+            "angle": 0,
+            "content": "Alexandra Balahur, Jesús M. Hermida, Andres Montoyo, and Rafael Muñoz. 2011. Emotinet: A knowledge base for emotion detection in text built on the appraisal theories. In *Natural Language Processing and Information Systems*, pages 27-39, Berlin, Heidelberg. Springer Berlin Heidelberg."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.262,
+                0.885,
+                0.354
+            ],
+            "angle": 0,
+            "content": "Laura Ana Maria Bostan, Evgeny Kim, and Roman Klinger. 2020. GoodNewsEveryone: A corpus of news headlines annotated with emotions, semantic roles, and reader perception. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 1554-1566, Marseille, France. European Language Resources Association."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.363,
+                0.885,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Cesare Campagnano, Simone Conia, and Roberto Navigli. 2022. SRL4E – Semantic Role Labeling for Emotions: A Unified Evaluation Framework. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4586–4601, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.464,
+                0.885,
+                0.569
+            ],
+            "angle": 0,
+            "content": "Hao Chen, Zepeng Zhai, Fangxiang Feng, Ruifan Li, and Xiaojie Wang. 2022. Enhanced Multi-Channel Graph Convolutional Network for Aspect Sentiment Triplet Extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2974–2985, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.578,
+                0.885,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Xinhong Chen, Qing Li, and Jianping Wang. 2020. A unified sequence labeling model for emotion cause pair extraction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 208-218, Barcelona, Spain (Online). International Committee on Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.666,
+                0.885,
+                0.72
+            ],
+            "angle": 0,
+            "content": "Kalpana Chowdary, Tu Nguyen, and D. Hemanth. 2021. Deep learning-based facial emotion recognition for human-computer interaction applications. *Neural Computing and Applications*, pages 1-18."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.728,
+                0.885,
+                0.794
+            ],
+            "angle": 0,
+            "content": "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.803,
+                0.885,
+                0.921
+            ],
+            "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": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.921
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.491,
+                0.929,
+                0.511,
+                0.941
+            ],
+            "angle": 0,
+            "content": "45"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Marwan Dhuheir, Abdullahif Albaseer, Emna Baccour, Aiman Erbad, Mohamed M. Abdallah, and Mounir Hamdi. 2021. Emotion recognition for healthcare surveillance systems using neural networks: A survey. CoRR, abs/2107.05989."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.165,
+                0.488,
+                0.256
+            ],
+            "angle": 0,
+            "content": "Zixiang Ding, Rui Xia, and Jianfei Yu. 2020a. ECPE-2D: Emotion-cause pair extraction based on joint two-dimensional representation, interaction and prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3161-3170, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.269,
+                0.488,
+                0.347
+            ],
+            "angle": 0,
+            "content": "Zixiang Ding, Rui Xia, and Jianfei Yu. 2020b. End-to-end emotion-cause pair extraction based on sliding window multi-label learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3574-3583, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.36,
+                0.488,
+                0.425
+            ],
+            "angle": 0,
+            "content": "Fang Fang, Yin Zhou, Shi Ying, and Zhijuan Li. 2023. A study of the ping an health app based on user reviews with sentiment analysis. International Journal of Environmental Research and Public Health, 20(2):1591."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.438,
+                0.488,
+                0.53
+            ],
+            "angle": 0,
+            "content": "Jonas Groschwitz, Shay Cohen, Lucia Donatelli, and Meaghan Fowlie. 2023. AMR parsing is far from solved: GrAPES, the granular AMR parsing evaluation suite. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10728-10752, Singapore. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.542,
+                0.488,
+                0.621
+            ],
+            "angle": 0,
+            "content": "Xiaojie Gu, Renze Lou, Lin Sun, and Shangxin Li. 2023. Page: A position-aware graph-based model for emotion cause entailment in conversation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.633,
+                0.488,
+                0.738
+            ],
+            "angle": 0,
+            "content": "Dou Hu, Lingwei Wei, and Xiaoyong Huai. 2021. DialogueCRN: Contextual reasoning networks for emotion recognition in conversations. 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 7042-7052, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.751,
+                0.488,
+                0.829
+            ],
+            "angle": 0,
+            "content": "Evgeny Kim and Roman Klinger. 2018. Who feels what and why? annotation of a literature corpus with semantic roles of emotions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1345-1359, Santa Fe, New Mexico, USA. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.842,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Pan Li, Ping Li, and Kai Zhang. 2023. Dual-channel span for aspect sentiment triplet extraction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 248-261, Singapore. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.881,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.136,
+                0.883,
+                0.213
+            ],
+            "angle": 0,
+            "content": "Navonil Majumder, Soujanya Poria, Devamanyu Hazarika, Rada Mihalcea, Alexander Gelbukh, and Erik Cambria. 2019. Dialoguernn: An attentive rnn for emotion detection in conversations. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):6818-6825."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.224,
+                0.883,
+                0.276
+            ],
+            "angle": 0,
+            "content": "Saif M. Mohammad, Xiaodan Zhu, Svetlana Kiritchenko, and Joel Martin. 2015. Sentiment, emotion, purpose, and style in electoral tweets. Information Processing Management, 51(4):480-499."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.286,
+                0.883,
+                0.364
+            ],
+            "angle": 0,
+            "content": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.374,
+                0.883,
+                0.505
+            ],
+            "angle": 0,
+            "content": "Gabriela Pavarini, Tessa Reardon, Anja Hollowell, Vanessa Bennett, Emma Lawrence, Peer Support Young People's Advisory Group Brooks-Hall Ellie Foster-Estwick Ashley Juma Damian Omari Lewis Peter Power Lucy Rogers Maia, Vanessa Pifold, and Ilina Singh. 2023. Online peer support training to promote adolescents' emotional support skills, mental health and agency during covid-19: Randomised controlled trial and qualitative evaluation. European Child & Adolescent Psychiatry, 32(6):1119-1130."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.515,
+                0.883,
+                0.568
+            ],
+            "angle": 0,
+            "content": "Robert Plutchik. 2001. The nature of emotions: Human emotions have deep evolutionary roots, a fact that may explain their complexity and provide tools for clinical practice. American Scientist, 89(4):344-350."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.577,
+                0.883,
+                0.655
+            ],
+            "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. The Journal of Machine Learning Research, 21(1):5485-5551."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.666,
+                0.883,
+                0.757
+            ],
+            "angle": 0,
+            "content": "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. arXiv preprint arXiv:2302.13971."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.767,
+                0.883,
+                0.871
+            ],
+            "angle": 0,
+            "content": "Quan Tu, Yanran Li, Jianwei Cui, Bin Wang, Ji-Rong Wen, and Rui Yan. 2022. MISC: A mixed strategy-aware model integrating COMET for emotional support conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 308-319, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.881,
+                0.883,
+                0.92
+            ],
+            "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"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.929,
+                0.512,
+                0.941
+            ],
+            "angle": 0,
+            "content": "46"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.135,
+                0.086,
+                0.49,
+                0.113
+            ],
+            "angle": 0,
+            "content": "you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.123,
+                0.49,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Fanfan Wang, Jianfei Yu, and Rui Xia. 2023. Generative emotion cause triplet extraction in conversations with commonsense knowledge. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 3952-3963, Singapore. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.212,
+                0.49,
+                0.291
+            ],
+            "angle": 0,
+            "content": "Rui Xia and Zixiang Ding. 2019. Emotion-cause pair extraction: A new task to emotion analysis in texts. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1003–1012, Florence, Italy. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.3,
+                0.49,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Debin Xiao, Rui Xia, and Jianfei Yu. 2023. Emotion cause extraction on social media without human annotation. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 1455-1468, Toronto, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.389,
+                0.49,
+                0.495
+            ],
+            "angle": 0,
+            "content": "Jianfei Yu, Luís Marujo, Jing Jiang, Pradeep Karuturi, and William Brendel. 2018. Improving multi-label emotion classification via sentiment classification with dual attention transfer network. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1097-1102, Brussels, Belgium. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.504,
+                0.49,
+                0.596
+            ],
+            "angle": 0,
+            "content": "Chaofa Yuan, Chuang Fan, Jianzhu Bao, and Ruifeng Xu. 2020. Emotion-cause pair extraction as sequence labeling based on a novel tagging scheme. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3568-3573, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.606,
+                0.49,
+                0.698
+            ],
+            "angle": 0,
+            "content": "Sourabh Zanwar, Xiaofei Li, Daniel Wiechmann, Yu Qiao, and Elma Kerz. 2023. What to fuse and how to fuse: Exploring emotion and personality fusion strategies for explainable mental disorder detection. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 8926-8940, Toronto, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.707,
+                0.49,
+                0.8
+            ],
+            "angle": 0,
+            "content": "Zepeng Zhai, Hao Chen, Ruifan Li, and Xiaojie Wang. 2023. USSA: A unified table filling scheme for structured sentiment analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14340-14353, Toronto, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.809,
+                0.49,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Xiaopeng Zheng, Zhiyue Liu, Zizhen Zhang, Zhaoyang Wang, and Jiahai Wang. 2022. UECA-prompt: Universal prompt for emotion cause analysis. In Proceedings of the 29th International Conference on Computational Linguistics, pages 7031-7041, Gyeongju, Republic of Korea. International Committee on Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.49,
+                0.929,
+                0.511,
+                0.94
+            ],
+            "angle": 0,
+            "content": "47"
+        }
+    ]
+]

2024/End-to-End Emotion Semantic Parsing/4ffbaf63-28e0-4983-95b1-695face241d4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd26c96b85ae2e4a97070455d33bb67d10c865b390e0dada239088ffa1143647
+size 639217

2024/End-to-End Emotion Semantic Parsing/full.md

@@ -0,0 +1,300 @@
+# End-to-End Emotion Semantic Parsing
+
+Xiaotong Jiang, Zhongqing Wang\*, and Guodong Zhou
+
+$^{1}$ Natural Language Processing Lab, Soochow University, Suzhou, China xtjiang21@stu.suda.edu.cn, {wangzq, gdzhou}@suda.edu.cn
+
+# Abstract
+
+Emotion detection is the task of automatically associating one or more emotions with a text. The emotions are experienced, targeted, and caused by different semantic constituents. Therefore, it is necessary to incorporate these semantic constituents into the process of emotion detection. In this study, we propose a new task called emotion semantic parsing which aims to parse the emotion and semantic constituents into an abstract semantic tree structure. In particular, we design an end-to-end generation model to capture the relations between emotion and all the semantic constituents, and to generate them jointly. Furthermore, we employ a task decomposition strategy to capture the semantic relation among these constituents in a more cognitive and structural way. Experimental results demonstrate the importance of the proposed task, and indicate the proposed model gives superior performance compared to other models.
+
+# 1 Introduction
+
+The goal of emotion detection is to automatically detect or categorize the emotional states of human according to some inputs. Nowadays, emotion detection can be found in a broad range of applications, including but not limited to emotional support (Tu et al., 2022; Pavarini et al., 2023), human-computer interaction (Chowdary et al., 2021; Alrowais et al., 2023) and healthcare surveillance (Dhuheir et al., 2021; Fang et al., 2023). Henceforth, emotion detection has attracted increasing attention from both research community and industry in recent years (Hu et al., 2021; Zanwar et al., 2023).
+
+In literature, most studies focus on detecting emotion from the given text. For example, feature-based works (Balahur et al., 2011) and deep learning approaches, including RNNs (Majumder et al.,
+
+![](images/70c77cfa0599318b2e305e2e78b8b7ecbe29fb3eb098f449c76f1f6e4ef3ac57.jpg)  
+He clenched his fists as he watched the bully taunt his little brother.  
+Figure 1: Example of abstract semantic tree.
+
+2019), Transformers (Yu et al., 2018), and pretrained language models (Zheng et al., 2022; Zanwar et al., 2023) are used in emotion detection. Nevertheless, over the past few years, numerous studies (Xia and Ding, 2019; Kim and Klinger, 2018; Campagnano et al., 2022) have indicated that the emotions may be experienced, targeted, and/or caused by different semantic constituents, and can be linked to form abstract semantic structures.
+
+However, most existing works only take "cause" into consideration, neglecting the intricate relationships between emotion constituents, such as the strong relationship between trigger and emotion, as well as the similarity of sentiment and emotion labels. This limitation makes acquiring a nuanced understanding of the interplay between emotion and semantic constituents challenging, thereby hindering the model's capability to thoroughly comprehend emotional expression. By addressing these complexities, we believe it can serve as a foundational component for improving results in downstream real-world tasks by providing a more comprehensive understanding of emotional dynamics and their influence.
+
+To tackle the above problem, we propose a new
+
+task called emotion semantic parsing which aims to parse the emotion and semantic constituents into an abstract semantic tree structure. In this work, we consider five emotion-related semantic constituents into this task, including emotion, sentiment, trigger, holder, and target. As shown in Figure 1, emotion and sentiment refer to the specific emotion and sentiment labels of the sentence; trigger refers to a trigger word or expression that describes an emotion; and the constituents of holder and target correspond to the person or entities towards whom or from whom the emotion is directed, respectively.
+
+Moreover, we use the form of the abstract semantic tree structure to model the relationship between these constituents, we use the Trigger node to represent the whole abstract semantic tree, and we use the virtual Feelings and Entities nodes to compass other constituent nodes. The tree structure can be considered as a semantic representation in order to better represent the structure of the emotion-related constituents. As shown in Figure 1, the tree structure models a sentence using a rooted directed acyclic graph, highlighting its main elements (e.g. emotion, triggers) and semantic relations. It can thus potentially reveal a more comprehensive and complete semantic structure for extracting these constituents. Furthermore, since the emotion and semantic constituents are correlated in the form of the abstract semantic tree structure, we design an end-to-end generative model to predict them jointly. Specifically, our proposed model begins by decomposing the original task into several sub-tasks, as shown in Figure 2. Subsequently, the proposed model utilizes template-based prompt learning to extract each semantic constituent and semantic constituents pair separately for each sub-tasks. Finally, the proposed model leverages the acquired knowledge from sub-tasks to generate all semantic constituents in the main task.
+
+We conducted extensive experiments on the unified benchmark. Experimental results demonstrate the importance of the proposed task, and indicate the proposed model gives superior performance compared to other models. Further analysis shows the effect of the proposed abstract semantic tree structure and decomposition strategies.
+
+# 2 Related Works
+
+In the domain of emotion analysis, previous studies have mainly focused on text-based emotion detection. Early feature-based work often relied on
+
+traditional machine learning algorithms and manually crafted word-level emotion lexicons (Balahure et al., 2011). As deep neural networks progressed, emotion detection methods began utilizing various networks, such as RNNs (Majumder et al., 2019) and Transformers (Vaswani et al., 2017; Yu et al., 2018), to learn the emotion-related representation for detection. Recently, pre-trained language models (Devlin et al., 2019; Raffel et al., 2020) have dominated this area with its flexibility and rich internal knowledge (Alvarez-Gonzalez et al., 2021; Zheng et al., 2022).
+
+The emotions are not expressed individually, and can be experienced, targeted, and caused by different semantic constituents (Xia and Ding, 2019; Xiao et al., 2023; Campagnano et al., 2022). Most recent studies focus on emotion cause extraction task (Xia and Ding, 2019; Wang et al., 2023; Gu et al., 2023), with the goal of extracting the potential pairs of emotions and corresponding causes in the document and conversation scenario. Many end-to-end approaches have been proposed. These works can be broadly categorized into two main groups: some focus on multi-task learning (Ding et al., 2020a,b; Chen et al., 2022), while others transform the original task into a unified sequence labeling problem (Yuan et al., 2020; Chen et al., 2020). Recently, Xiao et al. (2023) developed a rule-based system to extract emotion causes on social media from constituency parsing trees. Wang et al. (2023) proposed emotion cause triplet extraction to extract emotion-cause-category triplets for fine-grained emotion cause analysis.
+
+Different from prior studies, we introduce a novel task, termed emotion semantic parsing. This task seeks to comprehensively examine the impact of various semantic constituents on emotional expressions. To achieve this, we develop an end-to-end generative model specifically designed to capture the intricate relationships between emotion and a wide range of semantic constituents.
+
+# 3 Preliminaries
+
+As shown in Figure 2, we introduce a novel task termed emotion semantic parsing. This task focuses on the generation of an abstract semantic tree structure that encapsulates both emotion and semantic constituents. To achieve this, we develop an end-to-end generative model specifically tailored for the generation of the abstract semantic tree structure. Our proposed model adopts a
+
+decomposition-based strategy, breaking down the complex task into several more manageable subtasks. Specifically, our proposed model begins by decomposing the original task into several subtasks. Finally, the proposed model leverages the acquired knowledge from sub-tasks to generate all semantic constituents in the main task.
+
+In this section, we introduce two important issues of this task: abstract semantic tree construction and the abstract semantic tree parsing model. The decomposition-based strategy for abstract semantic tree parsing will be discussed in the next section.
+
+# 3.1 Abstract Semantic Tree Construction
+
+Inspired by the widespread impact of tree structures in parsing tasks (Li et al., 2023; Groschwitz et al., 2023), we represent the relationships among multiple constituents in a tree structure. This approach allows us to capture the complex dynamics inherent in emotional expression. We first introduce the constituents in the abstract semantic tree, then we describe the construction process of the tree.
+
+# Definitions of Semantic Constituents
+
+There are five types of semantic constituents, each serving as a node in the abstract semantic tree. Their definitions are provided below:
+
+- Emotion is the specific emotion label represented from the original text. In this study, we use Plutchik (2001)'s taxonomy of emotion.  
+- Sentiment is the specific sentiment label represented from the original text, including "positive", "negative", and "neutral".  
+- Trigger refers to a trigger word or expression that describes an emotion, even if it is implicit.  
+- Holder refers to a person or entity that feels or experiences the emotion identified by the trigger.  
+- Target refers to a person or entity towards whom/which the emotion identified by the trigger is directed.
+
+# Process of Tree Construction
+
+As shown in Figure 1, the abstract semantic tree represents a sentence through a rooted directed acyclic graph utilizing the aforementioned semantic constituents. The construction process is outlined below:
+
+- We take the "Trigger" node as the root node, since it is the most important constituent in the emotion parsing.  
+- The virtual nodes "Feelings" and "Entities" are linked to Trigger node.  
+- The "Sentiment" and "Emotion" are linked to "Feelings" as terminal nodes, while the "Holder" and "Target" are linked to "Entities" node.
+
+As Figure 1 shows, the trigger "clenched his fists" implies the author's emotion. The holder and target are "he" and "the bully taunt his little brother", respectively. By leveraging these emotion-related semantic constituents, the model can accurately predict the "anger" emotion and "negative" polarity of the sentence.
+
+# 3.2 Basic Semantic Tree Parsing
+
+After we construct the abstract semantic tree, we then employ a generative model to parse the emotion and all the semantic constituents into a linearized tree structure. As shown in Figure 2, given an input text, the output can be represented as a linearized semantic tree.
+
+We adopt the pre-trained language model T5 (Chung et al., 2022) as our transformer-based encoder-decoder architecture. Therefore, by formulating parsing tasks as a text generation problem, we can tackle them in a unified sequence-to-sequence framework without a task-specific model design.
+
+The main challenge of emotion semantic parsing is the complex structure inherently in the emotional expression. Specifically, the complex structure will bring out two challenges: First, the large number of semantic constituents raises the complexity to the parsing task. Consequently, the precise extraction of every constituent becomes challenging yet indispensable. Second, effectively modeling these internal relationships also presents a significant challenge in the task.
+
+# 4 Decomposition-based Semantic Tree Parsing Model
+
+In this study, we propose a decomposed-based generative model to tackle the above challenges. The motivation is to alleviate the complexity of the original task by decomposing it into several sub-tasks. As shown in Figure 2, our proposed model can be
+
+![](images/c8db31deb594d79748377026b359f71ac241c28c0a504b872213340c37690e9a.jpg)  
+Figure 2: Overview of the proposed model.
+
+![](images/23ffac2dbeaf8a802478e94064ac3283fe2a48403710f281026c73d1148deb2e.jpg)  
+Figure 3: Examples of task decomposition strategies. Due to the space limit, prompt texts in this figure are shortened. Original prompt texts are presented in Section 4.1.
+
+separated into two stages: decomposition and recomposition. Our model firstly breaks down the original abstract semantic parsing task into several sub-tasks, and then designs individual instruction-based templates for each task. All sub-tasks are fine-tuned by a generative backbone at the same time. Based on the model parameters from the last re-composition stage, our model further finetunes its parameters again with the original task, which involves jointly parsing all semantic constituents.
+
+# 4.1 Task Decomposition Strategy
+
+To implement a generalized framework that is applicable to all tasks, we design templates for each individual task. The input template for each task comprises the task-specific instruction, and the original text to be parsed. The output template for each task is the sub-tree containing the constituents to be parsed. The following templates use the example in Figure 2 as the Original Text.
+
+Mono Parsing includes trigger, emotion, sentiment, holder, and target parsing, individually. As shown in Figure 3(a), these sub-tasks aim to capture each specific constituent:
+
+//taking emotion parsing task as an example
+
+INPUT Detect the emotion label in the emotional expression: + Original Text
+
+Output [Emotion (Anger)]
+
+Trigger-related Pair Parsing includes trigger-emotion, trigger-sentiment, trigger-holder, trigger-target, trigger-feelings, and trigger-entities. As shown in Figure 3(b), these sub-tasks aim to capture the relationship between the trigger and each specific constituent:
+
+//taking trigger-holder parsing task as an example
+
+INPUT Extract both the trigger and holder in the emotional expression: + Original Text
+
+Output [Trigger (clench his fists), [Entities, [Holder (he)]]]
+
+Entities Pair Parsing aims to capture the relationship between holder and target, and can gain insights into the interpersonal dynamics and relational aspects that influence emotional experiences. The example is shown in Figure 3(c), with the template of prompt provided below:
+
+INPUT Extract both holder and target in the emotional expression: + Original Text
+
+Output [Entities, [Holder (he), Target (the bully taunt his little brother) ]]
+
+Feelings Pair Parsing aims to capture the relationship between sentiment and emotion. The example is shown in Figure 3(d), with the template of prompt provided below:
+
+INPUT Extract both the emotion and sentiment labels in the emotional expression: + Original Text
+
+Output [Feelings, [Emotion (Anger), Sentiment (Negative) ]]
+
+# 4.2 Task Re-composition Strategy
+
+After gaining knowledge from sub-tasks, the model has learned the pattern of the template and the relationships from different emotion-related semantic constituents. Thus it can boost the final end-to-end emotion semantic parsing task, which is to predict the emotion label and outputs all emotion-related semantic constituents in the form of an abstract semantic tree. The example is shown in Figure 2, and the template of prompt is the below:
+
+INPUT Extract emotion trigger, holder, target, and emotion and sentiment labels in the emotional expression: + Original Text
+
+Output [Trigger (clenched his fists), Feelings, [Emotion (Anger), Sentiment (Negative)]], [Entities, [Holder (he), Target (the bully taunt his little brother) ]]
+
+# 4.3 Objective Functions and Training
+
+In this subsection, we show the objective functions and training process of the proposed model.
+
+The goal is to maximize the output linearized tree  $X_{T}$  probability given the sentence  $X$ . Therefore, we optimize the negative log-likelihood loss function:
+
+<table><tr><td></td><td>Amount</td></tr><tr><td>#Sample</td><td>7,234</td></tr><tr><td>Avg. Sent Len.</td><td>26.3</td></tr><tr><td>#Emotion Labels</td><td>9</td></tr><tr><td>#Sentiment Label</td><td>3</td></tr><tr><td>#Trigger</td><td>7,111</td></tr><tr><td>#Holder</td><td>5,633</td></tr><tr><td>#Target</td><td>6,127</td></tr></table>
+
+Table 1: Distribution of dataset.
+
+$$
+\mathcal {L} = - \frac {1}{| \tau |} \sum_ {(X, X _ {T}) \in \tau} \log p (X _ {T} | X; \theta) \tag {1}
+$$
+
+where  $\theta$  is the model parameters, and  $(X,X_{T})$  is a (sentence, tree) pair in training set  $\tau$ , then
+
+$$
+\begin{array}{l} \log p (X _ {T} | X; \theta) \\ = \sum_ {i = 1} ^ {n} \log p \left(x _ {T} ^ {i} \mid x _ {T} ^ {1}, x _ {T} ^ {2}, \dots x _ {T} ^ {i - 1}, X; \theta\right) \tag {2} \\ \end{array}
+$$
+
+where  $p(x_{T}^{i}|x_{T}^{1},x_{T}^{2},\dots x_{T}^{i - 1},X;\theta)$  is calculated by the decoder.
+
+# 5 Experimentation
+
+In this section, we introduce the datasets used for evaluation and the baseline methods employed for comparison. We then report the experimental results conducted from different perspectives, and analyze the effectiveness of the proposed model with different factors.
+
+# 5.1 Settings
+
+We evaluate the effectiveness of our proposed method on a benchmark proposed in Campagnano et al. (2022), which combines several datasets. We specifically choose three datasets (Mohammad et al., 2015; Kim and Klinger, 2018; Bostan et al., 2020) from this benchmark because they are the only ones that encompass annotations for all constituents. We enhance the quality of the original dataset by deleting duplicate or irrelevant examples (e.g., those with all trigger, holder, and target are "null"), while ensuring the trigger is not "null". The distribution of dataset is shown in Table 1. We choose  $90\%$  of the data as the training set, and the remaining  $10\%$  of the data as the testing set. Additionally,  $10\%$  of the training data is randomly set aside for validation during the training process.
+
+<table><tr><td>Methods</td><td>Emotion</td><td>Sentiment</td><td>Trigger</td><td>Holder</td><td>Target</td><td>Micro F1.</td></tr><tr><td>BERT-CRF</td><td>40.3</td><td>58.6</td><td>19.0</td><td>23.3</td><td>24.0</td><td>33.1</td></tr><tr><td>BERT-MRC</td><td>42.6</td><td>60.3</td><td>15.2</td><td>22.4</td><td>19.3</td><td>30.8</td></tr><tr><td>TaskFusion</td><td>45.7</td><td>62.3</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>USSA</td><td>41.3</td><td>-</td><td>17.3</td><td>32.6</td><td>14.3</td><td>-</td></tr><tr><td>T5</td><td>40.2</td><td>59.4</td><td>21.0</td><td>47.2</td><td>32.1</td><td>38.2</td></tr><tr><td>ChatGPT</td><td>22.3</td><td>30.7</td><td>10.4</td><td>21.3</td><td>9.6</td><td>18.9</td></tr><tr><td>LLaMA</td><td>42.4</td><td>57.6</td><td>22.4</td><td>43.5</td><td>31.3</td><td>39.4</td></tr><tr><td>Ours</td><td>46.3</td><td>63.6</td><td>30.8</td><td>55.0</td><td>37.4</td><td>46.7</td></tr></table>
+
+Table 2: Comparison with baselines.
+
+Our proposed method utilizes the generative backbone Flan-T5 (Chung et al., 2022) from HuggingFace's hub. All experiments are conducted on a single NVIDIA Tesla V100 GPU with 32GB memory. The model is trained for 10 epochs in each of the two stages, using the AdamW (Loshchilov and Hutter) optimizer, with a maximum sequence length of 256, a batch size of 16, and a learning rate of 3e-4. And the average results over 5 runs with different random seeds are reported. We adopt F1-score as the metric for all experiments, the micro F1 is employed to evaluate the overall performance.
+
+# 5.2 Main Results
+
+To better evaluate the effectiveness of our proposed model, we compare its performance against several strong baseline models, which are detailed below,
+
+- BERT-CRF (Campagnano et al., 2022) employs bidirectional LSTM layers over BERT representation, then uses CRF to generate token-level labels to determine the span boundary of each semantic constituent.  
+- BERT-MRC (Devlin et al., 2019) is used to predict the start and end indexes of all required semantic constituents.  
+- TaskFusion (Zanwar et al., 2023) is a multi-task framework with a pre-trained language model for emotion-related tasks. We adopt it for the proposed emotion semantic parsing task.  
+- USSA (Zhai et al., 2023) is a joint learning framework, which transfers aspect-based sentiment analysis into a 2D table-filling scheme. We adopt it for the proposed emotion semantic parsing task.
+
+- T5 (Chung et al., 2022) is a pre-trained language model which naturally suits the sequence-to-sequence task. It also can be considered as a joint learning framework, which generates all the semantic constituents jointly.  
+- ChatGPT (Ouyang et al., 2022) uses dialog-based instructional data in the supervised and RL-based meta-training stages. We use ChatGPT API to generate all required semantic constituents.  
+- LLaMA (Touvron et al., 2023) is a series of large language models trained on trillions of tokens. We use Alpaca-LoRA² to fine-tune LLaMA-7B to generate required semantic constituents.
+
+As demonstrated in Table 2, several key observations can be made: 1) Aligning with the current trend favoring generative frameworks, models that adopt a generative paradigm (i.e., T5, LLaMA, and our proposed model) generally outpace models based on a classification paradigm (i.e., BERT-CRF, BERT-MRC, Task Fusion, USSA) in terms of performance. 2) A standout observation is the substantial performance disparity between ChatGPT and all other methodologies. This discrepancy might be attributed to our limited usage of ChatGPT's API, without fine-tuning its internal parameters using insights from our training dataset.
+
+In contrast, our proposed model consistently outperforms prior studies  $(p < 0.05)$  in all settings. This underscores the effectiveness of our model, employing task decomposition strategies and the abstract semantic tree structure.
+
+![](images/f1305e83c2f21d4ba51501bf115fedd5ad29d5eb119ba222d6355bbaf6b65239.jpg)  
+Figure 4: Effect of tree structures.
+
+<table><tr><td>Tuples</td><td>e</td><td>s</td><td>t</td><td>h</td><td>a</td></tr><tr><td>(e)</td><td>42.5</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>(t)</td><td>-</td><td>-</td><td>28.4</td><td>-</td><td>-</td></tr><tr><td>(e,s)</td><td>43.5</td><td>65.1</td><td>-</td><td>-</td><td>-</td></tr><tr><td>(e,t)</td><td>45.2</td><td>-</td><td>29.1</td><td>-</td><td>-</td></tr><tr><td>(e,s,t)</td><td>45.7</td><td>67.5</td><td>30.7</td><td>-</td><td>-</td></tr><tr><td>(t,h)</td><td>-</td><td>-</td><td>29.3</td><td>43.1</td><td>-</td></tr><tr><td>(t,a)</td><td>-</td><td>-</td><td>29.1</td><td>-</td><td>32.4</td></tr><tr><td>(t,h,a)</td><td>-</td><td>-</td><td>29.3</td><td>50.0</td><td>35.5</td></tr><tr><td>(e,t,h,a)</td><td>45.9</td><td>-</td><td>29.4</td><td>52.4</td><td>35.3</td></tr><tr><td>Ours</td><td>46.3</td><td>63.6</td><td>30.8</td><td>55.0</td><td>37.4</td></tr></table>
+
+Table 3: Correlations among semantic constituents.  
+
+<table><tr><td>Methods</td><td>Micro F1.</td></tr><tr><td>Ours</td><td>46.7</td></tr><tr><td>- Mono</td><td>39.8</td></tr><tr><td>- TriggerPair</td><td>40.3</td></tr><tr><td>- EntitiesPair</td><td>41.8</td></tr><tr><td>- FeelingsPair</td><td>44.3</td></tr></table>
+
+Table 4: Impact of decomposition strategies.
+
+# 5.3 Influence of Semantic Constituents
+
+We then analyze the influence of the semantic constituents and the correlations among them. As shown in Table 3,  $e$  denotes the emotion label,  $s$  is the sentiment polarity,  $t, h$  and  $a$  are trigger, holder, and target, separately. In addition, the rows in the table mean different combinations of emotion-related semantic constituents. For example,  $(e, s)$  means that we used the proposed model to generate the emotion label and the sentiment polarity jointly.
+
+From the table, it is evident that there is a clear correlation between emotion and sentiment polarity, as the combined performance of  $(e,s)$  surpasses that of generating emotion  $e$  alone. This underscores the importance of considering sentiment polarity when analyzing and interpreting emotions. Additionally, our findings reveal that the majority of semantic constituents contribute significantly to
+
+emotion detection, especially trigger  $t$ . For example, the combined performance of  $(e, t)$  surpasses that of generating emotion  $e$  alone by a large margin. This underscores the importance of these semantic elements in understanding and contextualizing emotions within a given text.
+
+Furthermore, our results highlight a correlation between the holder  $h$  and target  $a$  with the trigger  $t$ , as they often refer to the person or entity experiencing or feeling the emotion triggered by a specific event or situation. This emphasizes the interconnected nature of these semantic constituents and their collective role in shaping emotional expressions.
+
+Notably, we observe that incorporating all semantic constituents leads to the highest performance, highlighting the effectiveness of our proposed model with semantic tree parsing. This underscores the utility and relevance of these semantic constituents in enhancing emotion detection accuracy.
+
+# 5.4 Effect of Tree Structures
+
+To assess the impact of the abstract semantic tree structure, we evaluated our model's performance using four distinct tree structures in Figure 4.
+
+In our initial assessment, we removed the entire tree structure, treating the semantic constituents as a linear sequence (a). This modification resulted in a substantial decrease in performance, from  $46.7\%$  to  $40.7\%$ , emphasizing the importance of the tree structure in capturing the intricate relationships among semantic constituents.
+
+Subsequently, we constructed two simplified tree structures, with either Trigger (b) or Emotion (c) serving as the root node. Interestingly, both of these structures yielded comparable performance results, suggesting that both of them are pivotal components in the emotion parsing task, and their roles are equally significant. Furthermore, we introduced two virtual nodes, labeled as Feelings and
+
+<table><tr><td>Input</td><td>Model</td><td>Output (Linearized Abstract Semantic Tree)</td></tr><tr><td rowspan="2">Jean Valjean was sincere. This sincerity from the very grief that it caused him, rendered inquiries useless, and conferred authority on all that that man had said. Here, for Marius, there was a strange reversal of situations.</td><td>T5</td><td>[Trigger, sincerity, [Feelings, [Emotion, trust], [Sentiment, positive]], [Entity, [Holder, him], [Target, null]]]</td></tr><tr><td>Ours</td><td>[Trigger, grief, [Feelings, [Emotion, sadness], [Sentiment, negative]], [Entity, [Holder, him], [Target, null]]]</td></tr><tr><td rowspan="2">&quot;Do you think me so very proud?&quot; he suddenly asked. &quot;I think you very strange.&quot;</td><td>T5</td><td>[Trigger, so very proud, [Feelings, [Emotion, joy], [Sentiment, positive]], [Entity, [Holder, you], [Target, me]]]</td></tr><tr><td>Ours</td><td>[Trigger, suddenly, [Feelings, [Emotion, surprise], [Sentiment, other]], [Entity, [Holder, he], [Target, null]]]</td></tr><tr><td rowspan="2">Last year I was wild to marry you, and you wouldn&#x27;t look at me.</td><td>T5</td><td>[Trigger, wouldn&#x27;t look at me, [Feelings, [Emotion, anger], [Sentiment, negative]], [Entity, [Holder, I], [Target, you]]]</td></tr><tr><td>Ours</td><td>[Trigger, wouldn&#x27;t look at me, [Feelings, [Emotion, disgust], [Sentiment, negative]], [Entity, [Holder, you], [Target, I]]]</td></tr></table>
+
+Figure 5: Examples of case study. The text in green means an exact match, while text in red indicates inaccuracies. "Null" means the extracted constituent is missing.
+
+Entities into the tree structure (d). The inclusion of these nodes led to a noticeable improvement in performance, highlighting the efficacy of our proposed tree structures and reinforcing the importance of considering semantic constituents within a structured framework for accurate emotion parsing.
+
+# 5.5 Impact of Decomposition Strategies
+
+We further investigated the influence of various decomposition strategies, and the results are presented in Table 4, where Mono denotes Mono Parsing, TriggerPair denotes Trigger-related Pair Parsing, EntitiesPair denotes Entities Pair Parsing, and FeelingsPair denotes Feelings Pair Parsing. All of these can be found in Section 4.1.
+
+Our analysis reveals that among all the strategies, mono parsing exerts the most substantial impact on model performance, aligning with our initial expectations. Mono parsing can be regarded as the foundational and indivisible task that serves as the basis for other higher-level tasks. Moreover, our findings indicate that all the decomposition strategies employed have a positive influence on the task, emphasizing the importance of breaking down the original task into smaller sub-tasks and the efficacy of reintegrating these sub-tasks to address the main task.
+
+# 5.6 Case Study
+
+To highlight our model's significance, we've chosen several challenging and illustrative examples shown in Figure 5.
+
+These challenging sentences share the same feature: they express more than one emotion expres
+
+sion, each associated with its own trigger, holder, and target. However, in each case, there is only one overall emotion. The presence of secondary emotions makes it difficult to fully grasp the overall emotional tone of the sentence.
+
+In the first two examples, the mis-detection of emotion expressions by T5 underscores the intricate and nuanced nature of the task, as well as the significance of each element in constructing a precise semantic representation. However, even when the correct emotion expression is identified, modeling the complex relationship between emotion and other semantic components remains challenging. As illustrated in the final example, although T5 correctly identifies the expression and its trigger, its inability to accurately detect the holder and target leads to an incorrect prediction of emotion. This highlights the interconnected nature of these elements and underscores their collective importance in achieving accurate emotion detection.
+
+In contrast, our proposed model, by leveraging task decomposition, demonstrates enhanced efficacy in navigating complexities present in difficult scenarios such as negotiation, emotion shifts, and multiple possible emotion expressions. This not only facilitates accurate detection of individual constituents but also enables a more nuanced understanding of their interrelationships, leading to improved overall performance.
+
+# 6 Conclusion
+
+In this study, we introduce a novel task called emotion semantic parsing, which focuses on analyzing emotions and semantic elements to construct an ab
+
+tract semantic tree structure. To achieve this, we develop an end-to-end generation model capable of capturing intricate relationships between emotions and various semantic constituents, generating them collaboratively. Furthermore, our proposed model employs template-based prompt learning, allowing for the separate extraction of each semantic constituent and pairs of semantic constituents for individual sub-tasks. Experimental results not only underscore the significance of the introduced task but also showcase the superior performance of our proposed model.
+
+# Limitations
+
+Our work focuses on employing a unified framework for emotion semantic parsing. However, one of the limitations is the computational complexity of the proposed model, whose training speed is slower than the classification-based models. Secondly, more tasks could be further explored, including nested emotion parsing, cross-domain and cross-lingual emotion semantic parsing. Finally, a large language model (LLM) can achieve superior performance in nearly all downstream tasks including our task, attributed to their ability to introduce rich knowledge learned during the pretraining stage for our proposed task. For open-source large language models like LLaMA (Touvron et al., 2023), limitations include high computational costs, high-quality data costs, as well as the need to address LLM-related issues such as safety, hallucinations, and ethical concerns, which remain to be resolved in future work.
+
+# Acknowledgements
+
+We would like to thank Shufan Fu for her invaluable advice and assistance with the experiments during this work. We also gratefully thank the anonymous reviewers for their insightful and valuable comments. This work is supported by the National Natural Science Foundation of China (No. 62076175, No.61976146), and Jiangsu Innovation Doctor Plan.
+
+# References
+
+Fadwa Alrowais, Noha Negm, Majdi Khalid, Nabil Almalki, Radwa Marzouk, Abdullah Mohamed, Mesfer Al Duhayyim, and Amani a Alneil. 2023. Modified earthworm optimization with deep learning assisted emotion recognition for human computer interface. IEEE Access, 11:35089-35096.
+
+Nurudin Alvarez-Gonzalez, Andreas Kaltenbrunner, and Vicenç Gómez. 2021. Uncovering the limits of text-based emotion detection. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2560-2583, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Alexandra Balahur, Jesús M. Hermida, Andres Montoyo, and Rafael Muñoz. 2011. Emotinet: A knowledge base for emotion detection in text built on the appraisal theories. In *Natural Language Processing and Information Systems*, pages 27-39, Berlin, Heidelberg. Springer Berlin Heidelberg.  
+Laura Ana Maria Bostan, Evgeny Kim, and Roman Klinger. 2020. GoodNewsEveryone: A corpus of news headlines annotated with emotions, semantic roles, and reader perception. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 1554-1566, Marseille, France. European Language Resources Association.  
+Cesare Campagnano, Simone Conia, and Roberto Navigli. 2022. SRL4E – Semantic Role Labeling for Emotions: A Unified Evaluation Framework. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4586–4601, Dublin, Ireland. Association for Computational Linguistics.  
+Hao Chen, Zepeng Zhai, Fangxiang Feng, Ruifan Li, and Xiaojie Wang. 2022. Enhanced Multi-Channel Graph Convolutional Network for Aspect Sentiment Triplet Extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2974–2985, Dublin, Ireland. Association for Computational Linguistics.  
+Xinhong Chen, Qing Li, and Jianping Wang. 2020. A unified sequence labeling model for emotion cause pair extraction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 208-218, Barcelona, Spain (Online). International Committee on Computational Linguistics.  
+Kalpana Chowdary, Tu Nguyen, and D. Hemanth. 2021. Deep learning-based facial emotion recognition for human-computer interaction applications. *Neural Computing and Applications*, pages 1-18.  
+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, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.
+
+Marwan Dhuheir, Abdullahif Albaseer, Emna Baccour, Aiman Erbad, Mohamed M. Abdallah, and Mounir Hamdi. 2021. Emotion recognition for healthcare surveillance systems using neural networks: A survey. CoRR, abs/2107.05989.  
+Zixiang Ding, Rui Xia, and Jianfei Yu. 2020a. ECPE-2D: Emotion-cause pair extraction based on joint two-dimensional representation, interaction and prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3161-3170, Online. Association for Computational Linguistics.  
+Zixiang Ding, Rui Xia, and Jianfei Yu. 2020b. End-to-end emotion-cause pair extraction based on sliding window multi-label learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3574-3583, Online. Association for Computational Linguistics.  
+Fang Fang, Yin Zhou, Shi Ying, and Zhijuan Li. 2023. A study of the ping an health app based on user reviews with sentiment analysis. International Journal of Environmental Research and Public Health, 20(2):1591.  
+Jonas Groschwitz, Shay Cohen, Lucia Donatelli, and Meaghan Fowlie. 2023. AMR parsing is far from solved: GrAPES, the granular AMR parsing evaluation suite. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10728-10752, Singapore. Association for Computational Linguistics.  
+Xiaojie Gu, Renze Lou, Lin Sun, and Shangxin Li. 2023. Page: A position-aware graph-based model for emotion cause entailment in conversation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5.  
+Dou Hu, Lingwei Wei, and Xiaoyong Huai. 2021. DialogueCRN: Contextual reasoning networks for emotion recognition in conversations. 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 7042-7052, Online. Association for Computational Linguistics.  
+Evgeny Kim and Roman Klinger. 2018. Who feels what and why? annotation of a literature corpus with semantic roles of emotions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1345-1359, Santa Fe, New Mexico, USA. Association for Computational Linguistics.  
+Pan Li, Ping Li, and Kai Zhang. 2023. Dual-channel span for aspect sentiment triplet extraction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 248-261, Singapore. Association for Computational Linguistics.
+
+Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Navonil Majumder, Soujanya Poria, Devamanyu Hazarika, Rada Mihalcea, Alexander Gelbukh, and Erik Cambria. 2019. Dialoguernn: An attentive rnn for emotion detection in conversations. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):6818-6825.  
+Saif M. Mohammad, Xiaodan Zhu, Svetlana Kiritchenko, and Joel Martin. 2015. Sentiment, emotion, purpose, and style in electoral tweets. Information Processing Management, 51(4):480-499.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744.  
+Gabriela Pavarini, Tessa Reardon, Anja Hollowell, Vanessa Bennett, Emma Lawrence, Peer Support Young People's Advisory Group Brooks-Hall Ellie Foster-Estwick Ashley Juma Damian Omari Lewis Peter Power Lucy Rogers Maia, Vanessa Pifold, and Ilina Singh. 2023. Online peer support training to promote adolescents' emotional support skills, mental health and agency during covid-19: Randomised controlled trial and qualitative evaluation. European Child & Adolescent Psychiatry, 32(6):1119-1130.  
+Robert Plutchik. 2001. The nature of emotions: Human emotions have deep evolutionary roots, a fact that may explain their complexity and provide tools for clinical practice. American Scientist, 89(4):344-350.  
+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. The Journal of Machine Learning Research, 21(1):5485-5551.  
+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. arXiv preprint arXiv:2302.13971.  
+Quan Tu, Yanran Li, Jianwei Cui, Bin Wang, Ji-Rong Wen, and Rui Yan. 2022. MISC: A mixed strategy-aware model integrating COMET for emotional support conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 308-319, Dublin, Ireland. Association for Computational Linguistics.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all
+
+you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.  
+Fanfan Wang, Jianfei Yu, and Rui Xia. 2023. Generative emotion cause triplet extraction in conversations with commonsense knowledge. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 3952-3963, Singapore. Association for Computational Linguistics.  
+Rui Xia and Zixiang Ding. 2019. Emotion-cause pair extraction: A new task to emotion analysis in texts. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1003–1012, Florence, Italy. Association for Computational Linguistics.  
+Debin Xiao, Rui Xia, and Jianfei Yu. 2023. Emotion cause extraction on social media without human annotation. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 1455-1468, Toronto, Canada. Association for Computational Linguistics.  
+Jianfei Yu, Luís Marujo, Jing Jiang, Pradeep Karuturi, and William Brendel. 2018. Improving multi-label emotion classification via sentiment classification with dual attention transfer network. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1097-1102, Brussels, Belgium. Association for Computational Linguistics.  
+Chaofa Yuan, Chuang Fan, Jianzhu Bao, and Ruifeng Xu. 2020. Emotion-cause pair extraction as sequence labeling based on a novel tagging scheme. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3568-3573, Online. Association for Computational Linguistics.  
+Sourabh Zanwar, Xiaofei Li, Daniel Wiechmann, Yu Qiao, and Elma Kerz. 2023. What to fuse and how to fuse: Exploring emotion and personality fusion strategies for explainable mental disorder detection. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 8926-8940, Toronto, Canada. Association for Computational Linguistics.  
+Zepeng Zhai, Hao Chen, Ruifan Li, and Xiaojie Wang. 2023. USSA: A unified table filling scheme for structured sentiment analysis. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14340-14353, Toronto, Canada. Association for Computational Linguistics.  
+Xiaopeng Zheng, Zhiyue Liu, Zizhen Zhang, Zhaoyang Wang, and Jiahai Wang. 2022. UECA-prompt: Universal prompt for emotion cause analysis. In Proceedings of the 29th International Conference on Computational Linguistics, pages 7031-7041, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.

2024/End-to-End Emotion Semantic Parsing/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84c735d1300bb7be60bd1f7c74952172f0b35dfbdf2600b118752cd28de6dc91
+size 388522

2024/Enhanced Language Model Truthfulness with Learnable Intervention and Uncertainty Expression/1e32c8cd-be45-4789-ad01-c4457a07e0bd_content_list.json

@@ -0,0 +1,2088 @@
+[
+    {
+        "type": "text",
+        "text": "Enhanced Language Model Truthfulness with Learnable Intervention and Uncertainty Expression",
+        "text_level": 1,
+        "bbox": [
+            203,
+            89,
+            794,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Farima Fatahi Bayat, Xin Liu, H. V. Jagadish, Lu Wang",
+        "bbox": [
+            260,
+            154,
+            746,
+            172
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Michigan, Ann Arbor",
+        "bbox": [
+            357,
+            173,
+            645,
+            187
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{farimaf, liuxincs, jag, wangluxy} @umich.edu",
+        "bbox": [
+            312,
+            187,
+            690,
+            204
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs) can generate long-form and coherent text, yet they often hallucinate facts, which undermines their reliability. To mitigate this issue, inference-time methods steer LLM representations toward the \"truthful directions\" previously learned for truth elicitation. However, applying these truthful directions with the same intensity fails to generalize across different query contexts. We propose LITO, a Learnable Intervention method for Truthfulness Optimization that automatically identifies the optimal intervention intensity tailored to each specific context. LITO explores a sequence of model generations based on increasing levels of intervention intensities. It selects the most accurate response or refuses to answer when the predictions are highly uncertain. Experiments on multiple LLMs and question-answering datasets demonstrate that LITO improves truthfulness while preserving task accuracy. The adaptive nature of LITO counters the limitations of one-size-fits-all intervention methods, maximizing truthfulness by reflecting the model's internal knowledge only when it is confident. Our code is available at https://github.com/launchnlp/LITO.",
+        "bbox": [
+            142,
+            282,
+            460,
+            652
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            665,
+            258,
+            682
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite their impressive performance across a wide range of natural language processing (NLP) tasks, large language models (LLMs) still generate hallucinated outputs that lack real-world basis, limiting their reliability in critical applications that require truthful responses. Many promising directions are explored to overcome this challenge, such as developing methods to ground LLMs in external knowledge and incorporate credibility indicators into model outputs (Gao et al., 2023; Fatahi Bayat et al., 2023). Another class of methods states the presence of a linear representation of \"truth\" in model activations (Marks and Tegmark, 2023; Li et al., 2023; Burns et al., 2022). These methods",
+        "bbox": [
+            112,
+            694,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/cb775361bd164722d4ed0d9cf76bb8e88a39c280a374e05603e159da46fc0297.jpg",
+        "image_caption": [
+            "Figure 1: Model responses using the inference-time intervention method with intensities increasing from 5 to 25. For different queries, the model achieves correct responses at varying intensity levels, indicated by green (correct) and red (incorrect) colors. Darkness of color represents the model's confidence in its response."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            250,
+            880,
+            423
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "train linear probes on top of LLM's internal activations to identify truthful directions in their representation space. In particular, Burns et al. (2022) claims that the representation of the truth, amongst a few other features, satisfies a logical consistency structure. They learn a linear projection of hidden states under the consistency-based objective and associate it with the truthful direction. However, Farquhar et al. (2023) later shows that (1) arbitrary features satisfy the logical consistency property, and (2) unsupervised methods detect superficial features that do not represent the truth. This indicates that an unsupervised search for truthful directions overly relies on surface features without additional mechanisms to reveal truthfulness.",
+        "bbox": [
+            507,
+            532,
+            884,
+            772
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To avoid capturing irrelevant features, Li et al. (2023) proposed a supervised probe learning that directly identifies the truthful directions based on correct and incorrect statements in the TruthfulQA dataset (Lin et al., 2022). This method, called inference-time intervention (ITI), trains supervised linear probes on the output of each attention head, treating the resulting probe weights as truthful directions. Additionally, a scaling coefficient is tuned",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12388",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2024, pages 12388-12400",
+        "bbox": [
+            220,
+            945,
+            778,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "August 11-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            284,
+            959,
+            712,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "to determine the intensity at which each direction should be added to its respective head output at inference time. However, amplifying the truthful directions with a fixed single intensity does not generalize across all contexts. Figure 1 demonstrates this by showing the Llama2-Chat-7B (Touvron et al., 2023b) model's performance on answering various queries from the Natural Questions dataset (Kwiatkowski et al., 2019) after applying the ITI technique with gradually increasing truthful direction intensities. Interestingly, the model arrives at a correct response within different intensity ranges for different questions. This suggests the optimal intervention magnitude is context-dependent, varying across questions based on factors such as their topic, complexity, ambiguity levels, etc. Moreover, the truthful directions may not capture all aspects of truthfulness. Therefore, adjusting the intensity alone cannot guarantee accurate responses. For instance, consider the question \"What flag is red and has a gold star?\" in Figure 1. Intervening with varying strengths of truthful directions does not result in a correct answer. In such cases, the model should express uncertainty to stay truthful.",
+        "bbox": [
+            115,
+            84,
+            490,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the limitations of one-size-fits-all intervention solutions by prior methods, we propose a Learnable Intervention method for Truthfulness Optimization, LITO. LITO identifies truthful direction intensities that suit different contexts, e.g., different questions. Given a sequence of model generations at multiple levels of intervention intensities, we develop a method to maximize truthfulness, which we define as selecting factual responses when the model is highly confident and refusing to respond otherwise. To achieve this, we collect model responses, including textual outputs, hidden representations, and confidence values, at increasing levels of intervention intensity. We then train an LSTM-based classifier to assess the accuracy of these responses based on the sequence of hidden states. During inference, the system selects the most accurate response if any is deemed accurate by the classifier; otherwise, it outputs \"I have no comment\" to express uncertainty and refuse to answer.",
+        "bbox": [
+            115,
+            478,
+            489,
+            814
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We measure the performance of LITO and other methods in balancing truthfulness and accuracy, introducing a novel evaluation metric called the Truthfulness-Accuracy (TA) score. This metric evaluates the trade-off between truthfulness and task-specific accuracy by measuring how effec",
+        "bbox": [
+            115,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tively different methods produce truthful outputs that appropriately acknowledge uncertainty while also achieving high accuracy on the target task.",
+        "bbox": [
+            512,
+            84,
+            880,
+            131
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "LITO is a learnable intervention methodology agnostic to the specific intervention method used, as long as the method can identify and apply truthful directions to the model's internal representations. In this paper, we instantiate LITO using the ITI method and extend its application to unsupervised truthful directions detected through representation engineering (RepE) (Zou et al., 2023). We conduct comprehensive experiments across four datasets and two categories of language models: Llama and GPT-2. The results show that LITO significantly improves truthfulness while preserving high task accuracy across different intervention methods. For example, using the ITI method, LITO boosts the  $TA$  score of Llama2-Chat-7B by 9.6 points on the Natural Questions dataset. Additionally, we evaluate LITO in a cross-domain setting to demonstrate its transferability across tasks.",
+        "bbox": [
+            512,
+            133,
+            882,
+            422
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Problem Statement and Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            512,
+            437,
+            867,
+            451
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We consider the problem of mitigating hallucinations in large language models through truthfulness enhancement. Our approach involves methods that steer the model's activation space towards factuality. This work focuses on open-domain question-answering, where models are tasked with responding to real-world queries. We utilize a short prompt that contains task-specific instructions, five demonstrations, and the target question. The model is expected to provide an accurate response to each question or express uncertainty by stating \"I have no comment\" when the answer is unknown.",
+        "bbox": [
+            512,
+            464,
+            882,
+            656
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Inference-time Intervention (ITI)",
+        "text_level": 1,
+        "bbox": [
+            512,
+            671,
+            815,
+            686
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To enhance truthfulness, we adopt a supervised truth elicitation technique called inference-time intervention (ITI) (Li et al., 2023). This method employs probing to detect the model's internal representations of truth. ITI trains one probe per attention head (in each layer) that linearly associates each attention head's output with a true/false label. To collect data for training each probe, ITI prompts the model with question-answer pairs where the answer is correct (1) or incorrect (0). Next, for each prompt, it collects the attention activation  $x_{l}^{h}$ , per layer  $l$  and per head  $h$ , of the answer's last token along with its binary labels  $y$ . A linear probe  $p(x_{l}^{h}) = \\text{sigmoid}(\\langle d_{l}^{h}, x_{l}^{h} \\rangle)$  is then trained",
+        "bbox": [
+            512,
+            694,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12389",
+        "bbox": [
+            478,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "on each head, and a sparse set of heads with the highest validation accuracy is selected. ITI shifts each selected head's activation  $x_{l}^{h}$  towards its corresponding probe weights  $d_{l}^{h}$  presented as a truthful direction. To achieve this, ITI adds truthful directions, amplified by a tuned coefficient  $\\alpha$  (the intervention intensity), to their corresponding head activation for each next token prediction as:",
+        "bbox": [
+            112,
+            84,
+            489,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx _ {l} ^ {h} = x _ {l} ^ {h} + \\alpha d _ {l} ^ {h} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            238,
+            225,
+            487,
+            244
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Learnable Intervention for Truthfulness Optimization",
+        "text_level": 1,
+        "bbox": [
+            112,
+            259,
+            475,
+            292
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Figure 1, applying a single intervention direction to selected head activation does not yield truthful results. To overcome this, we introduce a learnable intervention technique that gathers model outputs when the model is directed toward truthful directions at multiple intensity levels. Given an LM with  $L$  layers and  $H$  attention heads per layer, we use the ITI method to identify truthful directions (probe weights) as  $\\mathcal{D} = \\{d_l^h | l \\in L', h \\in H'\\}$ , where  $L' \\subseteq L$  and  $H' \\subseteq H$  represent the subsets of layers and heads selected by ITI. We then apply directions  $\\mathcal{D}$  at  $k$  different intensity levels (denoted by  $\\alpha$  values) for each input prompt, collect responses  $A = \\{a_1, a_2,.., a_k\\}$  at each intensity level, and output the most truthful answer, if available, or express uncertainty. The following section describes our intervention approach in detail.",
+        "bbox": [
+            112,
+            298,
+            489,
+            588
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Approach",
+        "text_level": 1,
+        "bbox": [
+            112,
+            602,
+            235,
+            619
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this work, we develop an intervention technique that dynamically adjusts to optimal intensity value, enhancing truthfulness based on prompt characteristics. Specifically, we increase the intensity  $(\\alpha)$  of the truthful directions  $\\mathcal{D}$ , learned by ITI, across  $k$  iterations, maintaining uniform intensity levels across all selected directions. By targeting a small subset of attention heads, ITI minimizes its impact on the LM's overall performance. Thus, small increases in intensity can yield similar outputs. To generate distinct responses from the intervened language model, we apply intervention intensities in increments of 5, i.e.  $\\alpha \\in 5,10,\\dots ,5k$ . Let  $LLM_{\\alpha}$  denote the LLM intervened with intensity  $\\alpha$  and  $A = \\{a_{1},a_{2},\\ldots ,a_{k}\\}$  denotes the collection of model responses, where  $a_{i} = LLM_{\\alpha = 5i}(x)$ . Each response  $a_{i}$  contains (1) the textual model generation  $y_{i}$  which consists of  $N$  tokens, (2) the",
+        "bbox": [
+            112,
+            629,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "model's last-layer hidden states  $h_i$  for generated tokens, and (3) the confidence score  $p(y_i|x)$ . Following Liu et al. (2024), we compute the confidence score as geometric mean across the sequence of token probabilities:",
+        "bbox": [
+            507,
+            84,
+            884,
+            164
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np \\left(y _ {i} \\mid x\\right) = \\sqrt [ N ]{\\prod_ {t = 1} ^ {N} p \\left(y _ {i , t} \\mid x , y _ {i , <   t}\\right)} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            564,
+            178,
+            882,
+            204
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We collect the three output components for each of the  $k$  interventions and pass all outputs to our adaptive intervention system, LITO. Our system then assesses the accuracy of each response and outputs the most truthful response if one exists.",
+        "bbox": [
+            507,
+            233,
+            882,
+            313
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Training",
+        "text_level": 1,
+        "bbox": [
+            507,
+            326,
+            626,
+            343
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We start with the hidden states  $\\mathcal{H} = \\{h_1, \\dots, h_k\\}$  corresponding to  $k$  different responses  $A = \\{a_1, a_2, \\dots, a_k\\}$ . Each  $h_i \\in \\mathcal{H}$  represents the last-layer hidden states for  $N$  tokens in  $a_i \\in A$ . We aggregate these hidden states across all generated tokens by taking their mean:",
+        "bbox": [
+            507,
+            349,
+            882,
+            445
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh _ {i} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} h _ {i, j} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            631,
+            458,
+            882,
+            502
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We target hidden states from the last layer as it provides an informative representation that captures the generation history and current state of the model. These aggregated hidden states are then fed into a 1-layer Long Short-Term Memory (LSTM). This allows the recurrent model to take a holistic view of response patterns, rather than examining them individually. The LSTM can thus learn how the responses change over increasing levels of intervention, identifying transitions and breaking points, drops in confidence or fluency, and potentially viable intervention zones. We showcase the effectiveness of selecting LSTM in Section 6.2. The LSTM outputs a hidden representation denoted as  $h_{r,i}$  for each response representation  $h_i$ :",
+        "bbox": [
+            507,
+            523,
+            884,
+            765
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh _ {r, 1}, \\dots , h _ {r, k} = L S T M \\left(h _ {1}, \\dots , h _ {k}\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            564,
+            777,
+            882,
+            794
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Finally, the hidden outputs of the LSTM are passed through a fully connected layer, followed by a sigmoid nonlinearity, to obtain the factuality probability  $p_w(h_{r,i})$  for each response, defined as  $p_w(h_{r,i}) = \\delta(\\langle w, h_{r,i} \\rangle)$ , where  $w$  represents the learned parameters and  $h_{r,i}$  denotes the LSTM's hidden representation of response  $a_i \\in A$ .",
+        "bbox": [
+            507,
+            809,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12390",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/9da6741399a2b813ebd1253b6962194bf6f52be7227bbd4e82ab2e9afd2e5e02.jpg",
+        "image_caption": [
+            "Figure 2: Overview of LITO method. Given the input prompt  $x$  with the question \"Bacterial cell walls are made rigid by the presence of?\", our method first collects model-generated responses after applying ITI-identified directions at 5 intensities  $LLM_{\\alpha = 5k}(x)$  (Section 3). Each response contains the textual response, the model's confidence of the generated response (shown by darkness of color), and the aggregated hidden representations  $h_i$ , computed as the average across hidden states of response tokens. LITO predicts the accuracy of each response given its hidden representations and selects the accurate response (labeled as 1) with the highest confidence or indicates uncertainty."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            152,
+            80,
+            847,
+            221
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Inference",
+        "text_level": 1,
+        "bbox": [
+            112,
+            344,
+            235,
+            357
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "At inference time, we input the aggregated hidden state  $h_i$  corresponding to each answer  $a_i \\in A$  through our trained system to determine its accuracy label  $\\mathbb{I}(\\delta(\\langle w, h_{r,i} \\rangle) > 0.5)$  where  $\\mathbb{I}$  is the indicator function. If all responses are predicted as nonfactual, the system conveys its uncertainty by outputting \"I have no comment\". Otherwise, LITO outputs the response with the highest confidence value  $p(y_i|x)$ . Formally:",
+        "bbox": [
+            112,
+            366,
+            489,
+            512
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ni ^ {*} = \\arg \\max  (p \\left(y _ {i} | x\\right)) \\quad s. t. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            526,
+            487,
+            544
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\delta (\\langle w, h _ {r, i} \\rangle) > 0. 5\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            206,
+            546,
+            344,
+            564
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Therefore, the final output is  $y_{i}$  or \"I have no comment\" in case all predictions are zero (inaccurate). Figure 2 shows an overview of how LITO operates.",
+        "bbox": [
+            112,
+            579,
+            489,
+            627
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Datasets and Training Labels for LITO",
+        "text_level": 1,
+        "bbox": [
+            112,
+            642,
+            485,
+            659
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            670,
+            226,
+            684
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this work, we focus on open-domain question-answering (QA). To train and evaluate LITO, we select QA tasks that vary in response length, targeting datasets with phrase-level and sentence-level responses. For phrase-level openQA datasets, we use NaturalQuestions (NQ) (Kwiatkowski et al., 2019), SciQ (Welbl et al., 2017), and TriviaQA (Joshi et al., 2017), all of which include short responses (e.g., named entities). For sentence-level responses, we choose TruthfulQA (Lin et al., 2022) where model responses are complete sentences. All datasets employed are in English.",
+        "bbox": [
+            112,
+            693,
+            489,
+            885
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We adopt an in-domain truthful direction identification approach. To this end, we use the val",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ization set of NaturalQuestions  $(\\mathrm{NQ})^{1}$  and TriviaQA $^{2}$  datasets that contain correct answers, and GPT-4-generated incorrect answers to serve as an adversarial data point. We randomly select 1K samples from each dataset for ITI probe training and save the rest of the samples (2.4K) for testing our method. SciQ is a multi-choice science question-answering dataset. We use its 1K validation set for ITI probe training and 1K test set for final evaluation. In addition to ITI training data, we randomly sample 3K instances from the train set of these phrase-level datasets to train LITO. Given that there is no official training set for TruthfulQA, we randomly select 408 instances from the original validation set to train the ITI method and find the optimal direction. We use the same set to train LITO and use the rest of the data for evaluation.",
+        "bbox": [
+            507,
+            343,
+            884,
+            618
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Training Label Construction",
+        "text_level": 1,
+        "bbox": [
+            507,
+            629,
+            781,
+            644
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "First, we use the ITI method to identify truthful directions that can later be integrated into the model's representations with amplified intensity. Next, we utilize the curated training data to prompt variants of the LM, as depicted in Figure 2, collecting the textual response, confidence score, and final-layer representations for each resulting generation. To label each response for accuracy, a DeBERTa-large model (He et al., 2021), fine-tuned on the MultiNLI (Williams et al., 2018) task, annotates phrase-level outputs. It classifies each textual response as correct if it can be entailed from the reference answer. To annotate model generations on TruthfulQA, we",
+        "bbox": [
+            507,
+            649,
+            882,
+            859
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "'https://huggingface.co/datasets/OamPatel/iti_nq_open_val",
+        "bbox": [
+            507,
+            866,
+            880,
+            892
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://huggingface.co/datasets/OamPatel/iti_trivia_qa_val",
+        "bbox": [
+            509,
+            892,
+            880,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12391",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ask GPT-4 to assess response accuracy based on semantic equivalence to the reference.3",
+        "bbox": [
+            112,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            127,
+            321,
+            143
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Prompts",
+        "text_level": 1,
+        "bbox": [
+            112,
+            152,
+            228,
+            167
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We adopt the same prompt format for evaluating TruthfulQA. Specifically, the \"QA prompt\" consists of an instruction, 5 question-answer pairs as in-context learning examples, and the target question the model should answer. We use the following instruction in all experiments: \"Interpret each question literally and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply \"I have no comment.\" unless you are completely certain of the answer.\"",
+        "bbox": [
+            112,
+            173,
+            487,
+            348
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To elicit concise responses for phrase-level QA, we include five in-context learning examples from each dataset. The full set of prompts used for evaluating the LMs on the different datasets is provided in Appendix B.",
+        "bbox": [
+            112,
+            350,
+            489,
+            430
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.2 Metrics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            439,
+            220,
+            455
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The output response of an intervention method can be factually accurate, inaccurate, or indicate uncertainty by outputting \"I have no comment\". We measure truthfulness as the portion of accurate or uncertain responses. However, the language model or intervention approach could default to \"I have no comment.\" to maximize their truthfulness. Therefore, we also measure accuracy by computing task-specific accuracy. Note that aggregation-based methods cannot surpass the accuracy of individual model generations they operate on. Finally, to measure the balance between truthfulness and accuracy, we propose the TA score which computes the geometric mean of truthfulness and accuracy:",
+        "bbox": [
+            112,
+            461,
+            489,
+            687
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT A = \\sqrt {\\text {T r u t h f u l n e s s} \\times \\text {A c c u r a c y}} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            695,
+            487,
+            714
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A higher TA score indicates that a method better balances the trade-off between producing accurate responses on the target task and generating truthful outputs that appropriately reflect uncertainty.",
+        "bbox": [
+            112,
+            722,
+            487,
+            785
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.3 Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            796,
+            218,
+            809
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We test intervention methods on two families of models: (1) Llama models: Vicuna-7B (Chiang et al., 2023), Llama2-chat-7B, and Llama2-chat-13B (Touvron et al., 2023a), and (2) GPT-2 models: GPT2-large and GPT2-XL (Radford et al., 2019).",
+        "bbox": [
+            112,
+            816,
+            487,
+            897
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.4 Baseline Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            694,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We apply the ITI method by setting  $k = 5$  and intervening in each model with five distinct intensity values,  $\\alpha \\in \\{5,10,15,20,25\\}$  which establish ITI baselines. The optimal  $k$  value selection is detailed in Section 6.2.3. We independently compute the baseline performance at each intensity. An oracle strategy is then used to select the intensity at which ITI performs best, and we report the results. We also employ three answer selection methods, evaluating the model outputs across five intensities to generate a truthful response, as described below.",
+        "bbox": [
+            507,
+            105,
+            882,
+            282
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Majority Voting: Given the model outputs  $A = \\{a_{1}, a_{2}, \\dots, a_{5}\\}$ , this method chooses the most repeated answer by taking a majority vote among textual responses. In case of a tie, the answer with the highest confidence is chosen as the final answer. For sentence-level responses where repetition rarely happens, all responses have one occurrence (tie) and thus the response with the maximum confidence is chosen.",
+        "bbox": [
+            507,
+            291,
+            882,
+            435
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Maximum Confidence: This method chooses the answer to which the model has assigned the maximum confidence.",
+        "bbox": [
+            507,
+            445,
+            882,
+            492
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Maximum Confidence  $>T$ : The difference between this method and Maximum Confidence is that it only selects an answer if its confidence is above a certain threshold. If such an answer does not exist, the final output is: \"I have no comment.\" This approach effectively filters out low-confidence answers, ensuring the reflection of uncertainty. We set  $T = 0.6$  as it shows the best average performance across datasets and LMs.<sup>4</sup>",
+        "bbox": [
+            507,
+            502,
+            882,
+            646
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Experimental Results and Further Analyses",
+        "text_level": 1,
+        "bbox": [
+            507,
+            659,
+            838,
+            693
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we first compare LITO against several counterparts and present our findings in Section 6.1. Then, in Section 6.2, we assess LITO's generalization to another truthful elicitation method, explore its performance in cross-domain settings, and investigate our design choices.",
+        "bbox": [
+            507,
+            702,
+            882,
+            799
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6.1 Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            810,
+            613,
+            824
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6.1.1 Results Compared to Original LM and ITI Baseline",
+        "text_level": 1,
+        "bbox": [
+            507,
+            829,
+            873,
+            860
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1 shows the performance of different methods in terms of their  $TA$  score on 4 datasets and 5 LMs.",
+        "bbox": [
+            507,
+            866,
+            882,
+            897
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "GPT-4 prompt for measuring correctness in Appendix A.",
+        "bbox": [
+            134,
+            903,
+            485,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "4Implementation details are provided in Appendix C.",
+        "bbox": [
+            529,
+            903,
+            853,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12392",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/ca4831d23363e811ad1ab152bb2d1ef58c2b4d56fa81efdc55958535cbf82efc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Task</td><td>Model</td><td>Original LM</td><td>ITI (best of 5)</td><td>Maj. Vote</td><td>Max Conf.</td><td>Max Conf. &gt;T</td><td>LITO</td></tr><tr><td rowspan=\"5\">NQ</td><td>GPT2-large</td><td>12.2</td><td>15.4</td><td>12.9</td><td>15.0</td><td>14.2</td><td>27.2</td></tr><tr><td>GPT2-XL</td><td>15.5</td><td>17.7</td><td>16.5</td><td>18.6</td><td>22.0</td><td>29.1</td></tr><tr><td>Llama2-Chat-7B</td><td>29.2</td><td>31.7</td><td>31.7</td><td>31.3</td><td>33.5</td><td>38.8</td></tr><tr><td>Llama2-Chat-13B</td><td>32.7</td><td>33.9</td><td>34.2</td><td>33.4</td><td>38.9</td><td>41.5</td></tr><tr><td>Vicuna-7B</td><td>30.0</td><td>30.3</td><td>29.3</td><td>30.0</td><td>35.0</td><td>36.2</td></tr><tr><td rowspan=\"5\">SciQ</td><td>GPT2-large</td><td>39.4</td><td>40.0</td><td>39.7</td><td>40</td><td>27.8</td><td>47.0</td></tr><tr><td>GPT2-XL</td><td>40.5</td><td>41.5</td><td>41.2</td><td>41.3</td><td>36.9</td><td>46.8</td></tr><tr><td>Llama2-Chat-7B</td><td>65.4</td><td>66.1</td><td>64.8</td><td>64.9</td><td>65.8</td><td>66.2</td></tr><tr><td>Llama2-Chat-13B</td><td>71.4</td><td>72.1</td><td>71.0</td><td>70.7</td><td>70.6</td><td>71.6</td></tr><tr><td>Vicuna-7B</td><td>61.7</td><td>61.4</td><td>57.5</td><td>60.2</td><td>62.7</td><td>61.9</td></tr><tr><td rowspan=\"5\">TriviaQA</td><td>GPT2-large</td><td>32.3</td><td>50.4</td><td>38.2</td><td>44.5</td><td>39.7</td><td>59.2</td></tr><tr><td>GPT2-XL</td><td>31.3</td><td>41.5</td><td>36.1</td><td>40.5</td><td>39.6</td><td>49.6</td></tr><tr><td>Llama2-Chat-7B</td><td>70.0</td><td>70.7</td><td>70.7</td><td>72.1</td><td>72.3</td><td>74.0</td></tr><tr><td>Llama2-Chat-13B</td><td>76.1</td><td>76.2</td><td>75.5</td><td>74.9</td><td>75.5</td><td>76.6</td></tr><tr><td>Vicuna-7B</td><td>67.7</td><td>68.3</td><td>68.9</td><td>71.2</td><td>72.5</td><td>72.0</td></tr><tr><td rowspan=\"5\">TruthfulQA</td><td>GPT2-large</td><td>15.9</td><td>15.9</td><td>15.5</td><td>13.9</td><td>18.6</td><td>24.0</td></tr><tr><td>GPT2-XL</td><td>20.7</td><td>26.8</td><td>23.1</td><td>25.1</td><td>24.5</td><td>30.7</td></tr><tr><td>Llama2-Chat-7B</td><td>45.1</td><td>49.9</td><td>48.1</td><td>49.4</td><td>49.4</td><td>49.6</td></tr><tr><td>Llama2-Chat-13B</td><td>52.4</td><td>52.8</td><td>54.2</td><td>53.1</td><td>53.1</td><td>54.3</td></tr><tr><td>Vicuna-7B</td><td>43.1</td><td>41.5</td><td>42.9</td><td>41.2</td><td>40.9</td><td>45.0</td></tr></table>",
+        "bbox": [
+            157,
+            80,
+            840,
+            361
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Results of LITO and baselines across 4 benchmarks and 5 LMs in terms of TA score (presented in Equation 6). \"ITI (best of 5)\" represents the peak ITI performance across 5 intervention intensities  $(\\alpha)$  selected by an oracle. The best and second-best TA score per model and per dataset is in bold. We highlight numbers where LITO improves over both the original LM and all baselines in blue; when LITO outperforms the original LM, it is colored in green. LITO effectively improves truthfulness while preserving high accuracy, surpassing other counterparts.",
+        "bbox": [
+            112,
+            370,
+            882,
+            458
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "It also highlights the peak ITI performance across 5 intensities selected by an oracle strategy, providing a comparison against other methods. As illustrated, LITO consistently improves over the original LM's performance across all datasets, showing the effectiveness of our approach. Particularly, LITO outperforms the original GPT-2 language models by a large margin, achieving an average TA score improvement of  $+14.4$  for GPT2-large and  $+12.0$  for GPT2-XL. This improvement is due to a notable increase in truthfulness while maintaining accuracy levels. ITI exhibits slightly superior performance when applied to Llama2 models on the phrase-level SciQ  $(+0.5)$  and TruthfulQA  $(+0.3)$  tasks.",
+        "bbox": [
+            112,
+            482,
+            489,
+            708
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.1.2 Results Compared to Aggregation-based Methods",
+        "text_level": 1,
+        "bbox": [
+            112,
+            721,
+            485,
+            751
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our approach demonstrates consistent performance gains over other aggregation-based methods, as shown in Table 1. The Maximum Confidence  $> T$  baseline shows higher performance improvement compared to its counterparts, outperforming LITO trained on Vicuna-7B hidden representations on NQ and TriviaQA benchmarks. Our investigation reveals that the Maximum Confidence  $> T$  baseline preserves its input accuracy levels while enhancing truthfulness. In contrast, LITO sacrifices some",
+        "bbox": [
+            112,
+            758,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "degree of accuracy to achieve higher truthfulness.",
+        "bbox": [
+            507,
+            482,
+            878,
+            498
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 illustrates LITO's truthfulness and accuracy scores compared to other baselines. It ranks within the top 2 for the highest truthfulness scores across all datasets and LMs. Additionally, LITO maintains accuracy within  $5\\%$  of the ITI method in 16 out of 20 experiments, illustrating its effective balance between truthfulness and accuracy. This makes LITO particularly valuable in settings where response truthfulness is crucial. As mentioned in Section 5.4, we set  $T = 0.6$  for Maximum Confidence  $> T$ . However, unlike Maximum Confidence, this baseline exhibits low accuracy levels with GPT-2 models, suggesting that smaller models may suffer from poor calibration, as indicated by (Kadavath et al., 2022). Another key observation from Figure 3 is that the Majority Vote closely follows the ITI average, demonstrating its inability to significantly improve upon input responses.",
+        "bbox": [
+            505,
+            499,
+            885,
+            790
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.2 Further Analyses",
+        "text_level": 1,
+        "bbox": [
+            507,
+            799,
+            692,
+            814
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.2.1 LITO Generalizes across Intervention Techniques",
+        "text_level": 1,
+        "bbox": [
+            507,
+            819,
+            870,
+            852
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this work, we primarily instantiate and evaluate our proposed methodology using the inference-time intervention (ITI) method as the underlying truthful intervention technique. However, LITO is",
+        "bbox": [
+            507,
+            854,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12393",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/abd84367bef8cabf85753f021bf9ce360c6457ba98e5183fd759c1d5bfe48d6f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            84,
+            371,
+            139
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d69a627cae8dc4874fe21e554366af73b84f2ed400cd38e6799584cdac2745f2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            374,
+            85,
+            621,
+            139
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f0b08593bb5b51536877616793f54048e81356655e54c1075d10ed83179286e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            625,
+            85,
+            875,
+            139
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/197bf4454fa1fe6855a4b51b6c647e189409088b14cedd41113ab06f5c8e2eb1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            145,
+            371,
+            203
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/650cfb216bb28bdafbfa12a88335f1cf38316385b927b5406b0d8c63d1e13dc8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            374,
+            148,
+            623,
+            203
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8e950bf2433aa5090302e13403614bd18b23e3cb5dd2bf3aea961cb4a4e2d233.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            626,
+            149,
+            875,
+            203
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/694213214df5a7e889ebb1716f75d9a4f9f68a27503e390991beea7bc9d672b4.jpg",
+        "image_caption": [
+            "Figure 3: Truthfulness and accuracy scores per dataset on five LMs. ITI represents the average ITI performance across 5 intensities to demonstrate how closely the Majority Vote follows this baseline. In all experiments, LITO is ranked within the top 2 in terms of truthfulness while preserving accuracy, leading to its superior  $TA$  performance."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            248,
+            209,
+            497,
+            263
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/aaae9c00671c48a0da5777fe84832f552e5cc8f81044da4ee3053974890e554f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            209,
+            751,
+            263
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/84aa82a49d772af4255029016188f1fbf08cc9c1af430573b610a8ba2e8adce2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Task</td><td>Model</td><td>Original LM</td><td>RepE (best of 5)</td><td>Maj. Vote</td><td>Max Conf.</td><td>Max Conf. &gt;T</td><td>LITO</td></tr><tr><td rowspan=\"2\">NQ</td><td>GPT2-XL</td><td>15.5</td><td>15.6</td><td>15.6</td><td>15.5</td><td>16.9</td><td>27.5</td></tr><tr><td>Llama2-Chat-7B</td><td>29.2</td><td>29.4</td><td>29.4</td><td>28.4</td><td>29.5</td><td>35.5</td></tr><tr><td rowspan=\"2\">SciQ</td><td>GPT2-XL</td><td>40.5</td><td>40.8</td><td>40.5</td><td>40.2</td><td>35.2</td><td>46.5</td></tr><tr><td>Llama2-Chat-7B</td><td>65.4</td><td>66.1</td><td>65.9</td><td>65.9</td><td>65.1</td><td>65.5</td></tr><tr><td rowspan=\"2\">TriviaQA</td><td>GPT2-XL</td><td>31.3</td><td>31.6</td><td>31.4</td><td>30.0</td><td>31.4</td><td>40.2</td></tr><tr><td>Llama2-Chat-7B</td><td>70.0</td><td>70.6</td><td>70.1</td><td>70.3</td><td>71.8</td><td>71.92</td></tr><tr><td rowspan=\"2\">TruthfulQA</td><td>GPT2-XL</td><td>19.1</td><td>19.1</td><td>18.8</td><td>18.6</td><td>18.2</td><td>26.3</td></tr><tr><td>Llama2-Chat-7B</td><td>42.6</td><td>43.4</td><td>38.5</td><td>38.4</td><td>38.4</td><td>39.1</td></tr></table>",
+        "bbox": [
+            156,
+            332,
+            842,
+            468
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: The results of RepE-based LITO instantiation and various baselines across four benchmarks are detailed in terms of the TA score. RepE (best of 5) indicates the performance of RepE at its optimal intensity level determined by an oracle. Details on colors and fonts are provided in Table 1.",
+        "bbox": [
+            112,
+            479,
+            882,
+            521
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "compatible with any technique that enhances the language model truthfulness by identifying truthful directions in the model's representation space. To showcase this generalizability, we further instantiate LITO using the Representation Engineering (RepE) intervention method (Zou et al., 2023) as the underlying truthful intervention technique. A discussion on the selection of RepE is provided in Appendix D.",
+        "bbox": [
+            112,
+            546,
+            487,
+            690
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "RepE identifies truthful directions in a language model's representations by leveraging truthful/untruthful counterfactual pairs. It collects the layerwise representations produced by the model when prompted with these pairs. Then, it computes the representation differences between each counterfactual pair across layers. RepE employs unsupervised techniques, such as Principal Component Analysis (PCA), to isolate a single truthful direction per layer from these representation differences. During inference, these directions are multiplied by an intervention coefficient and added to their corresponding layer outputs in the language model. Given these characteristics, we can readily apply",
+        "bbox": [
+            112,
+            694,
+            489,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "LITO on top of the RepE intervention technique and evaluate the potential performance gains. To do so, we use the same setup mentioned in Section 5 for truthful direction identification LITO. We intervene by steering the representation of the layer that exhibits the highest truthfulness accuracy based on the learned directions. Following the experimental setting in Appendix C, we collect language model generations at 5 different intervention intensity values 1, 2, 3, 4, 5. Note that since the intervention is layer-level (as opposed to ITI which was at the activation head-level), we chose smaller intensity values compared to ITI to prevent nonsensical generations.",
+        "bbox": [
+            507,
+            546,
+            882,
+            772
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Our experiments instantiating LITO with the RepE intervention technique are shown in Table 2 across the 4 tasks and 2 different LM categories studied in this paper. Compared to best-performing setup of RepE across the 5 interventions, our results show an average performance gain of  $+9$ ,  $+2.5$ ,  $+4.9$ , and  $+1.5$  TA points on the NQ, SciQ, TriviaQA, and TruthfulQA datasets respectively, corroborating our findings on the",
+        "bbox": [
+            507,
+            774,
+            884,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12394",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ITI-based LITO improvements.",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            371,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.2.2 LITO Learns Task-agnostic Notions of Truth",
+        "text_level": 1,
+        "bbox": [
+            112,
+            116,
+            480,
+            147
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We developed an intervention method that adapts to different intensity levels and contexts. Next, we evaluate how well this method, trained on one task, can generalize to others. We trained and tested the ITI-based LITO instantiation on every dataset pair, highlighting the resulting transfer capabilities for the 5 large language models in Figure 4. Our method generally shows effective transferability, with LITO demonstrating strong performance when trained on one task and tested on others. Specifically, while LITO trained on the TruthfulQA dataset exhibits limited transferability, as noted by Li et al. (2023), it achieves near in-domain performance levels when trained on TriviaQA. This effectiveness could stem from TriviaQA's broad general knowledge base, which applies to more specialized domains (e.g. SciQ). Notably, for the NQ task, LITO even exceeds its in-domain performance on GPT2-Large. Overall, our adaptive intervention method consistently maintains  $TA$  scores across most out-of-domain scenarios, indicating minimal performance degradation.",
+        "bbox": [
+            115,
+            155,
+            489,
+            510
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.2.3 Design Choices",
+        "text_level": 1,
+        "bbox": [
+            112,
+            526,
+            294,
+            541
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we validate our design choices for using an LSTM as the core LITO component and determine the optimal number of interventions  $(k)$ .",
+        "bbox": [
+            112,
+            549,
+            489,
+            596
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "LSTM vs. MLP: We justify using a recurrent neural network to analyze patterns in sequences of interventions, rather than examining them individually. For this purpose, using the same experimental setup, we substitute our LSTM model with a fully connected layer followed by a ReLU nonlinearity. We measure the binary classification performance in terms of accuracy and F1 score across all 4 question-answering tasks, using the Llama2-Chat-7B model as the base LM. We denote the method that replaces the LSTM with a linear layer as LITO  $MLP$ . The results, presented in Table 3, show that the LSTM model substantially outperforms the baseline on phrase-level QA tasks. The F1 score on the TruthfuQA task shows a noticeable performance drop. However, TruthfulQA is a challenging task with limited training data, requiring the LSTM to have more examples to learn complex sequential patterns effectively.",
+        "bbox": [
+            115,
+            613,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/603f13b10e9cf92b850dd7a07a45dfd7d0036d6e032c0f18e127722d53d69dbe.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Task</td><td colspan=\"2\">LITO</td><td colspan=\"2\">LITOMLP</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>NQ</td><td>71.9</td><td>50.4</td><td>69.6</td><td>46.2</td></tr><tr><td>SciQ</td><td>66.5</td><td>71.9</td><td>65.1</td><td>71.8</td></tr><tr><td>TriviaQA</td><td>71.4</td><td>79.5</td><td>70.2</td><td>77.6</td></tr><tr><td>TruthfulQA</td><td>75.2</td><td>55.7</td><td>74.4</td><td>59.8</td></tr></table>",
+        "bbox": [
+            532,
+            80,
+            858,
+            178
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Comparing the classification accuracy and F1 score of LITO with LITO  $_{MLP}$ , with superior results highlighted in green. LITO outperforms LITO  $_{MLP}$  in short-form QA across both metrics.",
+        "bbox": [
+            507,
+            187,
+            882,
+            244
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "$k$  Tuning: Throughout our experiments, we set the number of responses  $k$  to 5. To investigate the impact of this choice, we evaluate our method's performance using different values of  $k$  across the validation sets<sup>5</sup> of all 4 datasets, with the Llama2-Chat-7B model. As shown in Figure 5, for the NQ dataset,  $k = 5$  achieves a significant performance improvement over  $k = 4$ , and increasing  $k$  beyond 5 yields negligible benefits for SciQ. Although  $k = 6$  provides marginal benefits for other datasets, the additional computational cost does not justify a higher  $k$ . Thus, we choose  $k = 5$  as the optimal balance between performance and efficiency, suitable for most applications. However, Figure 5 suggests that in scenarios where truthfulness is paramount and efficiency constraints are relaxed, collecting more language model generations can be beneficial.",
+        "bbox": [
+            507,
+            271,
+            884,
+            560
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            574,
+            665,
+            589
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7.1 Hallucination in LLMs",
+        "text_level": 1,
+        "bbox": [
+            507,
+            602,
+            737,
+            615
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Addressing hallucinations in LLMs can be classified into two categories: training methods and inference-time methods. Training methods include introducing faithfulness-based loss functions (Yoon et al., 2022; Qiu et al., 2023), and supervised finetuning to utilize the external knowledge graph (Ji et al., 2023; Fatahi Bayat et al., 2023), aiming to strengthen the factualness of LLMs. Despite their effectiveness, training or fine-tuning LLMs becomes impractical due to their parameter size. On the contrary, inference-time methods do not require tuning the LLM itself. For example, representative methods include prompt-based methods with model feedback (Si et al., 2023; Mündler et al., 2023; Lei et al., 2023). These methods prompt the model to provide feedback for its previous output",
+        "bbox": [
+            507,
+            624,
+            884,
+            882
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "We divide the training data into five parts for 5-fold cross-validation.",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12395",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3e9ad5b5415386cd72e765d743b7ed55eef1701dbd5fbb33dc058c0d65ba8b6c.jpg",
+        "image_caption": [
+            "Figure 4: Transfer results of ITI-based LITO, measured by  $TA$  score on 5 LMs. The y-axis corresponds to the training dataset, and the x-axis corresponds to the test dataset. Each cell represents the out-of-domain performance (ood) relative to its corresponding in-domain performance (id), computed as  $100 \\times (ood - id) / id$ . Across most datasets, LITO exhibits strong transfer capabilities (relative to in-domain setup)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            80,
+            880,
+            195
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/f5019481b320632ca78bf80893e67307de70b942cf2aaa4c357c819e1fcd2955.jpg",
+        "image_caption": [
+            "Figure 5: Performance of LITO on validation set of 4 datasets using different  $k$  values. As illustrated,  $k = 5$  provides a sweet spot between performance and computational overhead."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            285,
+            485,
+            448
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "and then instruct the model to predict better generation given the feedback. Moreover, researchers explored incorporating retrieved contexts to enhance factuality (Varshney et al., 2023; Cao et al., 2023). However, such methods require access to valid sources of knowledge which is challenging and causes delayed response. Recently, some methods propose to modify the hidden states or the prediction distribution during decoding, such as CAD (Shi et al., 2023) and DoLa (Chuang et al., 2023). The effect of such methods on other model characteristics is yet underexplored. To address these limitations, LITO collects model responses across varying intervention intensities and employs a learnable mechanism to output the most truthful response without adversely affecting other desirable model characteristics.",
+        "bbox": [
+            112,
+            542,
+            489,
+            815
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7.2 LLMs Intervention",
+        "text_level": 1,
+        "bbox": [
+            112,
+            831,
+            314,
+            845
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The intervention of LLMs involves generating directional vectors of truthfulness and integrating these vectors into the forward pass of LLMs, guiding them toward factual generations. For example,",
+        "bbox": [
+            112,
+            854,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "in ITI (Li et al., 2023), linear probing is employed to identify attention heads with distinct activation distributions for true and false statements, allowing intervention on these heads to guide the model toward generating truthful outputs. RepE (Zou et al., 2023) detects the per-layer truthful directions by prompting the language model with pairs of instructions with contrastive meanings and integrating these directions into each layer during decoding. Similarly, ActAdd (Turner et al., 2023) exploits activation differences from pairs of counterfactual prompts to control the generation process.",
+        "bbox": [
+            507,
+            288,
+            884,
+            481
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Yet, these methods apply directions amplified with a uniform intensity across all instances, causing insufficient or excessive intervention in many cases. Moreover, prior methods lack a principled refusal mechanism to selectively abstain from generating outputs when the model has low confidence. This shortcoming poses risks that can limit the use of these techniques in high-stakes applications and severely harm end-users. Instead, LITO selects the most accurate response among multiple generations with varying intervention intensities or refuses to respond if no such answer is found.",
+        "bbox": [
+            507,
+            482,
+            882,
+            675
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            686,
+            640,
+            700
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we introduced LITO, a novel learnable intervention method that adjusts the intensity of truthfulness directions based on the question context. Our approach explores generations at multiple intensities, selecting the most accurate output or expressing uncertainty when necessary. Comprehensive experiments demonstrate consistent improvements in balancing truthfulness and task accuracy over original LMs and existing inference-time techniques. An exciting future direction is developing mechanisms to dynamically determine the number and range of intensities to explore based on prompt characteristics.",
+        "bbox": [
+            507,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12396",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            278,
+            101
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "This work is supported in part by the National Science Foundation through grant IIS-2046016. We thank ARR reviewers for their helpful comments.",
+        "bbox": [
+            112,
+            111,
+            489,
+            160
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            173,
+            220,
+            187
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "This work has limitations that could be addressed in future research. First, we focused on short phrase-level and sentence-level responses, but the performance of our approach on longer text generation remains unknown. Second, LITO's accuracy relies on the quality of the truthful directions identified by the underlying inference-time intervention method. Enhancing the truthfulness signals provided as input could further improve results. Moreover, while adaptive intervention selection mitigates excessive intensities, it still requires multiple passes through the LLM, increasing the response time. Compared to the studied intervention techniques (ITI and RepE), LITO required  $k$  times more inference time as it queries the language model  $k$  times. Finally, the interpretability of LITO's selections could be deeply investigated. Visualizing the model's learned notions of uncertainty over intervention intensities may uncover interesting patterns. Nonetheless, this work demonstrates the promise of applying adaptive intervention to prevent model hallucinations.",
+        "bbox": [
+            115,
+            200,
+            490,
+            554
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            569,
+            265,
+            583
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "This work proposes a method aimed at improving factuality and reducing inaccurate responses in large language model question answering. As open-domain question-answering systems become more prevalent, enhancing truthfulness and reliability is crucial for safe deployment. However, our approach still relies on the capabilities of the underlying model architecture. Future work must continue addressing the potential harms of large generative models related to issues like bias, toxicity, and misinformation. Additionally, adaptive intervention techniques introduce potential downsides if misused. While eliciting factuality reveals the knowledge housed in models, bad actors could exploit similar methods to intentionally expose or induce false beliefs. Future research should explore protections against adversarial attacks alongside efforts to curb hallucination.",
+        "bbox": [
+            115,
+            596,
+            489,
+            884
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "On the positive side, reliable question-answering could broadly advance access to knowledge and",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "combat the viral spread of misinformation. But care must also be taken with any technology able to generate convincing false text. We believe methods that promote truthful AI while mitigating potential harms align with ethical priorities for language technology. This work marks an initial step, but ongoing progress requires interdisciplinary collaboration on the societal impacts of synthetic media.",
+        "bbox": [
+            507,
+            84,
+            885,
+            212
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            509,
+            240,
+            608,
+            255
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2022. Discovering latent knowledge in language models without supervision.",
+            "Heijing Cao, Zhenwei An, Jiazhan Feng, Kun Xu, Liwei Chen, and Dongyan Zhao. 2023. A step closer to comprehensive answers: Constrained multi-stage question decomposition with large language models. CoRR, abs/2311.07491.",
+            "Zhongzhi Chen, Xingwu Sun, Xianfeng Jiao, Fengzong Lian, Zhanhui Kang, Di Wang, and Chengzhong Xu. 2024. Truth forest: Toward multi-scale truthfulness in large language models through intervention without tuning. Proceedings of the AAAI Conference on Artificial Intelligence, 38(19):20967-20974.",
+            "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.",
+            "Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James R. Glass, and Pengcheng He. 2023. Dola: Decoding by contrasting layers improves factuality in large language models. CoRR, abs/2309.03883.",
+            "Sebastian Farquhar, Vikrant Varma, Zachary Kenton, Johannes Gasteiger, Vladimir Mikulik, and Rohin Shah. 2023. Challenges with unsupervised llm knowledge discovery.",
+            "Farima Fatahi Bayat, Kun Qian, Benjamin Han, Yisi Sang, Anton Belyy, Samira Khorshidi, Fei Wu, Ihab Ilyas, and Yunyao Li. 2023. FLEEK: Factual error detection and correction with evidence retrieved from external knowledge. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 124-130, Singapore. Association for Computational Linguistics.",
+            "Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. 2023. RARR: Researching and revising what language models say, using language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1:"
+        ],
+        "bbox": [
+            509,
+            263,
+            885,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12397",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Long Papers), pages 16477-16508, Toronto, Canada. Association for Computational Linguistics.",
+            "Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pretraining with gradient-disentangled embedding sharing. CoRR, abs/2111.09543.",
+            "Ziwei Ji, Zihan Liu, Nayeon Lee, Tiezheng Yu, Bryan Wilie, Min Zeng, and Pascale Fung. 2023. RHO: Reducing hallucination in open-domain dialogues with knowledge grounding. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4504-4522, Toronto, Canada. Association for Computational Linguistics.",
+            "Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601-1611, Vancouver, Canada. Association for Computational Linguistics.",
+            "Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav Fort, Deep Ganguli, Danny Hernandez, Josh Jacobson, Jackson Kernion, Shauna Kravec, Liane Lovitt, Kamal Ndousse, Catherine Olsson, Sam Ringer, Dario Amodei, Tom Brown, Jack Clark, Nicholas Joseph, Ben Mann, Sam McCandlish, Chris Olah, and Jared Kaplan. 2022. Language models (mostly) know what they know.",
+            "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural Questions: A Benchmark for Question Answering Research. Transactions of the Association for Computational Linguistics, 7:453-466.",
+            "Deren Lei, Yaxi Li, Mengya Hu, Mingyu Wang, Vincent Yun, Emily Ching, and Eslam Kamal. 2023. Chain of natural language inference for reducing large language model ungrounded hallucinations. CoRR, abs/2310.03951.",
+            "Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Inference-time intervention: Eliciting truthful answers from a language model.",
+            "Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xin Liu, Muhammad Khalifa, and Lu Wang. 2024. Litcab: Lightweight language model calibration over short- and long-form responses. In International Conference on Learning Representations (ICLR).",
+            "Samuel Marks and Max Tegmark. 2023. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets.",
+            "Niels Mündler, Jingxuan He, Slobodan Jenko, and Martin T. Vechev. 2023. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. CoRR, abs/2305.15852.",
+            "Yifu Qiu, Yftah Ziser, Anna Korhonen, Edoardo Ponti, and Shay Cohen. 2023. Detecting and mitigating hallucinations in multilingual summarisation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8914-8932, Singapore. Association for Computational Linguistics.",
+            "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.",
+            "Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen-tau Yih. 2023. Trusting your evidence: Hallucinate less with context-aware decoding. CoRR, abs/2305.14739.",
+            "Chenglei Si, Zhe Gan, Zhengyuan Yang, Shuohang Wang, Jianfeng Wang, Jordan Boyd-Graber, and Lijuan Wang. 2023. Prompting gpt-3 to be reliable. In International Conference on Learning Representations (ICLR).",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinez, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023a. Llama 2: Open foundation and finetuned chat models. CoRR, abs/2307.09288.",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "12398",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open foundation and fine-tuned chat models.",
+        "bbox": [
+            131,
+            85,
+            490,
+            347
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Alexander Matt Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Monte MacDiarmid. 2023. Activation addition: Steering language models without optimization. CoRR, abs/2308.10248.",
+        "bbox": [
+            115,
+            357,
+            489,
+            412
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Neeraj Varshney, Wenlin Yao, Hongming Zhang, Jianshu Chen, and Dong Yu. 2023. A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation. CoRR, abs/2307.03987.",
+        "bbox": [
+            115,
+            423,
+            489,
+            489
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. ArXiv, abs/1707.06209.",
+        "bbox": [
+            115,
+            500,
+            489,
+            541
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            552,
+            489,
+            670
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Sunjae Yoon, Eunseop Yoon, Hee Suk Yoon, Junyeong Kim, and Chang Yoo. 2022. Information-theoretic text hallucination reduction for video-grounded dialogue. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4182-4193, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            682,
+            489,
+            776
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J. Zico Kolter, and Dan Hendrycks. 2023. Representation engineering: A top-down approach to AI transparency. CoRR, abs/2310.01405.",
+        "bbox": [
+            115,
+            785,
+            490,
+            904
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A GPT4 Prompt for Evaluating Correctness",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            801,
+            115
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Are the following two answers to my question \"[Question]\" semantically equivalent? (Answer \"Yes\" or \"No\" first, and then explain your answer.)",
+        "bbox": [
+            519,
+            129,
+            867,
+            178
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. [Reference]",
+            "2. [Model Response]"
+        ],
+        "bbox": [
+            522,
+            179,
+            668,
+            202
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Prompts of Four Tasks",
+        "text_level": 1,
+        "bbox": [
+            509,
+            216,
+            744,
+            233
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In this section, we show the prompts of four datasets that we used throughout the experiments.",
+        "bbox": [
+            507,
+            241,
+            882,
+            273
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.1 NQ",
+        "text_level": 1,
+        "bbox": [
+            509,
+            282,
+            588,
+            298
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Interpret each question literally, and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply \"I have no comment\" unless you are completely certain of the answer.",
+        "bbox": [
+            521,
+            307,
+            853,
+            380
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Q: who plays the voice of joy in inside out?  \nA: Amy Poehler",
+            "Q: which two countries were the leading powers during the cold war? A: Soviet Union and United States",
+            "Q: when was the last province added to canada? A: April 1, 1999",
+            "Q: where is the saracens rugby union team based? A: London, England",
+            "Q: the lymphatic tissues at the back of the mouth are called? A: Tonsils"
+        ],
+        "bbox": [
+            521,
+            391,
+            875,
+            580
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.2 TruthfulQA",
+        "text_level": 1,
+        "bbox": [
+            510,
+            590,
+            653,
+            606
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Interpret each question literally, and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply \"I have no comment\" unless you are completely certain of the answer.",
+            "Q: What is human life expectancy in the United States?",
+            "A: Human life expectancy in the United States is 78 years.",
+            "Q: Who was president of the United States in 1955?  \nA: Dwight D. Eisenhower was president of the United States in 1955.",
+            "Q: What is the square root of banana?",
+            "A: I have no comment.",
+            "Q: How does a telescope work?",
+            "A: Telescopes use lenses or mirrors to focus light and make objects appear closer.",
+            "Q: Where were the 1992 Olympics held?",
+            "A: The 1992 Olympics were held in Barcelona, Spain."
+        ],
+        "bbox": [
+            521,
+            615,
+            889,
+            913
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "12399",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.3 SciQ",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            203,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Interpret each question literally, and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply \"I have no comment\" unless you are completely certain of the answer.",
+        "bbox": [
+            124,
+            115,
+            455,
+            186
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Q: What is the least dangerous radioactive decay?  \nA: alpha decay",
+            "Q: What is the number of electrons equal to in every electrically neutral atom? A: protons",
+            "Q: What happens to old oceanic crust at convergent boundaries?  \nA: destroyed",
+            "Q: Sexually reproducing organisms alternate between which stages?  \nA: haploid and diploid",
+            "Q: Motors are the most common application of magnetic force on current-carrying wires. motors have loops of wire in this?  \nA: magnetic field"
+        ],
+        "bbox": [
+            127,
+            198,
+            487,
+            423
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B.4 TriviaQA",
+        "text_level": 1,
+        "bbox": [
+            115,
+            442,
+            238,
+            458
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Interpret each question literally, and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply \"I have no comment\" unless you are completely certain of the answer.",
+        "bbox": [
+            126,
+            473,
+            455,
+            544
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Q: New York Yankees legend Lou Gehrig was known by what nickname?  \nA: Iron horse",
+            "Q: Which was the first European country to abolish capital punishment?  \nA: Norway",
+            "Q: A bone is joined to a muscle by what tough band of inelastic fibrous tissue?  \nA: Tendon",
+            "Q: In what language was the New Testament originally written?  \nA: In Greek",
+            "Q: Psychologist William Moulton Marston, inventor of the polygraph, or lie detector, also created a famous comic book heroine,. Who was she? A: Wonder Woman"
+        ],
+        "bbox": [
+            127,
+            556,
+            487,
+            791
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            810,
+            356,
+            828
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Using the ITI method, we intervene with 5 different intensity values  $\\alpha \\in \\{5,10,15,20,25\\}$  across all models and datasets. Our choice of small, equally-spaced intensity values allows us to collect distinct response changes from the LLMs while ensuring",
+        "bbox": [
+            112,
+            838,
+            489,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "minimal invasiveness.",
+        "bbox": [
+            509,
+            84,
+            673,
+            98
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "To collect model outputs for training our method, we conducted 100 experiments each taking 1-2 hours using one NVIDIA A40 GPU. To train our system, as shown in Figure 2, we set the size of the LSTM's output hidden state to  $1/8$ th of its input size, which is the LLM's hidden state dimension. For example, the LSTM output size of our trained method on Vicuna-7B is 512. We employ upsampling and downsampling techniques to balance the combinations of correct and incorrect responses fed to LITO. To mitigate overfitting, we apply L2 weight decay regularization with a coefficient of 0.001 to LITO's parameters during training. In total, we train our method 20 times, once per LLM model and dataset pair. We employ early stopping with a patience of 10 epochs and a maximum of 50 epochs, saving the model checkpoint with the highest F1 score. Each training run utilizes 64 CPU cores and completes within 3-5 minutes depending on the size of the training dataset and the dimension of LLM's hidden states.",
+        "bbox": [
+            507,
+            99,
+            884,
+            437
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D Choice of RepE Intervention Technique",
+        "text_level": 1,
+        "bbox": [
+            507,
+            449,
+            882,
+            467
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in Section 6.2.1, LITO is compatible with any intervention technique that improves language models' truthfulness by identifying truthful directions in the model's representation space. Among such techniques are Truth Forest (Chen et al., 2024), Contrast-Consistent Search (CCS) (Burns et al., 2022), and Representation Engineering (RepE) (Zou et al., 2023). We did not choose to evaluate Truth Forest since it is highly motivated by and has a similar basis as the Iterative Truth Intervention (ITI) method already covered in our paper. Additionally, we investigated the CCS method which trains unsupervised probes by incorporating the consistency structure of truthfulness into its loss function. Specifically, we explored how CCS can be used for intervention during text generation, as this method was originally only explored for the classification task. However, on the SciQ dataset, the CCS directions failed to effectively separate truthful from untruthful spaces (near chance accuracy), despite supervised logistic regression achieving around  $80\\%$  accuracy. Given this failure to correctly distinguish the two spaces, we did not rely on the CCS method for intervention. Lastly, we evaluated the RepE method, as its truthful directions showed promising accuracy on all four datasets studied in this paper.",
+        "bbox": [
+            507,
+            475,
+            884,
+            910
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "12400",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2024/Enhanced Language Model Truthfulness with Learnable Intervention and Uncertainty Expression/1e32c8cd-be45-4789-ad01-c4457a07e0bd_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:812b1098bfdaa65ba88bd7925dfe51496d9dff2d2aebaf7376125dc72f09225d
+size 870909

2024/Enhanced Language Model Truthfulness with Learnable Intervention and Uncertainty Expression/full.md

@@ -0,0 +1,383 @@
+# Enhanced Language Model Truthfulness with Learnable Intervention and Uncertainty Expression
+
+Farima Fatahi Bayat, Xin Liu, H. V. Jagadish, Lu Wang
+
+University of Michigan, Ann Arbor
+
+{farimaf, liuxincs, jag, wangluxy} @umich.edu
+
+# Abstract
+
+Large language models (LLMs) can generate long-form and coherent text, yet they often hallucinate facts, which undermines their reliability. To mitigate this issue, inference-time methods steer LLM representations toward the "truthful directions" previously learned for truth elicitation. However, applying these truthful directions with the same intensity fails to generalize across different query contexts. We propose LITO, a Learnable Intervention method for Truthfulness Optimization that automatically identifies the optimal intervention intensity tailored to each specific context. LITO explores a sequence of model generations based on increasing levels of intervention intensities. It selects the most accurate response or refuses to answer when the predictions are highly uncertain. Experiments on multiple LLMs and question-answering datasets demonstrate that LITO improves truthfulness while preserving task accuracy. The adaptive nature of LITO counters the limitations of one-size-fits-all intervention methods, maximizing truthfulness by reflecting the model's internal knowledge only when it is confident. Our code is available at https://github.com/launchnlp/LITO.
+
+# 1 Introduction
+
+Despite their impressive performance across a wide range of natural language processing (NLP) tasks, large language models (LLMs) still generate hallucinated outputs that lack real-world basis, limiting their reliability in critical applications that require truthful responses. Many promising directions are explored to overcome this challenge, such as developing methods to ground LLMs in external knowledge and incorporate credibility indicators into model outputs (Gao et al., 2023; Fatahi Bayat et al., 2023). Another class of methods states the presence of a linear representation of "truth" in model activations (Marks and Tegmark, 2023; Li et al., 2023; Burns et al., 2022). These methods
+
+![](images/cb775361bd164722d4ed0d9cf76bb8e88a39c280a374e05603e159da46fc0297.jpg)  
+Figure 1: Model responses using the inference-time intervention method with intensities increasing from 5 to 25. For different queries, the model achieves correct responses at varying intensity levels, indicated by green (correct) and red (incorrect) colors. Darkness of color represents the model's confidence in its response.
+
+train linear probes on top of LLM's internal activations to identify truthful directions in their representation space. In particular, Burns et al. (2022) claims that the representation of the truth, amongst a few other features, satisfies a logical consistency structure. They learn a linear projection of hidden states under the consistency-based objective and associate it with the truthful direction. However, Farquhar et al. (2023) later shows that (1) arbitrary features satisfy the logical consistency property, and (2) unsupervised methods detect superficial features that do not represent the truth. This indicates that an unsupervised search for truthful directions overly relies on surface features without additional mechanisms to reveal truthfulness.
+
+To avoid capturing irrelevant features, Li et al. (2023) proposed a supervised probe learning that directly identifies the truthful directions based on correct and incorrect statements in the TruthfulQA dataset (Lin et al., 2022). This method, called inference-time intervention (ITI), trains supervised linear probes on the output of each attention head, treating the resulting probe weights as truthful directions. Additionally, a scaling coefficient is tuned
+
+to determine the intensity at which each direction should be added to its respective head output at inference time. However, amplifying the truthful directions with a fixed single intensity does not generalize across all contexts. Figure 1 demonstrates this by showing the Llama2-Chat-7B (Touvron et al., 2023b) model's performance on answering various queries from the Natural Questions dataset (Kwiatkowski et al., 2019) after applying the ITI technique with gradually increasing truthful direction intensities. Interestingly, the model arrives at a correct response within different intensity ranges for different questions. This suggests the optimal intervention magnitude is context-dependent, varying across questions based on factors such as their topic, complexity, ambiguity levels, etc. Moreover, the truthful directions may not capture all aspects of truthfulness. Therefore, adjusting the intensity alone cannot guarantee accurate responses. For instance, consider the question "What flag is red and has a gold star?" in Figure 1. Intervening with varying strengths of truthful directions does not result in a correct answer. In such cases, the model should express uncertainty to stay truthful.
+
+To address the limitations of one-size-fits-all intervention solutions by prior methods, we propose a Learnable Intervention method for Truthfulness Optimization, LITO. LITO identifies truthful direction intensities that suit different contexts, e.g., different questions. Given a sequence of model generations at multiple levels of intervention intensities, we develop a method to maximize truthfulness, which we define as selecting factual responses when the model is highly confident and refusing to respond otherwise. To achieve this, we collect model responses, including textual outputs, hidden representations, and confidence values, at increasing levels of intervention intensity. We then train an LSTM-based classifier to assess the accuracy of these responses based on the sequence of hidden states. During inference, the system selects the most accurate response if any is deemed accurate by the classifier; otherwise, it outputs "I have no comment" to express uncertainty and refuse to answer.
+
+We measure the performance of LITO and other methods in balancing truthfulness and accuracy, introducing a novel evaluation metric called the Truthfulness-Accuracy (TA) score. This metric evaluates the trade-off between truthfulness and task-specific accuracy by measuring how effec
+
+tively different methods produce truthful outputs that appropriately acknowledge uncertainty while also achieving high accuracy on the target task.
+
+LITO is a learnable intervention methodology agnostic to the specific intervention method used, as long as the method can identify and apply truthful directions to the model's internal representations. In this paper, we instantiate LITO using the ITI method and extend its application to unsupervised truthful directions detected through representation engineering (RepE) (Zou et al., 2023). We conduct comprehensive experiments across four datasets and two categories of language models: Llama and GPT-2. The results show that LITO significantly improves truthfulness while preserving high task accuracy across different intervention methods. For example, using the ITI method, LITO boosts the  $TA$  score of Llama2-Chat-7B by 9.6 points on the Natural Questions dataset. Additionally, we evaluate LITO in a cross-domain setting to demonstrate its transferability across tasks.
+
+# 2 Problem Statement and Preliminaries
+
+We consider the problem of mitigating hallucinations in large language models through truthfulness enhancement. Our approach involves methods that steer the model's activation space towards factuality. This work focuses on open-domain question-answering, where models are tasked with responding to real-world queries. We utilize a short prompt that contains task-specific instructions, five demonstrations, and the target question. The model is expected to provide an accurate response to each question or express uncertainty by stating "I have no comment" when the answer is unknown.
+
+# 2.1 Inference-time Intervention (ITI)
+
+To enhance truthfulness, we adopt a supervised truth elicitation technique called inference-time intervention (ITI) (Li et al., 2023). This method employs probing to detect the model's internal representations of truth. ITI trains one probe per attention head (in each layer) that linearly associates each attention head's output with a true/false label. To collect data for training each probe, ITI prompts the model with question-answer pairs where the answer is correct (1) or incorrect (0). Next, for each prompt, it collects the attention activation  $x_{l}^{h}$ , per layer  $l$  and per head  $h$ , of the answer's last token along with its binary labels  $y$ . A linear probe  $p(x_{l}^{h}) = \text{sigmoid}(\langle d_{l}^{h}, x_{l}^{h} \rangle)$  is then trained
+
+on each head, and a sparse set of heads with the highest validation accuracy is selected. ITI shifts each selected head's activation  $x_{l}^{h}$  towards its corresponding probe weights  $d_{l}^{h}$  presented as a truthful direction. To achieve this, ITI adds truthful directions, amplified by a tuned coefficient  $\alpha$  (the intervention intensity), to their corresponding head activation for each next token prediction as:
+
+$$
+x _ {l} ^ {h} = x _ {l} ^ {h} + \alpha d _ {l} ^ {h} \tag {1}
+$$
+
+# 2.2 Learnable Intervention for Truthfulness Optimization
+
+As illustrated in Figure 1, applying a single intervention direction to selected head activation does not yield truthful results. To overcome this, we introduce a learnable intervention technique that gathers model outputs when the model is directed toward truthful directions at multiple intensity levels. Given an LM with  $L$  layers and  $H$  attention heads per layer, we use the ITI method to identify truthful directions (probe weights) as  $\mathcal{D} = \{d_l^h | l \in L', h \in H'\}$ , where  $L' \subseteq L$  and  $H' \subseteq H$  represent the subsets of layers and heads selected by ITI. We then apply directions  $\mathcal{D}$  at  $k$  different intensity levels (denoted by  $\alpha$  values) for each input prompt, collect responses  $A = \{a_1, a_2,.., a_k\}$  at each intensity level, and output the most truthful answer, if available, or express uncertainty. The following section describes our intervention approach in detail.
+
+# 3 Approach
+
+In this work, we develop an intervention technique that dynamically adjusts to optimal intensity value, enhancing truthfulness based on prompt characteristics. Specifically, we increase the intensity  $(\alpha)$  of the truthful directions  $\mathcal{D}$ , learned by ITI, across  $k$  iterations, maintaining uniform intensity levels across all selected directions. By targeting a small subset of attention heads, ITI minimizes its impact on the LM's overall performance. Thus, small increases in intensity can yield similar outputs. To generate distinct responses from the intervened language model, we apply intervention intensities in increments of 5, i.e.  $\alpha \in 5,10,\dots ,5k$ . Let  $LLM_{\alpha}$  denote the LLM intervened with intensity  $\alpha$  and  $A = \{a_{1},a_{2},\ldots ,a_{k}\}$  denotes the collection of model responses, where  $a_{i} = LLM_{\alpha = 5i}(x)$ . Each response  $a_{i}$  contains (1) the textual model generation  $y_{i}$  which consists of  $N$  tokens, (2) the
+
+model's last-layer hidden states  $h_i$  for generated tokens, and (3) the confidence score  $p(y_i|x)$ . Following Liu et al. (2024), we compute the confidence score as geometric mean across the sequence of token probabilities:
+
+$$
+p \left(y _ {i} \mid x\right) = \sqrt [ N ]{\prod_ {t = 1} ^ {N} p \left(y _ {i , t} \mid x , y _ {i , <   t}\right)} \tag {2}
+$$
+
+We collect the three output components for each of the  $k$  interventions and pass all outputs to our adaptive intervention system, LITO. Our system then assesses the accuracy of each response and outputs the most truthful response if one exists.
+
+# 3.1 Training
+
+We start with the hidden states  $\mathcal{H} = \{h_1, \dots, h_k\}$  corresponding to  $k$  different responses  $A = \{a_1, a_2, \dots, a_k\}$ . Each  $h_i \in \mathcal{H}$  represents the last-layer hidden states for  $N$  tokens in  $a_i \in A$ . We aggregate these hidden states across all generated tokens by taking their mean:
+
+$$
+h _ {i} = \frac {1}{N} \sum_ {j = 1} ^ {N} h _ {i, j} \tag {3}
+$$
+
+We target hidden states from the last layer as it provides an informative representation that captures the generation history and current state of the model. These aggregated hidden states are then fed into a 1-layer Long Short-Term Memory (LSTM). This allows the recurrent model to take a holistic view of response patterns, rather than examining them individually. The LSTM can thus learn how the responses change over increasing levels of intervention, identifying transitions and breaking points, drops in confidence or fluency, and potentially viable intervention zones. We showcase the effectiveness of selecting LSTM in Section 6.2. The LSTM outputs a hidden representation denoted as  $h_{r,i}$  for each response representation  $h_i$ :
+
+$$
+h _ {r, 1}, \dots , h _ {r, k} = L S T M \left(h _ {1}, \dots , h _ {k}\right) \tag {4}
+$$
+
+Finally, the hidden outputs of the LSTM are passed through a fully connected layer, followed by a sigmoid nonlinearity, to obtain the factuality probability  $p_w(h_{r,i})$  for each response, defined as  $p_w(h_{r,i}) = \delta(\langle w, h_{r,i} \rangle)$ , where  $w$  represents the learned parameters and  $h_{r,i}$  denotes the LSTM's hidden representation of response  $a_i \in A$ .
+
+![](images/9da6741399a2b813ebd1253b6962194bf6f52be7227bbd4e82ab2e9afd2e5e02.jpg)  
+Figure 2: Overview of LITO method. Given the input prompt  $x$  with the question "Bacterial cell walls are made rigid by the presence of?", our method first collects model-generated responses after applying ITI-identified directions at 5 intensities  $LLM_{\alpha = 5k}(x)$  (Section 3). Each response contains the textual response, the model's confidence of the generated response (shown by darkness of color), and the aggregated hidden representations  $h_i$ , computed as the average across hidden states of response tokens. LITO predicts the accuracy of each response given its hidden representations and selects the accurate response (labeled as 1) with the highest confidence or indicates uncertainty.
+
+# 3.2 Inference
+
+At inference time, we input the aggregated hidden state  $h_i$  corresponding to each answer  $a_i \in A$  through our trained system to determine its accuracy label  $\mathbb{I}(\delta(\langle w, h_{r,i} \rangle) > 0.5)$  where  $\mathbb{I}$  is the indicator function. If all responses are predicted as nonfactual, the system conveys its uncertainty by outputting "I have no comment". Otherwise, LITO outputs the response with the highest confidence value  $p(y_i|x)$ . Formally:
+
+$$
+i ^ {*} = \arg \max  (p \left(y _ {i} | x\right)) \quad s. t. \tag {5}
+$$
+
+$$
+\delta (\langle w, h _ {r, i} \rangle) > 0. 5
+$$
+
+Therefore, the final output is  $y_{i}$  or "I have no comment" in case all predictions are zero (inaccurate). Figure 2 shows an overview of how LITO operates.
+
+# 4 Datasets and Training Labels for LITO
+
+# 4.1 Datasets
+
+In this work, we focus on open-domain question-answering (QA). To train and evaluate LITO, we select QA tasks that vary in response length, targeting datasets with phrase-level and sentence-level responses. For phrase-level openQA datasets, we use NaturalQuestions (NQ) (Kwiatkowski et al., 2019), SciQ (Welbl et al., 2017), and TriviaQA (Joshi et al., 2017), all of which include short responses (e.g., named entities). For sentence-level responses, we choose TruthfulQA (Lin et al., 2022) where model responses are complete sentences. All datasets employed are in English.
+
+We adopt an in-domain truthful direction identification approach. To this end, we use the val
+
+ization set of NaturalQuestions  $(\mathrm{NQ})^{1}$  and TriviaQA $^{2}$  datasets that contain correct answers, and GPT-4-generated incorrect answers to serve as an adversarial data point. We randomly select 1K samples from each dataset for ITI probe training and save the rest of the samples (2.4K) for testing our method. SciQ is a multi-choice science question-answering dataset. We use its 1K validation set for ITI probe training and 1K test set for final evaluation. In addition to ITI training data, we randomly sample 3K instances from the train set of these phrase-level datasets to train LITO. Given that there is no official training set for TruthfulQA, we randomly select 408 instances from the original validation set to train the ITI method and find the optimal direction. We use the same set to train LITO and use the rest of the data for evaluation.
+
+# 4.2 Training Label Construction
+
+First, we use the ITI method to identify truthful directions that can later be integrated into the model's representations with amplified intensity. Next, we utilize the curated training data to prompt variants of the LM, as depicted in Figure 2, collecting the textual response, confidence score, and final-layer representations for each resulting generation. To label each response for accuracy, a DeBERTa-large model (He et al., 2021), fine-tuned on the MultiNLI (Williams et al., 2018) task, annotates phrase-level outputs. It classifies each textual response as correct if it can be entailed from the reference answer. To annotate model generations on TruthfulQA, we
+
+ask GPT-4 to assess response accuracy based on semantic equivalence to the reference.3
+
+# 5 Experimental Setup
+
+# 5.1 Prompts
+
+We adopt the same prompt format for evaluating TruthfulQA. Specifically, the "QA prompt" consists of an instruction, 5 question-answer pairs as in-context learning examples, and the target question the model should answer. We use the following instruction in all experiments: "Interpret each question literally and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply "I have no comment." unless you are completely certain of the answer."
+
+To elicit concise responses for phrase-level QA, we include five in-context learning examples from each dataset. The full set of prompts used for evaluating the LMs on the different datasets is provided in Appendix B.
+
+# 5.2 Metrics
+
+The output response of an intervention method can be factually accurate, inaccurate, or indicate uncertainty by outputting "I have no comment". We measure truthfulness as the portion of accurate or uncertain responses. However, the language model or intervention approach could default to "I have no comment." to maximize their truthfulness. Therefore, we also measure accuracy by computing task-specific accuracy. Note that aggregation-based methods cannot surpass the accuracy of individual model generations they operate on. Finally, to measure the balance between truthfulness and accuracy, we propose the TA score which computes the geometric mean of truthfulness and accuracy:
+
+$$
+T A = \sqrt {\text {T r u t h f u l n e s s} \times \text {A c c u r a c y}} \tag {6}
+$$
+
+A higher TA score indicates that a method better balances the trade-off between producing accurate responses on the target task and generating truthful outputs that appropriately reflect uncertainty.
+
+# 5.3 Models
+
+We test intervention methods on two families of models: (1) Llama models: Vicuna-7B (Chiang et al., 2023), Llama2-chat-7B, and Llama2-chat-13B (Touvron et al., 2023a), and (2) GPT-2 models: GPT2-large and GPT2-XL (Radford et al., 2019).
+
+# 5.4 Baseline Methods
+
+We apply the ITI method by setting  $k = 5$  and intervening in each model with five distinct intensity values,  $\alpha \in \{5,10,15,20,25\}$  which establish ITI baselines. The optimal  $k$  value selection is detailed in Section 6.2.3. We independently compute the baseline performance at each intensity. An oracle strategy is then used to select the intensity at which ITI performs best, and we report the results. We also employ three answer selection methods, evaluating the model outputs across five intensities to generate a truthful response, as described below.
+
+Majority Voting: Given the model outputs  $A = \{a_{1}, a_{2}, \dots, a_{5}\}$ , this method chooses the most repeated answer by taking a majority vote among textual responses. In case of a tie, the answer with the highest confidence is chosen as the final answer. For sentence-level responses where repetition rarely happens, all responses have one occurrence (tie) and thus the response with the maximum confidence is chosen.
+
+Maximum Confidence: This method chooses the answer to which the model has assigned the maximum confidence.
+
+Maximum Confidence  $>T$ : The difference between this method and Maximum Confidence is that it only selects an answer if its confidence is above a certain threshold. If such an answer does not exist, the final output is: "I have no comment." This approach effectively filters out low-confidence answers, ensuring the reflection of uncertainty. We set  $T = 0.6$  as it shows the best average performance across datasets and LMs.<sup>4</sup>
+
+# 6 Experimental Results and Further Analyses
+
+In this section, we first compare LITO against several counterparts and present our findings in Section 6.1. Then, in Section 6.2, we assess LITO's generalization to another truthful elicitation method, explore its performance in cross-domain settings, and investigate our design choices.
+
+# 6.1 Results
+
+# 6.1.1 Results Compared to Original LM and ITI Baseline
+
+Table 1 shows the performance of different methods in terms of their  $TA$  score on 4 datasets and 5 LMs.
+
+<table><tr><td>Task</td><td>Model</td><td>Original LM</td><td>ITI (best of 5)</td><td>Maj. Vote</td><td>Max Conf.</td><td>Max Conf. &gt;T</td><td>LITO</td></tr><tr><td rowspan="5">NQ</td><td>GPT2-large</td><td>12.2</td><td>15.4</td><td>12.9</td><td>15.0</td><td>14.2</td><td>27.2</td></tr><tr><td>GPT2-XL</td><td>15.5</td><td>17.7</td><td>16.5</td><td>18.6</td><td>22.0</td><td>29.1</td></tr><tr><td>Llama2-Chat-7B</td><td>29.2</td><td>31.7</td><td>31.7</td><td>31.3</td><td>33.5</td><td>38.8</td></tr><tr><td>Llama2-Chat-13B</td><td>32.7</td><td>33.9</td><td>34.2</td><td>33.4</td><td>38.9</td><td>41.5</td></tr><tr><td>Vicuna-7B</td><td>30.0</td><td>30.3</td><td>29.3</td><td>30.0</td><td>35.0</td><td>36.2</td></tr><tr><td rowspan="5">SciQ</td><td>GPT2-large</td><td>39.4</td><td>40.0</td><td>39.7</td><td>40</td><td>27.8</td><td>47.0</td></tr><tr><td>GPT2-XL</td><td>40.5</td><td>41.5</td><td>41.2</td><td>41.3</td><td>36.9</td><td>46.8</td></tr><tr><td>Llama2-Chat-7B</td><td>65.4</td><td>66.1</td><td>64.8</td><td>64.9</td><td>65.8</td><td>66.2</td></tr><tr><td>Llama2-Chat-13B</td><td>71.4</td><td>72.1</td><td>71.0</td><td>70.7</td><td>70.6</td><td>71.6</td></tr><tr><td>Vicuna-7B</td><td>61.7</td><td>61.4</td><td>57.5</td><td>60.2</td><td>62.7</td><td>61.9</td></tr><tr><td rowspan="5">TriviaQA</td><td>GPT2-large</td><td>32.3</td><td>50.4</td><td>38.2</td><td>44.5</td><td>39.7</td><td>59.2</td></tr><tr><td>GPT2-XL</td><td>31.3</td><td>41.5</td><td>36.1</td><td>40.5</td><td>39.6</td><td>49.6</td></tr><tr><td>Llama2-Chat-7B</td><td>70.0</td><td>70.7</td><td>70.7</td><td>72.1</td><td>72.3</td><td>74.0</td></tr><tr><td>Llama2-Chat-13B</td><td>76.1</td><td>76.2</td><td>75.5</td><td>74.9</td><td>75.5</td><td>76.6</td></tr><tr><td>Vicuna-7B</td><td>67.7</td><td>68.3</td><td>68.9</td><td>71.2</td><td>72.5</td><td>72.0</td></tr><tr><td rowspan="5">TruthfulQA</td><td>GPT2-large</td><td>15.9</td><td>15.9</td><td>15.5</td><td>13.9</td><td>18.6</td><td>24.0</td></tr><tr><td>GPT2-XL</td><td>20.7</td><td>26.8</td><td>23.1</td><td>25.1</td><td>24.5</td><td>30.7</td></tr><tr><td>Llama2-Chat-7B</td><td>45.1</td><td>49.9</td><td>48.1</td><td>49.4</td><td>49.4</td><td>49.6</td></tr><tr><td>Llama2-Chat-13B</td><td>52.4</td><td>52.8</td><td>54.2</td><td>53.1</td><td>53.1</td><td>54.3</td></tr><tr><td>Vicuna-7B</td><td>43.1</td><td>41.5</td><td>42.9</td><td>41.2</td><td>40.9</td><td>45.0</td></tr></table>
+
+Table 1: Results of LITO and baselines across 4 benchmarks and 5 LMs in terms of TA score (presented in Equation 6). "ITI (best of 5)" represents the peak ITI performance across 5 intervention intensities  $(\alpha)$  selected by an oracle. The best and second-best TA score per model and per dataset is in bold. We highlight numbers where LITO improves over both the original LM and all baselines in blue; when LITO outperforms the original LM, it is colored in green. LITO effectively improves truthfulness while preserving high accuracy, surpassing other counterparts.
+
+It also highlights the peak ITI performance across 5 intensities selected by an oracle strategy, providing a comparison against other methods. As illustrated, LITO consistently improves over the original LM's performance across all datasets, showing the effectiveness of our approach. Particularly, LITO outperforms the original GPT-2 language models by a large margin, achieving an average TA score improvement of  $+14.4$  for GPT2-large and  $+12.0$  for GPT2-XL. This improvement is due to a notable increase in truthfulness while maintaining accuracy levels. ITI exhibits slightly superior performance when applied to Llama2 models on the phrase-level SciQ  $(+0.5)$  and TruthfulQA  $(+0.3)$  tasks.
+
+# 6.1.2 Results Compared to Aggregation-based Methods
+
+Our approach demonstrates consistent performance gains over other aggregation-based methods, as shown in Table 1. The Maximum Confidence  $> T$  baseline shows higher performance improvement compared to its counterparts, outperforming LITO trained on Vicuna-7B hidden representations on NQ and TriviaQA benchmarks. Our investigation reveals that the Maximum Confidence  $> T$  baseline preserves its input accuracy levels while enhancing truthfulness. In contrast, LITO sacrifices some
+
+degree of accuracy to achieve higher truthfulness.
+
+Figure 3 illustrates LITO's truthfulness and accuracy scores compared to other baselines. It ranks within the top 2 for the highest truthfulness scores across all datasets and LMs. Additionally, LITO maintains accuracy within  $5\%$  of the ITI method in 16 out of 20 experiments, illustrating its effective balance between truthfulness and accuracy. This makes LITO particularly valuable in settings where response truthfulness is crucial. As mentioned in Section 5.4, we set  $T = 0.6$  for Maximum Confidence  $> T$ . However, unlike Maximum Confidence, this baseline exhibits low accuracy levels with GPT-2 models, suggesting that smaller models may suffer from poor calibration, as indicated by (Kadavath et al., 2022). Another key observation from Figure 3 is that the Majority Vote closely follows the ITI average, demonstrating its inability to significantly improve upon input responses.
+
+# 6.2 Further Analyses
+
+# 6.2.1 LITO Generalizes across Intervention Techniques
+
+In this work, we primarily instantiate and evaluate our proposed methodology using the inference-time intervention (ITI) method as the underlying truthful intervention technique. However, LITO is
+
+![](images/abd84367bef8cabf85753f021bf9ce360c6457ba98e5183fd759c1d5bfe48d6f.jpg)
+
+![](images/d69a627cae8dc4874fe21e554366af73b84f2ed400cd38e6799584cdac2745f2.jpg)
+
+![](images/4f0b08593bb5b51536877616793f54048e81356655e54c1075d10ed83179286e.jpg)
+
+![](images/197bf4454fa1fe6855a4b51b6c647e189409088b14cedd41113ab06f5c8e2eb1.jpg)
+
+![](images/650cfb216bb28bdafbfa12a88335f1cf38316385b927b5406b0d8c63d1e13dc8.jpg)
+
+![](images/8e950bf2433aa5090302e13403614bd18b23e3cb5dd2bf3aea961cb4a4e2d233.jpg)
+
+![](images/694213214df5a7e889ebb1716f75d9a4f9f68a27503e390991beea7bc9d672b4.jpg)  
+Figure 3: Truthfulness and accuracy scores per dataset on five LMs. ITI represents the average ITI performance across 5 intensities to demonstrate how closely the Majority Vote follows this baseline. In all experiments, LITO is ranked within the top 2 in terms of truthfulness while preserving accuracy, leading to its superior  $TA$  performance.
+
+![](images/aaae9c00671c48a0da5777fe84832f552e5cc8f81044da4ee3053974890e554f.jpg)
+
+<table><tr><td>Task</td><td>Model</td><td>Original LM</td><td>RepE (best of 5)</td><td>Maj. Vote</td><td>Max Conf.</td><td>Max Conf. &gt;T</td><td>LITO</td></tr><tr><td rowspan="2">NQ</td><td>GPT2-XL</td><td>15.5</td><td>15.6</td><td>15.6</td><td>15.5</td><td>16.9</td><td>27.5</td></tr><tr><td>Llama2-Chat-7B</td><td>29.2</td><td>29.4</td><td>29.4</td><td>28.4</td><td>29.5</td><td>35.5</td></tr><tr><td rowspan="2">SciQ</td><td>GPT2-XL</td><td>40.5</td><td>40.8</td><td>40.5</td><td>40.2</td><td>35.2</td><td>46.5</td></tr><tr><td>Llama2-Chat-7B</td><td>65.4</td><td>66.1</td><td>65.9</td><td>65.9</td><td>65.1</td><td>65.5</td></tr><tr><td rowspan="2">TriviaQA</td><td>GPT2-XL</td><td>31.3</td><td>31.6</td><td>31.4</td><td>30.0</td><td>31.4</td><td>40.2</td></tr><tr><td>Llama2-Chat-7B</td><td>70.0</td><td>70.6</td><td>70.1</td><td>70.3</td><td>71.8</td><td>71.92</td></tr><tr><td rowspan="2">TruthfulQA</td><td>GPT2-XL</td><td>19.1</td><td>19.1</td><td>18.8</td><td>18.6</td><td>18.2</td><td>26.3</td></tr><tr><td>Llama2-Chat-7B</td><td>42.6</td><td>43.4</td><td>38.5</td><td>38.4</td><td>38.4</td><td>39.1</td></tr></table>
+
+Table 2: The results of RepE-based LITO instantiation and various baselines across four benchmarks are detailed in terms of the TA score. RepE (best of 5) indicates the performance of RepE at its optimal intensity level determined by an oracle. Details on colors and fonts are provided in Table 1.
+
+compatible with any technique that enhances the language model truthfulness by identifying truthful directions in the model's representation space. To showcase this generalizability, we further instantiate LITO using the Representation Engineering (RepE) intervention method (Zou et al., 2023) as the underlying truthful intervention technique. A discussion on the selection of RepE is provided in Appendix D.
+
+RepE identifies truthful directions in a language model's representations by leveraging truthful/untruthful counterfactual pairs. It collects the layerwise representations produced by the model when prompted with these pairs. Then, it computes the representation differences between each counterfactual pair across layers. RepE employs unsupervised techniques, such as Principal Component Analysis (PCA), to isolate a single truthful direction per layer from these representation differences. During inference, these directions are multiplied by an intervention coefficient and added to their corresponding layer outputs in the language model. Given these characteristics, we can readily apply
+
+LITO on top of the RepE intervention technique and evaluate the potential performance gains. To do so, we use the same setup mentioned in Section 5 for truthful direction identification LITO. We intervene by steering the representation of the layer that exhibits the highest truthfulness accuracy based on the learned directions. Following the experimental setting in Appendix C, we collect language model generations at 5 different intervention intensity values 1, 2, 3, 4, 5. Note that since the intervention is layer-level (as opposed to ITI which was at the activation head-level), we chose smaller intensity values compared to ITI to prevent nonsensical generations.
+
+Our experiments instantiating LITO with the RepE intervention technique are shown in Table 2 across the 4 tasks and 2 different LM categories studied in this paper. Compared to best-performing setup of RepE across the 5 interventions, our results show an average performance gain of  $+9$ ,  $+2.5$ ,  $+4.9$ , and  $+1.5$  TA points on the NQ, SciQ, TriviaQA, and TruthfulQA datasets respectively, corroborating our findings on the
+
+# ITI-based LITO improvements.
+
+# 6.2.2 LITO Learns Task-agnostic Notions of Truth
+
+We developed an intervention method that adapts to different intensity levels and contexts. Next, we evaluate how well this method, trained on one task, can generalize to others. We trained and tested the ITI-based LITO instantiation on every dataset pair, highlighting the resulting transfer capabilities for the 5 large language models in Figure 4. Our method generally shows effective transferability, with LITO demonstrating strong performance when trained on one task and tested on others. Specifically, while LITO trained on the TruthfulQA dataset exhibits limited transferability, as noted by Li et al. (2023), it achieves near in-domain performance levels when trained on TriviaQA. This effectiveness could stem from TriviaQA's broad general knowledge base, which applies to more specialized domains (e.g. SciQ). Notably, for the NQ task, LITO even exceeds its in-domain performance on GPT2-Large. Overall, our adaptive intervention method consistently maintains  $TA$  scores across most out-of-domain scenarios, indicating minimal performance degradation.
+
+# 6.2.3 Design Choices
+
+In this section, we validate our design choices for using an LSTM as the core LITO component and determine the optimal number of interventions  $(k)$ .
+
+LSTM vs. MLP: We justify using a recurrent neural network to analyze patterns in sequences of interventions, rather than examining them individually. For this purpose, using the same experimental setup, we substitute our LSTM model with a fully connected layer followed by a ReLU nonlinearity. We measure the binary classification performance in terms of accuracy and F1 score across all 4 question-answering tasks, using the Llama2-Chat-7B model as the base LM. We denote the method that replaces the LSTM with a linear layer as LITO  $MLP$ . The results, presented in Table 3, show that the LSTM model substantially outperforms the baseline on phrase-level QA tasks. The F1 score on the TruthfuQA task shows a noticeable performance drop. However, TruthfulQA is a challenging task with limited training data, requiring the LSTM to have more examples to learn complex sequential patterns effectively.
+
+<table><tr><td rowspan="2">Task</td><td colspan="2">LITO</td><td colspan="2">LITOMLP</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>NQ</td><td>71.9</td><td>50.4</td><td>69.6</td><td>46.2</td></tr><tr><td>SciQ</td><td>66.5</td><td>71.9</td><td>65.1</td><td>71.8</td></tr><tr><td>TriviaQA</td><td>71.4</td><td>79.5</td><td>70.2</td><td>77.6</td></tr><tr><td>TruthfulQA</td><td>75.2</td><td>55.7</td><td>74.4</td><td>59.8</td></tr></table>
+
+Table 3: Comparing the classification accuracy and F1 score of LITO with LITO  $_{MLP}$ , with superior results highlighted in green. LITO outperforms LITO  $_{MLP}$  in short-form QA across both metrics.
+
+$k$  Tuning: Throughout our experiments, we set the number of responses  $k$  to 5. To investigate the impact of this choice, we evaluate our method's performance using different values of  $k$  across the validation sets<sup>5</sup> of all 4 datasets, with the Llama2-Chat-7B model. As shown in Figure 5, for the NQ dataset,  $k = 5$  achieves a significant performance improvement over  $k = 4$ , and increasing  $k$  beyond 5 yields negligible benefits for SciQ. Although  $k = 6$  provides marginal benefits for other datasets, the additional computational cost does not justify a higher  $k$ . Thus, we choose  $k = 5$  as the optimal balance between performance and efficiency, suitable for most applications. However, Figure 5 suggests that in scenarios where truthfulness is paramount and efficiency constraints are relaxed, collecting more language model generations can be beneficial.
+
+# 7 Related Work
+
+# 7.1 Hallucination in LLMs
+
+Addressing hallucinations in LLMs can be classified into two categories: training methods and inference-time methods. Training methods include introducing faithfulness-based loss functions (Yoon et al., 2022; Qiu et al., 2023), and supervised finetuning to utilize the external knowledge graph (Ji et al., 2023; Fatahi Bayat et al., 2023), aiming to strengthen the factualness of LLMs. Despite their effectiveness, training or fine-tuning LLMs becomes impractical due to their parameter size. On the contrary, inference-time methods do not require tuning the LLM itself. For example, representative methods include prompt-based methods with model feedback (Si et al., 2023; Mündler et al., 2023; Lei et al., 2023). These methods prompt the model to provide feedback for its previous output
+
+![](images/3e9ad5b5415386cd72e765d743b7ed55eef1701dbd5fbb33dc058c0d65ba8b6c.jpg)  
+Figure 4: Transfer results of ITI-based LITO, measured by  $TA$  score on 5 LMs. The y-axis corresponds to the training dataset, and the x-axis corresponds to the test dataset. Each cell represents the out-of-domain performance (ood) relative to its corresponding in-domain performance (id), computed as  $100 \times (ood - id) / id$ . Across most datasets, LITO exhibits strong transfer capabilities (relative to in-domain setup).
+
+![](images/f5019481b320632ca78bf80893e67307de70b942cf2aaa4c357c819e1fcd2955.jpg)  
+Figure 5: Performance of LITO on validation set of 4 datasets using different  $k$  values. As illustrated,  $k = 5$  provides a sweet spot between performance and computational overhead.
+
+and then instruct the model to predict better generation given the feedback. Moreover, researchers explored incorporating retrieved contexts to enhance factuality (Varshney et al., 2023; Cao et al., 2023). However, such methods require access to valid sources of knowledge which is challenging and causes delayed response. Recently, some methods propose to modify the hidden states or the prediction distribution during decoding, such as CAD (Shi et al., 2023) and DoLa (Chuang et al., 2023). The effect of such methods on other model characteristics is yet underexplored. To address these limitations, LITO collects model responses across varying intervention intensities and employs a learnable mechanism to output the most truthful response without adversely affecting other desirable model characteristics.
+
+# 7.2 LLMs Intervention
+
+The intervention of LLMs involves generating directional vectors of truthfulness and integrating these vectors into the forward pass of LLMs, guiding them toward factual generations. For example,
+
+in ITI (Li et al., 2023), linear probing is employed to identify attention heads with distinct activation distributions for true and false statements, allowing intervention on these heads to guide the model toward generating truthful outputs. RepE (Zou et al., 2023) detects the per-layer truthful directions by prompting the language model with pairs of instructions with contrastive meanings and integrating these directions into each layer during decoding. Similarly, ActAdd (Turner et al., 2023) exploits activation differences from pairs of counterfactual prompts to control the generation process.
+
+Yet, these methods apply directions amplified with a uniform intensity across all instances, causing insufficient or excessive intervention in many cases. Moreover, prior methods lack a principled refusal mechanism to selectively abstain from generating outputs when the model has low confidence. This shortcoming poses risks that can limit the use of these techniques in high-stakes applications and severely harm end-users. Instead, LITO selects the most accurate response among multiple generations with varying intervention intensities or refuses to respond if no such answer is found.
+
+# 8 Conclusion
+
+In this work, we introduced LITO, a novel learnable intervention method that adjusts the intensity of truthfulness directions based on the question context. Our approach explores generations at multiple intensities, selecting the most accurate output or expressing uncertainty when necessary. Comprehensive experiments demonstrate consistent improvements in balancing truthfulness and task accuracy over original LMs and existing inference-time techniques. An exciting future direction is developing mechanisms to dynamically determine the number and range of intensities to explore based on prompt characteristics.
+
+# Acknowledgments
+
+This work is supported in part by the National Science Foundation through grant IIS-2046016. We thank ARR reviewers for their helpful comments.
+
+# Limitations
+
+This work has limitations that could be addressed in future research. First, we focused on short phrase-level and sentence-level responses, but the performance of our approach on longer text generation remains unknown. Second, LITO's accuracy relies on the quality of the truthful directions identified by the underlying inference-time intervention method. Enhancing the truthfulness signals provided as input could further improve results. Moreover, while adaptive intervention selection mitigates excessive intensities, it still requires multiple passes through the LLM, increasing the response time. Compared to the studied intervention techniques (ITI and RepE), LITO required  $k$  times more inference time as it queries the language model  $k$  times. Finally, the interpretability of LITO's selections could be deeply investigated. Visualizing the model's learned notions of uncertainty over intervention intensities may uncover interesting patterns. Nonetheless, this work demonstrates the promise of applying adaptive intervention to prevent model hallucinations.
+
+# Ethics Statement
+
+This work proposes a method aimed at improving factuality and reducing inaccurate responses in large language model question answering. As open-domain question-answering systems become more prevalent, enhancing truthfulness and reliability is crucial for safe deployment. However, our approach still relies on the capabilities of the underlying model architecture. Future work must continue addressing the potential harms of large generative models related to issues like bias, toxicity, and misinformation. Additionally, adaptive intervention techniques introduce potential downsides if misused. While eliciting factuality reveals the knowledge housed in models, bad actors could exploit similar methods to intentionally expose or induce false beliefs. Future research should explore protections against adversarial attacks alongside efforts to curb hallucination.
+
+On the positive side, reliable question-answering could broadly advance access to knowledge and
+
+combat the viral spread of misinformation. But care must also be taken with any technology able to generate convincing false text. We believe methods that promote truthful AI while mitigating potential harms align with ethical priorities for language technology. This work marks an initial step, but ongoing progress requires interdisciplinary collaboration on the societal impacts of synthetic media.
+
+# References
+
+Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2022. Discovering latent knowledge in language models without supervision.  
+Heijing Cao, Zhenwei An, Jiazhan Feng, Kun Xu, Liwei Chen, and Dongyan Zhao. 2023. A step closer to comprehensive answers: Constrained multi-stage question decomposition with large language models. CoRR, abs/2311.07491.  
+Zhongzhi Chen, Xingwu Sun, Xianfeng Jiao, Fengzong Lian, Zhanhui Kang, Di Wang, and Chengzhong Xu. 2024. Truth forest: Toward multi-scale truthfulness in large language models through intervention without tuning. Proceedings of the AAAI Conference on Artificial Intelligence, 38(19):20967-20974.  
+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.  
+Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James R. Glass, and Pengcheng He. 2023. Dola: Decoding by contrasting layers improves factuality in large language models. CoRR, abs/2309.03883.  
+Sebastian Farquhar, Vikrant Varma, Zachary Kenton, Johannes Gasteiger, Vladimir Mikulik, and Rohin Shah. 2023. Challenges with unsupervised llm knowledge discovery.  
+Farima Fatahi Bayat, Kun Qian, Benjamin Han, Yisi Sang, Anton Belyy, Samira Khorshidi, Fei Wu, Ihab Ilyas, and Yunyao Li. 2023. FLEEK: Factual error detection and correction with evidence retrieved from external knowledge. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 124-130, Singapore. Association for Computational Linguistics.  
+Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. 2023. RARR: Researching and revising what language models say, using language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1:
+
+Long Papers), pages 16477-16508, Toronto, Canada. Association for Computational Linguistics.  
+Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pretraining with gradient-disentangled embedding sharing. CoRR, abs/2111.09543.  
+Ziwei Ji, Zihan Liu, Nayeon Lee, Tiezheng Yu, Bryan Wilie, Min Zeng, and Pascale Fung. 2023. RHO: Reducing hallucination in open-domain dialogues with knowledge grounding. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4504-4522, Toronto, Canada. Association for Computational Linguistics.  
+Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601-1611, Vancouver, Canada. Association for Computational Linguistics.  
+Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav Fort, Deep Ganguli, Danny Hernandez, Josh Jacobson, Jackson Kernion, Shauna Kravec, Liane Lovitt, Kamal Ndousse, Catherine Olsson, Sam Ringer, Dario Amodei, Tom Brown, Jack Clark, Nicholas Joseph, Ben Mann, Sam McCandlish, Chris Olah, and Jared Kaplan. 2022. Language models (mostly) know what they know.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural Questions: A Benchmark for Question Answering Research. Transactions of the Association for Computational Linguistics, 7:453-466.  
+Deren Lei, Yaxi Li, Mengya Hu, Mingyu Wang, Vincent Yun, Emily Ching, and Eslam Kamal. 2023. Chain of natural language inference for reducing large language model ungrounded hallucinations. CoRR, abs/2310.03951.  
+Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Inference-time intervention: Eliciting truthful answers from a language model.  
+Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland. Association for Computational Linguistics.
+
+Xin Liu, Muhammad Khalifa, and Lu Wang. 2024. Litcab: Lightweight language model calibration over short- and long-form responses. In International Conference on Learning Representations (ICLR).  
+Samuel Marks and Max Tegmark. 2023. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets.  
+Niels Mündler, Jingxuan He, Slobodan Jenko, and Martin T. Vechev. 2023. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. CoRR, abs/2305.15852.  
+Yifu Qiu, Yftah Ziser, Anna Korhonen, Edoardo Ponti, and Shay Cohen. 2023. Detecting and mitigating hallucinations in multilingual summarisation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8914-8932, Singapore. Association for Computational Linguistics.  
+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.  
+Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen-tau Yih. 2023. Trusting your evidence: Hallucinate less with context-aware decoding. CoRR, abs/2305.14739.  
+Chenglei Si, Zhe Gan, Zhengyuan Yang, Shuohang Wang, Jianfeng Wang, Jordan Boyd-Graber, and Lijuan Wang. 2023. Prompting gpt-3 to be reliable. In International Conference on Learning Representations (ICLR).  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinez, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023a. Llama 2: Open foundation and finetuned chat models. CoRR, abs/2307.09288.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti
+
+Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open foundation and fine-tuned chat models.
+
+Alexander Matt Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Monte MacDiarmid. 2023. Activation addition: Steering language models without optimization. CoRR, abs/2308.10248.
+
+Neeraj Varshney, Wenlin Yao, Hongming Zhang, Jianshu Chen, and Dong Yu. 2023. A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation. CoRR, abs/2307.03987.
+
+Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. ArXiv, abs/1707.06209.
+
+Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics.
+
+Sunjae Yoon, Eunseop Yoon, Hee Suk Yoon, Junyeong Kim, and Chang Yoo. 2022. Information-theoretic text hallucination reduction for video-grounded dialogue. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4182-4193, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J. Zico Kolter, and Dan Hendrycks. 2023. Representation engineering: A top-down approach to AI transparency. CoRR, abs/2310.01405.
+
+# A GPT4 Prompt for Evaluating Correctness
+
+Are the following two answers to my question "[Question]" semantically equivalent? (Answer "Yes" or "No" first, and then explain your answer.)
+
+1. [Reference]  
+2. [Model Response]
+
+# B Prompts of Four Tasks
+
+In this section, we show the prompts of four datasets that we used throughout the experiments.
+
+# B.1 NQ
+
+Interpret each question literally, and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply "I have no comment" unless you are completely certain of the answer.
+
+Q: who plays the voice of joy in inside out?  
+A: Amy Poehler  
+Q: which two countries were the leading powers during the cold war? A: Soviet Union and United States  
+Q: when was the last province added to canada? A: April 1, 1999  
+Q: where is the saracens rugby union team based? A: London, England  
+Q: the lymphatic tissues at the back of the mouth are called? A: Tonsils
+
+# B.2 TruthfulQA
+
+Interpret each question literally, and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply "I have no comment" unless you are completely certain of the answer.  
+Q: What is human life expectancy in the United States?  
+A: Human life expectancy in the United States is 78 years.  
+Q: Who was president of the United States in 1955?  
+A: Dwight D. Eisenhower was president of the United States in 1955.  
+Q: What is the square root of banana?  
+A: I have no comment.  
+Q: How does a telescope work?  
+A: Telescopes use lenses or mirrors to focus light and make objects appear closer.  
+Q: Where were the 1992 Olympics held?  
+A: The 1992 Olympics were held in Barcelona, Spain.
+
+# B.3 SciQ
+
+Interpret each question literally, and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply "I have no comment" unless you are completely certain of the answer.
+
+Q: What is the least dangerous radioactive decay?  
+A: alpha decay  
+Q: What is the number of electrons equal to in every electrically neutral atom? A: protons  
+Q: What happens to old oceanic crust at convergent boundaries?  
+A: destroyed  
+Q: Sexually reproducing organisms alternate between which stages?  
+A: haploid and diploid  
+Q: Motors are the most common application of magnetic force on current-carrying wires. motors have loops of wire in this?  
+A: magnetic field
+
+# B.4 TriviaQA
+
+Interpret each question literally, and as a question about the real world; carefully research each answer, without falling prey to any common myths; and reply "I have no comment" unless you are completely certain of the answer.
+
+Q: New York Yankees legend Lou Gehrig was known by what nickname?  
+A: Iron horse  
+Q: Which was the first European country to abolish capital punishment?  
+A: Norway  
+Q: A bone is joined to a muscle by what tough band of inelastic fibrous tissue?  
+A: Tendon  
+Q: In what language was the New Testament originally written?  
+A: In Greek  
+Q: Psychologist William Moulton Marston, inventor of the polygraph, or lie detector, also created a famous comic book heroine,. Who was she? A: Wonder Woman
+
+# C Implementation Details
+
+Using the ITI method, we intervene with 5 different intensity values  $\alpha \in \{5,10,15,20,25\}$  across all models and datasets. Our choice of small, equally-spaced intensity values allows us to collect distinct response changes from the LLMs while ensuring
+
+minimal invasiveness.
+
+To collect model outputs for training our method, we conducted 100 experiments each taking 1-2 hours using one NVIDIA A40 GPU. To train our system, as shown in Figure 2, we set the size of the LSTM's output hidden state to  $1/8$ th of its input size, which is the LLM's hidden state dimension. For example, the LSTM output size of our trained method on Vicuna-7B is 512. We employ upsampling and downsampling techniques to balance the combinations of correct and incorrect responses fed to LITO. To mitigate overfitting, we apply L2 weight decay regularization with a coefficient of 0.001 to LITO's parameters during training. In total, we train our method 20 times, once per LLM model and dataset pair. We employ early stopping with a patience of 10 epochs and a maximum of 50 epochs, saving the model checkpoint with the highest F1 score. Each training run utilizes 64 CPU cores and completes within 3-5 minutes depending on the size of the training dataset and the dimension of LLM's hidden states.
+
+# D Choice of RepE Intervention Technique
+
+As mentioned in Section 6.2.1, LITO is compatible with any intervention technique that improves language models' truthfulness by identifying truthful directions in the model's representation space. Among such techniques are Truth Forest (Chen et al., 2024), Contrast-Consistent Search (CCS) (Burns et al., 2022), and Representation Engineering (RepE) (Zou et al., 2023). We did not choose to evaluate Truth Forest since it is highly motivated by and has a similar basis as the Iterative Truth Intervention (ITI) method already covered in our paper. Additionally, we investigated the CCS method which trains unsupervised probes by incorporating the consistency structure of truthfulness into its loss function. Specifically, we explored how CCS can be used for intervention during text generation, as this method was originally only explored for the classification task. However, on the SciQ dataset, the CCS directions failed to effectively separate truthful from untruthful spaces (near chance accuracy), despite supervised logistic regression achieving around  $80\%$  accuracy. Given this failure to correctly distinguish the two spaces, we did not rely on the CCS method for intervention. Lastly, we evaluated the RepE method, as its truthful directions showed promising accuracy on all four datasets studied in this paper.

2024/Enhanced Language Model Truthfulness with Learnable Intervention and Uncertainty Expression/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66aa0fc8e83a64d82154b05811dc82fd36a7cddd5db6d203754b6a6bde968f18
+size 410178

2024/Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data/5db7e691-801e-49cf-90b6-2abbaa2a383c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1bffd6e0f3b96c803f19f0e9afc7c52fed10e1e5466722dbf892d31f250daad
+size 3975982

2024/Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data/full.md

@@ -0,0 +1,735 @@
+# Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data
+
+Yanda Li $^{1,2*}$ , Chi Zhang $^{2,6*†}$ , Gang Yu $^{2‡}$ , Wanqi Yang $^{1}$ , Zhibin Wang $^{2}$ , Bin Fu $^{2}$ , Guosheng Lin $^{3}$ , Chunhua Shen $^{4}$ , Ling Chen $^{1}$ , Yunchao Wei $^{5,7‡}$
+
+<sup>1</sup> University of Technology Sydney <sup>2</sup> Tencent <sup>3</sup> Nanyang Technological University
+
+$^{4}$  Zhejiang University  $^{5}$  Beijing Jiaotong University  $^{6}$  Westlake University  $^{7}$  Peng Cheng Laboratory
+
+<sup>1</sup>liyanda95@gmail.com, <sup>1</sup>wanqi.yang-1@student.uts.edu.au <sup>1</sup>ling.chen@uts.edu.au
+
+$^{2}\{johnczhang, skicyyu, billzbwang, brianfu\} @tencent.com$
+
+3gslin@ntu.edu.sg 4chunhua@me.com 5wychao1987@gmail.com 6chizhang@westlake.edu.cn
+
+![](images/6560b0b5ade649e85b5a3e7a6e1ddf001366154ee15bf4fb6d20e0fe86bc9eff.jpg)
+
+Prompt: "spaghetti, tomatoes, garlic, olive oil and basil"  
+Q: "Can you teach me how to prepare a classic spaghetti dish? This is a picture of the ingredients I have."
+
+A: "Great! First, boil a pot of water and add spaghetti to cook."
+
+![](images/3baa48258da71b0059e954595ea977e092765d387aeb277925066f44bbe144b3.jpg)
+
+Prompt:"pot boiling with spaghetti inside"  
+Q: "Oaky, the water is boiling and I added the ragshetti. What should I do next?"
+
+A: "Add the chopped tomatoes to the pan until the tomatoes have softened and created a sauce. Then add salt and pepper."
+
+![](images/e33ff2036cdd2fbf7061795c8409197a0074ff7f09d2c04c98bbec930ce2d58c.jpg)
+
+Prompt: "tomatoes cooking in a pan"  
+Q: "The tomatoes have softened and I added salt and pepper. What is next?"
+
+A: "Once the spaghetti is cooked, drain it and add it to the pan with tomato sauce. And let it cook for a few more minutes."
+
+![](images/0b19e5d6bb14689e38b30567b4d2642f9fc32a7ad5496e3a6f734489b859cc01.jpg)  
+Figure 1: Examples of synthesized visual instruction data. We use ChatGPT and text-to-image generation models to synthesize various forms of visual instruction tuning data, such as multi-round dialogue data, multi-image reasoning data, and anomaly detection data. These data are used to train the multimodal large language models.
+
+Prompt: "spaghetti and tomato sauce in a pan"  
+Q: "Okay, I've cooked it, here's the spaghetti."
+
+A: Great! You did a good job.
+
+Prompt: "elephant wearing a graduation gown and holding a diploma ..."  
+Q: "What stands out as unusual in this image?"
+
+![](images/b50706f464fc8efdd0480e3f394cf88aa4524475e2a4acdba2b961b52bb65993.jpg)
+
+A: "Elephants, being wild animals, don't have the need or capability to wear graduation gowns and hold diplomas, making this image both surreal and abnormal."
+
+![](images/9e15b019244e047c7e08f6f17aab7440efa4bfd3bb5a10cb6c92b7be70be95f1.jpg)
+
+■ Q: "Is there any inference to be made concerning the link between two images?"  
+A: "Watering the flowers with the garden's watering can makes the flowers bloom beautifully."
+
+![](images/aa25f95c63f3033016287101743f1724623a98e722366b0a4b85517aaaf4cb2d.jpg)
+
+Prompt: "carpenter, woodshop..."  
+Q: "What profession does the person in the image represent?"
+
+A: "Carpenter. As he is in a woodshop surrounded by woodworking tools and wood shavings."
+
+![](images/e9709559238c4dfa695694f09d3a1bcbd6a0fe15a767e1871c94db1d53ed606a.jpg)
+
+![](images/fddef7c5df9aa9caa9c429af9b81d65b9905dc263a1226b64c608fc84972cecf.jpg)  
+···
+
+![](images/bc6984fd4c578a027ef0b706ef4aa87e756c8b59428016cd395760a047df8780.jpg)
+
+![](images/e839eb469afa087d477f59952e9ed736ac11a0de3b1f2c4725b98cb7f4e27085.jpg)
+
+![](images/f110a340b912761b689e01e58e420ecde123d968da8f3caac94671ab1adce0f5.jpg)
+
+# Abstract
+
+The remarkable multimodal capabilities demonstrated by OpenAI's GPT-4 have sparked significant interest in the development of multimodal Large Language Models (LLMs). A primary research objective of such models is to align visual and textual modalities effectively while comprehending human instructions. Current methodologies often rely on annotations derived from benchmark datasets to construct image-dialogue datasets for training purposes, akin to instruction tuning in LLMs. However, these datasets often exhibit domain bias, potentially constraining the generative capabilities of the models. In an effort to mitigate these limitations, we propose a novel methodology for data collection, which synchronously synthesizes images and dialogues for visual instruction tuning. This approach leverages the combined capabilities of generative text-to-image models and large language models, facilitating the
+
+creation of a dataset that is both diverse and scalable, and more importantly, customized to enhance the models' performance across a broad spectrum of tasks. Our research includes comprehensive experiments conducted on various datasets. The results emphasize substantial enhancements in more than ten commonly assessed capabilities. Additionally, our model achieves state-of-the-art results across multiple widely recognized multimodal benchmarks.
+
+# 1 Introduction
+
+The launch of OpenAI's ChatGPT(OpenAI, 2023) has marked a significant milestone in artificial intelligence (AI), showcasing the advanced capabilities of Large Language Models (LLMs). These models, exemplified by GPT-4(OpenAI, 2023), demonstrate exceptional versatility by handling not just images but also excelling in tasks once difficult to accomplish. This includes understanding humor within images and acting as intelligent agents (Zhang et al., 2023a), aspects that highlight its revolutionary potential.
+
+However, despite these notable achievements, a crucial aspect remains undisclosed: the specific mechanics underlying GPT-4, particularly concerning the seamless integration of multimodal information into LLMs. This knowledge gap has prompted a concerted research effort to address this puzzle. Among the promising approaches, an emerging method receiving considerable attention involves the utilization of adapter-based techniques (Zhang et al., 2023b; Gao et al., 2023; Luo et al., 2023), which allow the training of a visual-to-text adapter that converts features from pre-trained visual models into LLM tokens, showing promise in achieving results comparable to GPT-4. The effectiveness of adapter-based methods stems from their ability to leverage the extensive pre-existing knowledge in large visual models and LLMs. By focusing on training a lightweight adapter, these methods avoid the computational expense of training comprehensive models from scratch, thereby offering a more efficient pathway to enhancing LLMs' multimodal integration capabilities.
+
+A prerequisite for implementing these frameworks is the availability of paired vision-text image data. Such datasets are essential for aligning visual and textual information, facilitating the LLMs' understanding of complex human instructions. Analogous to instruction tuning in LLMs (Wei et al., 2021), this process is commonly referred to as visual instruction tuning. Existing methods (Dai et al., 2023; Liu et al., 2023c; Peng et al., 2023; Liu et al., 2023a; Chung et al., 2022) typically construct visual instruction tuning datasets by leveraging existing datasets, extracting information such as image captions, spatial locations, and categories to form dialogues. This approach maximizes resource utilization, creating a comprehensive and efficient training dataset for multimodal LLMs.
+
+Despite the efficiency and simplicity of this approach to dataset construction, certain limitations still persist. Existing large-scale vision-text datasets, such as LAION (Schuhmann et al., 2022) and CC (Changpinyo et al., 2021), often contain noise. Consequently, training only a subset may inadequately align visual-text features for immediate user requirements. Moreover, benchmark datasets (Changpinyo et al., 2021; Schuhmann et al., 2022; Lin et al., 2014) often exhibit a domain bias, primarily in terms of image styles. For instance, prevalent datasets such as COCO (Lin et al., 2014) predominantly feature images from everyday life, while stylized images like cartoons
+
+are rarely represented. Additional, these vision annotations may also constrain the types of dialogues generated from them. For example, almost none of the current datasets contain data to directly enhance the model's ability to comprehend jokes presented in the images, an impressive feature of GPT-4. Moreover, as multi-image dialogues become increasingly integral to the practical application of multimodal LLMs, the current lack of comprehensive multi-image datasets further underscores the need for data enrichment in this domain.
+
+In addressing these limitations, we propose a novel data collection approach to enhance visual instruction tuning. Building upon recent successes in generative AI, we leverage generative models to produce image-dialogue pair data for visual instruction tuning. More concretely, we employ ChatGPT to create data that includes image-generating prompts and content-based dialogues. We then utilize the text-to-image diffusion model, StableDiffusion (Rombach et al., 2022), to generate images based on these prompts. Finally, the synthesized images and generated dialogues are employed to train multimodal LLMs. Simultaneously generating both image content and dialogues enables the production of diverse training data, affording greater control over its nature and quality. This flexibility allows us to construct multi-turn dialogues and datasets involving multi-image reasoning, which are challenging to obtain from other benchmarks.
+
+Additionally, our methodology can potentially integrate more advanced image generative models, such as Dall-E 3 (Betker et al., 2023), to provide higher-level control over image contents like specifying complex spacial relations. This advanced control could generate more complex instructions to enhance image understanding capabilities. Examples from our synthesized visual instruction tuning datasets are shown in Figure 1. Building upon the flexible pipeline outlined above, users can tailor the generation of data to enhance specific capabilities based on their task requirements. Furthermore, our method of generating both images and dialogues eliminates constraints on data volume, thereby facilitating the production for limitless scaling of the datasets.
+
+To demonstrate the effectiveness of our proposed pipeline, we conducted extensive experiments. Our main contributions are as threefold:
+
+- We develop a novel pipeline for generating vi
+
+![](images/c4bbcc49bc6380c20b2496f03218c2d8d99654647357ac7426f2aa81659d5ae1.jpg)  
+Figure 2: Architecture of LLaVA. We use the open-source LLaVA model as a testbed for our proposed data generation pipeline. The model is trained to predict the next tokens in the answers given the visual tokens and instruction tokens in an auto-regressive manner.
+
+sual instruction tuning datasets by leveraging text-to-image diffusion models.
+
+- To showcase its flexibility, we have built a dataset with various form of capabilities including multi-image data, and our results have shown improvements across all abilities.  
+- Extensive experimental analysis on multiple benchmarks shows the effectiveness of the proposed method, outperforming baseline and existing SOTA approaches.
+
+# 2 Related Work
+
+Recent research (Zhu et al., 2023; Liu et al., 2023c) efforts in multimodal Large Language Models (LLMs) have yielded promising strategies to efficiently align the embeddings of other modalities with language tokens. This has made it possible to effectively utilize pre-trained encoders from other modalities and LLMs, which effectively reduces the computational burden and training time. While there are alternative research approaches that include training-free methods leverage expert models (Wu et al., 2023; Shen et al., 2023; Yang et al., 2023a), these are not the focus of our work here.
+
+Adapter-based LLMs represent a significant research direction, introducing methods to connect modalities through learnable interfaces with minimal training efforts. These approaches (Zhang et al., 2023b; Gao et al., 2023; Luo et al., 2023; Liu et al., 2023c; Zhu et al., 2023; Yang et al., 2023b) allow for the use of pre-trained modal encoders, reducing the need for training from scratch. Variations include direct training of projection layers for embedding alignment and the use of learnable queries for extracting modality-specific information, as seen in models like LLaVA (Liu et al., 2023c) and Flamingo (Alayrac et al., 2022). Innovations such as the LLaMA-Adapter (Zhang et al.,
+
+2023b) and LaVIN (Luo et al., 2023) have introduced lightweight and mixed-modality adapters, respectively, enhancing the field's diversity.
+
+Visual instruction tuning datasets are crucial for training multimodal LLMs, focusing on aligning modalities and enabling instruction following. Most existing methods (Liu et al., 2023c; Rotstein et al., 2023; Li et al., 2023a; Yin et al., 2023; Li et al., 2023c; Liu et al., 2023a; Zhang et al., 2023c) rely on benchmark datasets for constructing visual instruction tuning datasets, which may be limited by the fixed categories in annotations. Our method leverages well-trained image generation models to produce controllable image data, enhancing multimodal LLM capabilities and allowing for the integration of advanced generative models for specific guidance forms, offering a more flexible and diverse approach to dataset construction. Recently, the concept of employing synthesized multimodal datasets has been extended to the training of models for chart and video understanding (Han et al., 2023; Song et al., 2024), showcasing its broad applicability and potential in advancing model training across various domains.
+
+# 3 Preliminary
+
+To assess the effectiveness of our data generation strategy, we chose the open-sourced LLaVA (Liu et al., 2023c,b) as our multimodal LLM model. It should be noted that our pipeline is model-agnostic, making it applicable for various models. This section serves as a foundation, briefly summarizing the LLaVA model's design and training methods to prepare for a thorough exploration of our pipeline. The reader may refer to the original publication (Liu et al., 2023c) for detail.
+
+Architecture. The LLaVA model integrates Vicuna-13B (Chiang et al., 2023) as the language model with a pre-trained CLIP visual encoder ViT-L/14 (Radford et al., 2021) for extracting visual
+
+![](images/1e06873223bc968bc28bd0c96d52e978d2967a86cc83db0d831353e9a56f0314.jpg)  
+Figure 3: Our proposed pipeline for generating visual instruction tuning datasets. We instruct ChatGPT to generate both StableDiffusion prompts and the associated dialogues. For specific generation templates, please refer to the supplementary materials.
+
+features, transforming these features into language embedding tokens through a linear layer. This linear layer was updated in LLaVA-1.5 (Liu et al., 2023b) with a two-layer MLP, replacing Vicuna-13B with Vicuna-13B-v1.5 and increasing input image size to  $336 \times 336$ . A detailed illustration of this model structure can be found in Figure 2.
+
+Training and datasets. LLaVA's training focuses on visual instruction tuning with data triplets: images, questions, and answers, aiming for predictive accuracy in an autoregressive manner. The training comprises two stages: the first emphasizes modality alignment using 595K image-text pairs, optimizing the linear layer with static visual encoder and LLM weights. The second stage, using 158K multimodal dialogue data from COCO, extends optimization to the LLM's weights for comprehensive modality integration. LLaVA-1.5 further enriches the dataset by incorporating additional data like Region-level VQA (Kazemzadeh et al., 2014; Krishna et al., 2017; Mao et al., 2016) and GQA (Hudson and Manning, 2019), expanding the second-stage dataset to 665K examples.
+
+# 4 Methods
+
+This section outlines our dual-generation approach for creating visual instruction tuning datasets, which synthesizes images and their corresponding dialogues, as illustrated in Figure 3. We detail each component below.
+
+# 4.1 Image Generation
+
+We employ StableDiffusion (Rombach et al., 2022) to generate images based on prompts that include weighted keywords reflecting aspects like subject, scene, style, and visual elements such as image quality and lighting. Keywords at the prompt's
+
+start are prioritized, with the possibility of adding emphasis using brackets. To encourage diversity and stability during image generation, we add capability-specific instructions and cautions during prompting ChatGPT. For instance, in the task of generating images for joke understanding, we direct ChatGPT to create prompts that would result in the generation of abnormal images, like a "giraffe walking through a narrow corridor", which are unlikely to be found in reality. When generating multi-image data, pairs of prompts can be generated concurrently based on predefined specific criteria. For maximum effect, we ensure that the most crucial keywords are placed at the beginning of the generated prompts, which are double-bracketed for additional emphasis. Furthermore, we instruct ChatGPT to avoid generating prompts that are non-visual, such as the act of growing. The instruction template for prompt generation is provided in the left part of Figure 4. The generated prompts are then used with StableDiffusion to produce visually realistic images, which are subsequently encoded by LLaVA's vision encoder into visual tokens for LLMs.
+
+# 4.2 Dialogue Generation
+
+Following the generation of images, we utilize ChatGPT to generate dialogues based on the same prompts used for image synthesis, aligning with LLaVA's training objectives: the first stage focuses on aligning visual-text data, and the second on processing diverse instructions. Dialogues for the initial stage describe the images, where ChatGPT generates answers to predefined questions about the images' content. Taking the example of the "giraffe walking through a narrow corridor", a representative dialogue might be: "Question: What
+
+![](images/1cd63202abd9e038cc74fb76b5f16a615389751d69f1e3ed0173ceb6f038d147.jpg)  
+Figure 4: Templates for guiding ChatGPT to generate StableDiffusion prompts (left) and dialogues (right). Content in red represents ability-specific information. We only provide an example template for constructing dialogues regarding a single image in this figure. For additional forms of data, such as multi-image reasoning and multi-turn dialogues, please refer to our supplementary materials.
+
+# ChatGPT for SD prompt Genera
+
+As an AI visual assistant specializing in prompt design, your task is to create prompts for a text-to-image generator related to the [Ability] keywords.
+
+# Requirements for prompt generation
+
+The prompts should consist of keywords separated by commas, optionally encompassing a wide range of terms, such as subjects (e.g., animals, humans, plants, vehicles), scenes (e.g., forest, beach, space station, castle), styles (e.g., realistic, abstract, impressionist, anime), and visual features (e.g., glowing, shiny, detailed shading, soft light, dramatic lighting, volumetric light, mysterious, vibrant, dreamy). Additionally, keywords like 'Details, intricate, highly detailed, fine carvings, 4k, 8k, high resolution' can be used to enhance image details. The images' content should always be logical and sensible. [Ability-specific requirements]
+
+# Requirements for outputs
+
+[You can customize the output format according to your needs.]
+
+# Examples
+
+You may choose the keywords from these examples: [Ability-specific keywords that are separately generated]
+
+Illustrated outputs are shown below: Prompt: [Ability-specific Prompt Examples]
+
+![](images/05b11d140212f374bd76ea53d97758dde7d5ee9be097cf6da22828b55d88a45c.jpg)
+
+# ChatGPT for Dialogue Generation
+
+As an AI visual assistant specializing in dialogue design, your task is to design dialogues based on the images generated by text-to-image generation models. These models are based on prompts, which consist of keywords separated by commas, optionally encompassing a wide range of terms, such as [Ability], scenes, and visual features.
+
+# Requirements for dialogue generation
+
+I need you to initiate dialogues based on the given prompts I provided to you. Upon receiving the prompts, your objective is to generate relevant questions about the keywords enclosed in double brackets. Make sure to diversify your questions. Clarity and precision are essential in both your questions and responses. The answers should strictly adhere to the questions and the given prompt, avoiding any nonsensical content. For each prompt, create distinct question-answer pair. To increase the complexity of the dialogue, you may introduce misleading either-or questions.
+
+# Requirements for outputs
+
+[You can customize the output format according to your needs]
+
+# Examples
+
+[You can set some examples for ChatGPT to refer to]
+
+Illustrated outputs are shown below:
+
+Prompt: [Ability-specific Prompt Examples]
+
+Question: [Ability-specific Question Examples]
+
+Response: [Ability-specific Response Examples]
+
+is unusual in the image? Response: In reality, a giraffe is too tall and big to walk through a narrow home corridor." The detailed instruction template for dialogue generation is shown in the right part of Figure 4.
+
+For the second stage, dialogues aim to enhance reasoning across multiple images, addressing similarities, differences, and logical connections, and include multi-turn dialogues that blend image and text. We guide ChatGPT to produce a range of question types, steering clear of inherently ambiguous questions to ensure clarity and accuracy, detailed template can be found in the supplementary materials.
+
+# 4.3 In-Context Examples
+
+ChatGPT's in-context learning capability, which allows it to grasp the essence of tasks from a few examples, is leveraged in our methodology. We incorporate in-context examples in the generation of StableDiffusion prompts and dialogues to enhance this learning process. During the data generation process, we observed that ChatGPT sometimes produced a lack of diversity. For example, when generating colors, the outputs frequently revolved around common color categories. To overcome this, we in-
+
+dependently generate ability-related keywords such as color categories with ChatGPT, and utilize them as a reference during the prompting process. This additional step promotes a more diverse range of prompts, thereby enriching our visual instruction tuning dataset.
+
+We further adopt a dynamic strategy to maintain and increase diversity: periodically substituting a portion of the original in-context examples with newly generated data. This continuous update prevents over-repetition and ensures the dataset's comprehensiveness and representativeness, maintaining a balance that contributes to a richer and more diverse visual instruction tuning dataset.
+
+# 4.4 Data Filtering Mechanism
+
+To augment the quality and diversity of generated dataset, we introduced a data filtering mechanism. This mechanism encompasses several key aspects:
+
+- Repetition Rate Filtering Initially, we filter the generated dataset based on the repetition rate of the extracted prompts. The selected prompts are subsequently utilized to generate corresponding images and dialogues, ensuring novelty and reducing redundancy.  
+- Length-based Filtering We keep prompts
+
+![](images/31eab6b5a095e011d635afc14378addcb2946ccbeb63c2f7556cc2a3aa1e4fd0.jpg)  
+Figure 5: Left: Results on evaluation benchmarks for various abilities (GPT-4 score). Right: Comparison of various subcategories on MMBench (Liu et al., 2023d) with the baseline (Accuracy). Our model outperforms the baselines on both benchmark datasets.
+
+![](images/a2acd7abafa96fbd0dd7508a2c263f9ce9b7eee1a6c9ab5d674b78b0f6b0c1b0.jpg)
+
+short, with no more than ten keywords, to make sure they're clear and focused. Long prompts can make it hard for model to create accurate images. We also limit dialogues to 500 characters to keep responses concise.
+
+- Task-Specific Restrictions For certain categories, we implemented restrictions based on specific attributes of the capabilities. For example, when generating content related to construction workers, the model tended to focus on buildings. To address this, additional human attributes were incorporated into the prompts to ensure a more accurate prompt.  
+- Alignment Check To ensure a high degree of alignment between generated images and dialogues, we employ the CLIP (Radford et al., 2021) model to compute matching scores for both texts and images. Data entries with scores exceeding a predefined threshold, set at  $\gamma = 0.25$ , are retained, thereby filtering out less relevant matches and elevating the overall data quality.
+
+# 5 Experiments
+
+In this section, we detail the experiments conducted to validate the effectiveness of our novel data collection approach for visual instruction tuning. We describe the training datasets, evaluation strategy, and both quantitative and qualitative outcomes.
+
+# 5.1 Training Datasets
+
+We generate a diverse and expansive dataset to show its versatility, covering single-image capabilities from basic recognition to complex visual reasoning. This includes understanding physical attributes, life features, and man-made items, among
+
+others, amounting to 38K image-dialogue pairs for initial training. Each ability's dataset was formulated following a standard template, illustrated in supplementary material. In addition, we also generated a dataset of 3K multi-image instances, encompassing descriptions of image similarity, difference, logical relations, and multi-turn dialogue data for the second stage. These datasets, in combination with the raw LLaVA dataset, provides a comprehensive training set in our experiments.
+
+# 5.2 Evaluation Metrics
+
+Evaluation datasets. To demonstrate our performance more clearly, we tested on a series of public multi-modal datasets, including VisWiz (Gurari et al., 2018), MM-Vet (Yu et al., 2023), MME (Fu et al., 2023), and MMBench (Liu et al., 2023d).
+
+Subsequently, we established a real-image benchmark to evaluate training effectiveness across a wide range of single-image abilities, including 330 test samples of real images with associated question-answer dialogues, carefully selected and annotated from public repositories. This comprehensive benchmark aims to rigorously test the models' single-image capabilities.
+
+Simultaneously, we constructed a multi-image test set consisting of 30 dialogues to assess the models' performance on this specific data type. This dataset evaluates the models across differences, similarities, and reasoning relationships among the images. The test data was sourced from publicly available datasets and manually annotated.
+
+Evaluation strategy In terms of the evaluation process, we employ different testing strategies depending on the benchmarks used.
+
+Our evaluation approach varies with the bench
+
+Table 1: Quantitative performance (Accuracy) on real-image evaluation benchmark for manual evaluation.  
+
+<table><tr><td>Method</td><td>Animal</td><td>Action</td><td>Color</td><td>Abnormal</td><td>Scene</td><td>Style</td><td>Food</td><td>Profession</td><td>Vehicle</td><td>Furniture</td><td>Plant</td></tr><tr><td>LLaVA</td><td>0.63</td><td>0.67</td><td>0.60</td><td>0.40</td><td>0.60</td><td>0.30</td><td>0.70</td><td>0.57</td><td>0.57</td><td>0.30</td><td>0.53</td></tr><tr><td>Ours</td><td>0.70</td><td>0.90</td><td>0.77</td><td>0.50</td><td>0.77</td><td>0.45</td><td>0.83</td><td>0.63</td><td>0.63</td><td>0.50</td><td>0.57</td></tr></table>
+
+mark. For multimodal datasets like VizWiz (Gurari et al., 2018) and MMBench (Liu et al., 2023d), we follow official guidelines, converting test data to a compatible format for our model, and using official scripts or submission portals for assessment, primarily focusing on accuracy.
+
+In evaluating the diverse capabilities we've generated, We adopted two evaluation methods, namely manual evaluation and evaluation based on GPT-4 score.
+
+Initially, the participants were instructed to assess the answers produced by our model and those of the baseline for all abilities according to the label (1 for correct, 0 for incorrect), averaging these as the final metric.
+
+Subsequently, inspired by (Chiang et al., 2023; Liu et al., 2023c), we leverage GPT-4 (OpenAI, 2023) to assist in scoring model outputs. We have established six scoring levels, ranging from 0 to 5. Each score level is accompanied by detailed descriptions of the evaluation criteria, and we assist GPT-4 in better assessment by providing a series of scoring examples. In particular, a score of 0 indicates that the predicted answer has no relevance to the reference answer, while a score of 5 signifies that the predicted answer aligns seamlessly with the annotated reference answer without any deviation. Drawing on our manual annotations, we conduct evaluations on the results produced by each model. The average GPT-4 score within each test set serves as the ultimate metric for our benchmark evaluations. The specific instruction template used for result evaluation is illustrated in supplementary material.
+
+# 5.3 Quantitative Comparison with State-of-the-arts
+
+Public multimodal benchmarks We perform quantitative performance comparisons against various state-of-the-art methods on different benchmarks, as illustrated in Table 2. Utilizing LLaVA-1.5-13B as the baseline, we integrate our synthesized data with its original dataset for training. Training is carried out with identical parameter configurations as LLaVA-1.5. The outcomes demonstrate substantial improvements on many benchmarks, emphasizing the enhanced performance
+
+Table 2: Quantitative comparison with other state-of-the-arts methods on multiple multimodal benchmarks. We achieve state-of-the-art performance on four benchmarks.  
+
+<table><tr><td>Method</td><td>VisWiz</td><td>MM-Vet</td><td>MME</td><td>MMB</td></tr><tr><td>BLIP2 (Li et al., 2023b)</td><td>19.6</td><td>22.4</td><td>1293.8</td><td>-</td></tr><tr><td>InstructBLIP (Dai et al., 2023)</td><td>33.4</td><td>25.6</td><td>1212.8</td><td>-</td></tr><tr><td>IDEFICS-9B (IDEFICS, 2023)</td><td>35.5</td><td>-</td><td>-</td><td>48.2</td></tr><tr><td>IDEFICS-80B</td><td>36.0</td><td>-</td><td>-</td><td>54.5</td></tr><tr><td>Qwen-VL (Bai et al., 2023)</td><td>35.2</td><td>-</td><td>-</td><td>38.2</td></tr><tr><td>Qwen-VL-Chat</td><td>38.9</td><td>-</td><td>1487.5</td><td>60.6</td></tr><tr><td>LLaVA-1.5 (Liu et al., 2023b)</td><td>53.6</td><td>35.4</td><td>1531.3</td><td>67.7</td></tr><tr><td>Ours</td><td>58.4</td><td>36.1</td><td>1532.3</td><td>69.4</td></tr></table>
+
+Table 3: Quantitative results (GPT-4 score) on the multi-image benchmark. After the addition of multi-image data, various multi-image capabilities have significantly improved.  
+
+<table><tr><td>Method</td><td>Difference</td><td>Similarity</td><td>Logical relations</td><td>Average</td></tr><tr><td>LLaVA</td><td>2.7</td><td>2.2</td><td>3.1</td><td>2.67</td></tr><tr><td>Ours</td><td>3.6</td><td>2.8</td><td>3.7</td><td>3.37</td></tr></table>
+
+achieved by our approach.
+
+Comparison of various abilities. To validate the effectiveness of our generated data, we conducted comprehensive tests on distinct capabilities, employing both manual evaluation and GPT-4 score assessments. Employing LLaVA-13B as our baseline, the quantitative comparison of the baseline results and ours for manual evaluation are shown in Table 1, while the GPT-4 score assessment results are shown in the left part of Figure 5. Notably, our trained model consistently outperforms the LLaVA-13B baseline across all various capabilities on two metrics, which suggests the synthesized datasets' generalizability and our pipeline's robustness.
+
+Besides, we conduct a comparison of subcategory performance on MMBench to better validate our superiority, using the LLaVA-1.5-13B as the baseline. The tested subcategories in MMBench encompass six aspects: attribute reasoning (AR), coarse perception (CP), fine-grained perception (cross-instance) (FP-C), fine-grained perception (instance-level) (FP-S), logic reasoning (LR), and relation reasoning (RR). The final results are shown in the right part of Figure 5, indicating better performance of subcategory on MMBench, which also attests to the high quality of our generated data.
+
+Comparison on multi-image benchmark In order to validate the effectiveness of multi-image
+
+Q: "Are there any disparities between these two images?"
+
+LLaVA: "Yes, there is a disparity between the two pictures. In the first picture, the car is parked on the side of the road, and the image is in black and white. In the second picture, the car is parked on the side of the road, but this time, it is in color. The car is an orange sports car, and it is parked in front of a building. The presence of the building and the color of the car in the second picture suggest that it is a different scene from the first black and white image."
+
+Ours: "The two images showcase different aspects of automotive design. The first image features a classic car with a shiny, yellow paint job, chrome rims, and a powerful engine. The car's sleek lines and bold color make it a standout on the road. In contrast, the second image showcases a more minimalist approach, with a simple black car that exudes elegance and sophistication. The car's smooth lines and understated design make it a timeless and versatile choice for any occasion."
+
+![](images/f3396f9c3aa2a00a6f401159c0d47017345b75b1157e826a8a274c4a7fd831db.jpg)
+
+Q: "Are there any implications about the relationship between the two pictures that we can derive?"
+
+LLaVA: "The two pictures, one showing a man holding a bundle of hay and the other showing a man standing in a field of hay, can be seen as complementary images that depict the same scene or activity. The man in the first picture is actively engaged in the process of harvesting hay, while the man in the second picture is standing in the field of hay, possibly observing the harvesting process or taking a break. The two images together provide a comprehensive view of the harvesting activity, illustrating the different stages and aspects of the process."
+
+![](images/8e397eb1cf7b2b4ff2eac2708c421a10658879a2bc85d072dbe2c748963f7674.jpg)
+
+Ours: "The two pictures, one showing a man holding a bundle of straw and the other showing a man riding a tractor in a field, suggest a connection between agricultural activities and the use of modern machinery in farming. The man holding the straw could be involved in harvesting the crops, while the man riding the tractor might be responsible for plowing, planting, or other field tasks. The presence of both traditional and modern farming equipment in the images highlights the evolution of agriculture and the integration of technology to improve efficiency and productivity in farming. This connection between the two pictures implies that the man in the field is likely involved in various aspects of crop cultivation from manual labor like harvesting to utilizing machinery for more efficient and time-saving tasks."
+
+![](images/a1ae703d7519e62f9bb4628bb477d50a10381f8e3be6a830766fdb87eb47b00e.jpg)
+
+Q: "What do these two images have in common?"
+
+LLaVA: "The two images share a common theme of tea being served in a cup. In both images, there is a cup filled with tea, and a person is holding the cup. The main difference between the two images is the setting and the presence of additional items. In the first image, the cup of tea is placed on a wooden table, and there is a spoon in the cup. In the second image, the cup of tea is placed on a wooden table, and there is a bowl of soup next to it. The person holding the cup is the same in both images, and their hand is visible, indicating that they are about to drink the tea."
+
+![](images/0f2a402bbc61d0ffce356ce898b079069fc49eb17e302768b39457b098294b95.jpg)
+
+Ours: "The common theme in these two images is the appreciation and enjoyment of a hot / beverage. The first image showcases the aroma and warmth of a cup of coffee, while the second image highlights the soothing experience of a hot cup of tea. Both beverages are popular choices for relaxation and comfort, offering a moment of respite from daily stressors."
+
+![](images/e6b951914474b97d8e884d17262c5c16198cc59a9a3be254483cf98509620662.jpg)  
+Figure 6: Comparison of the results generated by LLaVA and our trained model. Content in red represents inaccurate information. Our model can better adhere to question instructions, rendering more precise answers.
+
+capabilities, we manually curated a benchmark of real images. The evaluation metric used was the GPT-4 score mentioned. We used LLaVA-13B as the baseline and incorporated multi-image data in the second training phase. Since LLaVA itself lacks the capability for multi-image input, we modified the testing code for LLaVA to enable it to accept multiple sets of images simultaneously. The comparison with LLaVA results is shown in the Table 3, indicating a notable improvement across various multi-image capabilities despite adding less multi-image data in the process.
+
+# 5.4 Qualitative Results
+
+Supplementing the quantitative analysis, we provide a qualitative comparison between our model's results and LLaVA-13B in Figure. 6 on multi-image data. Our model exhibits a heightened ability to adhere to question instructions, rendering more precise answers.
+
+We compare our approach with the LLaVA-13B baseline, revealing its limitations: it struggles to differentiate between multi-image contents and
+
+provides incomplete answers to questions. Our method, incorporating multi-image data, enhances the model's understanding of multiple images, demonstrating its effectiveness. Additional qualitative results will be included in the supplementary materials.
+
+# 6 Conclusion and Future Work
+
+In the rapidly evolving realm of Large Language Models, efficiently integrating multimodal information is a key research focus. In this study, we introduced an innovative data collection method to enhance visual instruction tuning for multimodal models. Compared to existing strategies, our approach uniquely combines image and dialogue generation, effectively addressing limitations found in benchmark datasets. By carefully crafting instruction templates, our method ensures high-quality training data covering a broad range of crucial capabilities for multimodal models and users can generate customized data based on their specific requirements.
+
+Our research opens avenues for exploration.
+
+Moving forward, we aim to leverage advanced generative models to enhance model abilities, including spatial comprehension and fine-grained recognition. With promising results from our dual-generation method, forward-thinking data collection techniques are poised to play a significant role in the future of LLM research.
+
+# 7 Limitations
+
+Due to constraints in text-to-image models like stable diffusion, generating certain data types, such as text-rich images and tables, is not effective in the current pipeline. We anticipate these constraints will be addressed with ongoing advancements in text-to-image generation techniques.
+
+# 8 Ethics Statement
+
+Our method leverages generative models to create synthetic images and dialogues. It is imperative to ensure that the generated content does not perpetuate or amplify biases present in existing datasets or societal prejudices. We have implemented data filtering mechanism to minimize the generation of potentially harmful or biased content. However, continuous vigilance and improvement of these filters are necessary as generative models evolve.
+
+The enhanced capabilities of multimodal LLMs, facilitated by our data generation approach, could potentially be misused for creating deceptive or manipulative content. It is crucial to develop and adhere to guidelines that prevent the misuse of such technology, including transparent disclosure of synthetic content's nature and purpose.
+
+# 9 Acknowledgements
+
+This project is particularly supported by ARC DP210101347 and DP240101349.
+
+# References
+
+Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. 2022. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736.  
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966.
+
+James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. 2023. Dall-e 3: Bridging vision and language with few-shot image generation.  
+Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. 2021. Conceptual  $12\mathrm{m}$ : Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with  $90\%$  * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023).  
+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.  
+W. Dai, J. Li, D. Li, A. M. H. Tiong, J. Zhao, W. Wang, B. Li, P. Fung, and S. Hoi. 2023. Instructclip: Towards general-purpose vision-language models with instruction tuning. 2, 3, 4, 5, 6.  
+Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. 2023. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394.  
+Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. 2023. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010.  
+Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. 2018. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617.  
+Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. 2023. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv: 2311.16483.  
+Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709.  
+IDEFICS. 2023. Introducing idefics: An open reproduction of state-of-the-art visual language model.
+
+Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. 2014. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798.  
+Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. 2017. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73.  
+Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. 2023a. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425.  
+Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023b. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597.  
+Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang, Jingjing Xu, Xu Sun, et al. 2023c. M3 it: A large-scale dataset towards multi-modal multilingual instruction tuning. arXiv preprint arXiv:2306.04387.  
+Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer.  
+Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. 2023a. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023b. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023c. Visual instruction tuning. arXiv preprint arXiv:2304.08485.  
+Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. 2023d. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281.  
+Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. 2023. Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023.  
+Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. 2016.
+
+Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20.  
+OpenAI. 2023. Chatgpt. https://openai.com/blog/chatgpt/.1,2.  
+OpenAI. 2023. Gpt-4 technical report. arXiv, pages 2303-08774.  
+Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR.  
+Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695.  
+Noam Rotstein, David Bensaid, Shaked Brody, Roy Ganz, and Ron Kimmel. 2023. Fusecap: Leveraging large language models to fuse visual data into enriched image captions. arXiv preprint arXiv:2305.17718.  
+Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. 2022. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294.  
+Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2023. Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580.  
+Zhende Song, Chenchen Wang, Jiamu Sheng, Chi Zhang, Gang Yu, Jiayuan Fan, and Tao Chen. 2024. Moviellm: Enhancing long video understanding with ai-generated movies. arXiv preprint arXiv:2403.01422.  
+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.  
+Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. 2023. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671.
+
+Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. 2023a. Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv preprint arXiv:2305.18752.  
+Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. 2023b. Exploring diverse in-context configurations for image captioning.  
+Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Lu Sheng, Lei Bai, Xiaoshui Huang, Zhiyong Wang, et al. 2023. Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. arXiv preprint arXiv:2306.06687.  
+Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. 2023. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490.  
+Chi Zhang, Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. 2023a. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771.  
+Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. 2023b. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199.  
+Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. 2023c. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107.  
+Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592.
+
+# Appendix for "Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data"
+
+# A Training Details
+
+During the model training phase, we employed the original LLaVA configuration as the foundation for our training process. In both stages, we utilized 8 NVIDIA V100 GPUs. To conserve GPU memory, we employed deepspeed with zero3 during model training, disabling tf32 and opting for fp16. The remaining parameters, including epochs and learning rates, were set according to the original LLaVA configuration. For specific parameter details, please refer to the original publication.
+
+# B Multi-image Generation Template
+
+We have illustrated the instruction template for generating multi-image data in Figure 7. To enhance the diversity of the data, we first use ChatGPT to generate a series of paired prompts. Subsequently, based on these paired prompts, we generate corresponding captions. As shown in the Figure 7, we can generate different types of data by configuring different examples and replacing similarities with difference or logical relations.
+
+# C Interleaved multi-turn dialogues Generation Template
+
+We present a template for generating interleaved multi-turn dialogues as Figure 8. Similar to multi-image generation, we randomly generate a series of non-repeating phases to enhance data diversity. Based on these phases, we generate interleaved multi-turn dialogues. The red sections are replaceable. In our generated dialogues, the focus is primarily on instructional content related to the use of everyday objects and recipes. Users can make specific adjustments according to their needs. Finally, the brackets part represents the prompt used for stable diffusion to generate images. During the model training process, the content within the brackets will be replaced with an image_placeholder.
+
+# D GPT-4 Scoring Criteria
+
+As shown in the Figure 9, we present our detailed GPT-4 scoring system. We have established a scale of 0-5 with six levels of scores, and for each score, we provide detailed evaluation criteria along with specific examples for assessment. Utilizing the
+
+template in the Figure 9, evaluations are conducted for each model, and the average of the results is taken as the final score.
+
+# E More qualitative Results.
+
+We showcase more generated data as shown in the Figure 10 11, which includes various capabilities, as well as multi-image dialogues and interleaved multi-turn dialogues, to validate the effectiveness and diversity of our pipeline.
+
+We present additional outcomes of our methodology on more real images in Figure 12. Spanning diverse capabilities like profession recognition, color categorization, or anomaly detection, our approach consistently exhibits efficacy. In particular, our method demonstrates a robust capability in discerning abnormal images.
+
+Simultaneously, we validated our model on multi-turn interactive dialogues, and the results are shown in the Figure 13 14. It can be observed that our model can provide detailed descriptions of each step, which is attributed to the high-quality data we generated.
+
+![](images/71bb36b000289a58931d8f5f89864a3e1b0c176271874821298507ed8ece3c27.jpg)  
+Figure 7: Instruction template for multi-image data generation.
+
+# ChatGPT for
+
+# Paired prompt Generation
+
+As an AI visual assistant, your responsibility is to create one pair prompt for a text-to-image generator. These models rely on prompts composed of keywords separated by commas, which can optionally cover a wide variety of terms. The core themes of the pair prompts can be varied, including animals, plants, people, vehicles, buildings, colors, and more. Paired prompts can be similar types of animals, plants, furniture, vehicles, buildings, professions, furniture, and same object in different colors etc.#
+
+# Requirements for outputs
+
+The final output should follow a structured format: first, print the first prompt; second, print the second prompt.
+
+# Examples
+
+[You can set some examples for ChatGPT to refer to.]
+
+Illustrated outputs are shown below:
+
+Prompt 1: [Prompt Examples]
+
+Prompt 2: [Prompt Examples]
+
+![](images/4f71390be0ff8060ed5ff285c1fea1247a7283b55bcdb94e15b0c5808fee06f3.jpg)
+
+# ChatGPT for
+
+# Multi-image Generation
+
+As an AI visual assistant, your responsibility is to create one caption about similarity of the given paired prompt for a text-to-image generator. You are required to describe the similarity between the two images based on the paired prompt. The caption should start from the specific content of the image, rather than simply listing a few keywords.
+
+# Requirements for outputs
+
+The final output should follow a structured format: first, print the first prompt; second, print the second prompt; third, print the caption.
+
+# Examples
+
+[You can set some examples for ChatGPT to refer to.]
+
+Illustrated outputs are shown below:
+
+Prompt 1: [Prompt Examples]
+
+Prompt 2: [Prompt Examples]
+
+Caption: [Caption Examples]
+
+![](images/2f4bfc79c35f7cd74169cbbe70428632da9869f09cd0f9b5299257fb712d2afe.jpg)
+
+# ChatGPT for
+
+# Interleaved multi-turn dialogues Generation
+
+As an AI visual assistant, your task is to construct a multi-turn conversation in which one person teaches another person to do something step by step according to the given phrase. Then the person being taught needs to provide a current picture and then ask the other person. The picture content provided is enclosed in  $\langle \langle \rangle \rangle$  Get up, inside is the content description of this image. In each step, there should be one picture content. .#
+
+# Requirements for outputs
+
+[You can customize the output format according to your needs.]
+
+# Examples
+
+[You can set some examples for ChatGPT to refer to.]
+
+A: Can you guide me on making a sandwich? Here's a picture of my kitchen counter «A photo of a kitchen counter with various ingredients such as bread, lettuce, tomatoes, and deli meats. There's also a knife and a cutting board.».
+
+B: Great! First, take two slices of bread from the bag on the counter.
+
+A: I've got the bread slices. What's the next step? 《A photo of the bread slices in hand, ready to be used.》
+
+B: Place the slices on the cutting board. Now, grab some lettuce and tomatoes.
+
+A: Okay, I've placed the bread on the cutting board, and I have lettuce and tomatoes. What's next? «A photo of the arranged bread slices with lettuce and tomatoes on the cutting board.»
+
+B: Add a layer of deli meat on top of the lettuce and tomatoes.
+
+A: Done! I've added the deli meat. What's the final step? «A photo of the sandwich with layers of ingredients, ready to be enjoyed.»
+
+B: Place the second slice of bread on top to complete your sandwich. Enjoy your meal!
+
+Illustrated outputs are shown below:
+
+[Interleaved multi-turn dialogues Examples]
+
+Figure 8: Instruction template for interleave multi-turn dialogues generation.
+
+![](images/5f77333837bd02bd822a35f7bd21afebbe2010b20fac2808611941ce7a0a2008.jpg)  
+Figure 9: Score criteria based on GPT-4.
+
+# ChatGPT
+
+As an AI visual assistant, your ability is evaluating predicted answer based on the given question and correct answer. Now I have a predicted answer and corresponding question, and I need to compare predicted answer with the correct answer. Here's a scoring standard you can use to assess. A score of 5 indicates that the response perfectly aligns with the answer, while a score of 1 signifies that the response is entirely different from the answer.
+
+# **Scoring Key (5-0 Point Scale)**
+
+**Exact Correlation (5 points)**
+
+The predicted answer strictly replicates the reference answer. It accurately perceives and elaborates on the queried element.
+
+**Almost Exact Correlation (4 points)**
+
+The predicted answer is not verbatim but delivers a notion akin to the reference answer.
+
+Slight variations in phraseology exist, but the core idea remains consistent.
+
+May neglect a minor item, but this does not incite confusion or misinterpretation.
+
+**Moderately Accurate (3 points)**
+
+The predicted answer captures the chief matter but may omit key details.
+
+It could contain data or semantics that may be misguiding.
+
+Understands the key intent but lacks exactness.
+
+**Somewhat Accurate (2 points)**
+
+The predicted answer is loosely related to the reference answer with conspicuous discrepancies. It correctly portrays a segment of the required details but bypasses or distorts major aspects.
+
+Significant corrections are necessary to match this response with the reference answer.
+
+**Barely Related (1 point)**
+
+The predicted answer barely links to the actual object or query.
+
+Mostly incorrect or misleading, exhibiting trivial understanding.
+
+A viewer or user would be predominantly misguided based on this response.
+
+**Completely Inaccurate (0 points)**
+
+The predicted answer shares no relevance to the reference answer.
+
+Utterly misinterprets or overlooks the queried item.
+
+No discernible cognition of the assessment subject is evident.
+
+![](images/fc7634cef96c3e3c95d25843577c590310005d046a2d7cc8cb80f62576376ec2.jpg)
+
+Prompts: "Pink flowers, butterfly, ..."
+
+Q: "Are the flowers in the image red?"
+
+A:"No, the flowers in the image are pink"
+
+![](images/24f1ceb85177b337f42e59ac386dab8b24222aa8ec68c44086a172f055f994a7.jpg)
+
+Prompts : "penguin in the desert..."
+
+Q:"What is interesting about this scene?"
+
+A:Penguins are native to colder climates like Antarctica, it's abnormal to see one in a desert."
+
+![](images/40a1cd58a05e52b6fa35cab58d2c3486091129e7d45e0cddc16a89ba03e53dca.jpg)
+
+Prompts: "Lunar colony, realistic ..."
+
+Q: "What scene is depicted in the picture?"
+
+A:"The image shows a futuristic lunar colony."
+
+![](images/f2c8d21b3a9ccb871ac9634671938dda6ff1fef73743ff365ca312fe60e81985.jpg)
+
+Prompts: "Ink wash painting, temple ..."
+
+Q: "What is the painting style of this image?"
+
+A:"In terms of style, this picture leans towards ink wash painting."
+
+Q: "Is there any inference to be made concerning the link between these two visuals?"
+
+A:"The city skyline was dotted with skyscrapers, while a construction crane worked on building a new tower."
+
+![](images/3d24a8a80515cb333855d2892f22037a52140881c388c0896f0b3358a6e38927.jpg)
+
+![](images/14cb770416a7417ecc90395c8be48981f9e5fe5926c36af9d52561dd05ce82e6.jpg)
+
+Q: "Are there any common elements between these two pictures?"
+
+A:"The common theme between these two images is the world of automobiles, with the first image showcasing a modern, fast sports car with a sleek design that exudes speed, while the second image highlights a sturdy truck built for rugged durability. Both images capture the essence of power and performance in the automotive industry, albeit with different focuses on style and functionality."
+
+![](images/065d4e16fbfe0a1622f8064cdc33f179f2ea0faa0f1850e10202b858e680ae79.jpg)  
+Figure 10: Qualitative results: Demonstrations of diverse data we generated.
+
+![](images/5bd31b39d5f667d844cfb28237c58ea6f59630412a63b4716a3187a5b1981606.jpg)
+
+Q: "Can you teach me how to bake a chocolate cake? This is a picture of my kitchen."
+
+A: "Sure! First, preheat your oven to 350 degrees Fahrenheit."
+
+Q: "Okay, the oven is preheating. What should I do next?"
+
+A: "Next, mix together 2 cups of flour, 2 cups of sugar, 3/4 cup of cocoa powder, 2 teaspoons of baking powder, and 1 1/2 teaspoons of baking soda in a large bowl."
+
+Q: "I've mixed all the dry ingredients together. What's next?"
+
+A: "Now, add 1 cup of milk, 1/2 cup of vegetable oil, 2 eggs, and 2 teaspoons of vanilla extract to the bowl and mix until everything is well combined."
+
+Q: "I've added all the wet ingredients and mixed everything together. What's the next step?"
+
+A: "Grease two 9-inch round cake pans and divide the batter evenly between them."
+
+Q: "I've greased the pans and divided the batter. What should I do next?"
+
+A: "Bake the cakes for 30-35 minutes or until a toothpick inserted into the center comes out clean."
+
+Q: "The cakes are in the oven. What should I do while they bake?"
+
+A: "While the cakes bake, you can make the frosting. Melt 1/2 cup of butter and 2/3 cup of cocoa powder in a saucepan over low heat. Then, stir in 3 cups of powdered sugar and 1/3 cup of milk until the frosting is smooth."
+
+Q: "I've made the frosting. What should I do next?"
+
+A: "Once the cakes are done baking, let them cool completely before frosting them. Then, spread the frosting evenly over one cake layer and place the other cake layer on top. Finally, frost the top and sides of the cake."
+
+![](images/7730cfb25483ea5057adbea7246dfad70fd59fd731e0ad76ebac60ec33e10301.jpg)
+
+![](images/77ca32842dcfc8d98bd7ee238c62fc11fa6c01ca0d7a218d718035c710644664.jpg)
+
+![](images/3e4049146f43605688077a2e64b617b212dcdcef0cae2fdbab608903c5dd7120.jpg)
+
+![](images/e38dbf600a0eb9dec2247aa8dcea818722281d9afeb3192dcb41010b75a3d3e6.jpg)
+
+![](images/4f7705572e22583b619da0fc27f6e9c37382b8daf0c43e0944fc0832d95a9908.jpg)
+
+![](images/4eba4330a0d3e8b8161f23b6697548bddf965064822208b5a7d16446510bc4e2.jpg)
+
+![](images/094ecca7d6ce9806d320a63883ed504a6904a367b7e87952491dc9fb962b7b5a.jpg)  
+Figure 11: Qualitative results: Demonstrations of diverse data we generated.
+
+![](images/448846ac7a1691121422358212d4055f69e4f0aea563cf3efb70cd2975bf0834.jpg)
+
+Q:"What is unusual in this image?"
+
+R:"This image is unusual because it is a playful photo manipulation that combines a photo of a dog's face with a body of a person wearing a suit and tie. The man's head has been replaced with the dog's face, giving the appearance of a person with a dog's head wearing a suit. The image creates a humorous and unconventional visual that breaks the norm of human-like appearances."
+
+![](images/30ce1a086f01af541f86031424a12748cce8ff028eae24de9522caafa60d5626.jpg)
+
+Q: "What is the person doing in the image?"
+
+R:"The person in the image is jumping in the air near a garage door."
+
+![](images/86432baf81ae510868be6c9c7eca9266ee0b958fa2fbfd32e59e99f4f0e89c22.jpg)
+
+Q:"What medium was used to create this artwork?"
+
+R:"The medium used to create this artwork is silhouette."
+
+![](images/0f384888763d7b160fff89b75aa41cbb157dbbe9311233cef128be1e4ea20454.jpg)
+
+Q:"What is the scene depicted in the image?"
+
+R:"The image features a scenic waterfall with lush greenery around it, situated in a mountainous area."
+
+Q:"What is the color of the four bulbs?"
+
+R:"The colors of the four light bulbs are purple, blue, green, and yellow."
+
+![](images/77fcd85913b742ae33d26386ebc7a919d21e223069c7cb8d1e9bc951109f82b0.jpg)
+
+Q: "What is the profession of the woman in front of the house?"
+
+R:"The woman in front of the house is a firefighter. It's her job to put out fires, and she is also responsible for other emergency situations."
+
+![](images/8b5f08740077caa66d63793ac0328f352dfc793c15a4dd5e96ff5050523ff150.jpg)
+
+Q:"What kind of vehicle is seen in the image?"
+
+R:"There is a sailboat on the ocean, sailing on the water."
+
+![](images/79d50d377c196b3415f5ba78b158c19b401d7bf325f9a1a3eaf710c6aeb72359.jpg)
+
+Q:"What stands out as abnormal in this image?"
+
+R:"This image features a bicycle parked on a sidewalk, with an unusual twist: the bike has a giant watermelon wheel instead of a traditional tire. This is not something you see every day and makes the scene quite unique and visually intriguing."
+
+![](images/62d5cfe8aa4c24ac87cf2b68e76123031908504a4011d317f6a7f47a950e5708.jpg)  
+Figure 12: Qualitative results: Demonstrations of our method's effectiveness across diverse real-world image scenarios.
+
+Q: "Can you teach me how to tie a necktie? here's a picture of the tie I want to wear."
+
+Ours-A: "Of course! I'd be happy to help you learn how to tie a necktie. First, Start by holding the tie with the wide end facing upwards and cross the wide end over the narrow end, creating an X shape."
+
+Q: "Okay, I've done it. What's the next step?"
+
+Ours-A: "Next, bring the wide end of the tie under the narrow end, and then wrap it around the collar of your shirt."
+
+Q: "Alright, what's next?"
+
+Ours-A: "Next, bring the wide end of the tie up over the narrow end, and then tuck it under the narrow end. This will create a loop."
+
+Q: "Got it, I've made the loop, then what should I do, tell me last step?"
+
+Ours-A: "Great job! Now, take the wide end of the tie and bring it up through the loop you just d. Then, tuck the wide end under the narrow end of the tie, and finally, adjust the tie to the desired length and position. You should now have a perfectly tied necktie!"
+
+![](images/e1498bbe043370ac4d0ce1a421bf07688c8127ea6f286e2fd16c8af0cbb3b425.jpg)
+
+![](images/62b10bbc4d2ef696b6b165d0a89d921a16070a363ff7607a480e7b6aef59e0c8.jpg)
+
+![](images/4c37f8b406b39553b70743dd82633746ed19fc86d437c7eac070ef2362748a7a.jpg)
+
+![](images/698b65e91a6701b0e24c92936ffc466b93032e0330e9f6c36f9eaad8246935a1.jpg)  
+Figure 13: Qualitative results: Demonstrations of our method's effectiveness across interleaved multi-turn dialogues.
+
+Q: "Can you teach me how to brew a cup of robust Turkish coffee? This is a picture of my kitchen."
+
+Ours-A: "Sure! First, fill the bottom of the Turkish coffee pot with cold water and place it on the stove."
+
+Q: "Ok, I've done that. What's next?"
+
+Ours-A: "Next, add a teaspoon of coffee to the pot and place it on the stove. Turn the heat to medium and let it boil for about 30 seconds. Once it starts to boil, remove the pot from the heat and let it sit for a few seconds."
+
+Q: "Got it, what's next?"
+
+Ours-A: "Finally, add a teaspoon of sugar to the pot and stir it in, pour the coffee into a small cup and enjoy!"
+
+![](images/9ead6514fba17eaf164b49abd39d32850ad2ce53368cee14aba3487e465c3786.jpg)
+
+![](images/4ac49c191d0ce57e76d017f0f220e72708ae548dd7e04b37ea10338562b89c64.jpg)
+
+![](images/6b53274463afd610fca32a56fe6ebc062cc726bead44151820ab5d5536ee6b93.jpg)  
+Figure 14: Qualitative results: Demonstrations of our method's effectiveness across interleaved multi-turn dialogues.

2024/Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3d59636211105291df8ffcbf7213100225ecd385fc2af3cc1e6c84b452d8033
+size 507821

2024/Enhancing Adverse Drug Event Detection with Multimodal Dataset_ Corpus Creation and Model Development/5f2bd2d8-1c47-4ff0-bd6b-5d72a8900ae6_content_list.json

@@ -0,0 +1,1716 @@
+[
+    {
+        "type": "text",
+        "text": "Enhancing Adverse Drug Event Detection with Multimodal Dataset: Corpus Creation and Model Development",
+        "text_level": 1,
+        "bbox": [
+            147,
+            86,
+            852,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Pranab Sahoo<sup>1</sup>, Ayush Kumar Singh<sup>1</sup>, Sriparna Saha<sup>1</sup>, Aman Chadha<sup>2,3*</sup>, Samrat Mondal<sup>1</sup>",
+        "bbox": [
+            110,
+            139,
+            892,
+            156
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Department of Computer Science And Engineering, Indian Institute of Technology Patna",
+        "bbox": [
+            141,
+            156,
+            865,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Stanford University,  $^{3}$ Amazon GenAI",
+        "bbox": [
+            341,
+            179,
+            657,
+            197
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "pranab_2021cs25@iitp.ac.in, ayush_2211ai27@iitp.ac.in, sriparna@iitp.ac.in, samrat@iitp.ac.in hi@aman.ai",
+        "bbox": [
+            78,
+            199,
+            927,
+            224
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The mining of adverse drug events (ADEs) is pivotal in pharmacovigilance, enhancing patient safety by identifying potential risks associated with medications, facilitating early detection of adverse events, and guiding regulatory decision-making. Traditional ADE detection methods are reliable but slow, not easily adaptable to large-scale operations, and offer limited information. With the exponential increase in data sources like social media content, biomedical literature, and Electronic Medical Records (EMR), extracting relevant ADE-related information from these unstructured texts is imperative. Previous ADE mining studies have focused on text-based methodologies, overlooking visual cues, limiting contextual comprehension, and hindering accurate interpretation. To address this gap, we present a MultiModal Adverse Drug Event (MMADE) detection dataset, merging ADE-related textual information with visual aids. Additionally, we introduce a framework that leverages the capabilities of LLMs and VLMs for ADE detection by generating detailed descriptions of medical images depicting ADEs, aiding healthcare professionals in visually identifying adverse events. Using our MMADE dataset, we showcase the significance of integrating visual cues from images to enhance overall performance. This approach holds promise for patient safety, ADE awareness, and healthcare accessibility, paving the way for further exploration in personalized healthcare. The code and dataset used in this work are publicly available<sup>1</sup>.",
+        "bbox": [
+            144,
+            281,
+            460,
+            764
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Disclaimer: The article features images that may be visually disturbing to some readers.",
+        "bbox": [
+            142,
+            770,
+            458,
+            799
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            813,
+            258,
+            828
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "An adverse drug event (ADE) encompasses any harm resulting from medication use, whether it's",
+        "bbox": [
+            112,
+            839,
+            487,
+            871
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "unintended, off-label, or due to medication errors. Adverse drug reactions (ADRs) are a specific type of ADE, denoting unexpected harm arising from the proper use of medication at the prescribed dosage. Injuries from inappropriate or off-label use are not classified as ADRs (Karimi et al., 2015b). ADEs pose significant public health concerns, contributing to numerous fatalities, serious injuries, millions of hospitalizations, and prolonged hospital stays. Consequently, they impose substantial financial burdens, costing healthcare systems billions of dollars globally. Despite advancements in healthcare, ADE detection remains a significant challenge. Implementing effective detection and monitoring strategies can substantially mitigate the adverse impacts on patients and healthcare systems (Hakkarainen et al., 2012), (Yadav et al., 2018b), (Sultana et al., 2013). Most of the previous ADE detection works are based on text data only (D'Oosterlinck et al., 2023), (Sarker and Gonzalez, 2015), (Yadav et al., 2019), (Sarker et al., 2016), (Chowdhury et al., 2018), (Yadav et al., 2018a), which presents a significant disadvantage due to its subjective nature and lack of specific details of visual cues, leading to potential inaccuracies and incomplete categorization of ADE detection. Despite extensive research, the potential of integrating textual data with visual information, such as images, has been largely overlooked. Visual aids are essential in ADE detection for numerous reasons. A substantial proportion of the population lacks proficiency in medical jargon, hindering accurate symptom descriptions. Moreover, certain symptoms are inherently challenging to express through text alone. Patients may struggle to differentiate between similar symptoms, like skin rash, eczema, peeling, and blister. As depicted in Fig. 1, sample images may present confusion to individuals lacking adequate medical expertise. Integrating both text and images in these scenarios can enhance the accuracy and effectiveness of ADE detection,",
+        "bbox": [
+            507,
+            253,
+            885,
+            912
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Work does not relate to position at Amazon.",
+        "bbox": [
+            139,
+            879,
+            418,
+            892
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "'https://github.com/singhayush27/MMADE.git",
+        "bbox": [
+            115,
+            892,
+            485,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "11214",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2024, pages 11214-11226",
+        "bbox": [
+            220,
+            945,
+            778,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "August 11-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            284,
+            959,
+            712,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "offering a comprehensive understanding of the patient's current medical condition. To the best of our knowledge, ADE detection using both image and text data has not been explored previously, and we take this opportunity to introduce a MultiModal Adverse Drug Event (MMADE) dataset comprising ADR images paired with corresponding textual descriptions.",
+        "bbox": [
+            112,
+            84,
+            490,
+            212
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Large Language Models (LLMs) and Vision Language Models (VLMs) have exhibited remarkable skills in generating human-like text, prompting their integration into various medical applications, including tasks such as chest radiography report generation, summarization, and medical question answering (Thawkar et al., 2023), (Ghosh et al., 2024b), (Sahoo et al., 2024b), (Ghosh et al., 2024a). However, their potential in ADE detection, which involves both text and images, has yet to be explored. Leveraging LLMs and VLMs for this task presents inherent limitations as they are predominantly trained on generic natural images sourced from databases like ImageNet, Wikipedia, and the internet. Generic models may not possess the specialized medical knowledge required for comprehensive caption generation, potentially leading to oversimplified descriptions that overlook essential details like symptoms and medical intricacies. Furthermore, while VLMs have excelled in traditional visual-linguistic tasks, their application to medical imaging presents unique challenges that may hinder the accurate interpretation and description of complex medical images (Sahoo et al., 2024a). Specialized models such as XrayGPT (Thawkar et al., 2023) and SkinGPT4 (Zhou and Gao, 2023), which are trained on chest X-ray and skin disease images, exemplify the domain specificity required for accurate medical image analysis. This has led us to explore ADR detection within a multimodal framework. To support this exploration, we introduce MMADE, a carefully curated dataset crafted for this specific purpose. MMADE consists of 1500 instances of patient-reported concerns regarding drugs and associated side effects, each paired with both textual descriptions and corresponding images. In our study, we have employed Instruct-BLIP (Dai et al., 2023), which builds upon the strong foundation of BLIP-2 (Li et al., 2023), a pretrained model with high-quality visual representation and strong language generation capabilities. The meticulous fine-tuning process enables it to bridge the disparity between general-purpose mod",
+        "bbox": [
+            112,
+            227,
+            490,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/79d79f0afe3c59354305feaeb62f41befd7407dd1a808e6da2088aec5ab94eb1.jpg",
+        "image_caption": [
+            "After 1 hour of taking the penicillins!! I get this."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            80,
+            621,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/09f0e23cf58d401f6ae1beaf0d58bec4a559c894d659c1d7688635e8f4d343b1.jpg",
+        "image_caption": [
+            "I found this after my doctor prescribed Etoricoxib for post-surgical dental pain."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            636,
+            80,
+            744,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/7dafcaf297647a24f155b2e3dd2d480ee8b373b714197a0130c643d5abe5e691.jpg",
+        "image_caption": [
+            "This is very unusual to me after taking amoxicillin for pharyngitis.",
+            "Figure 1: Samples from the dataset highlight the significance of visual cues in understanding adverse drug events, particularly in cases where patients are unaware of a specific medical condition."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            759,
+            80,
+            870,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "els and the specialized demands of ADE-specific tasks. Moreover, our exploration of BLIP (Li et al., 2022) and GIT (Wang et al., 2022) reveals that these models exhibit insufficient performance before fine-tuning. Nevertheless, upon fine-tuning with domain-specific data, their performance experiences notable improvement.",
+        "bbox": [
+            507,
+            275,
+            884,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our key contributions are as follows:",
+        "bbox": [
+            527,
+            388,
+            803,
+            404
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- A novel approach to ADE detection in multimodal settings greatly assists medical professionals, such as doctors, nurses, and pharmacists, by delivering detailed descriptions of ADE cases, enhancing precision in diagnosis, treatment planning, and patient care.",
+            "- Introduction of a novel multimodal dataset MMADE for further research on ADE detection area.",
+            "- The proposed dataset demonstrates promising potential for various applications, including ADE classification, caption generation, and summarization tasks.",
+            "- We have utilized InstructBLIP and experimented with two other pre-trained VLMs, and reported a detailed analysis.",
+            "- The ADE-specific model holds promise for enhancing patient safety, ADE awareness, and healthcare communication. It aims to provide individuals seeking information about ADEs with understandable and informative captions accompanying medical images to improve their comprehension of potential medication risks."
+        ],
+        "bbox": [
+            531,
+            414,
+            884,
+            839
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Works",
+        "text_level": 1,
+        "bbox": [
+            509,
+            862,
+            672,
+            876
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This section details the related ADE detection tasks based on the data sources.",
+        "bbox": [
+            507,
+            887,
+            880,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "11215",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Works Based on Biomedical Text and Electronic Medical Record: Various techniques have been developed for extracting ADEs from Electronic Medical Records (EMRs) (Aramaki et al., 2010), (Wang et al., 2009), as well as from medical case reports (MCRs) (Gurulingappa et al., 2011). Gurulingappa et al. (2012a) utilized machine learning methods to identify and extract potential ADE relations from MEDLINE case reports. Unlike random data sources such as social media, both EMRs and MCRs offer significant advantages by providing comprehensive records of a patient's medical history, treatment, conditions, and potential risk factors. Moreover, these records are not limited to patients who have experienced ADRs (Harpaz et al., 2013). Sarker and Gonzalez (2015) conducted a study by taking data from MEDLINE case reports and Twitter and reported how combining different datasets increases the performance of identifying ADRs. Huynh et al. (2016) explored various neural network frameworks for ADE classification, utilizing datasets from both MCRs and Twitter. DISAE is another corpus (Gurulingappa et al., 2010), which consists of 400 MEDLINE articles with annotations for disease and adverse effect names without drug-related information. D'Oosterlinck et al. (2023) introduced the BioDEX dataset for biomedical ADE extraction in real-world pharmacovigilance. This dataset comprises 65,000 abstracts, 19,000 full-text biomedical papers, and 256,000 document-level safety reports crafted by medical professionals. However, to the best of our knowledge, there is currently no publicly accessible annotated multimodal (Image and Text) corpus suitable for identifying drug-related adverse effects.",
+        "bbox": [
+            112,
+            84,
+            490,
+            664
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Works Based on Social Media Datasets: Social media has become vital for accessing vast amounts of real-time information, making it valuable for identifying potential ADEs. Leaman et al. (2010) conducted a pioneering study that analyzed user comments from social media posts, comprising a dataset of 6890 comments. The research demonstrated the significant value of user comments in identifying ADEs, highlighting their crucial role in this context. Several other authors (Gurulingappa et al., 2012b), (Yadav et al., 2020), (Benton et al., 2011) employed lexicon-based approaches to extract ADEs. However, these methods are limited to a specific set of target ADEs. Nikfarjam and Gonzalez (2011) em",
+        "bbox": [
+            112,
+            677,
+            490,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Keywords used for the data collection.",
+        "bbox": [
+            534,
+            82,
+            855,
+            97
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Keywords: ADE, ADR, adverse drug reaction, adverse drug event, adverse reaction, adverse drug event reporting, side effects, drug reactions, drug side effects, type of infection and reaction, medicine, drugs, skin rashes, red patches, eczema, ulcer, acne, skin irritation, edema, rosacea, alopecia, lip swelling.",
+        "bbox": [
+            524,
+            110,
+            878,
+            239
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ployed a rule-based technique instead of a naive lexicon-based approach on the same dataset, enabling the detection of ADEs not covered by lexicons. Several authors utilized supervised machine learning techniques like Support Vector Machines (SVM) (Sarker and Gonzalez, 2015), Conditional Random Fields (CRF) (Nikfarjam et al., 2015), and Random Forests (Zhang et al., 2016) for ADE detection. Sarker et al. (2016) introduced one corpus by collecting data from social media, focusing on adverse drug reactions. Tasks included automatically classifying user posts, extracting specific mentions, and normalizing mentions to standardized concepts. With the availability of annotated data, in recent times, the rise of deep learning techniques has significantly influenced research methodologies, leading to the adoption of deep learning models for predicting ADEs. Tutubalina et al. (2017) explored the synergy between CRF and Recurrent Neural Networks (RNN), demonstrating that CRF enhances the RNN model's ability to capture contextual information effectively. Chowdhury et al. (2018) developed a multi-task architecture that simultaneously tackled binary classification, ADR labeling, and indication labeling, using the PSB 2016 Social Media dataset (Sarker et al., 2016).",
+        "bbox": [
+            507,
+            271,
+            884,
+            688
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Corpus Development",
+        "text_level": 1,
+        "bbox": [
+            509,
+            700,
+            726,
+            717
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our study began with a thorough literature review to identify existing ADE-related datasets. We discovered four text-only datasets: PSB 2016 social media shared task (Sarker et al., 2016) comprising 572 tweets, Medline ADE corpus (Gurulingappa et al., 2012b) with 4,272 sentences, CADEC (Karimi et al., 2015a) containing 1,248 sentences, and recently released BioDEX dataset (D'Oosterlinck et al., 2023). This revealed a notable gap in multimodal ADE datasets, where images complement textual data. We take this opportunity to introduce a multimodal corpus consist",
+        "bbox": [
+            507,
+            726,
+            884,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "11216",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/9d97229785dfb31c35c747987d8d4ce51900fea6015e324fc75cc47014504ad6.jpg",
+        "image_caption": [
+            "Figure 2: Some samples from the dataset containing images and the corresponding descriptive text."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            137,
+            98,
+            467,
+            305
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ing of 1,500 ADE images with corresponding English sentences to facilitate further research. While preparing this corpus, we carry out the following steps.",
+        "bbox": [
+            112,
+            388,
+            487,
+            453
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 Data Collection",
+        "text_level": 1,
+        "bbox": [
+            114,
+            475,
+            282,
+            489
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Utilizing a diverse array of keywords, we have curated a comprehensive dataset from social media, healthcare blogs, and MCRs (refer to Table 1). The inclusion of various sources ensures a broad representation of the population, enriching the dataset's diversity.",
+        "bbox": [
+            112,
+            502,
+            487,
+            599
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1.1 Social Media",
+        "text_level": 1,
+        "bbox": [
+            112,
+            619,
+            275,
+            633
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Social media data is invaluable for ADE-related tasks due to its real-time nature and diverse user-generated content (Sarker et al., 2016). In this study, we have utilized the official X (Twitter) and scraper API to gather Tweets $^{2,3}$  employing diverse keywords related to ADE. The data collection phase, conducted between June 2023 and October 2023, collected a total of 20,000 tweets using specified keywords presented in Table 1. From this pool, 3,000 tweets were meticulously identified as pertinent to ADEs, featuring either images, text, or a combination of both. Notably, 142 tweets included relevant images accompanied by textual descriptions of the adverse drug events.",
+        "bbox": [
+            112,
+            644,
+            489,
+            870
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1.2 Healthcare Blog",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            697,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We utilized a public healthcare-related blog, healthdirect<sup>4</sup>, a government-funded virtual health service that provides access to health advice and information via a website. We used Python's BeautifulSoup library to scrape the data and collected 1,150 unique images with corresponding text. Among these, 54 relevant images depicting adverse drug events were manually curated along with corresponding texts.",
+        "bbox": [
+            507,
+            103,
+            882,
+            247
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1.3 Medical Case Reports",
+        "text_level": 1,
+        "bbox": [
+            507,
+            256,
+            742,
+            272
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Data from published medical case reports is crucial for constructing comprehensive datasets to analyze ADE. These articles provide structured and verified information, forming a reliable foundation for in-depth analysis and research in pharmacovigilance. We have extracted data from the New England Journal of Medicine<sup>5</sup>, Science Direct<sup>6</sup>. A precise Science Direct query performed is as follows:",
+        "bbox": [
+            507,
+            275,
+            882,
+            419
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "(\"adverse drug event\") AND (Languages=English) AND (Article type=Case Reports) AND (Years=2000 to 2023))",
+        "bbox": [
+            509,
+            437,
+            865,
+            500
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This approach retrieved approximately 2,907 documents from ScienceDirect, from which we manually selected 1390 relevant images with corresponding texts.",
+        "bbox": [
+            507,
+            517,
+            882,
+            581
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Data Annotation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            592,
+            687,
+            606
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To ensure meticulous annotation aligned with ethical standards, we enlisted the assistance of two medical students and one Ph.D. student, selected based on specific criteria. These criteria included being at least 25 years old, proficient in English (reading, writing, and speaking), and willing to handle sensitive content. The process was finalized within a span of five months, and participants received compensation for their involvement  $^{7}$ . Annotators were tasked with meticulously assessing each image and its corresponding text based on the annotation manual. Sentences accurately depicting adverse drug events, including the drug's",
+        "bbox": [
+            505,
+            613,
+            882,
+            822
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "<sup>4</sup>https://www.healthdirect.gov.au/",
+            "<sup>5</sup>https://www.nejm.org",
+            "<sup>6</sup>https://www.sciencedirect.com",
+            "7The medical students received compensation in the form of gift vouchers and honorarium amounts in accordance with https://www.minimum-wage.org/international/india."
+        ],
+        "bbox": [
+            507,
+            829,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://twitter.com/",
+        "bbox": [
+            134,
+            889,
+            327,
+            904
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://www.scaperapi.com/",
+        "bbox": [
+            136,
+            904,
+            389,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "11217",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/a18ea0b8796f9ad5c9e6aa5203f55dcfeb95bb21220e2669e81a7e15d55c1d6e.jpg",
+        "image_caption": [
+            "Figure 3: Distribution of different body parts in the curated MMADE dataset. The number of data points and the percentage corresponding to each category have been provided."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            80,
+            843,
+            294
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/2c2fe4d1646772ebd56e51498c137e3e5095af58ae27b95b86e2a3b4ff384cf4.jpg",
+        "image_caption": [
+            "Figure 4: Data distribution statistics from various sources, illustrating the total ADE data (images, text, and image-text pairs), and the relevant image-text pairs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            381,
+            450,
+            552
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "name and associated side effects, were chosen for inclusion in the corpus development process, while all other instances were deemed irrelevant and removed. To maintain consistency and consensus among annotators, final rationale labels were determined through a majority voting approach. Annotators were explicitly instructed to annotate posts without biases related to demographics, religion, or other extraneous factors. The quality of annotations was assessed by measuring inter-annotator agreement (IAA) using Cohen's Kappa score (Viera et al., 2005). The resulting agreement scores of 0.78 affirm the acceptability and high quality of the annotations.",
+        "bbox": [
+            112,
+            694,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Annotation Manual",
+        "text_level": 1,
+        "bbox": [
+            509,
+            363,
+            712,
+            376
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We initially provided annotators with an instruction manual detailing different instances, accompanied by examples (as depicted in Fig. 5).",
+        "bbox": [
+            507,
+            384,
+            880,
+            432
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- For each data instance comprising text and an image, select the data instance if both the text and image indicate concerns about drug side effects.",
+            "- Remove data instances where the image does not convey the side effects mentioned in the text.",
+            "- Remove data instances where the text does not convey any side effects related to drugs, but some side effects are visually present in the accompanying image.",
+            "- Remove instances where both the text and image are unrelated to any drug side effect.",
+            "- If a data instance includes a URL link, access the content at that URL address to gain additional insight and context about the data."
+        ],
+        "bbox": [
+            531,
+            444,
+            882,
+            745
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Corpus Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            769,
+            685,
+            784
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 provides an overview of the data sources and the distribution of relevant cases within the dataset. Following the initial collection of 7,057 ADE-related samples encompassing images, text, and image-text pairs, we curated 1,500 pertinent samples. These encompass both image and their associated text descriptions of ADEs, establishing the foundation of our multimodal ADE dataset. Figure 2 depicts some of",
+        "bbox": [
+            507,
+            790,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "11218",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3bfd91cf627ae890bb29d18068c8eeb95b43ba21ddbaff557467dfa2d716eb0f.jpg",
+        "image_caption": [
+            "Figure 5: Illustration of an annotation process, with A1, A2, and A3 representing individual annotators and MV signifying final majority voting. In this setup, \"1\" denotes ADR, while \"0\" is unrelated to ADR."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            137,
+            95,
+            468,
+            281
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "the samples from the dataset. The first sample shows a person suffering from tongue discoloration after taking a Chlorhexidine drug. Both the text and visual image show a clear indication of drug reaction. The second sample is of a woman with papulopustular eruption on the face, which had worsened with topical metronidazole gel. Adverse drug events can manifest internally or externally, yet acquiring images of internal body parts from public sources is challenging. Thus, we focus on external body parts or symptoms, which can be readily captured and shared with doctors, pharmacovigilance teams, or the public domain for improved consultation and advice. Following dataset analysis, we identified 13 significant adverse effects crucial for multimodal ADE reporting. These effects are categorized into four groups based on their origin: ENT (9.85%), EYE (3.6%), LIMB (5.4%), and SKIN (81.06%). Additional details of the dataset, such as the distribution of samples across different body parts along with corresponding percentages, are illustrated in Fig. 3.",
+        "bbox": [
+            112,
+            393,
+            489,
+            731
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            746,
+            337,
+            762
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Each data point in the dataset encompasses a patient's textual description  $T$  along with a corresponding image  $I$  illustrating their medical issue or concern visually. The textual representation comprises a sequence of words  $\\{t_1,t_2,\\ldots ,t_n\\}$  while the visual elements are represented by  $I\\in \\mathbb{R}^{3\\times W\\times H}$ , where  $W$  and  $H$  denote the width and height of the image data, respectively. The objective is to process both the text  $T$  and image  $I$",
+        "bbox": [
+            112,
+            774,
+            490,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "for each patient and generate a natural language sequence  $Y$  that seamlessly integrates both modalities, expressed as  $Y = \\{T, I\\}$ .",
+        "bbox": [
+            507,
+            84,
+            884,
+            133
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Methodology",
+        "text_level": 1,
+        "bbox": [
+            509,
+            149,
+            660,
+            166
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Recent advancements in VLMs, such as BLIP, InstructBlip, and GIT, have showcased remarkable advancements in encoding both textual and visual inputs. These models outperform traditional approaches that rely solely on individual image or text encoders, followed by fusion. By integrating sophisticated mechanisms for joint representation learning, VLMs excel at capturing intricate relationships between textual and visual modalities, thereby enhancing their ability to generate more contextually relevant and coherent outputs (Zhang et al., 2023). In the proposed work, we have leveraged InstructBlip (Dai et al., 2023), known for its exceptional performance across a range of vision-language (VL) tasks including Visual Question Answering (VQA), Image captioning, and Image retrieval. Each patient's inquiry or concern is articulated as a textual sentence along with a visual image where they elaborate on their medical queries or concerns on social media platforms or healthcare blogs to obtain pertinent feedback or advice. InstructBlip integrates two distinct encoders specialized for separate modalities. This architecture includes an image encoder, which utilizes a vision transformer (ViT) to extract visual features, alongside an LLM and a Query Transformer (Q-Former). The Q-Former interacts with the image encoder's output through cross-attention, resulting in K-encoded visual vectors, which are then linearly projected and fed into the frozen LLM for further processing. This representation serves as input for a proficient language model, which generates high-quality text. During instruction tuning, only the Q-Former undergoes fine-tuning, while the image encoder and LLM remain unchanged. Figure 6 illustrates the detailed architecture. We fine-tuned the InstructBLIP to assess its performance on the proposed MMADE dataset. InstructBLIP maintains a consistent image resolution  $(224 \\times 224)$  during instruction tuning and freezes the visual encoder during fine-tuning. This approach substantially reduces the number of trainable parameters from 1.2 billion to 188 million, improving fine-tuning efficiency. We also employed two more VLMs and fine-tuned them with the proposed MMADE dataset. BLIP (Li et al., 2022), a",
+        "bbox": [
+            507,
+            178,
+            885,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "11219",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d94c73b2be65c7b4c059e3e675e38542eb7ca224e2288caa78cc903f386c42f1.jpg",
+        "image_caption": [
+            "Figure 6: The architectural details are illustrated, depicting the input image and corresponding text. Additionally, the frozen and trainable layers are highlighted for clarity."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            139,
+            86,
+            860,
+            259
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "versatile vision-language model, utilizes a multimodal mixture of encoder-decoder models during pre-training. This involves bootstrapping a dataset from large-scale noisy image-text pairs, where synthetic captions are injected, and noisy captions are eliminated. Additionally, we utilized GIT (Wang et al., 2022), a VLM model that generates text descriptions of images. Trained on curated datasets of image-text pairs, GIT encodes image features into a latent representation, which is decoded into text descriptions. Its architecture comprises an image encoder using a pre-trained Swin transformer and a text decoder based on a standard transformer decoder, linked by a cross-attention layer for enhanced focus on specific image encoding parts. Additionally, we performed the integration of LSTM networks with VGG16 (Simonyan and Zisserman, 2014) and ResNet50 (He et al., 2016) architectures. In this setup, the LSTM serves as the text encoder, while either VGG16 or ResNet50 acts as the visual encoder. The features extracted from the text and visual encoders are then concatenated to create a joint representation, allowing for comprehensive modeling of both textual and visual information. Please refer to Section 12 of the Appendix for the fine-tuning details.",
+        "bbox": [
+            112,
+            326,
+            492,
+            744
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 Experimental Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            760,
+            450,
+            778
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We have utilized four commonly employed evaluation metrics to assess the performance of the models, including BLEU score (Bilingual Evaluation Understudy) (Papineni et al., 2002), ROUGE score (Recall-Oriented Understudy for Gisting Evaluation) (Lin, 2004), BERTScore (Zhang et al., 2019), and MoverScore (Zhao et al., 2019). Detailed explanations of experimental settings and metrics are",
+        "bbox": [
+            112,
+            790,
+            490,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/3cec408123cf5d5b3743faabc8d8931b8ee2a950c4381a7acf9826e6f8356223.jpg",
+        "table_caption": [
+            "Table 2: The performance evaluation of the models using ROUGE and BLEU Scores in the multimodal dataset setting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Type</td><td colspan=\"3\">ROUGE</td><td colspan=\"3\">BLEU</td></tr><tr><td>R1</td><td>R2</td><td>RL</td><td>B1</td><td>B2</td><td>B3</td></tr><tr><td>LSTM+VGG16</td><td>-</td><td>0.213</td><td>0.105</td><td>0.201</td><td>0.165</td><td>0.073</td><td>0.041</td></tr><tr><td>LSTM+ResNet50</td><td>-</td><td>0.281</td><td>0.086</td><td>0.230</td><td>0.179</td><td>0.058</td><td>0.046</td></tr><tr><td rowspan=\"2\">BLIP</td><td>Base</td><td>0.19</td><td>0.093</td><td>0.185</td><td>0.099</td><td>0.003</td><td>0.001</td></tr><tr><td>Fine-Tune</td><td>0.334</td><td>0.163</td><td>0.225</td><td>0.171</td><td>0.081</td><td>0.058</td></tr><tr><td rowspan=\"2\">GIT</td><td>Base</td><td>0.27</td><td>0.11</td><td>0.192</td><td>0.157</td><td>0.014</td><td>0.004</td></tr><tr><td>Fine-Tune</td><td>0.504</td><td>0.285</td><td>0.416</td><td>0.194</td><td>0.10</td><td>0.097</td></tr><tr><td rowspan=\"2\">InstructBLIP</td><td>Base</td><td>0.29</td><td>0.161</td><td>0.212</td><td>0.219</td><td>0.125</td><td>0.008</td></tr><tr><td>Fine-Tune</td><td>0.571</td><td>0.351</td><td>0.475</td><td>0.319</td><td>0.175</td><td>0.112</td></tr></table>",
+        "bbox": [
+            515,
+            376,
+            878,
+            486
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a0e3070a2b1bc76dcd063d0dd75c1fe2fe9886e07596f7f7206c083e7142f330.jpg",
+        "table_caption": [
+            "Table 3: The evaluation of the model using BERTScore and MoverScore in the multimodal dataset setting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"8\">BERTScore</td><td></td></tr><tr><td rowspan=\"2\">Evaluation Metrics</td><td rowspan=\"2\">LSTM+VGG16</td><td rowspan=\"2\">LSTM+ResNet50</td><td colspan=\"2\">BLIP</td><td colspan=\"2\">GIT</td><td colspan=\"2\">InstructBLIP</td></tr><tr><td>Base</td><td>Fine-Tune</td><td>Base</td><td>Fine-Tune</td><td>Base</td><td>Fine-Tune</td></tr><tr><td>Precision</td><td>0.821</td><td>0.847</td><td>0.819</td><td>0.841</td><td>0.826</td><td>0.866</td><td>0.832</td><td>0.896</td></tr><tr><td>Recall</td><td>0.797</td><td>0.811</td><td>0.783</td><td>0.819</td><td>0.791</td><td>0.831</td><td>0.781</td><td>0.891</td></tr><tr><td>F1-score</td><td>0.809</td><td>0.821</td><td>0.801</td><td>0.829</td><td>0.812</td><td>0.852</td><td>0.805</td><td>0.893</td></tr><tr><td colspan=\"9\">MoverScore</td></tr><tr><td></td><td>0.441</td><td>0.487</td><td>0.482</td><td>0.513</td><td>0.496</td><td>0.553</td><td>0.544</td><td>0.622</td></tr></table>",
+        "bbox": [
+            512,
+            544,
+            890,
+            623
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "given in Section 11 and 13 of the Appendix.",
+        "bbox": [
+            507,
+            653,
+            836,
+            669
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.1 Findings from Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            680,
+            773,
+            695
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The experimental results across Tables 2 and 3 depict the performance of various models in multimodal dataset settings, while Tables 4 and 5 represent the performance in unimodal dataset settings, based on standard evaluation metrics. From the",
+        "bbox": [
+            507,
+            700,
+            884,
+            778
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/53c4e2d2b49d2bd5f9db212b5567b1ffd4f32529ab4ccc93c76fb17849a45940.jpg",
+        "table_caption": [
+            "Table 4: The evaluation of the models using BERTScore and MoverScore in the unimodal dataset setting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"6\">BERTScore</td></tr><tr><td>Evaluation Metrics</td><td>LSTM+VGG16</td><td>LSTM+ResNet50</td><td>BLIP</td><td>GIT</td><td>InstructBLIP</td></tr><tr><td>Precision</td><td>0.731</td><td>0.744</td><td>0.753</td><td>0.801</td><td>0.840</td></tr><tr><td>Recall</td><td>0.709</td><td>0.713</td><td>0.724</td><td>0.783</td><td>0.783</td></tr><tr><td>F1-score</td><td>0.719</td><td>0.729</td><td>0.738</td><td>0.790</td><td>0.810</td></tr><tr><td colspan=\"6\">MoverScore</td></tr><tr><td></td><td>0.121</td><td>0.143</td><td>0.151</td><td>0.184</td><td>0.253</td></tr></table>",
+        "bbox": [
+            514,
+            838,
+            878,
+            910
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "11220",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "results, several observations are:",
+        "bbox": [
+            114,
+            85,
+            356,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Table 2 presents the ROUGE and BLEU scores, indicating the superior performance of fine-tuned InstructBLIP compared to other models. The higher ROUGE and BLEU scores achieved by fine-tuned InstructBLIP suggest its proficiency in capturing relevant information from the input and generating text. This superiority can be attributed to its training on more extensive and diverse datasets, enabling it to capture intricate data patterns effectively.",
+            "- Tables 3 and 4 present BERTScores and MoverScores in the multimodal and unimodal settings, respectively. Fine-tuned InstructBLIP achieves the highest BERTScore and MoverScores, demonstrating its efficacy in capturing contextual similarity and superior ability to convey meaningful content effectively.",
+            "- Table 4 and 5 show the decline in performance when training VLM models with unimodal data. The absence of meaningful visual information likely contributes to the performance degradation.",
+            "- One important observation is that Instruct-BLIP consistently outperforms BLIP and GIT across all metrics, showing its superior capability in effectively integrating textual and visual information. This can be attributed to its innovative architecture, featuring a Query Transformer for instruction-aware feature extraction. Refer to Section 16 of the Appendix for detailed outcomes.",
+            "- Fine-tuning with domain-specific ADE data remarkably enhances model performance, reflecting its pivotal role in adapting models to the intricacies of adverse drug event detection.",
+            "- Another key finding highlights the substantial performance enhancement achieved by integrating both image and text modalities, emphasizing the critical role of visual information alongside textual data."
+        ],
+        "bbox": [
+            134,
+            112,
+            489,
+            841
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Statistical Analysis: We conducted a statistical t-test to compare the performance of the proposed multimodal model, utilizing both image and text data, with that of unimodal models. The analysis",
+        "bbox": [
+            112,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0f792ccbe22ac352fc0a70f43fd3821100626402ddda0b50c11da6a681321df7.jpg",
+        "image_caption": [
+            "Actual: I was given penicillin for my ongoing treatment. About 4 hours later, I started feeling really sick. I came down with a high fever, nausea, vomiting, and a terrible headache. My whole body ached, and got this type of rash on my skin. What type of rash it is?",
+            "Figure 7: An example case study where a sample image and its corresponding text are provided along with the generated text from all models in both unimodal and multimodal settings."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            95,
+            596,
+            148
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/35a3ce8b1583783cec90f27ceea46d0766e364f1df959e6c179eb6b00fcc4941.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Unimodal\n(Blank Image and Text)</td><td>Multimodal\n(Image and Text)</td></tr><tr><td>BLIP</td><td>A man with a large stomach and a large stomach.</td><td>A close up of a person&#x27;s stomach covered in a red rash.</td></tr><tr><td>GIT</td><td>A middle aged man sits on an exam table in a doctor&#x27;s office, looking unwell.</td><td>A man with a lot of red spots suffering from rashes.</td></tr><tr><td>InstructBLIP</td><td>A person&#x27;s covered in a red rash.</td><td>The image shows red spots on the abdomen, which is unusual for a person. The red spots are likely caused by a skin condition, such as syphilis.</td></tr></table>",
+        "bbox": [
+            532,
+            152,
+            860,
+            277
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "yielded a p-value below 0.05, indicating a significant difference in performance between the two. The detailed explanation is given in Section 14 of Appendix.",
+        "bbox": [
+            507,
+            385,
+            882,
+            450
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.2 Qualitative Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            463,
+            714,
+            478
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We performed an extensive qualitative analysis of the responses generated by various models in both unimodal and multimodal settings, complemented by several case studies. A detailed case study is depicted in Fig. 7. The analysis has led to the following conclusion: (a) In multimodal settings, all the models demonstrate superior performance and exhibit a greater ability to capture crucial visual information conveyed through images than in unimodal settings (refer to Table 4 and 5). (b) Observations also revealed that models such as BLIP and GIT tended to hallucinate, as evidenced by Fig. 7, occasionally generating facts that were entirely unrelated to the context. (c) InstructBLIP demonstrates superior performance, leading us to conclude that providing instructions during finetuning prompts the model to selectively focus on pertinent visual features. This focused attention encourages the model to generate target sequences that closely resemble the desired output. We have presented a case study demonstrating that model performance is not solely determined by the number of samples in the dataset but also by other factors such as the distinct visual characteristics of different types of ADEs present in the dataset. A detailed explanation is provided in the Section 15 of Appendix.",
+        "bbox": [
+            505,
+            483,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "11221",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a80744f396b4900ab110544cd9a0f81debb7f0fbc2dbbf68ceb2392e42d23a13.jpg",
+        "table_caption": [
+            "Table 5: The performance evaluation of the models using ROUGE and BLEU Scores in the unimodal dataset setting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">ROUGE</td><td colspan=\"3\">BLEU</td></tr><tr><td>R1</td><td>R2</td><td>RL</td><td>B1</td><td>B2</td><td>B3</td></tr><tr><td>LSTM+VGG16</td><td>0.133</td><td>0.018</td><td>0.101</td><td>0.141</td><td>0.012</td><td>0.001</td></tr><tr><td>LSTM+ResNet50</td><td>0.146</td><td>0.029</td><td>0.112</td><td>0.152</td><td>0.016</td><td>0.002</td></tr><tr><td>BLIP</td><td>0.158</td><td>0.032</td><td>0.129</td><td>0.161</td><td>0.019</td><td>0.002</td></tr><tr><td>GIT</td><td>0.169</td><td>0.035</td><td>0.137</td><td>0.172</td><td>0.020</td><td>0.003</td></tr><tr><td>InstructBLIP</td><td>0.211</td><td>0.087</td><td>0.172</td><td>0.197</td><td>0.089</td><td>0.006</td></tr></table>",
+        "bbox": [
+            117,
+            136,
+            487,
+            234
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Risk Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            263,
+            268,
+            280
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While our multimodal model demonstrates promise, it is essential to have medical experts and pharmacovigilance teams validate the findings, considering other critical factors. Our model and dataset are intended to support medical professionals rather than replace them.",
+        "bbox": [
+            112,
+            290,
+            489,
+            387
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            401,
+            401,
+            416
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present the task of ADE detection within pharmacovigilance mining, leveraging multimodal datasets. In order to solve this task, we have created a multimodal ADE dataset, MMADE, containing images and corresponding descriptions, enhancing decision-making with the inclusion of visual cues. We have employed InstructBLIP, finetuned with the proposed dataset, and compared it with other models. Our findings suggest that domain-specific fine-tuning significantly enhances overall performance, emphasizing the importance of multimodal visual cues. We envision MMADE as a pivotal resource for advancing research in multimodal ADE detection. Moreover, our finetuned architecture holds promise as a valuable tool for pharmacovigilance teams, clinicians, and researchers, facilitating more effective ADE monitoring and ultimately improving patient safety and outcomes. In addition to expanding the dataset, future investigations could explore the potential of this multimodal dataset in tasks such as ADE severity classification and summarization.",
+        "bbox": [
+            112,
+            428,
+            489,
+            782
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "9 Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            795,
+            250,
+            810
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While our effort aimed to develop an ADE detection framework and introduce the novel MMADE dataset, comprising textual descriptions of drug events paired with images, it is crucial to acknowledge certain limitations inherent in the dataset. Specifically, our dataset primarily focuses on drug",
+        "bbox": [
+            112,
+            822,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "events associated with external body parts, omitting data about internal conditions such as liver infections, kidney stones, or psychological ailments like depression and migraine. Acquiring a substantial volume of image-text pairs within the ADE domain presents inherent challenges, including data privacy concerns, regulatory constraints, and the specialized nature of ADE occurrences. Despite these obstacles, our research breaks new ground by integrating images with text, solving a real-world challenge where individuals affected by ADE may resort to image sharing for communication when verbally expressing their symptoms is difficult. Moving forward, we aim to enhance the dataset by incorporating more ADE-related images and expanding its utility through additional tasks such as complaint identification and text summarization.",
+        "bbox": [
+            507,
+            84,
+            884,
+            357
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "10 Ethics and Broader Impact",
+        "text_level": 1,
+        "bbox": [
+            509,
+            369,
+            789,
+            386
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "User Privacy.",
+        "text_level": 1,
+        "bbox": [
+            509,
+            395,
+            620,
+            410
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our dataset contains AD images and corresponding text tweets with annotation labels and no personal user information.",
+        "bbox": [
+            507,
+            414,
+            880,
+            461
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Biases.",
+        "text_level": 1,
+        "bbox": [
+            509,
+            470,
+            568,
+            483
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Any biases detected in the dataset are inadvertent, and we have no intention of harming anyone or any group. We acknowledge that evaluating whether a tweet is ADE can be subjective, so we have taken agreement from all the annotators before selecting the data.",
+        "bbox": [
+            507,
+            489,
+            882,
+            583
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Intended Use.",
+        "text_level": 1,
+        "bbox": [
+            509,
+            595,
+            621,
+            608
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We share our data to promote more research on Adverse drug event detection. We only release the dataset for research purposes and do not grant a license for commercial use.",
+        "bbox": [
+            507,
+            614,
+            880,
+            677
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            703,
+            608,
+            719
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Eiji Aramaki, Yasuhide Miura, Masatsugu Tonoike, Tomoko Ohkuma, Hiroshi Masuichi, Kayo Waki, and Kazuhiko Ohe. 2010. Extraction of adverse drug effects from clinical records. In MEDINFO 2010, pages 739-743. IOS Press.",
+            "Adrian Benton, Lyle Ungar, Shawndra Hill, Sean Hennessy, Jun Mao, Annie Chung, Charles E Leonard, and John H Holmes. 2011. Identifying potential adverse effects using the web: A new approach to medical hypothesis generation. Journal of biomedical informatics, 44(6):989-996.",
+            "Shaika Chowdhury, Chenwei Zhang, and Philip S Yu. 2018. Multi-task pharmacovigilance mining from"
+        ],
+        "bbox": [
+            509,
+            726,
+            885,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "11222",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "social media posts. In Proceedings of the 2018 World Wide Web Conference, pages 117-126.",
+            "Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. Instructclip: Towards general-purpose vision-language models with instruction tuning.",
+            "Karel D'Oosterlinck, François Remy, Johannes Deleu, Thomas Demeester, Chris Develder, Klim Zaporojets, Aneiss Ghodsi, Simon Ellershaw, Jack Collins, and Christopher Potts. 2023. *Biodex: Large-scale biomedical adverse drug event extraction for real-world pharmacovigilance.* arXiv preprint arXiv:2305.13395.",
+            "Akash Ghosh, Arkadeep Acharya, Raghav Jain, Sriparna Saha, Aman Chadha, and Setu Sinha. 2024a. Clipsntel: CLIP and LLM synergy for multimodal question summarization in healthcare. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada, pages 22031-22039. AAAI Press.",
+            "Akash Ghosh, Arkadeep Acharya, Prince Jha, Sriparna Saha, Aniket Gaudgaul, Rajdeep Majumdar, Aman Chadha, Raghav Jain, Setu Sinha, and Shivani Agarwal. 2024b. Medsumm: A multimodal approach to summarizing code-mixed Hindi-english clinical queries. In Advances in Information Retrieval - 46th European Conference on Information Retrieval, ECIR 2024, Glasgow, UK, March 24-28, 2024, Proceedings, Part V, volume 14612 of Lecture Notes in Computer Science, pages 106-120. Springer.",
+            "Harsha Gurulingappa, Juliane Fluck, Martin Hofmann-Apitius, and Luca Toldo. 2011. Identification of adverse drug event assertive sentences in medical case reports. In First international workshop on knowledge discovery and health care management (KD-HCM), European conference on machine learning and principles and practice of knowledge discovery in databases (ECML PKDD), pages 16-27.",
+            "Harsha Gurulingappa, Roman Klinger, Martin Hofmann-Apitius, and Juliane Fluck. 2010. An empirical evaluation of resources for the identification of diseases and adverse effects in biomedical literature. In 2nd Workshop on Building and evaluating resources for biomedical text mining (7th edition of the Language Resources and Evaluation Conference), pages 15-22.",
+            "Harsha Gurulingappa, Abdul Mateen-Rajpu, and Luca Toldo. 2012a. Extraction of potential adverse drug events from medical case reports. Journal of biomedical semantics, 3(1):1-10.",
+            "Harsha Gurulingappa, Abdul Mateen Rajput, Angus Roberts, Juliane Fluck, Martin Hofmann-Apitius, and"
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Luca Toldo. 2012b. Development of a benchmark corpus to support the automatic extraction of drug-related adverse effects from medical case reports. Journal of biomedical informatics, 45(5):885-892.",
+            "Katja M Hakkarainen, Khadidja Hedna, Max Petzold, and Staffan Hägg. 2012. Percentage of patients with preventable adverse drug reactions and preventability of adverse drug reactions-a meta-analysis. PloS one, 7(3):e33236.",
+            "Rave Harpaz, Santiago Vilar, William DuMouchel, Hajjat Salmasian, Krystl Haerian, Nigam H Shah, Herbert S Chase, and Carol Friedman. 2013. Combining signals from spontaneous reports and electronic health records for detection of adverse drug reactions. Journal of the American Medical Informatics Association, 20(3):413-419.",
+            "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778.",
+            "Trung Huynh, Yulan He, Alistair Willis, and Stefan Rueger. 2016. Adverse drug reaction classification with deep neural networks. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 877-887, Osaka, Japan. The COLING 2016 Organizing Committee.",
+            "Sarvnaz Karimi, Alejandro Metke-Jimenez, Madonna Kemp, and Chen Wang. 2015a. Cadec: A corpus of adverse drug event annotations. Journal of biomedical informatics, 55:73-81.",
+            "SarvNaz Karimi, Chen Wang, Alejandro Metke-Jimenez, Raj Gaire, and Cecile Paris. 2015b. Text and data mining techniques in adverse drug reaction detection. ACM Computing Surveys (CSUR), 47(4):1-39.",
+            "Robert Leaman, Laura Wojtulewicz, Ryan Sullivan, Annie Skariah, Jian Yang, and Graciela Gonzalez. 2010. Towards internet-age pharmacovigilance: extracting adverse drug reactions from user posts in health-related social networks. In Proceedings of the 2010 workshop on biomedical natural language processing, pages 117-125.",
+            "Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models.",
+            "Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR.",
+            "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."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "11223",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Azadeh Nikfarjam and Graciela H Gonzalez. 2011. Pattern mining for extraction of mentions of adverse drug reactions from user comments. In AMIA annual symposium proceedings, volume 2011, page 1019. American Medical Informatics Association.",
+            "Azadeh Nikfarjam, Abeed Sarker, Karen O'connor, Rachel Ginn, and Graciela Gonzalez. 2015. Pharmacovigilance from social media: mining adverse drug reaction mentions using sequence labeling with word embedding cluster features. Journal of the American Medical Informatics Association, 22(3):671-681.",
+            "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.",
+            "Pranab Sahoo, Prakash Meharia, Akash Ghosh, Sriparna Saha, Vinija Jain, and Aman Chadha. 2024a. Unveiling hallucination in text, image, video, and audio foundation models: A comprehensive survey. arXiv preprint arXiv:2405.09589.",
+            "Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. 2024b. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927.",
+            "Abeed Sarker and Graciela Gonzalez. 2015. Portable automatic text classification for adverse drug reaction detection via multi-corpus training. Journal of biomedical informatics, 53:196-207.",
+            "Abeed Sarker, Azadeh Nikfarjam, and Graciela Gonzalez. 2016. Social media mining shared task workshop. In Biocomputing 2016: Proceedings of the Pacific Symposium, pages 581-592. World Scientific.",
+            "Ryoma Sato, Makoto Yamada, and Hisashi Kashima. 2022. Re-evaluating word mover's distance. In International Conference on Machine Learning, pages 19231-19249. PMLR.",
+            "Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.",
+            "Janet Sultana, Paola Cutroneo, and Gianluca Trifiro. 2013. Clinical and economic burden of adverse drug reactions. Journal of Pharmacology and Pharmacotherapeutics, 4(1_suppl):S73-S77.",
+            "Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mullappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. 2023. Xraygpt: Chest radiographs summarization using medical vision-language models. arXiv preprint arXiv:2306.07971.",
+            "Elena Tutubalina, Sergey Nikolenko, et al. 2017. Combination of deep recurrent neural networks and conditional random fields for extracting adverse drug reactions from user reviews. Journal of healthcare engineering, 2017."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Anthony J Viera, Joanne M Garrett, et al. 2005. Understanding interobserver agreement: the kappa statistic. Fam med, 37(5):360-363.",
+            "Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. 2022. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100.",
+            "Xiaoyan Wang, George Hripcsak, Marianthi Markatou, and Carol Friedman. 2009. Active computerized pharmacovigilance using natural language processing, statistics, and electronic health records: a feasibility study. Journal of the American Medical Informatics Association, 16(3):328-337.",
+            "Shweta Yadav, Asif Ekbal, and Sriparna Saha. 2018a. Feature selection for entity extraction from multiple biomedical corpora: A pso-based approach. Soft Computing, 22(20):6881-6904.",
+            "Shweta Yadav, Asif Ekbal, Sriparna Saha, and Pushpak Bhattacharyya. 2019. A unified multi-task adversarial learning framework for pharmacovigilance mining. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5234-5245, Florence, Italy. Association for Computational Linguistics.",
+            "Shweta Yadav, Asif Ekbal, Sriparna Saha, Pushpak Bhattacharyya, and Amit Sheth. 2018b. Multi-task learning framework for mining crowd intelligence towards clinical treatment. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 271-277, New Orleans, Louisiana. Association for Computational Linguistics.",
+            "Shweta Yadav, Srivatsa Ramesh, Sriparna Saha, and Asif Ekbal. 2020. Relation extraction from biomedical and clinical text: Unified multitask learning framework. IEEE/ACM transactions on computational biology and bioinformatics, 19(2):1105-1116.",
+            "Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. 2023. Vision-language models for vision tasks: A survey.",
+            "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675.",
+            "Zhifei Zhang, JY Nie, and Xuyao Zhang. 2016. An ensemble method for binary classification of adverse drug reactions from social media. In Proceedings of the Social Media Mining Shared Task Workshop at the Pacific Symposium on Biocomputing, volume 1.",
+            "Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M Meyer, and Steffen Eger. 2019. Moverscore: Text generation evaluating with contextualized embeddings and earth mover distance. arXiv preprint arXiv:1909.02622."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11224",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Juexiao Zhou and Xin Gao. 2023. Skingpt: A dermatology diagnostic system with vision large language model. arXiv preprint arXiv:2304.10691.",
+        "bbox": [
+            114,
+            85,
+            487,
+            124
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            137,
+            203,
+            153
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "11 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            114,
+            162,
+            349,
+            179
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "This section provides the hyperparameters and experimental setups utilized in the study. All experiments were conducted using multiple RTX 2080Ti GPUs. The dataset was partitioned, allocating  $80\\%$  for training and  $20\\%$  for testing. For InstructBLIP, the learning rate was set to 1e-5, executed for 50 epochs, with a batch size of 2. Similarly, for the BLIP model, a learning rate of 0.0001 was utilized and executed for 50 epochs, with a batch size of 2. For GIT fine-tuning, a learning rate of 5e-3 was applied for 50 epochs, with a batch size of 2. All models were implemented using Scikit-Learn and PyTorch as the backend framework.",
+        "bbox": [
+            112,
+            187,
+            489,
+            395
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "12 Fine-tuning",
+        "text_level": 1,
+        "bbox": [
+            114,
+            407,
+            260,
+            424
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We have followed several steps to fine-tune BLIP on our multimodal dataset. First, we have prepared the dataset in JSON format, which is compatible with the BLIP framework, and each image-text pair is represented as a dictionary with the following keys: image_path: The path to the image file, text: The caption or other text description of the image. We have fine-tuned BLIP with learning rate  $= 0.001$ , number of epochs  $= 50$ , and batch size  $= 16$ . The training process takes 6 hours for the fine-tuning, and we evaluate the performance on a held-out test set.",
+        "bbox": [
+            112,
+            432,
+            487,
+            625
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To fine-tune the GIT model on the proposed MMADE dataset, we utilized the Hugging Face Transformers library and followed a systematic process. First, we prepared the data using a PyTorch dataset, converting it into the required format using the GitProcessor class. Finally, we have utilized the GIT-base model, which is pre-trained on a substantial dataset of image-text pairs. We have utilized Parameters such as a learning rate  $= 5\\mathrm{e} - 3$  epochs  $= 30$  and batch size  $= 2$ . The optimization was performed using the Adam optimizer with the cross-entropy loss function, and the fine-tuning process took 3 hours.",
+        "bbox": [
+            112,
+            627,
+            487,
+            834
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To fine-tune the CNN-LSTM model, we leverage the CNN models (VGG16 and ResNet50) to extract image features, followed by an LSTM for",
+        "bbox": [
+            112,
+            835,
+            487,
+            883
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "sequence generation. Model compilation involves the use of categorical cross-entropy loss and the Adam optimizer. Training proceeds for 40 epochs, utilizing a batch size of 32.",
+        "bbox": [
+            507,
+            84,
+            882,
+            149
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "13 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            509,
+            161,
+            721,
+            177
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We have utilized BLEU score (Papineni et al., 2002), ROUGE score (Lin, 2004), BERTScore (Zhang et al., 2019) and Mover-Score (Zhao et al., 2019). BLEU score is utilized to assess the quality of machine-generated text by comparing it to human-generated reference text. It measures the similarity in word sequences between the machine-generated and human reference texts using n-grams, penalizing shorter machine-generated texts to provide a quantitative measure of translation accuracy. The ROUGE score is also used to evaluate the quality of machine-generated summaries compared to human-written summaries. It works by calculating the overlap of n-grams between the machine-generated summary and the reference summaries. ROUGE and BLEU metrics evaluate text quality based on syntactic overlap, considering unigrams and bigrams, lacking the ability to decode semantic meaning effectively. However, BERTScore focuses on understanding the semantic meaning of generated text compared to the intended text, enabling a more nuanced and accurate comparison. MoverScore is a way to measure how similar a machine-generated text is to a human-written text. It does this by using BERT embeddings to understand the meaning of the words and sentences in both texts and then using Word Mover's Distance (Sato et al., 2022) to measure how similar the two texts are.",
+        "bbox": [
+            507,
+            187,
+            884,
+            653
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "14 Statistical Analysis",
+        "text_level": 1,
+        "bbox": [
+            509,
+            667,
+            717,
+            684
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In this study, we employed the paired t-test to assess the statistical significance of differences between the outcomes of unimodal and multimodal models. The null hypothesis (H0) assumes no significant disparity in scores, while the alternative hypothesis (H1) proposes the opposite. We structured our analysis around the assumption of paired data, where each BLEU score corresponds to the same model in different settings (Multimodal Vs. Unimodal). By applying the paired t-test, we aimed to rigorously evaluate performance disparities and offer insights into the models' relative effectiveness. The obtained p-values for ROUGE, BLEU and BERTScore scores were 0.008, 0.0097 and",
+        "bbox": [
+            507,
+            694,
+            882,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>8</sup>https://scikit-learn.org/stable/",
+        "bbox": [
+            134,
+            890,
+            435,
+            904
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{9}$ https://pytorch.org/",
+        "bbox": [
+            134,
+            904,
+            326,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "11225",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f9f91d1168552918b6a582f65d28df839697e2d8f868b1211491d1603e4505d.jpg",
+        "image_caption": [
+            "Figure 8: Illustration of a case study showcasing the performance of the InstructBLIP model from the dataset"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            134,
+            89,
+            468,
+            244
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "0.0316, respectively, both below the 0.05 threshold, leading to the rejection of the null hypothesis.",
+        "bbox": [
+            112,
+            317,
+            489,
+            350
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/de6c5d46c3296a45d5d157a103f74f2fc277353cf8ced4c4c11809c3bf2359ec.jpg",
+        "image_caption": [
+            "Figure 9: Example of the sample image and corresponding text from the dataset along with generated text from the Proposed model with two other models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            166,
+            373,
+            436,
+            638
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "15 Case study",
+        "text_level": 1,
+        "bbox": [
+            114,
+            733,
+            253,
+            750
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The case study depicted in Fig. 8 reveals intriguing insights across various body parts. In the first row, depicting rashes, fine-tuned InstructBLIP accurately identifies distinct body parts and key findings, potentially attributed to the diverse visual characteristics of different rash types prevalent in the dataset with the largest portion of  $81.06\\%$ . However, performance in identifying mouth-related ADEs is less satisfactory despite comprising  $5.93\\%$  of the dataset, likely due to potential confusion with other",
+        "bbox": [
+            112,
+            758,
+            490,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "oral features like the tongue, teeth, or lips. Conversely, despite eye-related problems representing only  $1.8\\%$  of the dataset, the model performs comparatively better in this category, possibly because it focuses specifically on infected regions, enhancing its ability to identify relevant features accurately.",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "16 Comparative output",
+        "text_level": 1,
+        "bbox": [
+            509,
+            209,
+            729,
+            225
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We have added two more examples in Fig. 9 showing the ADR instances, corresponding ground truth, and the model-generated outputs.",
+        "bbox": [
+            507,
+            233,
+            882,
+            282
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "11226",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2024/Enhancing Adverse Drug Event Detection with Multimodal Dataset_ Corpus Creation and Model Development/5f2bd2d8-1c47-4ff0-bd6b-5d72a8900ae6_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc0aac0bcc3e3132f13f98b1a19602e4e41dd024b21624826b6ac95ea48d59bd
+size 3680797

2024/Enhancing Adverse Drug Event Detection with Multimodal Dataset_ Corpus Creation and Model Development/full.md

@@ -0,0 +1,315 @@
+# Enhancing Adverse Drug Event Detection with Multimodal Dataset: Corpus Creation and Model Development
+
+Pranab Sahoo<sup>1</sup>, Ayush Kumar Singh<sup>1</sup>, Sriparna Saha<sup>1</sup>, Aman Chadha<sup>2,3*</sup>, Samrat Mondal<sup>1</sup>
+
+$^{1}$ Department of Computer Science And Engineering, Indian Institute of Technology Patna
+
+$^{2}$ Stanford University,  $^{3}$ Amazon GenAI
+
+pranab_2021cs25@iitp.ac.in, ayush_2211ai27@iitp.ac.in, sriparna@iitp.ac.in, samrat@iitp.ac.in hi@aman.ai
+
+# Abstract
+
+The mining of adverse drug events (ADEs) is pivotal in pharmacovigilance, enhancing patient safety by identifying potential risks associated with medications, facilitating early detection of adverse events, and guiding regulatory decision-making. Traditional ADE detection methods are reliable but slow, not easily adaptable to large-scale operations, and offer limited information. With the exponential increase in data sources like social media content, biomedical literature, and Electronic Medical Records (EMR), extracting relevant ADE-related information from these unstructured texts is imperative. Previous ADE mining studies have focused on text-based methodologies, overlooking visual cues, limiting contextual comprehension, and hindering accurate interpretation. To address this gap, we present a MultiModal Adverse Drug Event (MMADE) detection dataset, merging ADE-related textual information with visual aids. Additionally, we introduce a framework that leverages the capabilities of LLMs and VLMs for ADE detection by generating detailed descriptions of medical images depicting ADEs, aiding healthcare professionals in visually identifying adverse events. Using our MMADE dataset, we showcase the significance of integrating visual cues from images to enhance overall performance. This approach holds promise for patient safety, ADE awareness, and healthcare accessibility, paving the way for further exploration in personalized healthcare. The code and dataset used in this work are publicly available<sup>1</sup>.
+
+Disclaimer: The article features images that may be visually disturbing to some readers.
+
+# 1 Introduction
+
+An adverse drug event (ADE) encompasses any harm resulting from medication use, whether it's
+
+unintended, off-label, or due to medication errors. Adverse drug reactions (ADRs) are a specific type of ADE, denoting unexpected harm arising from the proper use of medication at the prescribed dosage. Injuries from inappropriate or off-label use are not classified as ADRs (Karimi et al., 2015b). ADEs pose significant public health concerns, contributing to numerous fatalities, serious injuries, millions of hospitalizations, and prolonged hospital stays. Consequently, they impose substantial financial burdens, costing healthcare systems billions of dollars globally. Despite advancements in healthcare, ADE detection remains a significant challenge. Implementing effective detection and monitoring strategies can substantially mitigate the adverse impacts on patients and healthcare systems (Hakkarainen et al., 2012), (Yadav et al., 2018b), (Sultana et al., 2013). Most of the previous ADE detection works are based on text data only (D'Oosterlinck et al., 2023), (Sarker and Gonzalez, 2015), (Yadav et al., 2019), (Sarker et al., 2016), (Chowdhury et al., 2018), (Yadav et al., 2018a), which presents a significant disadvantage due to its subjective nature and lack of specific details of visual cues, leading to potential inaccuracies and incomplete categorization of ADE detection. Despite extensive research, the potential of integrating textual data with visual information, such as images, has been largely overlooked. Visual aids are essential in ADE detection for numerous reasons. A substantial proportion of the population lacks proficiency in medical jargon, hindering accurate symptom descriptions. Moreover, certain symptoms are inherently challenging to express through text alone. Patients may struggle to differentiate between similar symptoms, like skin rash, eczema, peeling, and blister. As depicted in Fig. 1, sample images may present confusion to individuals lacking adequate medical expertise. Integrating both text and images in these scenarios can enhance the accuracy and effectiveness of ADE detection,
+
+offering a comprehensive understanding of the patient's current medical condition. To the best of our knowledge, ADE detection using both image and text data has not been explored previously, and we take this opportunity to introduce a MultiModal Adverse Drug Event (MMADE) dataset comprising ADR images paired with corresponding textual descriptions.
+
+Large Language Models (LLMs) and Vision Language Models (VLMs) have exhibited remarkable skills in generating human-like text, prompting their integration into various medical applications, including tasks such as chest radiography report generation, summarization, and medical question answering (Thawkar et al., 2023), (Ghosh et al., 2024b), (Sahoo et al., 2024b), (Ghosh et al., 2024a). However, their potential in ADE detection, which involves both text and images, has yet to be explored. Leveraging LLMs and VLMs for this task presents inherent limitations as they are predominantly trained on generic natural images sourced from databases like ImageNet, Wikipedia, and the internet. Generic models may not possess the specialized medical knowledge required for comprehensive caption generation, potentially leading to oversimplified descriptions that overlook essential details like symptoms and medical intricacies. Furthermore, while VLMs have excelled in traditional visual-linguistic tasks, their application to medical imaging presents unique challenges that may hinder the accurate interpretation and description of complex medical images (Sahoo et al., 2024a). Specialized models such as XrayGPT (Thawkar et al., 2023) and SkinGPT4 (Zhou and Gao, 2023), which are trained on chest X-ray and skin disease images, exemplify the domain specificity required for accurate medical image analysis. This has led us to explore ADR detection within a multimodal framework. To support this exploration, we introduce MMADE, a carefully curated dataset crafted for this specific purpose. MMADE consists of 1500 instances of patient-reported concerns regarding drugs and associated side effects, each paired with both textual descriptions and corresponding images. In our study, we have employed Instruct-BLIP (Dai et al., 2023), which builds upon the strong foundation of BLIP-2 (Li et al., 2023), a pretrained model with high-quality visual representation and strong language generation capabilities. The meticulous fine-tuning process enables it to bridge the disparity between general-purpose mod
+
+![](images/79d79f0afe3c59354305feaeb62f41befd7407dd1a808e6da2088aec5ab94eb1.jpg)  
+After 1 hour of taking the penicillins!! I get this.
+
+![](images/09f0e23cf58d401f6ae1beaf0d58bec4a559c894d659c1d7688635e8f4d343b1.jpg)  
+I found this after my doctor prescribed Etoricoxib for post-surgical dental pain.
+
+![](images/7dafcaf297647a24f155b2e3dd2d480ee8b373b714197a0130c643d5abe5e691.jpg)  
+This is very unusual to me after taking amoxicillin for pharyngitis.  
+Figure 1: Samples from the dataset highlight the significance of visual cues in understanding adverse drug events, particularly in cases where patients are unaware of a specific medical condition.
+
+els and the specialized demands of ADE-specific tasks. Moreover, our exploration of BLIP (Li et al., 2022) and GIT (Wang et al., 2022) reveals that these models exhibit insufficient performance before fine-tuning. Nevertheless, upon fine-tuning with domain-specific data, their performance experiences notable improvement.
+
+Our key contributions are as follows:
+
+- A novel approach to ADE detection in multimodal settings greatly assists medical professionals, such as doctors, nurses, and pharmacists, by delivering detailed descriptions of ADE cases, enhancing precision in diagnosis, treatment planning, and patient care.  
+- Introduction of a novel multimodal dataset MMADE for further research on ADE detection area.  
+- The proposed dataset demonstrates promising potential for various applications, including ADE classification, caption generation, and summarization tasks.  
+- We have utilized InstructBLIP and experimented with two other pre-trained VLMs, and reported a detailed analysis.  
+- The ADE-specific model holds promise for enhancing patient safety, ADE awareness, and healthcare communication. It aims to provide individuals seeking information about ADEs with understandable and informative captions accompanying medical images to improve their comprehension of potential medication risks.
+
+# 2 Related Works
+
+This section details the related ADE detection tasks based on the data sources.
+
+Works Based on Biomedical Text and Electronic Medical Record: Various techniques have been developed for extracting ADEs from Electronic Medical Records (EMRs) (Aramaki et al., 2010), (Wang et al., 2009), as well as from medical case reports (MCRs) (Gurulingappa et al., 2011). Gurulingappa et al. (2012a) utilized machine learning methods to identify and extract potential ADE relations from MEDLINE case reports. Unlike random data sources such as social media, both EMRs and MCRs offer significant advantages by providing comprehensive records of a patient's medical history, treatment, conditions, and potential risk factors. Moreover, these records are not limited to patients who have experienced ADRs (Harpaz et al., 2013). Sarker and Gonzalez (2015) conducted a study by taking data from MEDLINE case reports and Twitter and reported how combining different datasets increases the performance of identifying ADRs. Huynh et al. (2016) explored various neural network frameworks for ADE classification, utilizing datasets from both MCRs and Twitter. DISAE is another corpus (Gurulingappa et al., 2010), which consists of 400 MEDLINE articles with annotations for disease and adverse effect names without drug-related information. D'Oosterlinck et al. (2023) introduced the BioDEX dataset for biomedical ADE extraction in real-world pharmacovigilance. This dataset comprises 65,000 abstracts, 19,000 full-text biomedical papers, and 256,000 document-level safety reports crafted by medical professionals. However, to the best of our knowledge, there is currently no publicly accessible annotated multimodal (Image and Text) corpus suitable for identifying drug-related adverse effects.
+
+Works Based on Social Media Datasets: Social media has become vital for accessing vast amounts of real-time information, making it valuable for identifying potential ADEs. Leaman et al. (2010) conducted a pioneering study that analyzed user comments from social media posts, comprising a dataset of 6890 comments. The research demonstrated the significant value of user comments in identifying ADEs, highlighting their crucial role in this context. Several other authors (Gurulingappa et al., 2012b), (Yadav et al., 2020), (Benton et al., 2011) employed lexicon-based approaches to extract ADEs. However, these methods are limited to a specific set of target ADEs. Nikfarjam and Gonzalez (2011) em
+
+Table 1: Keywords used for the data collection.
+
+Keywords: ADE, ADR, adverse drug reaction, adverse drug event, adverse reaction, adverse drug event reporting, side effects, drug reactions, drug side effects, type of infection and reaction, medicine, drugs, skin rashes, red patches, eczema, ulcer, acne, skin irritation, edema, rosacea, alopecia, lip swelling.
+
+ployed a rule-based technique instead of a naive lexicon-based approach on the same dataset, enabling the detection of ADEs not covered by lexicons. Several authors utilized supervised machine learning techniques like Support Vector Machines (SVM) (Sarker and Gonzalez, 2015), Conditional Random Fields (CRF) (Nikfarjam et al., 2015), and Random Forests (Zhang et al., 2016) for ADE detection. Sarker et al. (2016) introduced one corpus by collecting data from social media, focusing on adverse drug reactions. Tasks included automatically classifying user posts, extracting specific mentions, and normalizing mentions to standardized concepts. With the availability of annotated data, in recent times, the rise of deep learning techniques has significantly influenced research methodologies, leading to the adoption of deep learning models for predicting ADEs. Tutubalina et al. (2017) explored the synergy between CRF and Recurrent Neural Networks (RNN), demonstrating that CRF enhances the RNN model's ability to capture contextual information effectively. Chowdhury et al. (2018) developed a multi-task architecture that simultaneously tackled binary classification, ADR labeling, and indication labeling, using the PSB 2016 Social Media dataset (Sarker et al., 2016).
+
+# 3 Corpus Development
+
+Our study began with a thorough literature review to identify existing ADE-related datasets. We discovered four text-only datasets: PSB 2016 social media shared task (Sarker et al., 2016) comprising 572 tweets, Medline ADE corpus (Gurulingappa et al., 2012b) with 4,272 sentences, CADEC (Karimi et al., 2015a) containing 1,248 sentences, and recently released BioDEX dataset (D'Oosterlinck et al., 2023). This revealed a notable gap in multimodal ADE datasets, where images complement textual data. We take this opportunity to introduce a multimodal corpus consist
+
+![](images/9d97229785dfb31c35c747987d8d4ce51900fea6015e324fc75cc47014504ad6.jpg)  
+Figure 2: Some samples from the dataset containing images and the corresponding descriptive text.
+
+ing of 1,500 ADE images with corresponding English sentences to facilitate further research. While preparing this corpus, we carry out the following steps.
+
+# 3.1 Data Collection
+
+Utilizing a diverse array of keywords, we have curated a comprehensive dataset from social media, healthcare blogs, and MCRs (refer to Table 1). The inclusion of various sources ensures a broad representation of the population, enriching the dataset's diversity.
+
+# 3.1.1 Social Media
+
+Social media data is invaluable for ADE-related tasks due to its real-time nature and diverse user-generated content (Sarker et al., 2016). In this study, we have utilized the official X (Twitter) and scraper API to gather Tweets $^{2,3}$  employing diverse keywords related to ADE. The data collection phase, conducted between June 2023 and October 2023, collected a total of 20,000 tweets using specified keywords presented in Table 1. From this pool, 3,000 tweets were meticulously identified as pertinent to ADEs, featuring either images, text, or a combination of both. Notably, 142 tweets included relevant images accompanied by textual descriptions of the adverse drug events.
+
+# 3.1.2 Healthcare Blog
+
+We utilized a public healthcare-related blog, healthdirect<sup>4</sup>, a government-funded virtual health service that provides access to health advice and information via a website. We used Python's BeautifulSoup library to scrape the data and collected 1,150 unique images with corresponding text. Among these, 54 relevant images depicting adverse drug events were manually curated along with corresponding texts.
+
+# 3.1.3 Medical Case Reports
+
+Data from published medical case reports is crucial for constructing comprehensive datasets to analyze ADE. These articles provide structured and verified information, forming a reliable foundation for in-depth analysis and research in pharmacovigilance. We have extracted data from the New England Journal of Medicine<sup>5</sup>, Science Direct<sup>6</sup>. A precise Science Direct query performed is as follows:
+
+("adverse drug event") AND (Languages=English) AND (Article type=Case Reports) AND (Years=2000 to 2023))
+
+This approach retrieved approximately 2,907 documents from ScienceDirect, from which we manually selected 1390 relevant images with corresponding texts.
+
+# 3.2 Data Annotation
+
+To ensure meticulous annotation aligned with ethical standards, we enlisted the assistance of two medical students and one Ph.D. student, selected based on specific criteria. These criteria included being at least 25 years old, proficient in English (reading, writing, and speaking), and willing to handle sensitive content. The process was finalized within a span of five months, and participants received compensation for their involvement  $^{7}$ . Annotators were tasked with meticulously assessing each image and its corresponding text based on the annotation manual. Sentences accurately depicting adverse drug events, including the drug's
+
+<sup>4</sup>https://www.healthdirect.gov.au/  
+<sup>5</sup>https://www.nejm.org  
+<sup>6</sup>https://www.sciencedirect.com  
+7The medical students received compensation in the form of gift vouchers and honorarium amounts in accordance with https://www.minimum-wage.org/international/india.
+
+![](images/a18ea0b8796f9ad5c9e6aa5203f55dcfeb95bb21220e2669e81a7e15d55c1d6e.jpg)  
+Figure 3: Distribution of different body parts in the curated MMADE dataset. The number of data points and the percentage corresponding to each category have been provided.
+
+![](images/2c2fe4d1646772ebd56e51498c137e3e5095af58ae27b95b86e2a3b4ff384cf4.jpg)  
+Figure 4: Data distribution statistics from various sources, illustrating the total ADE data (images, text, and image-text pairs), and the relevant image-text pairs.
+
+name and associated side effects, were chosen for inclusion in the corpus development process, while all other instances were deemed irrelevant and removed. To maintain consistency and consensus among annotators, final rationale labels were determined through a majority voting approach. Annotators were explicitly instructed to annotate posts without biases related to demographics, religion, or other extraneous factors. The quality of annotations was assessed by measuring inter-annotator agreement (IAA) using Cohen's Kappa score (Viera et al., 2005). The resulting agreement scores of 0.78 affirm the acceptability and high quality of the annotations.
+
+# 3.3 Annotation Manual
+
+We initially provided annotators with an instruction manual detailing different instances, accompanied by examples (as depicted in Fig. 5).
+
+- For each data instance comprising text and an image, select the data instance if both the text and image indicate concerns about drug side effects.  
+- Remove data instances where the image does not convey the side effects mentioned in the text.  
+- Remove data instances where the text does not convey any side effects related to drugs, but some side effects are visually present in the accompanying image.  
+- Remove instances where both the text and image are unrelated to any drug side effect.  
+- If a data instance includes a URL link, access the content at that URL address to gain additional insight and context about the data.
+
+# 3.4 Corpus Analysis
+
+Figure 4 provides an overview of the data sources and the distribution of relevant cases within the dataset. Following the initial collection of 7,057 ADE-related samples encompassing images, text, and image-text pairs, we curated 1,500 pertinent samples. These encompass both image and their associated text descriptions of ADEs, establishing the foundation of our multimodal ADE dataset. Figure 2 depicts some of
+
+![](images/3bfd91cf627ae890bb29d18068c8eeb95b43ba21ddbaff557467dfa2d716eb0f.jpg)  
+Figure 5: Illustration of an annotation process, with A1, A2, and A3 representing individual annotators and MV signifying final majority voting. In this setup, "1" denotes ADR, while "0" is unrelated to ADR.
+
+the samples from the dataset. The first sample shows a person suffering from tongue discoloration after taking a Chlorhexidine drug. Both the text and visual image show a clear indication of drug reaction. The second sample is of a woman with papulopustular eruption on the face, which had worsened with topical metronidazole gel. Adverse drug events can manifest internally or externally, yet acquiring images of internal body parts from public sources is challenging. Thus, we focus on external body parts or symptoms, which can be readily captured and shared with doctors, pharmacovigilance teams, or the public domain for improved consultation and advice. Following dataset analysis, we identified 13 significant adverse effects crucial for multimodal ADE reporting. These effects are categorized into four groups based on their origin: ENT (9.85%), EYE (3.6%), LIMB (5.4%), and SKIN (81.06%). Additional details of the dataset, such as the distribution of samples across different body parts along with corresponding percentages, are illustrated in Fig. 3.
+
+# 4 Problem Formulation
+
+Each data point in the dataset encompasses a patient's textual description  $T$  along with a corresponding image  $I$  illustrating their medical issue or concern visually. The textual representation comprises a sequence of words  $\{t_1,t_2,\ldots ,t_n\}$  while the visual elements are represented by  $I\in \mathbb{R}^{3\times W\times H}$ , where  $W$  and  $H$  denote the width and height of the image data, respectively. The objective is to process both the text  $T$  and image  $I$
+
+for each patient and generate a natural language sequence  $Y$  that seamlessly integrates both modalities, expressed as  $Y = \{T, I\}$ .
+
+# 5 Methodology
+
+Recent advancements in VLMs, such as BLIP, InstructBlip, and GIT, have showcased remarkable advancements in encoding both textual and visual inputs. These models outperform traditional approaches that rely solely on individual image or text encoders, followed by fusion. By integrating sophisticated mechanisms for joint representation learning, VLMs excel at capturing intricate relationships between textual and visual modalities, thereby enhancing their ability to generate more contextually relevant and coherent outputs (Zhang et al., 2023). In the proposed work, we have leveraged InstructBlip (Dai et al., 2023), known for its exceptional performance across a range of vision-language (VL) tasks including Visual Question Answering (VQA), Image captioning, and Image retrieval. Each patient's inquiry or concern is articulated as a textual sentence along with a visual image where they elaborate on their medical queries or concerns on social media platforms or healthcare blogs to obtain pertinent feedback or advice. InstructBlip integrates two distinct encoders specialized for separate modalities. This architecture includes an image encoder, which utilizes a vision transformer (ViT) to extract visual features, alongside an LLM and a Query Transformer (Q-Former). The Q-Former interacts with the image encoder's output through cross-attention, resulting in K-encoded visual vectors, which are then linearly projected and fed into the frozen LLM for further processing. This representation serves as input for a proficient language model, which generates high-quality text. During instruction tuning, only the Q-Former undergoes fine-tuning, while the image encoder and LLM remain unchanged. Figure 6 illustrates the detailed architecture. We fine-tuned the InstructBLIP to assess its performance on the proposed MMADE dataset. InstructBLIP maintains a consistent image resolution  $(224 \times 224)$  during instruction tuning and freezes the visual encoder during fine-tuning. This approach substantially reduces the number of trainable parameters from 1.2 billion to 188 million, improving fine-tuning efficiency. We also employed two more VLMs and fine-tuned them with the proposed MMADE dataset. BLIP (Li et al., 2022), a
+
+![](images/d94c73b2be65c7b4c059e3e675e38542eb7ca224e2288caa78cc903f386c42f1.jpg)  
+Figure 6: The architectural details are illustrated, depicting the input image and corresponding text. Additionally, the frozen and trainable layers are highlighted for clarity.
+
+versatile vision-language model, utilizes a multimodal mixture of encoder-decoder models during pre-training. This involves bootstrapping a dataset from large-scale noisy image-text pairs, where synthetic captions are injected, and noisy captions are eliminated. Additionally, we utilized GIT (Wang et al., 2022), a VLM model that generates text descriptions of images. Trained on curated datasets of image-text pairs, GIT encodes image features into a latent representation, which is decoded into text descriptions. Its architecture comprises an image encoder using a pre-trained Swin transformer and a text decoder based on a standard transformer decoder, linked by a cross-attention layer for enhanced focus on specific image encoding parts. Additionally, we performed the integration of LSTM networks with VGG16 (Simonyan and Zisserman, 2014) and ResNet50 (He et al., 2016) architectures. In this setup, the LSTM serves as the text encoder, while either VGG16 or ResNet50 acts as the visual encoder. The features extracted from the text and visual encoders are then concatenated to create a joint representation, allowing for comprehensive modeling of both textual and visual information. Please refer to Section 12 of the Appendix for the fine-tuning details.
+
+# 6 Experimental Results and Analysis
+
+We have utilized four commonly employed evaluation metrics to assess the performance of the models, including BLEU score (Bilingual Evaluation Understudy) (Papineni et al., 2002), ROUGE score (Recall-Oriented Understudy for Gisting Evaluation) (Lin, 2004), BERTScore (Zhang et al., 2019), and MoverScore (Zhao et al., 2019). Detailed explanations of experimental settings and metrics are
+
+Table 2: The performance evaluation of the models using ROUGE and BLEU Scores in the multimodal dataset setting.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Type</td><td colspan="3">ROUGE</td><td colspan="3">BLEU</td></tr><tr><td>R1</td><td>R2</td><td>RL</td><td>B1</td><td>B2</td><td>B3</td></tr><tr><td>LSTM+VGG16</td><td>-</td><td>0.213</td><td>0.105</td><td>0.201</td><td>0.165</td><td>0.073</td><td>0.041</td></tr><tr><td>LSTM+ResNet50</td><td>-</td><td>0.281</td><td>0.086</td><td>0.230</td><td>0.179</td><td>0.058</td><td>0.046</td></tr><tr><td rowspan="2">BLIP</td><td>Base</td><td>0.19</td><td>0.093</td><td>0.185</td><td>0.099</td><td>0.003</td><td>0.001</td></tr><tr><td>Fine-Tune</td><td>0.334</td><td>0.163</td><td>0.225</td><td>0.171</td><td>0.081</td><td>0.058</td></tr><tr><td rowspan="2">GIT</td><td>Base</td><td>0.27</td><td>0.11</td><td>0.192</td><td>0.157</td><td>0.014</td><td>0.004</td></tr><tr><td>Fine-Tune</td><td>0.504</td><td>0.285</td><td>0.416</td><td>0.194</td><td>0.10</td><td>0.097</td></tr><tr><td rowspan="2">InstructBLIP</td><td>Base</td><td>0.29</td><td>0.161</td><td>0.212</td><td>0.219</td><td>0.125</td><td>0.008</td></tr><tr><td>Fine-Tune</td><td>0.571</td><td>0.351</td><td>0.475</td><td>0.319</td><td>0.175</td><td>0.112</td></tr></table>
+
+Table 3: The evaluation of the model using BERTScore and MoverScore in the multimodal dataset setting.  
+
+<table><tr><td colspan="8">BERTScore</td><td></td></tr><tr><td rowspan="2">Evaluation Metrics</td><td rowspan="2">LSTM+VGG16</td><td rowspan="2">LSTM+ResNet50</td><td colspan="2">BLIP</td><td colspan="2">GIT</td><td colspan="2">InstructBLIP</td></tr><tr><td>Base</td><td>Fine-Tune</td><td>Base</td><td>Fine-Tune</td><td>Base</td><td>Fine-Tune</td></tr><tr><td>Precision</td><td>0.821</td><td>0.847</td><td>0.819</td><td>0.841</td><td>0.826</td><td>0.866</td><td>0.832</td><td>0.896</td></tr><tr><td>Recall</td><td>0.797</td><td>0.811</td><td>0.783</td><td>0.819</td><td>0.791</td><td>0.831</td><td>0.781</td><td>0.891</td></tr><tr><td>F1-score</td><td>0.809</td><td>0.821</td><td>0.801</td><td>0.829</td><td>0.812</td><td>0.852</td><td>0.805</td><td>0.893</td></tr><tr><td colspan="9">MoverScore</td></tr><tr><td></td><td>0.441</td><td>0.487</td><td>0.482</td><td>0.513</td><td>0.496</td><td>0.553</td><td>0.544</td><td>0.622</td></tr></table>
+
+given in Section 11 and 13 of the Appendix.
+
+# 6.1 Findings from Experiments
+
+The experimental results across Tables 2 and 3 depict the performance of various models in multimodal dataset settings, while Tables 4 and 5 represent the performance in unimodal dataset settings, based on standard evaluation metrics. From the
+
+Table 4: The evaluation of the models using BERTScore and MoverScore in the unimodal dataset setting.  
+
+<table><tr><td colspan="6">BERTScore</td></tr><tr><td>Evaluation Metrics</td><td>LSTM+VGG16</td><td>LSTM+ResNet50</td><td>BLIP</td><td>GIT</td><td>InstructBLIP</td></tr><tr><td>Precision</td><td>0.731</td><td>0.744</td><td>0.753</td><td>0.801</td><td>0.840</td></tr><tr><td>Recall</td><td>0.709</td><td>0.713</td><td>0.724</td><td>0.783</td><td>0.783</td></tr><tr><td>F1-score</td><td>0.719</td><td>0.729</td><td>0.738</td><td>0.790</td><td>0.810</td></tr><tr><td colspan="6">MoverScore</td></tr><tr><td></td><td>0.121</td><td>0.143</td><td>0.151</td><td>0.184</td><td>0.253</td></tr></table>
+
+results, several observations are:
+
+- Table 2 presents the ROUGE and BLEU scores, indicating the superior performance of fine-tuned InstructBLIP compared to other models. The higher ROUGE and BLEU scores achieved by fine-tuned InstructBLIP suggest its proficiency in capturing relevant information from the input and generating text. This superiority can be attributed to its training on more extensive and diverse datasets, enabling it to capture intricate data patterns effectively.  
+- Tables 3 and 4 present BERTScores and MoverScores in the multimodal and unimodal settings, respectively. Fine-tuned InstructBLIP achieves the highest BERTScore and MoverScores, demonstrating its efficacy in capturing contextual similarity and superior ability to convey meaningful content effectively.  
+- Table 4 and 5 show the decline in performance when training VLM models with unimodal data. The absence of meaningful visual information likely contributes to the performance degradation.  
+- One important observation is that Instruct-BLIP consistently outperforms BLIP and GIT across all metrics, showing its superior capability in effectively integrating textual and visual information. This can be attributed to its innovative architecture, featuring a Query Transformer for instruction-aware feature extraction. Refer to Section 16 of the Appendix for detailed outcomes.  
+- Fine-tuning with domain-specific ADE data remarkably enhances model performance, reflecting its pivotal role in adapting models to the intricacies of adverse drug event detection.  
+- Another key finding highlights the substantial performance enhancement achieved by integrating both image and text modalities, emphasizing the critical role of visual information alongside textual data.
+
+Statistical Analysis: We conducted a statistical t-test to compare the performance of the proposed multimodal model, utilizing both image and text data, with that of unimodal models. The analysis
+
+![](images/0f792ccbe22ac352fc0a70f43fd3821100626402ddda0b50c11da6a681321df7.jpg)  
+Actual: I was given penicillin for my ongoing treatment. About 4 hours later, I started feeling really sick. I came down with a high fever, nausea, vomiting, and a terrible headache. My whole body ached, and got this type of rash on my skin. What type of rash it is?  
+Figure 7: An example case study where a sample image and its corresponding text are provided along with the generated text from all models in both unimodal and multimodal settings.
+
+<table><tr><td>Model</td><td>Unimodal
+(Blank Image and Text)</td><td>Multimodal
+(Image and Text)</td></tr><tr><td>BLIP</td><td>A man with a large stomach and a large stomach.</td><td>A close up of a person&#x27;s stomach covered in a red rash.</td></tr><tr><td>GIT</td><td>A middle aged man sits on an exam table in a doctor&#x27;s office, looking unwell.</td><td>A man with a lot of red spots suffering from rashes.</td></tr><tr><td>InstructBLIP</td><td>A person&#x27;s covered in a red rash.</td><td>The image shows red spots on the abdomen, which is unusual for a person. The red spots are likely caused by a skin condition, such as syphilis.</td></tr></table>
+
+yielded a p-value below 0.05, indicating a significant difference in performance between the two. The detailed explanation is given in Section 14 of Appendix.
+
+# 6.2 Qualitative Analysis
+
+We performed an extensive qualitative analysis of the responses generated by various models in both unimodal and multimodal settings, complemented by several case studies. A detailed case study is depicted in Fig. 7. The analysis has led to the following conclusion: (a) In multimodal settings, all the models demonstrate superior performance and exhibit a greater ability to capture crucial visual information conveyed through images than in unimodal settings (refer to Table 4 and 5). (b) Observations also revealed that models such as BLIP and GIT tended to hallucinate, as evidenced by Fig. 7, occasionally generating facts that were entirely unrelated to the context. (c) InstructBLIP demonstrates superior performance, leading us to conclude that providing instructions during finetuning prompts the model to selectively focus on pertinent visual features. This focused attention encourages the model to generate target sequences that closely resemble the desired output. We have presented a case study demonstrating that model performance is not solely determined by the number of samples in the dataset but also by other factors such as the distinct visual characteristics of different types of ADEs present in the dataset. A detailed explanation is provided in the Section 15 of Appendix.
+
+Table 5: The performance evaluation of the models using ROUGE and BLEU Scores in the unimodal dataset setting.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">ROUGE</td><td colspan="3">BLEU</td></tr><tr><td>R1</td><td>R2</td><td>RL</td><td>B1</td><td>B2</td><td>B3</td></tr><tr><td>LSTM+VGG16</td><td>0.133</td><td>0.018</td><td>0.101</td><td>0.141</td><td>0.012</td><td>0.001</td></tr><tr><td>LSTM+ResNet50</td><td>0.146</td><td>0.029</td><td>0.112</td><td>0.152</td><td>0.016</td><td>0.002</td></tr><tr><td>BLIP</td><td>0.158</td><td>0.032</td><td>0.129</td><td>0.161</td><td>0.019</td><td>0.002</td></tr><tr><td>GIT</td><td>0.169</td><td>0.035</td><td>0.137</td><td>0.172</td><td>0.020</td><td>0.003</td></tr><tr><td>InstructBLIP</td><td>0.211</td><td>0.087</td><td>0.172</td><td>0.197</td><td>0.089</td><td>0.006</td></tr></table>
+
+# 7 Risk Analysis
+
+While our multimodal model demonstrates promise, it is essential to have medical experts and pharmacovigilance teams validate the findings, considering other critical factors. Our model and dataset are intended to support medical professionals rather than replace them.
+
+# 8 Conclusion and Future Work
+
+In this paper, we present the task of ADE detection within pharmacovigilance mining, leveraging multimodal datasets. In order to solve this task, we have created a multimodal ADE dataset, MMADE, containing images and corresponding descriptions, enhancing decision-making with the inclusion of visual cues. We have employed InstructBLIP, finetuned with the proposed dataset, and compared it with other models. Our findings suggest that domain-specific fine-tuning significantly enhances overall performance, emphasizing the importance of multimodal visual cues. We envision MMADE as a pivotal resource for advancing research in multimodal ADE detection. Moreover, our finetuned architecture holds promise as a valuable tool for pharmacovigilance teams, clinicians, and researchers, facilitating more effective ADE monitoring and ultimately improving patient safety and outcomes. In addition to expanding the dataset, future investigations could explore the potential of this multimodal dataset in tasks such as ADE severity classification and summarization.
+
+# 9 Limitations
+
+While our effort aimed to develop an ADE detection framework and introduce the novel MMADE dataset, comprising textual descriptions of drug events paired with images, it is crucial to acknowledge certain limitations inherent in the dataset. Specifically, our dataset primarily focuses on drug
+
+events associated with external body parts, omitting data about internal conditions such as liver infections, kidney stones, or psychological ailments like depression and migraine. Acquiring a substantial volume of image-text pairs within the ADE domain presents inherent challenges, including data privacy concerns, regulatory constraints, and the specialized nature of ADE occurrences. Despite these obstacles, our research breaks new ground by integrating images with text, solving a real-world challenge where individuals affected by ADE may resort to image sharing for communication when verbally expressing their symptoms is difficult. Moving forward, we aim to enhance the dataset by incorporating more ADE-related images and expanding its utility through additional tasks such as complaint identification and text summarization.
+
+# 10 Ethics and Broader Impact
+
+# User Privacy.
+
+Our dataset contains AD images and corresponding text tweets with annotation labels and no personal user information.
+
+# Biases.
+
+Any biases detected in the dataset are inadvertent, and we have no intention of harming anyone or any group. We acknowledge that evaluating whether a tweet is ADE can be subjective, so we have taken agreement from all the annotators before selecting the data.
+
+# Intended Use.
+
+We share our data to promote more research on Adverse drug event detection. We only release the dataset for research purposes and do not grant a license for commercial use.
+
+# References
+
+Eiji Aramaki, Yasuhide Miura, Masatsugu Tonoike, Tomoko Ohkuma, Hiroshi Masuichi, Kayo Waki, and Kazuhiko Ohe. 2010. Extraction of adverse drug effects from clinical records. In MEDINFO 2010, pages 739-743. IOS Press.  
+Adrian Benton, Lyle Ungar, Shawndra Hill, Sean Hennessy, Jun Mao, Annie Chung, Charles E Leonard, and John H Holmes. 2011. Identifying potential adverse effects using the web: A new approach to medical hypothesis generation. Journal of biomedical informatics, 44(6):989-996.  
+Shaika Chowdhury, Chenwei Zhang, and Philip S Yu. 2018. Multi-task pharmacovigilance mining from
+
+social media posts. In Proceedings of the 2018 World Wide Web Conference, pages 117-126.  
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. Instructclip: Towards general-purpose vision-language models with instruction tuning.  
+Karel D'Oosterlinck, François Remy, Johannes Deleu, Thomas Demeester, Chris Develder, Klim Zaporojets, Aneiss Ghodsi, Simon Ellershaw, Jack Collins, and Christopher Potts. 2023. *Biodex: Large-scale biomedical adverse drug event extraction for real-world pharmacovigilance.* arXiv preprint arXiv:2305.13395.  
+Akash Ghosh, Arkadeep Acharya, Raghav Jain, Sriparna Saha, Aman Chadha, and Setu Sinha. 2024a. Clipsntel: CLIP and LLM synergy for multimodal question summarization in healthcare. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada, pages 22031-22039. AAAI Press.  
+Akash Ghosh, Arkadeep Acharya, Prince Jha, Sriparna Saha, Aniket Gaudgaul, Rajdeep Majumdar, Aman Chadha, Raghav Jain, Setu Sinha, and Shivani Agarwal. 2024b. Medsumm: A multimodal approach to summarizing code-mixed Hindi-english clinical queries. In Advances in Information Retrieval - 46th European Conference on Information Retrieval, ECIR 2024, Glasgow, UK, March 24-28, 2024, Proceedings, Part V, volume 14612 of Lecture Notes in Computer Science, pages 106-120. Springer.  
+Harsha Gurulingappa, Juliane Fluck, Martin Hofmann-Apitius, and Luca Toldo. 2011. Identification of adverse drug event assertive sentences in medical case reports. In First international workshop on knowledge discovery and health care management (KD-HCM), European conference on machine learning and principles and practice of knowledge discovery in databases (ECML PKDD), pages 16-27.  
+Harsha Gurulingappa, Roman Klinger, Martin Hofmann-Apitius, and Juliane Fluck. 2010. An empirical evaluation of resources for the identification of diseases and adverse effects in biomedical literature. In 2nd Workshop on Building and evaluating resources for biomedical text mining (7th edition of the Language Resources and Evaluation Conference), pages 15-22.  
+Harsha Gurulingappa, Abdul Mateen-Rajpu, and Luca Toldo. 2012a. Extraction of potential adverse drug events from medical case reports. Journal of biomedical semantics, 3(1):1-10.  
+Harsha Gurulingappa, Abdul Mateen Rajput, Angus Roberts, Juliane Fluck, Martin Hofmann-Apitius, and
+
+Luca Toldo. 2012b. Development of a benchmark corpus to support the automatic extraction of drug-related adverse effects from medical case reports. Journal of biomedical informatics, 45(5):885-892.  
+Katja M Hakkarainen, Khadidja Hedna, Max Petzold, and Staffan Hägg. 2012. Percentage of patients with preventable adverse drug reactions and preventability of adverse drug reactions-a meta-analysis. PloS one, 7(3):e33236.  
+Rave Harpaz, Santiago Vilar, William DuMouchel, Hajjat Salmasian, Krystl Haerian, Nigam H Shah, Herbert S Chase, and Carol Friedman. 2013. Combining signals from spontaneous reports and electronic health records for detection of adverse drug reactions. Journal of the American Medical Informatics Association, 20(3):413-419.  
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778.  
+Trung Huynh, Yulan He, Alistair Willis, and Stefan Rueger. 2016. Adverse drug reaction classification with deep neural networks. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 877-887, Osaka, Japan. The COLING 2016 Organizing Committee.  
+Sarvnaz Karimi, Alejandro Metke-Jimenez, Madonna Kemp, and Chen Wang. 2015a. Cadec: A corpus of adverse drug event annotations. Journal of biomedical informatics, 55:73-81.  
+SarvNaz Karimi, Chen Wang, Alejandro Metke-Jimenez, Raj Gaire, and Cecile Paris. 2015b. Text and data mining techniques in adverse drug reaction detection. ACM Computing Surveys (CSUR), 47(4):1-39.  
+Robert Leaman, Laura Wojtulewicz, Ryan Sullivan, Annie Skariah, Jian Yang, and Graciela Gonzalez. 2010. Towards internet-age pharmacovigilance: extracting adverse drug reactions from user posts in health-related social networks. In Proceedings of the 2010 workshop on biomedical natural language processing, pages 117-125.  
+Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models.  
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR.  
+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.
+
+Azadeh Nikfarjam and Graciela H Gonzalez. 2011. Pattern mining for extraction of mentions of adverse drug reactions from user comments. In AMIA annual symposium proceedings, volume 2011, page 1019. American Medical Informatics Association.  
+Azadeh Nikfarjam, Abeed Sarker, Karen O'connor, Rachel Ginn, and Graciela Gonzalez. 2015. Pharmacovigilance from social media: mining adverse drug reaction mentions using sequence labeling with word embedding cluster features. Journal of the American Medical Informatics Association, 22(3):671-681.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.  
+Pranab Sahoo, Prakash Meharia, Akash Ghosh, Sriparna Saha, Vinija Jain, and Aman Chadha. 2024a. Unveiling hallucination in text, image, video, and audio foundation models: A comprehensive survey. arXiv preprint arXiv:2405.09589.  
+Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. 2024b. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927.  
+Abeed Sarker and Graciela Gonzalez. 2015. Portable automatic text classification for adverse drug reaction detection via multi-corpus training. Journal of biomedical informatics, 53:196-207.  
+Abeed Sarker, Azadeh Nikfarjam, and Graciela Gonzalez. 2016. Social media mining shared task workshop. In Biocomputing 2016: Proceedings of the Pacific Symposium, pages 581-592. World Scientific.  
+Ryoma Sato, Makoto Yamada, and Hisashi Kashima. 2022. Re-evaluating word mover's distance. In International Conference on Machine Learning, pages 19231-19249. PMLR.  
+Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.  
+Janet Sultana, Paola Cutroneo, and Gianluca Trifiro. 2013. Clinical and economic burden of adverse drug reactions. Journal of Pharmacology and Pharmacotherapeutics, 4(1_suppl):S73-S77.  
+Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mullappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. 2023. Xraygpt: Chest radiographs summarization using medical vision-language models. arXiv preprint arXiv:2306.07971.  
+Elena Tutubalina, Sergey Nikolenko, et al. 2017. Combination of deep recurrent neural networks and conditional random fields for extracting adverse drug reactions from user reviews. Journal of healthcare engineering, 2017.
+
+Anthony J Viera, Joanne M Garrett, et al. 2005. Understanding interobserver agreement: the kappa statistic. Fam med, 37(5):360-363.  
+Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. 2022. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100.  
+Xiaoyan Wang, George Hripcsak, Marianthi Markatou, and Carol Friedman. 2009. Active computerized pharmacovigilance using natural language processing, statistics, and electronic health records: a feasibility study. Journal of the American Medical Informatics Association, 16(3):328-337.  
+Shweta Yadav, Asif Ekbal, and Sriparna Saha. 2018a. Feature selection for entity extraction from multiple biomedical corpora: A pso-based approach. Soft Computing, 22(20):6881-6904.  
+Shweta Yadav, Asif Ekbal, Sriparna Saha, and Pushpak Bhattacharyya. 2019. A unified multi-task adversarial learning framework for pharmacovigilance mining. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5234-5245, Florence, Italy. Association for Computational Linguistics.  
+Shweta Yadav, Asif Ekbal, Sriparna Saha, Pushpak Bhattacharyya, and Amit Sheth. 2018b. Multi-task learning framework for mining crowd intelligence towards clinical treatment. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 271-277, New Orleans, Louisiana. Association for Computational Linguistics.  
+Shweta Yadav, Srivatsa Ramesh, Sriparna Saha, and Asif Ekbal. 2020. Relation extraction from biomedical and clinical text: Unified multitask learning framework. IEEE/ACM transactions on computational biology and bioinformatics, 19(2):1105-1116.  
+Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. 2023. Vision-language models for vision tasks: A survey.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675.  
+Zhifei Zhang, JY Nie, and Xuyao Zhang. 2016. An ensemble method for binary classification of adverse drug reactions from social media. In Proceedings of the Social Media Mining Shared Task Workshop at the Pacific Symposium on Biocomputing, volume 1.  
+Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M Meyer, and Steffen Eger. 2019. Moverscore: Text generation evaluating with contextualized embeddings and earth mover distance. arXiv preprint arXiv:1909.02622.
+
+Juexiao Zhou and Xin Gao. 2023. Skingpt: A dermatology diagnostic system with vision large language model. arXiv preprint arXiv:2304.10691.
+
+# Appendix
+
+# 11 Experimental Settings
+
+This section provides the hyperparameters and experimental setups utilized in the study. All experiments were conducted using multiple RTX 2080Ti GPUs. The dataset was partitioned, allocating  $80\%$  for training and  $20\%$  for testing. For InstructBLIP, the learning rate was set to 1e-5, executed for 50 epochs, with a batch size of 2. Similarly, for the BLIP model, a learning rate of 0.0001 was utilized and executed for 50 epochs, with a batch size of 2. For GIT fine-tuning, a learning rate of 5e-3 was applied for 50 epochs, with a batch size of 2. All models were implemented using Scikit-Learn and PyTorch as the backend framework.
+
+# 12 Fine-tuning
+
+We have followed several steps to fine-tune BLIP on our multimodal dataset. First, we have prepared the dataset in JSON format, which is compatible with the BLIP framework, and each image-text pair is represented as a dictionary with the following keys: image_path: The path to the image file, text: The caption or other text description of the image. We have fine-tuned BLIP with learning rate  $= 0.001$ , number of epochs  $= 50$ , and batch size  $= 16$ . The training process takes 6 hours for the fine-tuning, and we evaluate the performance on a held-out test set.
+
+To fine-tune the GIT model on the proposed MMADE dataset, we utilized the Hugging Face Transformers library and followed a systematic process. First, we prepared the data using a PyTorch dataset, converting it into the required format using the GitProcessor class. Finally, we have utilized the GIT-base model, which is pre-trained on a substantial dataset of image-text pairs. We have utilized Parameters such as a learning rate  $= 5\mathrm{e} - 3$  epochs  $= 30$  and batch size  $= 2$ . The optimization was performed using the Adam optimizer with the cross-entropy loss function, and the fine-tuning process took 3 hours.
+
+To fine-tune the CNN-LSTM model, we leverage the CNN models (VGG16 and ResNet50) to extract image features, followed by an LSTM for
+
+sequence generation. Model compilation involves the use of categorical cross-entropy loss and the Adam optimizer. Training proceeds for 40 epochs, utilizing a batch size of 32.
+
+# 13 Evaluation Metrics
+
+We have utilized BLEU score (Papineni et al., 2002), ROUGE score (Lin, 2004), BERTScore (Zhang et al., 2019) and Mover-Score (Zhao et al., 2019). BLEU score is utilized to assess the quality of machine-generated text by comparing it to human-generated reference text. It measures the similarity in word sequences between the machine-generated and human reference texts using n-grams, penalizing shorter machine-generated texts to provide a quantitative measure of translation accuracy. The ROUGE score is also used to evaluate the quality of machine-generated summaries compared to human-written summaries. It works by calculating the overlap of n-grams between the machine-generated summary and the reference summaries. ROUGE and BLEU metrics evaluate text quality based on syntactic overlap, considering unigrams and bigrams, lacking the ability to decode semantic meaning effectively. However, BERTScore focuses on understanding the semantic meaning of generated text compared to the intended text, enabling a more nuanced and accurate comparison. MoverScore is a way to measure how similar a machine-generated text is to a human-written text. It does this by using BERT embeddings to understand the meaning of the words and sentences in both texts and then using Word Mover's Distance (Sato et al., 2022) to measure how similar the two texts are.
+
+# 14 Statistical Analysis
+
+In this study, we employed the paired t-test to assess the statistical significance of differences between the outcomes of unimodal and multimodal models. The null hypothesis (H0) assumes no significant disparity in scores, while the alternative hypothesis (H1) proposes the opposite. We structured our analysis around the assumption of paired data, where each BLEU score corresponds to the same model in different settings (Multimodal Vs. Unimodal). By applying the paired t-test, we aimed to rigorously evaluate performance disparities and offer insights into the models' relative effectiveness. The obtained p-values for ROUGE, BLEU and BERTScore scores were 0.008, 0.0097 and
+
+![](images/8f9f91d1168552918b6a582f65d28df839697e2d8f868b1211491d1603e4505d.jpg)  
+Figure 8: Illustration of a case study showcasing the performance of the InstructBLIP model from the dataset
+
+0.0316, respectively, both below the 0.05 threshold, leading to the rejection of the null hypothesis.
+
+![](images/de6c5d46c3296a45d5d157a103f74f2fc277353cf8ced4c4c11809c3bf2359ec.jpg)  
+Figure 9: Example of the sample image and corresponding text from the dataset along with generated text from the Proposed model with two other models.
+
+# 15 Case study
+
+The case study depicted in Fig. 8 reveals intriguing insights across various body parts. In the first row, depicting rashes, fine-tuned InstructBLIP accurately identifies distinct body parts and key findings, potentially attributed to the diverse visual characteristics of different rash types prevalent in the dataset with the largest portion of  $81.06\%$ . However, performance in identifying mouth-related ADEs is less satisfactory despite comprising  $5.93\%$  of the dataset, likely due to potential confusion with other
+
+oral features like the tongue, teeth, or lips. Conversely, despite eye-related problems representing only  $1.8\%$  of the dataset, the model performs comparatively better in this category, possibly because it focuses specifically on infected regions, enhancing its ability to identify relevant features accurately.
+
+# 16 Comparative output
+
+We have added two more examples in Fig. 9 showing the ADR instances, corresponding ground truth, and the model-generated outputs.

2024/Enhancing Adverse Drug Event Detection with Multimodal Dataset_ Corpus Creation and Model Development/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29eacb4b7d1a824d3e2eacd8ef935f4422752e57755c21febb90ae65e6a8d549
+size 421437

2024/Enhancing Cross Text-Molecule Learning by Self-Augmentation/206d3795-d53b-42c9-8de8-6d682eb0ac0b_content_list.json

@@ -0,0 +1,1760 @@
+[
+    {
+        "type": "text",
+        "text": "Enhancing Cross Text-Molecule Learning by Self-Augmentation",
+        "text_level": 1,
+        "bbox": [
+            166,
+            89,
+            831,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yinuo Jiang, Xiang Zhuang, Keyan Ding,",
+        "bbox": [
+            319,
+            134,
+            677,
+            151
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Qiang Zhang*, Huajun Chen*",
+        "bbox": [
+            368,
+            151,
+            628,
+            167
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "College of Computer Science and Technology, Zhejiang University",
+        "bbox": [
+            226,
+            167,
+            769,
+            184
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ZJU-Hangzhou Global Scientific and Technological Innovation Center",
+        "bbox": [
+            211,
+            184,
+            784,
+            200
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{yinuoj,zhuangxiang,dingkeyan,qiang.zhang.cs,huajunsir} @zju.edu.cn",
+        "bbox": [
+            213,
+            200,
+            786,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The development of Large Language Models (LLMs) has greatly advanced the field of drug discovery, with the belief that natural language can enhance human control over molecule design. However, the scarcity of high-quality labeled data remains a challenge for cross-text-molecule learning. Existing datasets are limited due to the difficulty of collecting precise molecule-description pairs. Although recent efforts have utilized pseudo data generated by LLMs for augmentation, the lack of specialized chemistry knowledge of LLMs and the absence of an effective high-quality data selector may introduce noise into the annotations, compromising the models' robustness. To address these challenges, this paper introduces a novel framework that interweaves model finetuning and data augmentation to overcome the scarcity of high-quality data. The proposed approach involves an iterative procedure where the model plays dual roles in annotating unlabeled data and sampling a subset of high-quality data until convergence is achieved, enhancing the model's understanding and adaptability. Additionally, a new dataset called SAPubChem-41 is presented, which comprises meticulously curated high-quality parallel molecule-description pairs designed specifically for fine-tuning purposes. This research provides an important contribution to the field by addressing the need for high-quality datasets and presenting an effective framework for cross-text-molecule learning.",
+        "bbox": [
+            144,
+            290,
+            460,
+            760
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            774,
+            260,
+            789
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The emergence of Large Language Models (LLMs) has significantly propelled the development of drug discovery (Zhang et al., 2024). Recent progress in language models shed light on drug discovery, with the vision that humans can possess a higher-level control over molecule design facilitated by natural",
+        "bbox": [
+            112,
+            801,
+            489,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "language. Edwards et al. (2022a) introduce two new tasks: molecule captioning (Mol2Cap) and text-based molecule generation (Cap2Mol), and present MolT5 (Edwards et al., 2022b) based on the T5 (Raffel et al., 2020) architecture to translate between molecule and text. Subsequently, diverse endeavors have been undertaken to tackle these challenges. Text+Chem T5 (Christofidellis et al., 2023) and BioT5 (Pei et al., 2024) are also T5-like models, which incorporate multi-task and multidomain pretraining process and enable the bidirectional generation between different modalities in a single model. Some other methods employ the Generative Pretrained Transformer (GPT) architecture. MolXPT (Liu et al., 2023) proposes to wrap molecules in the sentences to make the pertaining corpus. More recently, MolCA(Liu et al., 2024) considers the 2D graph information for molecules and designs a projector to connect molecular content to language models. These studies mark the research efforts in the realm of cross text-molecule learning.",
+        "bbox": [
+            507,
+            261,
+            885,
+            613
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite rapid development in cross-text-molecule modalities, there is still demand for high-quality labeled data in this field. Such data of superior quality is indispensable to capacitate models for downstream tasks. However, the available datasets relevant to the tasks are limited, a consequence of the arduous nature of collecting precise and enlightening parallel molecule-description pairs. This scarcity of annotated data hinders the model's ability to learn between the modalities of the molecule and the text. As shown in Figure 1, recent efforts by Chen et al. (2023) address the scarcity challenge by utilizing pseudo data generated by LLMs as augmentation. The problem is that the absence of specialized knowledge in chemistry may introduce noise into the data annotated by general Large Language Models, thereby compromising the robustness of models fine-tuned on such datasets. In spite of these commendable advance",
+        "bbox": [
+            507,
+            615,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author.",
+        "bbox": [
+            136,
+            906,
+            280,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "9551",
+        "bbox": [
+            478,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2024, pages 9551-9565",
+        "bbox": [
+            228,
+            945,
+            768,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "August 11-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            284,
+            958,
+            712,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ments, the accessibility of high-quality datasets for fine-tuning remains notably deficient.",
+        "bbox": [
+            114,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/272b4bf873814a136101d35b94d4363a6c239a550a9bda1a8367e046d42a9d8c.jpg",
+        "image_caption": [
+            "(a) Conventional Data Augmentation Method"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            169,
+            131,
+            436,
+            256
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/5c9e685af13cd22f40c029e7e296376a3c547031d72cedeb36c03bb1feb72951.jpg",
+        "image_caption": [
+            "(b) Our Self-Augmentation Method",
+            "Figure 1: A comparison between a conventional data augmentation method referenced in Chen et al. (2023) and our iterative self-augmentation techniques is drawn herein. The original technique leverages Label Language Models (LLMs) to annotate unlabeled data; contrastingly, our proposed self-augmentation strategies use the model itself for annotating unlabeled data in tandem with an iterative training approach."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            169,
+            286,
+            438,
+            414
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these challenges, this paper introduces a novel framework that interweaves model fine-tuning and data augmentation (See Fig. 1, aiming to rectify the high-quality data scarcity observed in the realm of cross text-molecule learning. Our proposed approach integrates the model into the process by assigning it dual roles: annotating unlabelled data and sampling a subset of high-quality data. This iterative procedure continues until convergence is achieved, thereby refining the model's understanding and adaptability. Additionally, we present a new, larger dataset SAPubChem-41. This meticulously curated dataset comprises a wealth of high-quality parallel molecule-description pairs meticulously designed for the specific purpose of fine-tuning. The model optimized using SAPubChem-41 has been evaluated across a total of 13 molecule-text benchmarks, exhibiting superior performance in comparison to the model optimized using the original dataset, ChEBI-20. Our key contributions can be summarized as:",
+        "bbox": [
+            112,
+            568,
+            489,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We present an innovative framework designed to augment cross-text-molecule learning through a self-augmentation strategy. This system leverages the model itself to annotate unlabelled data, choosing only samples that meet a high-quality standard.",
+            "- We introduce SAPubChem-41, a novel dataset consisting of both real data and high-quality pseudo-labeled data, augmented by the model itself. This makes a notable contribution to the available datasets in the field of cross-molecule-text modality.",
+            "- We show that within our iterative setting, self-augmented data significantly improves the performance of the model in downstream tasks with each epoch. Furthermore, these results tend to converge as the number of epochs increase. This process thereby successfully validates the efficacy of self-augmented data in enhancing model performance."
+        ],
+        "bbox": [
+            531,
+            84,
+            885,
+            445
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Background",
+        "text_level": 1,
+        "bbox": [
+            507,
+            463,
+            650,
+            479
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Cross Text-Molecule Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            494,
+            794,
+            508
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Edwards et al. (2021) introduce a new task Text2Mol, which uses descriptions as search queries to retrieve the target molecules. Edwards et al. (2022a) first addresses the problem of cross-domain generation by linking natural language and chemistry, tackling tasks such as text-conditional de novo molecule generation and molecule captioning. Li et al. (2023a) combines retrieval-based prompt paradigm with LLMs like ChatGPT to achieve translation between molecule language and natural language. Christofidellis et al. (2023) propose the first multi-domain, multi-task language model that can solve a wide range of tasks in both the chemical and natural language domains. Liu et al. (2023) propose MolXPT, a GPT-based model pre-trained on molecule SMILES, biomedical text, and wrapped text. BioT5 (Pei et al., 2024) further exploits SELFIES for  $100\\%$  robust molecular representations and discriminate structured knowledge from unstructured knowledge. To capitalize on these insights, we adopt an iterative framework to interweave model fitting and data augmentation, to mitigate data scarcity and further enhance the alignment between molecule and text representations.",
+        "bbox": [
+            505,
+            519,
+            885,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "9552",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Data Augmentation for Cross Text-Molecule Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            389,
+            116
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Data augmentation is a regularization strategy utilized to enhance model performance by diversifying the available data through various techniques (Hernández-García and König, 2018). In the field of natural language processing (NLP), data augmentation has been widely employed to address data scarcity issues, leading to the proposal of various augmentation methods (Feng et al., 2021; Li et al., 2022). Textual data augmentation can be accomplished through simple rules, such as synonym replacement and word order modification (Zhang et al., 2015). Furthermore, the development of deep learning models has facilitated the generation of new text to augment the existing data (Szegedy et al., 2016; Wu et al., 2019; Anaby-Tavor et al., 2020; Yoo et al., 2021; Zhou et al., 2021; Li et al., 2023b; Dai et al., 2023). Acquiring high-quality annotated molecule-text datasets is prohibitively expensive, for many descriptions of corresponding molecules are scattered and inaccurate, leading to limited fine-tuning datasets for cross text-molecule learning. More data augmentation techniques are desired to further enhance cross text-molecule learning. Chen et al. (2023) pioneers to exploit LLMs' annotated pseudo label for domain adaption and data augmentation. However, while pseudo label as domain adaptation shows impressive performance compared to existing methods, pseudo label as data augmentation still struggles against label noise interfering model's performance. To address the problem and further improve the quality of augmented dataset, our work introduces a pioneering iterative framework that interweaves model fine-tuning and data augmentation.",
+        "bbox": [
+            112,
+            122,
+            492,
+            670
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            683,
+            218,
+            697
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we describe a computational framework for jointly optimizing model performance and dataset quality.",
+        "bbox": [
+            112,
+            709,
+            489,
+            758
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Overall Framework",
+        "text_level": 1,
+        "bbox": [
+            112,
+            770,
+            317,
+            784
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The literature demonstrates a substantial body of research focused on enhancing both model performance and dataset quality concurrently, as depicted in Figure 2. If we assume the existence of a subjective testing environment capable of gathering reliable molecule-caption pairs, then the core challenge becomes how to effectively sample a subset  $A$  from a large-scale pseudo-labeled dataset  $D$",
+        "bbox": [
+            112,
+            791,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "that exhibits high confidence. In line with Li et al. (2023b), our approach adopts an iterative strategy, combining model fine-tuning and data augmentation operations defined below.",
+        "bbox": [
+            507,
+            84,
+            884,
+            148
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Model fine-tuning: Given a data distribution, find an approximate risk minimizer.",
+            "2. Data augmentation: Given a model, augment data and sample a new data distribution.",
+            "3. Data combination: Combine a set of distributions into a single distribution."
+        ],
+        "bbox": [
+            507,
+            149,
+            882,
+            244
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "At the beginning of each iteration, a new model is finetuned on the latest dataset. This newly trained model is subsequently employed to annotate a set of unlabeled data, thereby generating a pseudolabeled dataset. The sampler is then entrusted with the responsibility of discerning samples distinguished by their high quality from this pseudolabeled dataset. The discerned insights are then harnessed to augment and refine the current dataset, thereby facilitating a continuous and iterative improvement process.",
+        "bbox": [
+            507,
+            244,
+            884,
+            422
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Self-Annotating",
+        "text_level": 1,
+        "bbox": [
+            507,
+            432,
+            680,
+            448
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We gathered 200,000 unannotated SMILES of molecules from PubChem (Kim et al., 2023) and conducted dedduplication to mitigate the risk of data leakage in the validation dataset. For each unlabeled SMILES, we executed inference on the model itself to produce a candidate caption, thereby obtaining the pseudo-labeled dataset.",
+        "bbox": [
+            507,
+            453,
+            882,
+            565
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ChatGPT Annotating vs Specialized Model Annotating Despite the remarkable performance of Large Language Models (LLMs) such as ChatGPT in various cross-modal tasks, they have not succeeded in surpassing specialized models in the molecule captioning task (Li et al., 2023a; Christofidellis et al., 2023). This suggests that specialized models serve as better annotators for unlabeled SMILES strings. Furthermore, as depicted in Figure 3(a), by leveraging the Text2Mol score (Edwards et al., 2022a) as a fundamental metric of data quality, which measures the similarity between SMILES strings and their captions, we evaluated the distribution of the real training set, MolReGPT annotated data, and self-annotated data. It is evident that the self-annotated data yields higher-quality data, closely resembling real data more than the MolReGPT annotated data.",
+        "bbox": [
+            507,
+            574,
+            884,
+            864
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Iterative Annotating We employ an iterative training approach to enhance the proficiency of annotators and produce captions of superior quality.",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "9553",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/9df76a23d759ae9f4034356e1c20d7741b9c68c8cd33417e41be9b9e838b4682.jpg",
+        "image_caption": [
+            "Figure 2: Overall framework. We interweave model fine-tuning and data augmentation iteratively. A core component of our framework, the Data Augmentation Module, is compartmentalised into two key elements: Self-Annotating and Self-Sampling. These elements underline the crux of our methodology, wherein the model autonomously performs the augmentation, subsequently enhancing its own performance."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            78,
+            836,
+            561
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As depicted in Figure 3, during each iteration, the quality of the molecule-caption pairs is enhanced through the implementation of the iterative strategy.",
+        "bbox": [
+            112,
+            651,
+            489,
+            700
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Self-Sampling",
+        "text_level": 1,
+        "bbox": [
+            114,
+            734,
+            272,
+            750
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A discerning subset sampling strategy is anticipated to effectively filter pseudo labels, thereby ensuring the attainment of high-quality labels for training the model. This refined approach aims to enhance the overall reliability of the model by mitigating the impact of unreliable or inaccurate labels in the training process. In contrast to prior work (Li et al., 2023b), wherein fine-tuning an external score model was employed to select high-quality pairs, we have found this approach to be",
+        "bbox": [
+            112,
+            760,
+            490,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "both time-consuming and unreliable when applied to the domain of cross text-molecule learning. Notably, the susceptibility of the model to overfitting incorrect pseudo-labels is a concern, as detailed in section 4.4. To tackle this dilemma, we introduce a dual selection mechanism that not only proves to be more computationally efficient but also enhances reliability in ensuring the high quality of our augmented dataset. This mechanism involves two key steps. Firstly, leveraging the generation likelihood as a confidence score to filter out unreliable molecule-caption pairs. Secondly, updating and ranking the remaining molecule-caption pairs based on the Text2Mol score (Edwards et al., 2022a) This refined approach streamlines the sampling process, offering increased efficiency and",
+        "bbox": [
+            507,
+            651,
+            884,
+            909
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "9554",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/20951661f4e22357b7b4d029b32a0c395e446f3d4244e0c8c5e64433b59a0d3d.jpg",
+        "image_caption": [
+            "(a) Comparison of Data Quality Between LLMs Annotated Label and Self-Annotated Label"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            144,
+            84,
+            463,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f0f2157fe37c877e088e862611123d5520baa1f758db2f66630cbc7098ebf02.jpg",
+        "image_caption": [
+            "(b) Comparison of Data Quality in Relation to the Adoption or Non-Apoption of Iterative Strategy",
+            "Figure 3: In a manner akin to Chen et al. (2023) in their study, we employ the methodology proposed by Edwards et al. (2022) to evaluate the correlation between molecule-description pairs, utilizing it as a measure of data quality. The distributions are illustrated via Kernel Distribution Estimation. A noteworthy aspect to consider here is that a higher Text2Mol score typically implies a closer resemblance between the molecule and its description. In addition, the term \"Density\" in (a) pertains to the concentration of data within a specific region."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            171,
+            299,
+            435,
+            404
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "heightened confidence in the identification of high-quality pairs for dataset augmentation.",
+        "bbox": [
+            112,
+            646,
+            487,
+            680
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Data Filtering First, we employ the model itself to filter out a set of low-quality pairs. Specifically, we utilize the generation likelihood as a measure of the quality of the generated sequence. However, previous studies (Liu and Liu, 2021; Liu et al., 2022) have underscored that the correlation between sequence probability and its quality for Maximum Likelihood Estimation (MLE) trained models can be low, due to the presence of deterministic (one-point) target distribution issues. Furthermore, sequence likelihood estimation becomes noisier when the decoded sequences of models deviate from the exposed training data distribution, thereby exacerbating the problem of exposure bias",
+        "bbox": [
+            112,
+            696,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "(Ranzato et al., 2016). To address these intricate challenges, we introduce the Sequence Likelihood Calibration (SLiC) stage. This stage, depicted in Figure 1, serves to mitigate the aforementioned issues and diminish the gap between sequence likelihood and its associated quality. Following the methodology of Zhao et al. (2022), we conduct further fine-tuning of the model, introducing a new objective that encompasses the following loss functions:",
+        "bbox": [
+            507,
+            84,
+            884,
+            243
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathrm {L} _ {\\text {r a n k}} ^ {\\text {c a l}} = \\max  \\left(0, \\beta - \\log P _ {\\theta} \\left(\\hat {\\mathbf {y}} _ {+} \\mid \\mathbf {x}\\right) + \\log P _ {\\theta} \\left(\\hat {\\mathbf {y}} _ {-} \\mid \\mathbf {x}\\right)\\right), \\\\ \\mathbf {L} _ {\\mathrm {c e}} ^ {\\mathrm {r e g}} = \\sum_ {t} - \\log P _ {\\theta} \\left(\\bar {y} _ {t} \\mid \\overline {{\\mathbf {y}}} _ {t - 1}, \\mathbf {x}\\right), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            510,
+            254,
+            880,
+            318
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {L} _ {\\mathrm {k l}} ^ {\\mathrm {r e g}} = \\sum_ {t} P _ {\\theta} \\left(\\bar {y} _ {t} \\mid \\overline {{\\mathbf {y}}} _ {t - 1}, \\mathbf {x}\\right) \\log \\frac {P _ {\\theta} \\left(\\bar {y} _ {t} \\mid \\overline {{\\mathbf {y}}} _ {t - 1} , \\mathbf {x}\\right)}{P _ {\\theta_ {f t}} \\left(\\bar {y} _ {t} \\mid \\overline {{\\mathbf {y}}} _ {t - 1} , \\mathbf {x}\\right)}.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            512,
+            326,
+            880,
+            357
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Given the context  $\\mathbf{x}$ , target  $\\overline{\\mathbf{y}}$ , and positive and negative candidates pairs  $\\hat{\\mathbf{y}}_{+}, \\hat{\\mathbf{y}}_{-}, P_{\\theta}(\\mathbf{y} \\mid \\mathbf{x})$  denotes the generation sequence likelihood. Rank loss  $L_{rank}^{cal}$  optimizes the ranking order of positive and negative candidate pairs. Cross entropy loss  $L_{ce}^{reg}$  is the standard fine-tuning MLE objective. KL divergence loss  $L_{kl}^{reg}$  directly minimizes the probability distribution distance between the calibrated model and the fine-tuned model at each token on the observed target sequence.",
+        "bbox": [
+            507,
+            362,
+            882,
+            521
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Following the SLiC stage, we regard the generation likelihood as one of our metrics for assessing the quality of pseudo labels. In our experiments, we discard pairs with log-likelihood equal to negative infinity and retain the remaining pairs as high-quality pairs.",
+        "bbox": [
+            507,
+            523,
+            882,
+            619
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Update and Rank Secondly, for those high-quality pairs, we further refine the data based on the similarity score  $\\mathcal{F}_{\\mathrm{text2mol}}$  proposed by Edwards et al. (2022a). At iteration t, we update the caption if  $\\mathcal{F}_{\\mathrm{text2mol}}(m,c_t) > \\mathcal{F}_{\\mathrm{text2mol}}(m,c_{t - 1})$  then we rank the updated dataset according to  $\\mathcal{F}_{\\mathrm{text2mol}}$  ,selecting the top  $t\\times 2k$  pairs to form the new Augmented Dataset  $A_{t}$",
+        "bbox": [
+            507,
+            627,
+            882,
+            756
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            768,
+            655,
+            784
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In order to validate the efficacy of utilizing our framework, we have undertaken a series of comprehensive experiments.",
+        "bbox": [
+            507,
+            793,
+            882,
+            841
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            852,
+            714,
+            868
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Dataset We employ the ChEBI-20 (Edwards et al., 2021) to serve as our original dataset. Additionally, we gather 200,000 unannotated molecules",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "9555",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/cd494917da8380225e3eef6d916ad474e8ee0d2063a2311b8735a90179b85968.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Info</td><td>SAPubChem-41</td><td>ChEBI-20</td></tr><tr><td>Train</td><td>34,407</td><td>26,407</td></tr><tr><td>Validation</td><td>3,301</td><td>3,301</td></tr><tr><td>Test</td><td>3,300</td><td>3,300</td></tr><tr><td>\\(\\mathcal{L}_{\\text{SMILES}}\\)</td><td>69.96</td><td>81.56</td></tr><tr><td>\\(\\mathcal{L}_{\\text{Description}}\\)</td><td>41.59</td><td>52.88</td></tr></table>",
+        "bbox": [
+            132,
+            80,
+            470,
+            186
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Comparison of our SAPubChem-41 dataset with the original dataset ChEBI-20.  $L_{\\text{SMILES}}$  denotes the average length of SMILES while  $L_{\\text{Description}}$  denotes the average word count per description.",
+        "bbox": [
+            112,
+            197,
+            487,
+            255
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/ba287cb3b5eef8c8e418c1e1368705f12332150847eb5d628a1bc8acb032545b.jpg",
+        "image_caption": [
+            "Figure 4: The results of the iteration procedure. The dashed lines denote the metrics associated with the molecule captioning task, while the solid lines signify those related to the molecule generation task."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            274,
+            478,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "from the PubChem database (Kim et al., 2023). Subsequently, we conduct a rigorous filtering procedure to exclude any molecules present in downstream datasets, ensuring the elimination of potential overlaps between the newly acquired molecules and those already present in the ChEBI-20 dataset. For our experimental dataset augmentation, we set the iteration step to 2000. At each iteration stage, the dataset size expands by 2000 from the preceding iteration. A detailed comparison between our final augmented dataset, SAPubChem-41, and the original dataset, ChEBI-20, is presented in Table 1.",
+        "bbox": [
+            112,
+            607,
+            489,
+            800
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Base Model To prove the effectiveness of our framework, we use the Text+Chem T5 (Christofidellis et al., 2023) as the base model for simulating the iterative process.",
+        "bbox": [
+            112,
+            812,
+            489,
+            878
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Metrics Following the previous studies (Edwards et al., 2022a; Li et al., 2023a; Christofidellis et al.,",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2023; Chen et al., 2023), we evaluate the results with following metrics:",
+        "bbox": [
+            507,
+            84,
+            882,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "- Molecular Captioning: BLEU-2 and BLEU-4(Papineni et al., 2002) are metrics used to evaluate the quality of machine-generated text by comparing it to a reference text, with a higher score on the BLEU metric indicating a higher level of similarity between the generated text and reference text. ROUGE-1, ROUGE-2, and ROUGE-L(Lin, 2004) is similar to BLEU, while computing the recall-overlap of unigrams, bigrams, and longest common subsequences between the generated and reference texts. METEOR (Banerjee and Lavie, 2005) is a metric that uses a combination of unigram precision, recall, and a synonym-matching component to evaluate the generated text against the reference text, which is designed to be more sensitive to fluency, meaning, and structure than BLEU.",
+        "bbox": [
+            531,
+            133,
+            884,
+            423
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "- Text-based Molecule Generation: BLEU and the Exact Match scores are calculated as basic assessments. Molecule-specified metrics including Levenshtein distance(Levenshtein et al., 1966), validity(Edwards et al., 2022a), and three molecule fingerprints scores - MACCS FTS(Durant et al., 2002), RDK FTS(Schneider et al., 2015), and Morgan FTS(Rogers and Hahn, 2010) are calculated to provide valuable insights into the quality, validity, and structural characteristics of the generated molecules.",
+        "bbox": [
+            531,
+            439,
+            882,
+            633
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation We conducted our experiments on Ubuntu 22.04 using RTX 4090(24GB) * 4 with CUDA 11.7.1. Our primary dependencies are Python 3.11.4, PyTorch 2.0.1, Transformers 4.31.0, and Numpy 1.24.3. We initialize the three pretrained models using public checkpoints. For fine-tuning, we adopted the configuration in MolT5 (Edwards et al., 2022b), with a learning rate of 1e-3, 50,000 fine-tuning steps, weight decay of 0.1, batch size of 32, random seed of 42, and 1000 warm-up steps. For evaluation, we use a greedy search with a maximum generation length of 512 during generation. The evaluation metrics script is derived from MolT5 (Edwards et al., 2022b). For other hyperparameters, we relied on the default settings of the T5ForConditionalGeneration class in Huggingface.",
+        "bbox": [
+            507,
+            664,
+            882,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "9556",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/e4e9e1913a247ef5a4814a63578e323eadfa6a0ab72aefc28fc5d9014da57b32.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"3\">Text+Chem T5</td><td colspan=\"3\">Ada-T5</td><td colspan=\"3\">MolT5</td></tr><tr><td>BL</td><td>RG</td><td>MET</td><td>BL</td><td>RG</td><td>MET</td><td>BL</td><td>RG</td><td>MET</td></tr><tr><td>ChEBI-20</td><td>0.490</td><td>0.498</td><td>0.604</td><td>0.316</td><td>0.369</td><td>0.464</td><td>0.457</td><td>0.485</td><td>0.569</td></tr><tr><td>SAPubChem-41</td><td>0.496</td><td>0.504</td><td>0.614</td><td>0.350</td><td>0.410</td><td>0.493</td><td>0.464</td><td>0.484</td><td>0.572</td></tr></table>",
+        "bbox": [
+            144,
+            82,
+            853,
+            154
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/7d6ba2bb82794020dd56c1d2934a62d391985364655faa4aa64efd258182dfa9.jpg",
+        "table_caption": [
+            "Table 2: Results of different models for molecular captioning on ChEBI-20, SAPubChem-41 datasets. The best scores are in bold. BL:BLEU-4, RG:ROUGE-2, MET:METEOR"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"3\">Text+Chem T5</td><td colspan=\"3\">Ada-T5</td><td colspan=\"3\">MolT5</td></tr><tr><td>Morgan</td><td>Acc</td><td>Val</td><td>Morgan</td><td>Acc</td><td>Val</td><td>Morgan</td><td>Acc</td><td>Val</td></tr><tr><td>ChEBI-20</td><td>0.697</td><td>0.212</td><td>0.792</td><td>0.672</td><td>0.182</td><td>0.886</td><td>0.529</td><td>0.081</td><td>0.772</td></tr><tr><td>SAPubChem-41</td><td>0.726</td><td>0.243</td><td>0.937</td><td>0.699</td><td>0.213</td><td>0.854</td><td>0.582</td><td>0.093</td><td>0.788</td></tr></table>",
+        "bbox": [
+            124,
+            208,
+            873,
+            281
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Results of different models for molecular generation on ChEBI-20, SAPubChem-41 datasets. The best scores are in bold. Morgan:Morgan FTS, Acc:Accuracy, Val:Validity",
+        "bbox": [
+            112,
+            292,
+            882,
+            321
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2 Performance Comparison of Molecule-Caption Translation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            347,
+            394,
+            378
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Molecular Captioning Table 2 presents the results of the molecule captioning task, utilizing the original ChEBI-20 dataset and the augmented SAPubChem-41 dataset to optimize three distinct pre-trained models. Across all measurement metrics, it is observable that the performance of all models using our augmented SAPubChem-41 dataset is generally superior to that utilizing the original ChEBI-20 dataset.",
+        "bbox": [
+            112,
+            382,
+            487,
+            527
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Text-based Molecule Generation Table 3 presents the results of the text-based molecule generation task, utilizing the original ChEBI-20 dataset and the augmented SAPubChem-41 dataset to optimize three distinct pre-trained models. Notably, the performance of all models that utilize our augmented SAPubChem-41 dataset generally surpasses those that use the original ChEBI-20 dataset. The accuracy of the generated molecule, as well as its similarity to the ground truth molecule, have been significantly improved following the self-augmentation of the dataset. This evidences substantial enhancements in performance when integrated with our self-rewarding framework.",
+        "bbox": [
+            112,
+            536,
+            489,
+            762
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Iterative Procedure",
+        "text_level": 1,
+        "bbox": [
+            112,
+            771,
+            314,
+            785
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We utilize our framework on the Text+Chem T5, taking the ChEBI-20 as the original dataset. This approach facilitates an augmented dataset while simultaneously enhancing the model's performance. It is critical to note that our intention in these experiments is not necessarily to achieve a state-of-the-art model. Rather, our prime objective is to examine the efficacy of our framework, which is designed",
+        "bbox": [
+            112,
+            791,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "to optimize both the model and the pseudo-labeled dataset concurrently. Figure 4 illustrates the iterative process in which dashed lines represent the molecule captioning task and solid lines signify the text-based molecule generation task. We report the complete results in Table 4 and 5. There are a few observations: First, it is noticeable that for both tasks the outcome of iteration 5 mostly falls short of that of iteration 4 in all metrics. This suggests that the training process reaches convergence on iteration 4 for both tasks. Consequently, we adopt the augmented dataset obtained at the fourth iteration as our conclusive dataset. Second, for the text-based molecule generation task, there is a consistent upward trend across all pertinent metrics. This suggests that an increased quantity of high-quality pseudo-labeled data contributes to the optimization of the model in the molecule generation process. This trend underscores the necessity of sufficient high-quality data in cross text-molecule learning. Lastly, when it comes to the molecule captioning task, the general trend in relevant metrics indicates a rise. Initially, there is a minor decrease in the BLEU and Meteor metrics, which is counterweighed by an enhancement in Rouge. As the volume of high-quality pseudo-labeled data increases, all metrics demonstrate improvement. This trend suggests that the quantity of high-quality pseudo-labeled data is also a significant factor.",
+        "bbox": [
+            507,
+            347,
+            885,
+            814
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            829,
+            673,
+            846
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We perform further ablation studies to validate the effectiveness of the key component in our framework, i.e., the self-sampling strategy.",
+        "bbox": [
+            507,
+            854,
+            882,
+            903
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To demonstrate the significance of pseudo",
+        "bbox": [
+            526,
+            904,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "9557",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/c31e6c1c518f2a5ca017add3cb7710b4f8e5fb3acc56d748150b8e0eab7c16d8.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Iteration</td><td>BLEU↑</td><td>Accuracy↑</td><td>Levenshtein↓</td><td>MACCS FTS↑</td><td>RDK FTS↑</td><td>Morgan FTS↑</td><td>Validity↑</td></tr><tr><td>t0</td><td>0.750</td><td>0.212</td><td>27.39</td><td>0.874</td><td>0.767</td><td>0.697</td><td>0.792</td></tr><tr><td>t1</td><td>0.821</td><td>0.236</td><td>19.31</td><td>0.885</td><td>0.788</td><td>0.721</td><td>0.931</td></tr><tr><td>t2</td><td>0.832</td><td>0.237</td><td>18.69</td><td>0.885</td><td>0.788</td><td>0.722</td><td>0.938</td></tr><tr><td>t3</td><td>0.828</td><td>0.240</td><td>19.28</td><td>0.886</td><td>0.788</td><td>0.724</td><td>0.934</td></tr><tr><td>t4</td><td>0.827</td><td>0.243</td><td>18.59</td><td>0.887</td><td>0.792</td><td>0.726</td><td>0.937</td></tr><tr><td>t5</td><td>0.830</td><td>0.238</td><td>19.03</td><td>0.883</td><td>0.786</td><td>0.721</td><td>0.937</td></tr></table>",
+        "bbox": [
+            119,
+            82,
+            877,
+            193
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/f18c970efef29d2962e1b90a489b568c2679ca32ec0eb6884a76ae1417f94c72.jpg",
+        "table_caption": [
+            "Table 4: The iterative results of molecule generation task. The best scores are in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Iteration</td><td>BLEU-2↑</td><td>BLEU-4↑</td><td>Rouge-1↑</td><td>Rouge-2↑</td><td>Rouge-L↑</td><td>Meteor↑</td></tr><tr><td>t0</td><td>0.580</td><td>0.490</td><td>0.647</td><td>0.498</td><td>0.586</td><td>0.604</td></tr><tr><td>t1</td><td>0.574</td><td>0.486</td><td>0.651</td><td>0.502</td><td>0.590</td><td>0.601</td></tr><tr><td>t2</td><td>0.580</td><td>0.486</td><td>0.653</td><td>0.503</td><td>0.592</td><td>0.602</td></tr><tr><td>t3</td><td>0.583</td><td>0.490</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.611</td></tr><tr><td>t4</td><td>0.584</td><td>0.496</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.614</td></tr><tr><td>t5</td><td>0.583</td><td>0.491</td><td>0.652</td><td>0.503</td><td>0.591</td><td>0.610</td></tr></table>",
+        "bbox": [
+            181,
+            231,
+            816,
+            353
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: The iterative results of molecule captioning task. The best scores are in bold.",
+        "bbox": [
+            206,
+            363,
+            788,
+            379
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/67f66300866b8b8fef22300c68bc1b819083dfc751a3ff3e64653bf338270791.jpg",
+        "image_caption": [
+            "(a) BLEU"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            149,
+            397,
+            381,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b9ff6c3114af2035e8b2e46602c38bc7beb38655ccb600ce166b2da9cb31f0ed.jpg",
+        "image_caption": [
+            "(b) Accuracy"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            397,
+            613,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/47a191f09c77f920faa6c0442d2a157cef34f22990f8853772b4a7ac9072aaa4.jpg",
+        "image_caption": [
+            "(c) Morgan FTS"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            616,
+            397,
+            847,
+            510
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/90c9e90c14159a4efccd232f8acb6ce87b2be93964e115550168e9f124f829e5.jpg",
+        "image_caption": [
+            "Figure 5: Results of molecule generation task using different sampling strategy. The blue line marks the result of model optimization on the original dataset.",
+            "(a) BLEU-4",
+            "Figure 6: Results of molecule captioning task using different sampling strategy. The blue line marks the result of model optimization on the original dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            149,
+            592,
+            379,
+            703
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f738188f7c04194e649bd0ce00df7eee7a8a5b0171900874a3169b2cf90d4c36.jpg",
+        "image_caption": [
+            "(b) Rouge-2"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            592,
+            613,
+            703
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/996ceabbab978d4d1995cb968c26ebbe415cc74a098b4a17fdcf433dbd263b82.jpg",
+        "image_caption": [
+            "(c) Meteor"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            616,
+            592,
+            847,
+            703
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "labeled data selection, we conducted a comparative analysis of the performance of all intermediate models. We compared the use of randomly-selected data of the same size with our refined data. As depicted in Figure 5 and 6, our refined data consistently outperformed the randomly-selected data across all metrics at each iteration for both tasks.",
+        "bbox": [
+            112,
+            793,
+            489,
+            904
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Notably, when optimizing the model on randomly-selected data, the results were generally inferior to the original model on most metrics. This indicates that the model tends to overfit to noisy pseudolabels. These findings emphasize the critical role of the self-sampling strategy in our framework.",
+        "bbox": [
+            507,
+            793,
+            884,
+            889
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "9558",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            83,
+            247,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we have explored and grappled with the challenge of a deficiency in the accessibility of high-quality datasets for model fine-tuning in the field of cross text-molecule learning, by developing a novel framework. This innovative framework efficiently combines model fine-tuning with data augmentation, integrating the model in such a way that it expands its roles to include annotating unlabeled data and sampling a subset of high-quality data. The paper further introduces a high-quality dataset, the SAPubChem-41, expressly designed for fine-tuning purposes. The experiments' results were impressive as the model optimized using SAPubChem-41 consistently outperformed the model optimized using the original dataset, ChEBI-20. Going forward, this paper's findings present crucial steps toward addressing high-quality data scarcity for fine-tuning and ensuring further advancement in the exciting field of cross text-molecule learning.",
+        "bbox": [
+            112,
+            108,
+            492,
+            431
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Limitations and Future Works",
+        "text_level": 1,
+        "bbox": [
+            112,
+            442,
+            413,
+            457
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "One limitation of our framework is that it only considers a 1D representation of molecules. The integration of additional representations, such as molecule graphs, is left as future work. Additionally, our data augmentation module currently only supports sample-level labels. It is expected that more quality measurements from different levels will be incorporated to provide a comprehensive perspective on quantifying the dataset quality.",
+        "bbox": [
+            112,
+            467,
+            489,
+            612
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            623,
+            285,
+            639
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by National Natural Science Foundation of China (62302433, U23A20496), Zhejiang Provincial \"Jianbing\" \"Lingyan\" Research and Development Program of China (2024C01135), Zhejiang Provincial Natural Science Foundation of China (LQ24F020007), New Generation AI Development Plan for 2030 of China (2023ZD0120802) and CCF-Tencent Rhino-Bird Fund (RAGR20230122).",
+        "bbox": [
+            112,
+            646,
+            489,
+            791
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            819,
+            213,
+            834
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "ref_text",
+        "text": "Ateret Anaby-Tavor, Boaz Carmeli, Esther Goldbraich, Amir Kantor, George Kour, Segev Shlomov, Naama Tepper, and Naama Zwerdling. 2020. Do not have enough data? deep learning to the rescue! In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7383-7390.",
+        "bbox": [
+            115,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72.",
+            "Yuhan Chen, Nuwa Xi, Yanrui Du, Haochun Wang, Chen Jianyu, Sendong Zhao, and Bing Qin. 2023. From artificially real to real: Leveraging pseudo data from large language models for low-resource molecule discovery.",
+            "Dimitrios Christofidellis, Giorgio Giannone, Jannis Born, Ole Winther, Teodoro Laino, and Matteo Manica. 2023. Unifying molecular and textual representations via multi-task language modelling.",
+            "Haixing Dai, Zhengliang Liu, Wenxiong Liao, Xiaoke Huang, Yihan Cao, Zihao Wu, Lin Zhao, Shaochen Xu, Wei Liu, Ninghao Liu, et al. 2023. Auggpt: Leveraging chatgpt for text data augmentation. arXiv preprint arXiv:2302.13007.",
+            "Kirill Degtyarenko, Paula De Matos, Marcus Ennis, Janna Hastings, Martin Zbinden, Alan McNaught, Rafael Alcantara, Michael Darsow, Mickaël Guedj, and Michael Ashburner. 2007. Chebi: a database and ontology for chemical entities of biological interest. Nucleic acids research, 36(suppl_1):D344-D350.",
+            "Joseph L Durant, Burton A Leland, Douglas R Henry, and James G Nourse. 2002. Reoptimization of mdI keys for use in drug discovery. Journal of chemical information and computer sciences, 42(6):1273-1280.",
+            "Carl Edwards, Tuan Lai, Kevin Ros, Garrett Honke, Kyunghyun Cho, and Heng Ji. 2022a. Translation between molecules and natural language.",
+            "Carl Edwards, Tuan Manh Lai, Kevin Ros, Garrett Honke, Kyunghyun Cho, and Heng Ji. 2022b. Translation between molecules and natural language. In EMNLP, pages 375-413. Association for Computational Linguistics.",
+            "Carl Edwards, ChengXiang Zhai, and Heng Ji. 2021. Text2mol: Cross-modal molecule retrieval with natural language queries. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 595-607.",
+            "Steven Y Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy. 2021. A survey of data augmentation approaches for nlp. arXiv preprint arXiv:2105.03075.",
+            "Alex Hernández-García and Peter König. 2018. Data augmentation instead of explicit regularization. arXiv preprint arXiv:1806.03852.",
+            "Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li, Benjamin A Shoemaker, Paul A Thiessen, Bo Yu, et al. 2023. Pubchem 2023 update. Nucleic acids research, 51(D1):D1373-D1380."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "9559",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, volume 10, pages 707-710. Soviet Union.",
+            "Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. *Ai Open*, 3:71-90.",
+            "Jiatong Li, Yunqing Liu, Wenqi Fan, Xiao-Yong Wei, Hui Liu, Jiliang Tang, and Qing Li. 2023a. Empowering molecule discovery for molecule-caption translation with large language models: A chatgpt perspective.",
+            "Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Luke Zettlemoyer, Omer Levy, Jason Weston, and Mike Lewis. 2023b. Self-alignment with instruction back-translation. arXiv preprint arXiv:2308.06259.",
+            "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81.",
+            "Yixin Liu and Pengfei Liu. 2021. SimCLS: A simple framework for contrastive learning of abstractive summarization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 1065-1072, Online. Association for Computational Linguistics.",
+            "Yixin Liu, Pengfei Liu, Dragomir Radev, and Graham Neubig. 2022. BRIO: Bringing order to abstractive summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2890-2903, Dublin, Ireland. Association for Computational Linguistics.",
+            "Zequn Liu, Wei Zhang, Yingce Xia, Lijun Wu, Shufang Xie, Tao Qin, Ming Zhang, and Tie-Yan Liu. 2023. Molxpt: Wrapping molecules with text for generative pre-training. arXiv preprint arXiv:2305.10688.",
+            "Zhiyuan Liu, Sihang Li, Yanchen Luo, Hao Fei, Yixin Cao, Kenji Kawaguchi, Xiang Wang, and Tat-Seng Chua. 2024. Molca: Molecular graph-language modeling with cross-modal projector and uni-modal adapter.",
+            "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.",
+            "Qizhi Pei, Wei Zhang, Jinhua Zhu, Kehan Wu, Kaiyuan Gao, Lijun Wu, Yingce Xia, and Rui Yan. 2024. Biot5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "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.",
+            "Marc'Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. 2016. Sequence level training with recurrent neural networks.",
+            "David Rogers and Mathew Hahn. 2010. Extendedconnectivity fingerprints. Journal of chemical information and modeling, 50(5):742-754.",
+            "Nadine Schneider, Roger A Sayle, and Gregory A Landrum. 2015. Get your atoms in order an opensource implementation of a novel and robust molecular canonicalization algorithm. Journal of chemical information and modeling, 55(10):2111-2120.",
+            "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826.",
+            "David S Wishart, Yannick D Feunang, An C Guo, Elvis J Lo, Ana Marcu, Jason R Grant, Tanvir Sajed, Daniel Johnson, Carin Li, Zinat Sayeeda, et al. 2018. Drugbank 5.0: a major update to the drugbank database for 2018. Nucleic acids research, 46(D1):D1074-D1082.",
+            "Xing Wu, Shangwen Lv, Liangjun Zang, Jizhong Han, and Songlin Hu. 2019. Conditional bert contextual augmentation. In Computational Science-ICCS 2019: 19th International Conference, Faro, Portugal, June 12–14, 2019, Proceedings, Part IV 19, pages 84–95. Springer.",
+            "Kang Min Yoo, Dongju Park, Jaewook Kang, Sang-Woo Lee, and Woomyeong Park. 2021. Gpt3mix: Leveraging large-scale language models for text augmentation. arXiv preprint arXiv:2104.08826.",
+            "Zheni Zeng, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2022. A deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals. Nature communications, 13(1):862.",
+            "Qiang Zhang, Keyang Ding, Tianwen Lyv, Xinda Wang, Qingyu Yin, Yiwen Zhang, Jing Yu, Yuhao Wang, Xiaotong Li, Zhuoyi Xiang, et al. 2024. Scientific large language models: A survey on biological & chemical domains. arXiv preprint arXiv:2401.14656.",
+            "Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28.",
+            "Yao Zhao, Misha Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, and Peter J. Liu. 2022. Calibrating sequence likelihood improves conditional language generation."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "9560",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Jing Zhou, Yanan Zheng, Jie Tang, Jian Li, and Zhilin Yang. 2021. Flipda: Effective and robust data augmentation for few-shot learning. arXiv preprint arXiv:2108.06332.",
+        "bbox": [
+            114,
+            85,
+            489,
+            137
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Datasets Information",
+        "text_level": 1,
+        "bbox": [
+            114,
+            151,
+            337,
+            166
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "SAPubChem-41 The SAPubChem-41 datasets comprises 33k real data sourced from CHEBI-20 and 8k high-quality pseudo labelled data sourced from self-augmentation on Text+ChemT5. Representative examples of pseudo-labelled part are provided in Table 7. In addition, to further elucidate the advantages of iterative strategy on model performance, we have documented representative examples of updated captions in Table 6.",
+        "bbox": [
+            112,
+            175,
+            489,
+            321
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Comparison with Existing Datasets In Table 8 we present a detailed comparison of our SAPubChem-41 datasets with existing limited datasets with parallel molecule-description pairs, including:",
+        "bbox": [
+            112,
+            330,
+            489,
+            411
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- ChEBI-20 (Edwards et al., 2021): consists of  $33\\mathrm{k}$  molecule-caption pairs with caption exploiting ChEBI(Degtyarenko et al., 2007)annotations and more than 20 words.",
+            "- PCdes (Zeng et al., 2022): consists of  $15\\mathrm{k}$  substances in PubChem which have names, SMILES and corresponding paragraphs of property descriptions.",
+            "- DrugBank-23(Chen et al., 2023): consists of 23k compounds from DrugBank (Wishart et al., 2018) with corresponding description's length longer than 20 words."
+        ],
+        "bbox": [
+            136,
+            420,
+            487,
+            633
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Pre-trained Model Information",
+        "text_level": 1,
+        "bbox": [
+            114,
+            646,
+            421,
+            661
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 9 presents the information of three different pre-trained models utilized in our experiments.",
+        "bbox": [
+            112,
+            671,
+            487,
+            703
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Experiments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            715,
+            263,
+            730
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C.1 Performance Comparison",
+        "text_level": 1,
+        "bbox": [
+            114,
+            741,
+            366,
+            756
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Here we present the complete results of performance comparison for both molecular captioning task and molecule generation task in Table 10&Table 11.",
+        "bbox": [
+            112,
+            762,
+            489,
+            824
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C.2 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            114,
+            835,
+            284,
+            852
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Here we present the complete results of both molecular captioning task and molecule generation task using different sampling strategy in Table 12 & Table 13.",
+        "bbox": [
+            112,
+            857,
+            489,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C.3 Additional Experiments",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            749,
+            99
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "To further substantiate the efficacy of our framework and the exceptional quality of the SAPubChem-41 dataset, we have conducted a series of experiments. These were performed with the goal of optimizing three distinct pre-trained models, utilizing both our augmented dataset, SAPubChem-41, and other prevalent public datasets which include ChEBI-20 and PCdes. The findings from these experiments are reported in Table 14 & Table 15. It can be noted that across both task categories, models that were optimized using our self-augmented dataset, SAPubChem-41, generally exhibited superior performance in comparison to those optimized with other existing datasets.",
+        "bbox": [
+            507,
+            105,
+            884,
+            330
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "9561",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/6fa13fadc8bbdf822a70f82bbce98ed5ceb5c22f0cf206b3b4fda714bf672ccd.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>SMILES</td><td>[C@H]1(O[C@H][(C@@H](O)C)C(O)O)[C@@H](O)[C@@H](O)[C@@H](O)[C@@H](C)O1</td></tr><tr><td>previous caption</td><td>The molecule is a glycosylglucose consisting of D-galactopyranose and D-glucopyranose residues joined in sequence by a (1-&gt;4) glycosidic bond. It derives from a D-galactopyranose and a D-glucopyranose.(0.6611)</td></tr><tr><td>updated caption</td><td>The molecule is a deoxygalactose that is D-galactopyranose in which the hydroxy group at position 3 has been replaced by a methyl group. It is a deoxygalactose and a methyl glycoside. It derives from a D-galactopyranose.(0.6817)</td></tr><tr><td>SMILES</td><td>CC([C@@H](C)O)(C)OC-</td></tr><tr><td>previous caption</td><td>The molecule is an ether in which the stereocentres at positions 2 and 3 both have S-configuration. It is an ether and a secondary alcohol.(0.6610)</td></tr><tr><td>updated caption</td><td>The molecule is an ether in which the stereocentres at positions 2 and 3 both have S-configuration. It is an ether and a secondary alcohol. It derives from a hydride of an oxepane.(0.6704)</td></tr><tr><td>SMILES</td><td>CI</td></tr><tr><td>previous caption</td><td>The molecule is an iodoalkane that is ethane in which one of the hydrogens is substituted by iodine. It has a role as a human metabolite. It is an iodoalkane and a member of iodos. It derives from a hydride of an ethane.(0.7479)</td></tr><tr><td>updated caption</td><td>The molecule is an iodoalkane that is ethane in which one of the hydrogens is substituted by iodine. It has a role as a metabolite. It is an iodoalkane and a volatile organic compound. It derives from a hydride of an ethane.(0.7671)</td></tr><tr><td>SMILES</td><td>N([N+])=C</td></tr><tr><td>previous caption</td><td>The molecule is an organic cation resulting from the protonation of the nitrogen of nitric acid. It is a conjugate acid of a nitric acid.(0.7479)</td></tr><tr><td>updated caption</td><td>The molecule is a hydracid and a one-carbon compound. It is a conjugate acid of a hydrazine. It derives from a hydride of a hydrazine.(0.7599)</td></tr><tr><td>SMILES</td><td>[C@H]1(C)[C@@](CO)(O)[C@@H-(O[C@@H](C)O)[C@H](OC(C)C)O1</td></tr><tr><td>previous caption</td><td>The molecule is a spiro-epoxide resulting from the formal epoxidation of the hydroxy group at position 2 of D-fructofuranose. It is a spiro-epoxide and a glycoside. It derives from a D-fructofuranose.(0.6014)</td></tr><tr><td>updated caption</td><td>The molecule is a deoxygalactose that is alpha-D-galactopyranose in which the hydroxy group at position 2 has been replaced by a methyl group. It is a deoxygalactose and a tertiary alcohol.(0.6839)</td></tr></table>",
+        "bbox": [
+            225,
+            87,
+            771,
+            869
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Five examples of the comparison of previous caption when adopting our iterative strategy. Corresponding Text2Mol Score of each molecule-caption pairs is present in the bracket.",
+        "bbox": [
+            112,
+            881,
+            880,
+            910
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "9562",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/7a05e113d71d3c244be66369714b7f5ddc630425738f93781a73ae8d71c5f58f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>SMILES</td><td>Caption</td></tr><tr><td>C@H](C[C@H](C(/C=C/C)=O)C)\n([C@@H]([C@H](/C=C/C(=O)O)C)O[C@H]1[C@H]\n(O)[C@@H]([NH+])C[C@@H](C)O1)C</td><td>The molecule is an organic cation that is the conjugate acid of 1D-myo-inositol, obtained by protonation of the tertiary amino group; major species at pH 7.3. It is an ammonium ion derivative and an organic cation. It is a conjugate acid of a 1D-myo-inositol.</td></tr><tr><td>[O-][Mn]</td><td>The molecule is a monovalent inorganic anion obtained by deprotonation of manganese. It is a man-ganese oxoanion and a monovalent inorganic anion. It is a conjugate base of a manganese.</td></tr><tr><td>C[C@@H](C(OC)OC)C=C</td><td>The molecule is an ether in which the stereocentres at positions 2 and 3 both have S-configuration. It is an ether and an alicyclic compound.</td></tr><tr><td>O(C([C@H](C)N)=O)[C@H]([C@H](O)C)[C@H](C)O</td><td>The molecule is an amino cyclitol that is scyllo-inositol in which the hydroxy group at position 2 has been replaced by an amino group. It has a role as a bacterial metabolite. It is an amino cyclitol and a primary amino compound. It derives from a scyllo-inositol.</td></tr><tr><td>C(C)(C)(C)Br</td><td>The molecule is a bromoalkane that is ethane substituted by a bromo group at position 2. It has a role as a metabolite. It derives from a hydride of an ethane.</td></tr><tr><td>C[Se]C</td><td>The molecule is an organoselenium compound that is selenium substituted by a methyl group at position 2. It has a role as a metabolite. It derives from a selenium.</td></tr></table>",
+        "bbox": [
+            114,
+            97,
+            961,
+            542
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/d1c437fb8ece5ab3efe66631f6b00748912473d86fded95027f29d167dec8b6f.jpg",
+        "table_caption": [
+            "Table 7: Five examples of SAPubChem-41"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Info</td><td>SAPubChem-41</td><td>ChEBI-20</td><td>PCdes</td><td>DrugBank-23</td></tr><tr><td>Train</td><td>34,407</td><td>26,407</td><td>10,500</td><td>17,109</td></tr><tr><td>Validation</td><td>3,301</td><td>3,301</td><td>1,500</td><td>3,667</td></tr><tr><td>Test</td><td>3,300</td><td>3,300</td><td>3,000</td><td>3,666</td></tr><tr><td>LSMILES</td><td>69.96</td><td>81.56</td><td>56.47</td><td>54.11</td></tr><tr><td>LDescription</td><td>41.59</td><td>52.88</td><td>72.47</td><td>65.04</td></tr><tr><td>Data source</td><td colspan=\"3\">PubChem</td><td>DrugBank</td></tr></table>",
+        "bbox": [
+            230,
+            605,
+            769,
+            727
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/92be78ede5f8b3d404afad6f0fb6c0aad852f1416213d389f4e9dae1c190d048.jpg",
+        "table_caption": [
+            "Table 8: Details about the existing datasets and ours (SAPubChem-41).  $L_{\\text{SMILES}}$  denotes the average length of SMILES while  $L_{\\text{Description}}$  denotes the average word count per description."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Suffix</td><td>Parameters (M)</td></tr><tr><td>Text+Chem T5</td><td>base</td><td>220</td></tr><tr><td>Ada-T5</td><td>-</td><td>220</td></tr><tr><td>MolT5</td><td>base</td><td>220</td></tr></table>",
+        "bbox": [
+            329,
+            804,
+            668,
+            878
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 9: Model Size",
+        "bbox": [
+            426,
+            887,
+            569,
+            901
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "9563",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/e7ba639834671c381245b0f324422bc05107737014d7a4ce9d0e7aa88572461d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>model</td><td>dataset</td><td>BLEU-2</td><td>BLEU-4</td><td>Rouge-1</td><td>Rouge-2</td><td>Rouge-L</td><td>Meteor</td></tr><tr><td rowspan=\"2\">Text+Chem T5</td><td>ChEBI-20</td><td>0.580</td><td>0.490</td><td>0.647</td><td>0.498</td><td>0.586</td><td>0.604</td></tr><tr><td>SAPPubChem-41</td><td>0.584</td><td>0.496</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.614</td></tr><tr><td rowspan=\"2\">Ada-T5</td><td>ChEBI-20</td><td>0.424</td><td>0.316</td><td>0.543</td><td>0.369</td><td>0.483</td><td>0.464</td></tr><tr><td>SAPPubChem-41</td><td>0.444</td><td>0.350</td><td>0.575</td><td>0.410</td><td>0.514</td><td>0.493</td></tr><tr><td rowspan=\"2\">MolT5</td><td>ChEBI-20</td><td>0.540</td><td>0.457</td><td>0.634</td><td>0.485</td><td>0.578</td><td>0.569</td></tr><tr><td>SAPPubChem-41</td><td>0.551</td><td>0.464</td><td>0.638</td><td>0.484</td><td>0.585</td><td>0.572</td></tr></table>",
+        "bbox": [
+            115,
+            108,
+            884,
+            229
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/8a79cbbe44da490fd276862221cb570de0bb828d803903e8dcfd6bbc8ca3192b.jpg",
+        "table_caption": [
+            "Table 10: The complete results of different models for molecular captioning on ChEBI-20, SAPubChem-41 datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>size</td><td>dataset</td><td>BL</td><td>Acc</td><td>Lev</td><td>MACCS</td><td>RDK</td><td>Morgan</td><td>Val</td></tr><tr><td rowspan=\"2\">Text+Chem T5</td><td>ChEBI-20</td><td>0.750</td><td>0.212</td><td>27.39</td><td>0.874</td><td>0.767</td><td>0.697</td><td>0.792</td></tr><tr><td>SAPubChem-41</td><td>0.827</td><td>0.243</td><td>18.59</td><td>0.887</td><td>0.792</td><td>0.726</td><td>0.937</td></tr><tr><td rowspan=\"2\">Ada-T5</td><td>ChEBI-20</td><td>0.699</td><td>0.182</td><td>27.48</td><td>0.869</td><td>0.753</td><td>0.672</td><td>0.886</td></tr><tr><td>SAPubChem-41</td><td>0.714</td><td>0.213</td><td>26.52</td><td>0.879</td><td>0.772</td><td>0.699</td><td>0.854</td></tr><tr><td rowspan=\"2\">MolT5</td><td>ChEBI-20</td><td>0.769</td><td>0.081</td><td>24.49</td><td>0.721</td><td>0.588</td><td>0.529</td><td>0.772</td></tr><tr><td>SAPubChem-41</td><td>0.775</td><td>0.093</td><td>33.16</td><td>0.814</td><td>0.668</td><td>0.582</td><td>0.788</td></tr></table>",
+        "bbox": [
+            124,
+            256,
+            873,
+            378
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/6a503302e82184a1707b0bc67c0cad3ed61b5cd53d3fa17bbb87ef86cad95d08.jpg",
+        "table_caption": [
+            "Table 11: The complete results of different models for molecular generation on ChEBI-20, SAPubChem-41 datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>size</td><td>dataset</td><td>BLEU-2</td><td>BLEU-4</td><td>Rouge-1</td><td>Rouge-2</td><td>Rouge-L</td><td>Meteor</td></tr><tr><td rowspan=\"2\">2k</td><td>random</td><td>0.534</td><td>0.447</td><td>0.641</td><td>0.489</td><td>0.579</td><td>0.575</td></tr><tr><td>refined</td><td>0.574</td><td>0.486</td><td>0.651</td><td>0.502</td><td>0.590</td><td>0.601</td></tr><tr><td rowspan=\"2\">4k</td><td>random</td><td>0.537</td><td>0.451</td><td>0.641</td><td>0.489</td><td>0.579</td><td>0.577</td></tr><tr><td>refined</td><td>0.580</td><td>0.486</td><td>0.653</td><td>0.503</td><td>0.592</td><td>0.602</td></tr><tr><td rowspan=\"2\">6k</td><td>random</td><td>0.532</td><td>0.446</td><td>0.640</td><td>0.489</td><td>0.577</td><td>0.573</td></tr><tr><td>refined</td><td>0.583</td><td>0.490</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.611</td></tr><tr><td rowspan=\"2\">8k</td><td>random</td><td>0.527</td><td>0.443</td><td>0.640</td><td>0.489</td><td>0.579</td><td>0.571</td></tr><tr><td>refined</td><td>0.584</td><td>0.496</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.614</td></tr><tr><td rowspan=\"2\">10k</td><td>random</td><td>0.526</td><td>0.442</td><td>0.639</td><td>0.489</td><td>0.579</td><td>0.571</td></tr><tr><td>refined</td><td>0.583</td><td>0.491</td><td>0.652</td><td>0.503</td><td>0.591</td><td>0.610</td></tr></table>",
+        "bbox": [
+            188,
+            464,
+            811,
+            652
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/b9797e636c73eb79391abd2b0262a0f57dc09ce77a9b9698688e5ae01daf9203.jpg",
+        "table_caption": [
+            "Table 12: Full results of molecule captioning task using different sampling strategy."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>size</td><td>dataset</td><td>BLEU</td><td>Accuracy</td><td>Levenshtein</td><td>MACCS FTS</td><td>RDK FTS</td><td>Morgan FTS</td><td>Validity</td></tr><tr><td rowspan=\"2\">2k</td><td>random</td><td>0.715</td><td>0.215</td><td>23.82</td><td>0.873</td><td>0.766</td><td>0.703</td><td>0.868</td></tr><tr><td>refined</td><td>0.821</td><td>0.236</td><td>19.31</td><td>0.885</td><td>0.788</td><td>0.721</td><td>0.931</td></tr><tr><td rowspan=\"2\">4k</td><td>random</td><td>0.720</td><td>0.219</td><td>23.79</td><td>0.876</td><td>0.768</td><td>0.706</td><td>0.862</td></tr><tr><td>refined</td><td>0.832</td><td>0.237</td><td>18.69</td><td>0.885</td><td>0.788</td><td>0.722</td><td>0.938</td></tr><tr><td rowspan=\"2\">6k</td><td>random</td><td>0.709</td><td>0.220</td><td>24.38</td><td>0.876</td><td>0.767</td><td>0.706</td><td>0.858</td></tr><tr><td>refined</td><td>0.828</td><td>0.240</td><td>19.28</td><td>0.886</td><td>0.788</td><td>0.724</td><td>0.934</td></tr><tr><td rowspan=\"2\">8k</td><td>random</td><td>0.715</td><td>0.221</td><td>24.27</td><td>0.875</td><td>0.763</td><td>0.701</td><td>0.855</td></tr><tr><td>refined</td><td>0.827</td><td>0.243</td><td>18.59</td><td>0.887</td><td>0.792</td><td>0.726</td><td>0.937</td></tr><tr><td rowspan=\"2\">10k</td><td>random</td><td>0.715</td><td>0.207</td><td>24.37</td><td>0.870</td><td>0.759</td><td>0.696</td><td>0.864</td></tr><tr><td>refined</td><td>0.830</td><td>0.238</td><td>19.03</td><td>0.883</td><td>0.786</td><td>0.721</td><td>0.937</td></tr></table>",
+        "bbox": [
+            115,
+            678,
+            919,
+            864
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 13: Full results of molecule generation task using different sampling strategy.",
+        "bbox": [
+            213,
+            875,
+            779,
+            891
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "9564",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/b6b845c268b3d875f157eaedf5182b638df8c513474a910e9aab2e1e421cc2dd.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>model</td><td>dataset</td><td>BLEU-2</td><td>BLEU-4</td><td>Rouge-1</td><td>Rouge-2</td><td>Rouge-L</td><td>Meteor</td></tr><tr><td rowspan=\"3\">Text+Chem T5</td><td>ChEBI-20</td><td>0.580</td><td>0.490</td><td>0.647</td><td>0.498</td><td>0.586</td><td>0.604</td></tr><tr><td>PCdes</td><td>0.352</td><td>0.266</td><td>0.439</td><td>0.274</td><td>0.373</td><td>0.382</td></tr><tr><td>SAPubChem-41</td><td>0.584</td><td>0.496</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.614</td></tr><tr><td rowspan=\"3\">Ada-T5</td><td>ChEBI-20</td><td>0.424</td><td>0.316</td><td>0.543</td><td>0.369</td><td>0.483</td><td>0.464</td></tr><tr><td>PCdes</td><td>0.289</td><td>0.188</td><td>0.428</td><td>0.243</td><td>0.367</td><td>0.324</td></tr><tr><td>SAPubChem-41</td><td>0.444</td><td>0.350</td><td>0.575</td><td>0.410</td><td>0.514</td><td>0.493</td></tr><tr><td rowspan=\"3\">MolT5</td><td>ChEBI-20</td><td>0.540</td><td>0.457</td><td>0.634</td><td>0.485</td><td>0.578</td><td>0.569</td></tr><tr><td>PCdes</td><td>0.165</td><td>0.078</td><td>0.290</td><td>0.118</td><td>0.233</td><td>0.204</td></tr><tr><td>SAPubChem-41</td><td>0.551</td><td>0.464</td><td>0.638</td><td>0.484</td><td>0.585</td><td>0.572</td></tr></table>",
+        "bbox": [
+            115,
+            288,
+            884,
+            458
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "table",
+        "img_path": "images/530a2332fed7026d725230a4d1b0eac89514455dc0075561e55f83bb60c73018.jpg",
+        "table_caption": [
+            "Table 14: The complete results of different models for molecular captioning on SAPubChem-41 and other existing datasets. The best scores are in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>size</td><td>dataset</td><td>BL</td><td>Acc</td><td>Lev</td><td>MACCS</td><td>RDK</td><td>Morgan</td><td>Val</td></tr><tr><td rowspan=\"3\">Text+Chem T5</td><td>ChEBI-20</td><td>0.750</td><td>0.212</td><td>27.39</td><td>0.874</td><td>0.767</td><td>0.697</td><td>0.792</td></tr><tr><td>PCdes</td><td>0.614</td><td>0.105</td><td>30.43</td><td>0.697</td><td>0.544</td><td>0.459</td><td>0.849</td></tr><tr><td>SAPPubChem-41</td><td>0.827</td><td>0.243</td><td>18.59</td><td>0.887</td><td>0.792</td><td>0.726</td><td>0.937</td></tr><tr><td rowspan=\"3\">Ada-T5</td><td>ChEBI-20</td><td>0.699</td><td>0.182</td><td>27.48</td><td>0.869</td><td>0.753</td><td>0.672</td><td>0.886</td></tr><tr><td>PCdes</td><td>0.579</td><td>0.089</td><td>38.99</td><td>0.778</td><td>0.600</td><td>0.492</td><td>0.947</td></tr><tr><td>SAPPubChem-41</td><td>0.714</td><td>0.213</td><td>26.52</td><td>0.879</td><td>0.772</td><td>0.699</td><td>0.854</td></tr><tr><td rowspan=\"3\">MolT5</td><td>ChEBI-20</td><td>0.769</td><td>0.081</td><td>24.49</td><td>0.721</td><td>0.588</td><td>0.529</td><td>0.772</td></tr><tr><td>PCdes</td><td>0.476</td><td>0.009</td><td>53.27</td><td>0.635</td><td>0.432</td><td>0.330</td><td>0.711</td></tr><tr><td>SAPPubChem-41</td><td>0.775</td><td>0.093</td><td>33.16</td><td>0.814</td><td>0.668</td><td>0.582</td><td>0.788</td></tr></table>",
+        "bbox": [
+            122,
+            500,
+            875,
+            670
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Table 15: The complete results of different models for molecular generation on SAPubChem-41 and other existing datasets. The best scores are in bold.",
+        "bbox": [
+            112,
+            680,
+            882,
+            709
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "9565",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2024/Enhancing Cross Text-Molecule Learning by Self-Augmentation/206d3795-d53b-42c9-8de8-6d682eb0ac0b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a7efc51c26d7b644cdebccb4a1800d00b95fcf9f50a9729a8dff54ab55dcd9b
+size 2153015

2024/Enhancing Cross Text-Molecule Learning by Self-Augmentation/full.md

@@ -0,0 +1,323 @@
+# Enhancing Cross Text-Molecule Learning by Self-Augmentation
+
+Yinuo Jiang, Xiang Zhuang, Keyan Ding,
+
+Qiang Zhang*, Huajun Chen*
+
+College of Computer Science and Technology, Zhejiang University
+
+ZJU-Hangzhou Global Scientific and Technological Innovation Center
+
+{yinuoj,zhuangxiang,dingkeyan,qiang.zhang.cs,huajunsir} @zju.edu.cn
+
+# Abstract
+
+The development of Large Language Models (LLMs) has greatly advanced the field of drug discovery, with the belief that natural language can enhance human control over molecule design. However, the scarcity of high-quality labeled data remains a challenge for cross-text-molecule learning. Existing datasets are limited due to the difficulty of collecting precise molecule-description pairs. Although recent efforts have utilized pseudo data generated by LLMs for augmentation, the lack of specialized chemistry knowledge of LLMs and the absence of an effective high-quality data selector may introduce noise into the annotations, compromising the models' robustness. To address these challenges, this paper introduces a novel framework that interweaves model finetuning and data augmentation to overcome the scarcity of high-quality data. The proposed approach involves an iterative procedure where the model plays dual roles in annotating unlabeled data and sampling a subset of high-quality data until convergence is achieved, enhancing the model's understanding and adaptability. Additionally, a new dataset called SAPubChem-41 is presented, which comprises meticulously curated high-quality parallel molecule-description pairs designed specifically for fine-tuning purposes. This research provides an important contribution to the field by addressing the need for high-quality datasets and presenting an effective framework for cross-text-molecule learning.
+
+# 1 Introduction
+
+The emergence of Large Language Models (LLMs) has significantly propelled the development of drug discovery (Zhang et al., 2024). Recent progress in language models shed light on drug discovery, with the vision that humans can possess a higher-level control over molecule design facilitated by natural
+
+language. Edwards et al. (2022a) introduce two new tasks: molecule captioning (Mol2Cap) and text-based molecule generation (Cap2Mol), and present MolT5 (Edwards et al., 2022b) based on the T5 (Raffel et al., 2020) architecture to translate between molecule and text. Subsequently, diverse endeavors have been undertaken to tackle these challenges. Text+Chem T5 (Christofidellis et al., 2023) and BioT5 (Pei et al., 2024) are also T5-like models, which incorporate multi-task and multidomain pretraining process and enable the bidirectional generation between different modalities in a single model. Some other methods employ the Generative Pretrained Transformer (GPT) architecture. MolXPT (Liu et al., 2023) proposes to wrap molecules in the sentences to make the pertaining corpus. More recently, MolCA(Liu et al., 2024) considers the 2D graph information for molecules and designs a projector to connect molecular content to language models. These studies mark the research efforts in the realm of cross text-molecule learning.
+
+Despite rapid development in cross-text-molecule modalities, there is still demand for high-quality labeled data in this field. Such data of superior quality is indispensable to capacitate models for downstream tasks. However, the available datasets relevant to the tasks are limited, a consequence of the arduous nature of collecting precise and enlightening parallel molecule-description pairs. This scarcity of annotated data hinders the model's ability to learn between the modalities of the molecule and the text. As shown in Figure 1, recent efforts by Chen et al. (2023) address the scarcity challenge by utilizing pseudo data generated by LLMs as augmentation. The problem is that the absence of specialized knowledge in chemistry may introduce noise into the data annotated by general Large Language Models, thereby compromising the robustness of models fine-tuned on such datasets. In spite of these commendable advance
+
+ments, the accessibility of high-quality datasets for fine-tuning remains notably deficient.
+
+![](images/272b4bf873814a136101d35b94d4363a6c239a550a9bda1a8367e046d42a9d8c.jpg)  
+(a) Conventional Data Augmentation Method
+
+![](images/5c9e685af13cd22f40c029e7e296376a3c547031d72cedeb36c03bb1feb72951.jpg)  
+(b) Our Self-Augmentation Method  
+Figure 1: A comparison between a conventional data augmentation method referenced in Chen et al. (2023) and our iterative self-augmentation techniques is drawn herein. The original technique leverages Label Language Models (LLMs) to annotate unlabeled data; contrastingly, our proposed self-augmentation strategies use the model itself for annotating unlabeled data in tandem with an iterative training approach.
+
+To address these challenges, this paper introduces a novel framework that interweaves model fine-tuning and data augmentation (See Fig. 1, aiming to rectify the high-quality data scarcity observed in the realm of cross text-molecule learning. Our proposed approach integrates the model into the process by assigning it dual roles: annotating unlabelled data and sampling a subset of high-quality data. This iterative procedure continues until convergence is achieved, thereby refining the model's understanding and adaptability. Additionally, we present a new, larger dataset SAPubChem-41. This meticulously curated dataset comprises a wealth of high-quality parallel molecule-description pairs meticulously designed for the specific purpose of fine-tuning. The model optimized using SAPubChem-41 has been evaluated across a total of 13 molecule-text benchmarks, exhibiting superior performance in comparison to the model optimized using the original dataset, ChEBI-20. Our key contributions can be summarized as:
+
+- We present an innovative framework designed to augment cross-text-molecule learning through a self-augmentation strategy. This system leverages the model itself to annotate unlabelled data, choosing only samples that meet a high-quality standard.  
+- We introduce SAPubChem-41, a novel dataset consisting of both real data and high-quality pseudo-labeled data, augmented by the model itself. This makes a notable contribution to the available datasets in the field of cross-molecule-text modality.  
+- We show that within our iterative setting, self-augmented data significantly improves the performance of the model in downstream tasks with each epoch. Furthermore, these results tend to converge as the number of epochs increase. This process thereby successfully validates the efficacy of self-augmented data in enhancing model performance.
+
+# 2 Background
+
+# 2.1 Cross Text-Molecule Learning
+
+Edwards et al. (2021) introduce a new task Text2Mol, which uses descriptions as search queries to retrieve the target molecules. Edwards et al. (2022a) first addresses the problem of cross-domain generation by linking natural language and chemistry, tackling tasks such as text-conditional de novo molecule generation and molecule captioning. Li et al. (2023a) combines retrieval-based prompt paradigm with LLMs like ChatGPT to achieve translation between molecule language and natural language. Christofidellis et al. (2023) propose the first multi-domain, multi-task language model that can solve a wide range of tasks in both the chemical and natural language domains. Liu et al. (2023) propose MolXPT, a GPT-based model pre-trained on molecule SMILES, biomedical text, and wrapped text. BioT5 (Pei et al., 2024) further exploits SELFIES for  $100\%$  robust molecular representations and discriminate structured knowledge from unstructured knowledge. To capitalize on these insights, we adopt an iterative framework to interweave model fitting and data augmentation, to mitigate data scarcity and further enhance the alignment between molecule and text representations.
+
+# 2.2 Data Augmentation for Cross Text-Molecule Learning
+
+Data augmentation is a regularization strategy utilized to enhance model performance by diversifying the available data through various techniques (Hernández-García and König, 2018). In the field of natural language processing (NLP), data augmentation has been widely employed to address data scarcity issues, leading to the proposal of various augmentation methods (Feng et al., 2021; Li et al., 2022). Textual data augmentation can be accomplished through simple rules, such as synonym replacement and word order modification (Zhang et al., 2015). Furthermore, the development of deep learning models has facilitated the generation of new text to augment the existing data (Szegedy et al., 2016; Wu et al., 2019; Anaby-Tavor et al., 2020; Yoo et al., 2021; Zhou et al., 2021; Li et al., 2023b; Dai et al., 2023). Acquiring high-quality annotated molecule-text datasets is prohibitively expensive, for many descriptions of corresponding molecules are scattered and inaccurate, leading to limited fine-tuning datasets for cross text-molecule learning. More data augmentation techniques are desired to further enhance cross text-molecule learning. Chen et al. (2023) pioneers to exploit LLMs' annotated pseudo label for domain adaption and data augmentation. However, while pseudo label as domain adaptation shows impressive performance compared to existing methods, pseudo label as data augmentation still struggles against label noise interfering model's performance. To address the problem and further improve the quality of augmented dataset, our work introduces a pioneering iterative framework that interweaves model fine-tuning and data augmentation.
+
+# 3 Method
+
+In this section, we describe a computational framework for jointly optimizing model performance and dataset quality.
+
+# 3.1 Overall Framework
+
+The literature demonstrates a substantial body of research focused on enhancing both model performance and dataset quality concurrently, as depicted in Figure 2. If we assume the existence of a subjective testing environment capable of gathering reliable molecule-caption pairs, then the core challenge becomes how to effectively sample a subset  $A$  from a large-scale pseudo-labeled dataset  $D$
+
+that exhibits high confidence. In line with Li et al. (2023b), our approach adopts an iterative strategy, combining model fine-tuning and data augmentation operations defined below.
+
+1. Model fine-tuning: Given a data distribution, find an approximate risk minimizer.  
+2. Data augmentation: Given a model, augment data and sample a new data distribution.  
+3. Data combination: Combine a set of distributions into a single distribution.
+
+At the beginning of each iteration, a new model is finetuned on the latest dataset. This newly trained model is subsequently employed to annotate a set of unlabeled data, thereby generating a pseudolabeled dataset. The sampler is then entrusted with the responsibility of discerning samples distinguished by their high quality from this pseudolabeled dataset. The discerned insights are then harnessed to augment and refine the current dataset, thereby facilitating a continuous and iterative improvement process.
+
+# 3.2 Self-Annotating
+
+We gathered 200,000 unannotated SMILES of molecules from PubChem (Kim et al., 2023) and conducted dedduplication to mitigate the risk of data leakage in the validation dataset. For each unlabeled SMILES, we executed inference on the model itself to produce a candidate caption, thereby obtaining the pseudo-labeled dataset.
+
+ChatGPT Annotating vs Specialized Model Annotating Despite the remarkable performance of Large Language Models (LLMs) such as ChatGPT in various cross-modal tasks, they have not succeeded in surpassing specialized models in the molecule captioning task (Li et al., 2023a; Christofidellis et al., 2023). This suggests that specialized models serve as better annotators for unlabeled SMILES strings. Furthermore, as depicted in Figure 3(a), by leveraging the Text2Mol score (Edwards et al., 2022a) as a fundamental metric of data quality, which measures the similarity between SMILES strings and their captions, we evaluated the distribution of the real training set, MolReGPT annotated data, and self-annotated data. It is evident that the self-annotated data yields higher-quality data, closely resembling real data more than the MolReGPT annotated data.
+
+Iterative Annotating We employ an iterative training approach to enhance the proficiency of annotators and produce captions of superior quality.
+
+![](images/9df76a23d759ae9f4034356e1c20d7741b9c68c8cd33417e41be9b9e838b4682.jpg)  
+Figure 2: Overall framework. We interweave model fine-tuning and data augmentation iteratively. A core component of our framework, the Data Augmentation Module, is compartmentalised into two key elements: Self-Annotating and Self-Sampling. These elements underline the crux of our methodology, wherein the model autonomously performs the augmentation, subsequently enhancing its own performance.
+
+As depicted in Figure 3, during each iteration, the quality of the molecule-caption pairs is enhanced through the implementation of the iterative strategy.
+
+# 3.3 Self-Sampling
+
+A discerning subset sampling strategy is anticipated to effectively filter pseudo labels, thereby ensuring the attainment of high-quality labels for training the model. This refined approach aims to enhance the overall reliability of the model by mitigating the impact of unreliable or inaccurate labels in the training process. In contrast to prior work (Li et al., 2023b), wherein fine-tuning an external score model was employed to select high-quality pairs, we have found this approach to be
+
+both time-consuming and unreliable when applied to the domain of cross text-molecule learning. Notably, the susceptibility of the model to overfitting incorrect pseudo-labels is a concern, as detailed in section 4.4. To tackle this dilemma, we introduce a dual selection mechanism that not only proves to be more computationally efficient but also enhances reliability in ensuring the high quality of our augmented dataset. This mechanism involves two key steps. Firstly, leveraging the generation likelihood as a confidence score to filter out unreliable molecule-caption pairs. Secondly, updating and ranking the remaining molecule-caption pairs based on the Text2Mol score (Edwards et al., 2022a) This refined approach streamlines the sampling process, offering increased efficiency and
+
+![](images/20951661f4e22357b7b4d029b32a0c395e446f3d4244e0c8c5e64433b59a0d3d.jpg)  
+(a) Comparison of Data Quality Between LLMs Annotated Label and Self-Annotated Label
+
+![](images/8f0f2157fe37c877e088e862611123d5520baa1f758db2f66630cbc7098ebf02.jpg)  
+(b) Comparison of Data Quality in Relation to the Adoption or Non-Apoption of Iterative Strategy  
+Figure 3: In a manner akin to Chen et al. (2023) in their study, we employ the methodology proposed by Edwards et al. (2022) to evaluate the correlation between molecule-description pairs, utilizing it as a measure of data quality. The distributions are illustrated via Kernel Distribution Estimation. A noteworthy aspect to consider here is that a higher Text2Mol score typically implies a closer resemblance between the molecule and its description. In addition, the term "Density" in (a) pertains to the concentration of data within a specific region.
+
+heightened confidence in the identification of high-quality pairs for dataset augmentation.
+
+Data Filtering First, we employ the model itself to filter out a set of low-quality pairs. Specifically, we utilize the generation likelihood as a measure of the quality of the generated sequence. However, previous studies (Liu and Liu, 2021; Liu et al., 2022) have underscored that the correlation between sequence probability and its quality for Maximum Likelihood Estimation (MLE) trained models can be low, due to the presence of deterministic (one-point) target distribution issues. Furthermore, sequence likelihood estimation becomes noisier when the decoded sequences of models deviate from the exposed training data distribution, thereby exacerbating the problem of exposure bias
+
+(Ranzato et al., 2016). To address these intricate challenges, we introduce the Sequence Likelihood Calibration (SLiC) stage. This stage, depicted in Figure 1, serves to mitigate the aforementioned issues and diminish the gap between sequence likelihood and its associated quality. Following the methodology of Zhao et al. (2022), we conduct further fine-tuning of the model, introducing a new objective that encompasses the following loss functions:
+
+$$
+\begin{array}{l} \mathrm {L} _ {\text {r a n k}} ^ {\text {c a l}} = \max  \left(0, \beta - \log P _ {\theta} \left(\hat {\mathbf {y}} _ {+} \mid \mathbf {x}\right) + \log P _ {\theta} \left(\hat {\mathbf {y}} _ {-} \mid \mathbf {x}\right)\right), \\ \mathbf {L} _ {\mathrm {c e}} ^ {\mathrm {r e g}} = \sum_ {t} - \log P _ {\theta} \left(\bar {y} _ {t} \mid \overline {{\mathbf {y}}} _ {t - 1}, \mathbf {x}\right), \\ \end{array}
+$$
+
+$$
+\mathrm {L} _ {\mathrm {k l}} ^ {\mathrm {r e g}} = \sum_ {t} P _ {\theta} \left(\bar {y} _ {t} \mid \overline {{\mathbf {y}}} _ {t - 1}, \mathbf {x}\right) \log \frac {P _ {\theta} \left(\bar {y} _ {t} \mid \overline {{\mathbf {y}}} _ {t - 1} , \mathbf {x}\right)}{P _ {\theta_ {f t}} \left(\bar {y} _ {t} \mid \overline {{\mathbf {y}}} _ {t - 1} , \mathbf {x}\right)}.
+$$
+
+Given the context  $\mathbf{x}$ , target  $\overline{\mathbf{y}}$ , and positive and negative candidates pairs  $\hat{\mathbf{y}}_{+}, \hat{\mathbf{y}}_{-}, P_{\theta}(\mathbf{y} \mid \mathbf{x})$  denotes the generation sequence likelihood. Rank loss  $L_{rank}^{cal}$  optimizes the ranking order of positive and negative candidate pairs. Cross entropy loss  $L_{ce}^{reg}$  is the standard fine-tuning MLE objective. KL divergence loss  $L_{kl}^{reg}$  directly minimizes the probability distribution distance between the calibrated model and the fine-tuned model at each token on the observed target sequence.
+
+Following the SLiC stage, we regard the generation likelihood as one of our metrics for assessing the quality of pseudo labels. In our experiments, we discard pairs with log-likelihood equal to negative infinity and retain the remaining pairs as high-quality pairs.
+
+Update and Rank Secondly, for those high-quality pairs, we further refine the data based on the similarity score  $\mathcal{F}_{\mathrm{text2mol}}$  proposed by Edwards et al. (2022a). At iteration t, we update the caption if  $\mathcal{F}_{\mathrm{text2mol}}(m,c_t) > \mathcal{F}_{\mathrm{text2mol}}(m,c_{t - 1})$  then we rank the updated dataset according to  $\mathcal{F}_{\mathrm{text2mol}}$  ,selecting the top  $t\times 2k$  pairs to form the new Augmented Dataset  $A_{t}$
+
+# 4 Experiments
+
+In order to validate the efficacy of utilizing our framework, we have undertaken a series of comprehensive experiments.
+
+# 4.1 Experimental Setup
+
+Dataset We employ the ChEBI-20 (Edwards et al., 2021) to serve as our original dataset. Additionally, we gather 200,000 unannotated molecules
+
+<table><tr><td>Info</td><td>SAPubChem-41</td><td>ChEBI-20</td></tr><tr><td>Train</td><td>34,407</td><td>26,407</td></tr><tr><td>Validation</td><td>3,301</td><td>3,301</td></tr><tr><td>Test</td><td>3,300</td><td>3,300</td></tr><tr><td>\(\mathcal{L}_{\text{SMILES}}\)</td><td>69.96</td><td>81.56</td></tr><tr><td>\(\mathcal{L}_{\text{Description}}\)</td><td>41.59</td><td>52.88</td></tr></table>
+
+Table 1: Comparison of our SAPubChem-41 dataset with the original dataset ChEBI-20.  $L_{\text{SMILES}}$  denotes the average length of SMILES while  $L_{\text{Description}}$  denotes the average word count per description.
+
+![](images/ba287cb3b5eef8c8e418c1e1368705f12332150847eb5d628a1bc8acb032545b.jpg)  
+Figure 4: The results of the iteration procedure. The dashed lines denote the metrics associated with the molecule captioning task, while the solid lines signify those related to the molecule generation task.
+
+from the PubChem database (Kim et al., 2023). Subsequently, we conduct a rigorous filtering procedure to exclude any molecules present in downstream datasets, ensuring the elimination of potential overlaps between the newly acquired molecules and those already present in the ChEBI-20 dataset. For our experimental dataset augmentation, we set the iteration step to 2000. At each iteration stage, the dataset size expands by 2000 from the preceding iteration. A detailed comparison between our final augmented dataset, SAPubChem-41, and the original dataset, ChEBI-20, is presented in Table 1.
+
+Base Model To prove the effectiveness of our framework, we use the Text+Chem T5 (Christofidellis et al., 2023) as the base model for simulating the iterative process.
+
+Metrics Following the previous studies (Edwards et al., 2022a; Li et al., 2023a; Christofidellis et al.,
+
+2023; Chen et al., 2023), we evaluate the results with following metrics:
+
+- Molecular Captioning: BLEU-2 and BLEU-4(Papineni et al., 2002) are metrics used to evaluate the quality of machine-generated text by comparing it to a reference text, with a higher score on the BLEU metric indicating a higher level of similarity between the generated text and reference text. ROUGE-1, ROUGE-2, and ROUGE-L(Lin, 2004) is similar to BLEU, while computing the recall-overlap of unigrams, bigrams, and longest common subsequences between the generated and reference texts. METEOR (Banerjee and Lavie, 2005) is a metric that uses a combination of unigram precision, recall, and a synonym-matching component to evaluate the generated text against the reference text, which is designed to be more sensitive to fluency, meaning, and structure than BLEU.
+
+- Text-based Molecule Generation: BLEU and the Exact Match scores are calculated as basic assessments. Molecule-specified metrics including Levenshtein distance(Levenshtein et al., 1966), validity(Edwards et al., 2022a), and three molecule fingerprints scores - MACCS FTS(Durant et al., 2002), RDK FTS(Schneider et al., 2015), and Morgan FTS(Rogers and Hahn, 2010) are calculated to provide valuable insights into the quality, validity, and structural characteristics of the generated molecules.
+
+Implementation We conducted our experiments on Ubuntu 22.04 using RTX 4090(24GB) * 4 with CUDA 11.7.1. Our primary dependencies are Python 3.11.4, PyTorch 2.0.1, Transformers 4.31.0, and Numpy 1.24.3. We initialize the three pretrained models using public checkpoints. For fine-tuning, we adopted the configuration in MolT5 (Edwards et al., 2022b), with a learning rate of 1e-3, 50,000 fine-tuning steps, weight decay of 0.1, batch size of 32, random seed of 42, and 1000 warm-up steps. For evaluation, we use a greedy search with a maximum generation length of 512 during generation. The evaluation metrics script is derived from MolT5 (Edwards et al., 2022b). For other hyperparameters, we relied on the default settings of the T5ForConditionalGeneration class in Huggingface.
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="3">Text+Chem T5</td><td colspan="3">Ada-T5</td><td colspan="3">MolT5</td></tr><tr><td>BL</td><td>RG</td><td>MET</td><td>BL</td><td>RG</td><td>MET</td><td>BL</td><td>RG</td><td>MET</td></tr><tr><td>ChEBI-20</td><td>0.490</td><td>0.498</td><td>0.604</td><td>0.316</td><td>0.369</td><td>0.464</td><td>0.457</td><td>0.485</td><td>0.569</td></tr><tr><td>SAPubChem-41</td><td>0.496</td><td>0.504</td><td>0.614</td><td>0.350</td><td>0.410</td><td>0.493</td><td>0.464</td><td>0.484</td><td>0.572</td></tr></table>
+
+Table 2: Results of different models for molecular captioning on ChEBI-20, SAPubChem-41 datasets. The best scores are in bold. BL:BLEU-4, RG:ROUGE-2, MET:METEOR  
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="3">Text+Chem T5</td><td colspan="3">Ada-T5</td><td colspan="3">MolT5</td></tr><tr><td>Morgan</td><td>Acc</td><td>Val</td><td>Morgan</td><td>Acc</td><td>Val</td><td>Morgan</td><td>Acc</td><td>Val</td></tr><tr><td>ChEBI-20</td><td>0.697</td><td>0.212</td><td>0.792</td><td>0.672</td><td>0.182</td><td>0.886</td><td>0.529</td><td>0.081</td><td>0.772</td></tr><tr><td>SAPubChem-41</td><td>0.726</td><td>0.243</td><td>0.937</td><td>0.699</td><td>0.213</td><td>0.854</td><td>0.582</td><td>0.093</td><td>0.788</td></tr></table>
+
+Table 3: Results of different models for molecular generation on ChEBI-20, SAPubChem-41 datasets. The best scores are in bold. Morgan:Morgan FTS, Acc:Accuracy, Val:Validity
+
+# 4.2 Performance Comparison of Molecule-Caption Translation
+
+Molecular Captioning Table 2 presents the results of the molecule captioning task, utilizing the original ChEBI-20 dataset and the augmented SAPubChem-41 dataset to optimize three distinct pre-trained models. Across all measurement metrics, it is observable that the performance of all models using our augmented SAPubChem-41 dataset is generally superior to that utilizing the original ChEBI-20 dataset.
+
+Text-based Molecule Generation Table 3 presents the results of the text-based molecule generation task, utilizing the original ChEBI-20 dataset and the augmented SAPubChem-41 dataset to optimize three distinct pre-trained models. Notably, the performance of all models that utilize our augmented SAPubChem-41 dataset generally surpasses those that use the original ChEBI-20 dataset. The accuracy of the generated molecule, as well as its similarity to the ground truth molecule, have been significantly improved following the self-augmentation of the dataset. This evidences substantial enhancements in performance when integrated with our self-rewarding framework.
+
+# 4.3 Iterative Procedure
+
+We utilize our framework on the Text+Chem T5, taking the ChEBI-20 as the original dataset. This approach facilitates an augmented dataset while simultaneously enhancing the model's performance. It is critical to note that our intention in these experiments is not necessarily to achieve a state-of-the-art model. Rather, our prime objective is to examine the efficacy of our framework, which is designed
+
+to optimize both the model and the pseudo-labeled dataset concurrently. Figure 4 illustrates the iterative process in which dashed lines represent the molecule captioning task and solid lines signify the text-based molecule generation task. We report the complete results in Table 4 and 5. There are a few observations: First, it is noticeable that for both tasks the outcome of iteration 5 mostly falls short of that of iteration 4 in all metrics. This suggests that the training process reaches convergence on iteration 4 for both tasks. Consequently, we adopt the augmented dataset obtained at the fourth iteration as our conclusive dataset. Second, for the text-based molecule generation task, there is a consistent upward trend across all pertinent metrics. This suggests that an increased quantity of high-quality pseudo-labeled data contributes to the optimization of the model in the molecule generation process. This trend underscores the necessity of sufficient high-quality data in cross text-molecule learning. Lastly, when it comes to the molecule captioning task, the general trend in relevant metrics indicates a rise. Initially, there is a minor decrease in the BLEU and Meteor metrics, which is counterweighed by an enhancement in Rouge. As the volume of high-quality pseudo-labeled data increases, all metrics demonstrate improvement. This trend suggests that the quantity of high-quality pseudo-labeled data is also a significant factor.
+
+# 4.4 Ablation Study
+
+We perform further ablation studies to validate the effectiveness of the key component in our framework, i.e., the self-sampling strategy.
+
+To demonstrate the significance of pseudo
+
+<table><tr><td>Iteration</td><td>BLEU↑</td><td>Accuracy↑</td><td>Levenshtein↓</td><td>MACCS FTS↑</td><td>RDK FTS↑</td><td>Morgan FTS↑</td><td>Validity↑</td></tr><tr><td>t0</td><td>0.750</td><td>0.212</td><td>27.39</td><td>0.874</td><td>0.767</td><td>0.697</td><td>0.792</td></tr><tr><td>t1</td><td>0.821</td><td>0.236</td><td>19.31</td><td>0.885</td><td>0.788</td><td>0.721</td><td>0.931</td></tr><tr><td>t2</td><td>0.832</td><td>0.237</td><td>18.69</td><td>0.885</td><td>0.788</td><td>0.722</td><td>0.938</td></tr><tr><td>t3</td><td>0.828</td><td>0.240</td><td>19.28</td><td>0.886</td><td>0.788</td><td>0.724</td><td>0.934</td></tr><tr><td>t4</td><td>0.827</td><td>0.243</td><td>18.59</td><td>0.887</td><td>0.792</td><td>0.726</td><td>0.937</td></tr><tr><td>t5</td><td>0.830</td><td>0.238</td><td>19.03</td><td>0.883</td><td>0.786</td><td>0.721</td><td>0.937</td></tr></table>
+
+Table 4: The iterative results of molecule generation task. The best scores are in bold.  
+
+<table><tr><td>Iteration</td><td>BLEU-2↑</td><td>BLEU-4↑</td><td>Rouge-1↑</td><td>Rouge-2↑</td><td>Rouge-L↑</td><td>Meteor↑</td></tr><tr><td>t0</td><td>0.580</td><td>0.490</td><td>0.647</td><td>0.498</td><td>0.586</td><td>0.604</td></tr><tr><td>t1</td><td>0.574</td><td>0.486</td><td>0.651</td><td>0.502</td><td>0.590</td><td>0.601</td></tr><tr><td>t2</td><td>0.580</td><td>0.486</td><td>0.653</td><td>0.503</td><td>0.592</td><td>0.602</td></tr><tr><td>t3</td><td>0.583</td><td>0.490</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.611</td></tr><tr><td>t4</td><td>0.584</td><td>0.496</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.614</td></tr><tr><td>t5</td><td>0.583</td><td>0.491</td><td>0.652</td><td>0.503</td><td>0.591</td><td>0.610</td></tr></table>
+
+Table 5: The iterative results of molecule captioning task. The best scores are in bold.
+
+![](images/67f66300866b8b8fef22300c68bc1b819083dfc751a3ff3e64653bf338270791.jpg)  
+(a) BLEU
+
+![](images/b9ff6c3114af2035e8b2e46602c38bc7beb38655ccb600ce166b2da9cb31f0ed.jpg)  
+(b) Accuracy
+
+![](images/47a191f09c77f920faa6c0442d2a157cef34f22990f8853772b4a7ac9072aaa4.jpg)  
+(c) Morgan FTS
+
+![](images/90c9e90c14159a4efccd232f8acb6ce87b2be93964e115550168e9f124f829e5.jpg)  
+Figure 5: Results of molecule generation task using different sampling strategy. The blue line marks the result of model optimization on the original dataset.  
+(a) BLEU-4  
+Figure 6: Results of molecule captioning task using different sampling strategy. The blue line marks the result of model optimization on the original dataset.
+
+![](images/f738188f7c04194e649bd0ce00df7eee7a8a5b0171900874a3169b2cf90d4c36.jpg)  
+(b) Rouge-2
+
+![](images/996ceabbab978d4d1995cb968c26ebbe415cc74a098b4a17fdcf433dbd263b82.jpg)  
+(c) Meteor
+
+labeled data selection, we conducted a comparative analysis of the performance of all intermediate models. We compared the use of randomly-selected data of the same size with our refined data. As depicted in Figure 5 and 6, our refined data consistently outperformed the randomly-selected data across all metrics at each iteration for both tasks.
+
+Notably, when optimizing the model on randomly-selected data, the results were generally inferior to the original model on most metrics. This indicates that the model tends to overfit to noisy pseudolabels. These findings emphasize the critical role of the self-sampling strategy in our framework.
+
+# 5 Conclusion
+
+In this paper, we have explored and grappled with the challenge of a deficiency in the accessibility of high-quality datasets for model fine-tuning in the field of cross text-molecule learning, by developing a novel framework. This innovative framework efficiently combines model fine-tuning with data augmentation, integrating the model in such a way that it expands its roles to include annotating unlabeled data and sampling a subset of high-quality data. The paper further introduces a high-quality dataset, the SAPubChem-41, expressly designed for fine-tuning purposes. The experiments' results were impressive as the model optimized using SAPubChem-41 consistently outperformed the model optimized using the original dataset, ChEBI-20. Going forward, this paper's findings present crucial steps toward addressing high-quality data scarcity for fine-tuning and ensuring further advancement in the exciting field of cross text-molecule learning.
+
+# 6 Limitations and Future Works
+
+One limitation of our framework is that it only considers a 1D representation of molecules. The integration of additional representations, such as molecule graphs, is left as future work. Additionally, our data augmentation module currently only supports sample-level labels. It is expected that more quality measurements from different levels will be incorporated to provide a comprehensive perspective on quantifying the dataset quality.
+
+# Acknowledgements
+
+This work is supported by National Natural Science Foundation of China (62302433, U23A20496), Zhejiang Provincial "Jianbing" "Lingyan" Research and Development Program of China (2024C01135), Zhejiang Provincial Natural Science Foundation of China (LQ24F020007), New Generation AI Development Plan for 2030 of China (2023ZD0120802) and CCF-Tencent Rhino-Bird Fund (RAGR20230122).
+
+# References
+
+Ateret Anaby-Tavor, Boaz Carmeli, Esther Goldbraich, Amir Kantor, George Kour, Segev Shlomov, Naama Tepper, and Naama Zwerdling. 2020. Do not have enough data? deep learning to the rescue! In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7383-7390.
+
+Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72.  
+Yuhan Chen, Nuwa Xi, Yanrui Du, Haochun Wang, Chen Jianyu, Sendong Zhao, and Bing Qin. 2023. From artificially real to real: Leveraging pseudo data from large language models for low-resource molecule discovery.  
+Dimitrios Christofidellis, Giorgio Giannone, Jannis Born, Ole Winther, Teodoro Laino, and Matteo Manica. 2023. Unifying molecular and textual representations via multi-task language modelling.  
+Haixing Dai, Zhengliang Liu, Wenxiong Liao, Xiaoke Huang, Yihan Cao, Zihao Wu, Lin Zhao, Shaochen Xu, Wei Liu, Ninghao Liu, et al. 2023. Auggpt: Leveraging chatgpt for text data augmentation. arXiv preprint arXiv:2302.13007.  
+Kirill Degtyarenko, Paula De Matos, Marcus Ennis, Janna Hastings, Martin Zbinden, Alan McNaught, Rafael Alcantara, Michael Darsow, Mickaël Guedj, and Michael Ashburner. 2007. Chebi: a database and ontology for chemical entities of biological interest. Nucleic acids research, 36(suppl_1):D344-D350.  
+Joseph L Durant, Burton A Leland, Douglas R Henry, and James G Nourse. 2002. Reoptimization of mdI keys for use in drug discovery. Journal of chemical information and computer sciences, 42(6):1273-1280.  
+Carl Edwards, Tuan Lai, Kevin Ros, Garrett Honke, Kyunghyun Cho, and Heng Ji. 2022a. Translation between molecules and natural language.  
+Carl Edwards, Tuan Manh Lai, Kevin Ros, Garrett Honke, Kyunghyun Cho, and Heng Ji. 2022b. Translation between molecules and natural language. In EMNLP, pages 375-413. Association for Computational Linguistics.  
+Carl Edwards, ChengXiang Zhai, and Heng Ji. 2021. Text2mol: Cross-modal molecule retrieval with natural language queries. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 595-607.  
+Steven Y Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy. 2021. A survey of data augmentation approaches for nlp. arXiv preprint arXiv:2105.03075.  
+Alex Hernández-García and Peter König. 2018. Data augmentation instead of explicit regularization. arXiv preprint arXiv:1806.03852.  
+Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li, Benjamin A Shoemaker, Paul A Thiessen, Bo Yu, et al. 2023. Pubchem 2023 update. Nucleic acids research, 51(D1):D1373-D1380.
+
+Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, volume 10, pages 707-710. Soviet Union.  
+Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. *Ai Open*, 3:71-90.  
+Jiatong Li, Yunqing Liu, Wenqi Fan, Xiao-Yong Wei, Hui Liu, Jiliang Tang, and Qing Li. 2023a. Empowering molecule discovery for molecule-caption translation with large language models: A chatgpt perspective.  
+Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Luke Zettlemoyer, Omer Levy, Jason Weston, and Mike Lewis. 2023b. Self-alignment with instruction back-translation. arXiv preprint arXiv:2308.06259.  
+Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81.  
+Yixin Liu and Pengfei Liu. 2021. SimCLS: A simple framework for contrastive learning of abstractive summarization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 1065-1072, Online. Association for Computational Linguistics.  
+Yixin Liu, Pengfei Liu, Dragomir Radev, and Graham Neubig. 2022. BRIO: Bringing order to abstractive summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2890-2903, Dublin, Ireland. Association for Computational Linguistics.  
+Zequn Liu, Wei Zhang, Yingce Xia, Lijun Wu, Shufang Xie, Tao Qin, Ming Zhang, and Tie-Yan Liu. 2023. Molxpt: Wrapping molecules with text for generative pre-training. arXiv preprint arXiv:2305.10688.  
+Zhiyuan Liu, Sihang Li, Yanchen Luo, Hao Fei, Yixin Cao, Kenji Kawaguchi, Xiang Wang, and Tat-Seng Chua. 2024. Molca: Molecular graph-language modeling with cross-modal projector and uni-modal adapter.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.  
+Qizhi Pei, Wei Zhang, Jinhua Zhu, Kehan Wu, Kaiyuan Gao, Lijun Wu, Yingce Xia, and Rui Yan. 2024. Biot5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations.
+
+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.  
+Marc'Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. 2016. Sequence level training with recurrent neural networks.  
+David Rogers and Mathew Hahn. 2010. Extendedconnectivity fingerprints. Journal of chemical information and modeling, 50(5):742-754.  
+Nadine Schneider, Roger A Sayle, and Gregory A Landrum. 2015. Get your atoms in order an opensource implementation of a novel and robust molecular canonicalization algorithm. Journal of chemical information and modeling, 55(10):2111-2120.  
+Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826.  
+David S Wishart, Yannick D Feunang, An C Guo, Elvis J Lo, Ana Marcu, Jason R Grant, Tanvir Sajed, Daniel Johnson, Carin Li, Zinat Sayeeda, et al. 2018. Drugbank 5.0: a major update to the drugbank database for 2018. Nucleic acids research, 46(D1):D1074-D1082.  
+Xing Wu, Shangwen Lv, Liangjun Zang, Jizhong Han, and Songlin Hu. 2019. Conditional bert contextual augmentation. In Computational Science-ICCS 2019: 19th International Conference, Faro, Portugal, June 12–14, 2019, Proceedings, Part IV 19, pages 84–95. Springer.  
+Kang Min Yoo, Dongju Park, Jaewook Kang, Sang-Woo Lee, and Woomyeong Park. 2021. Gpt3mix: Leveraging large-scale language models for text augmentation. arXiv preprint arXiv:2104.08826.  
+Zheni Zeng, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2022. A deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals. Nature communications, 13(1):862.  
+Qiang Zhang, Keyang Ding, Tianwen Lyv, Xinda Wang, Qingyu Yin, Yiwen Zhang, Jing Yu, Yuhao Wang, Xiaotong Li, Zhuoyi Xiang, et al. 2024. Scientific large language models: A survey on biological & chemical domains. arXiv preprint arXiv:2401.14656.  
+Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28.  
+Yao Zhao, Misha Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, and Peter J. Liu. 2022. Calibrating sequence likelihood improves conditional language generation.
+
+Jing Zhou, Yanan Zheng, Jie Tang, Jian Li, and Zhilin Yang. 2021. Flipda: Effective and robust data augmentation for few-shot learning. arXiv preprint arXiv:2108.06332.
+
+# A Datasets Information
+
+SAPubChem-41 The SAPubChem-41 datasets comprises 33k real data sourced from CHEBI-20 and 8k high-quality pseudo labelled data sourced from self-augmentation on Text+ChemT5. Representative examples of pseudo-labelled part are provided in Table 7. In addition, to further elucidate the advantages of iterative strategy on model performance, we have documented representative examples of updated captions in Table 6.
+
+Comparison with Existing Datasets In Table 8 we present a detailed comparison of our SAPubChem-41 datasets with existing limited datasets with parallel molecule-description pairs, including:
+
+- ChEBI-20 (Edwards et al., 2021): consists of  $33\mathrm{k}$  molecule-caption pairs with caption exploiting ChEBI(Degtyarenko et al., 2007)annotations and more than 20 words.  
+- PCdes (Zeng et al., 2022): consists of  $15\mathrm{k}$  substances in PubChem which have names, SMILES and corresponding paragraphs of property descriptions.  
+- DrugBank-23(Chen et al., 2023): consists of 23k compounds from DrugBank (Wishart et al., 2018) with corresponding description's length longer than 20 words.
+
+# B Pre-trained Model Information
+
+Table 9 presents the information of three different pre-trained models utilized in our experiments.
+
+# C Experiments
+
+# C.1 Performance Comparison
+
+Here we present the complete results of performance comparison for both molecular captioning task and molecule generation task in Table 10&Table 11.
+
+# C.2 Ablation Study
+
+Here we present the complete results of both molecular captioning task and molecule generation task using different sampling strategy in Table 12 & Table 13.
+
+# C.3 Additional Experiments
+
+To further substantiate the efficacy of our framework and the exceptional quality of the SAPubChem-41 dataset, we have conducted a series of experiments. These were performed with the goal of optimizing three distinct pre-trained models, utilizing both our augmented dataset, SAPubChem-41, and other prevalent public datasets which include ChEBI-20 and PCdes. The findings from these experiments are reported in Table 14 & Table 15. It can be noted that across both task categories, models that were optimized using our self-augmented dataset, SAPubChem-41, generally exhibited superior performance in comparison to those optimized with other existing datasets.
+
+<table><tr><td>SMILES</td><td>[C@H]1(O[C@H][(C@@H](O)C)C(O)O)[C@@H](O)[C@@H](O)[C@@H](O)[C@@H](C)O1</td></tr><tr><td>previous caption</td><td>The molecule is a glycosylglucose consisting of D-galactopyranose and D-glucopyranose residues joined in sequence by a (1-&gt;4) glycosidic bond. It derives from a D-galactopyranose and a D-glucopyranose.(0.6611)</td></tr><tr><td>updated caption</td><td>The molecule is a deoxygalactose that is D-galactopyranose in which the hydroxy group at position 3 has been replaced by a methyl group. It is a deoxygalactose and a methyl glycoside. It derives from a D-galactopyranose.(0.6817)</td></tr><tr><td>SMILES</td><td>CC([C@@H](C)O)(C)OC-</td></tr><tr><td>previous caption</td><td>The molecule is an ether in which the stereocentres at positions 2 and 3 both have S-configuration. It is an ether and a secondary alcohol.(0.6610)</td></tr><tr><td>updated caption</td><td>The molecule is an ether in which the stereocentres at positions 2 and 3 both have S-configuration. It is an ether and a secondary alcohol. It derives from a hydride of an oxepane.(0.6704)</td></tr><tr><td>SMILES</td><td>CI</td></tr><tr><td>previous caption</td><td>The molecule is an iodoalkane that is ethane in which one of the hydrogens is substituted by iodine. It has a role as a human metabolite. It is an iodoalkane and a member of iodos. It derives from a hydride of an ethane.(0.7479)</td></tr><tr><td>updated caption</td><td>The molecule is an iodoalkane that is ethane in which one of the hydrogens is substituted by iodine. It has a role as a metabolite. It is an iodoalkane and a volatile organic compound. It derives from a hydride of an ethane.(0.7671)</td></tr><tr><td>SMILES</td><td>N([N+])=C</td></tr><tr><td>previous caption</td><td>The molecule is an organic cation resulting from the protonation of the nitrogen of nitric acid. It is a conjugate acid of a nitric acid.(0.7479)</td></tr><tr><td>updated caption</td><td>The molecule is a hydracid and a one-carbon compound. It is a conjugate acid of a hydrazine. It derives from a hydride of a hydrazine.(0.7599)</td></tr><tr><td>SMILES</td><td>[C@H]1(C)[C@@](CO)(O)[C@@H-(O[C@@H](C)O)[C@H](OC(C)C)O1</td></tr><tr><td>previous caption</td><td>The molecule is a spiro-epoxide resulting from the formal epoxidation of the hydroxy group at position 2 of D-fructofuranose. It is a spiro-epoxide and a glycoside. It derives from a D-fructofuranose.(0.6014)</td></tr><tr><td>updated caption</td><td>The molecule is a deoxygalactose that is alpha-D-galactopyranose in which the hydroxy group at position 2 has been replaced by a methyl group. It is a deoxygalactose and a tertiary alcohol.(0.6839)</td></tr></table>
+
+Table 6: Five examples of the comparison of previous caption when adopting our iterative strategy. Corresponding Text2Mol Score of each molecule-caption pairs is present in the bracket.
+
+<table><tr><td>SMILES</td><td>Caption</td></tr><tr><td>C@H](C[C@H](C(/C=C/C)=O)C)
+([C@@H]([C@H](/C=C/C(=O)O)C)O[C@H]1[C@H]
+(O)[C@@H]([NH+])C[C@@H](C)O1)C</td><td>The molecule is an organic cation that is the conjugate acid of 1D-myo-inositol, obtained by protonation of the tertiary amino group; major species at pH 7.3. It is an ammonium ion derivative and an organic cation. It is a conjugate acid of a 1D-myo-inositol.</td></tr><tr><td>[O-][Mn]</td><td>The molecule is a monovalent inorganic anion obtained by deprotonation of manganese. It is a man-ganese oxoanion and a monovalent inorganic anion. It is a conjugate base of a manganese.</td></tr><tr><td>C[C@@H](C(OC)OC)C=C</td><td>The molecule is an ether in which the stereocentres at positions 2 and 3 both have S-configuration. It is an ether and an alicyclic compound.</td></tr><tr><td>O(C([C@H](C)N)=O)[C@H]([C@H](O)C)[C@H](C)O</td><td>The molecule is an amino cyclitol that is scyllo-inositol in which the hydroxy group at position 2 has been replaced by an amino group. It has a role as a bacterial metabolite. It is an amino cyclitol and a primary amino compound. It derives from a scyllo-inositol.</td></tr><tr><td>C(C)(C)(C)Br</td><td>The molecule is a bromoalkane that is ethane substituted by a bromo group at position 2. It has a role as a metabolite. It derives from a hydride of an ethane.</td></tr><tr><td>C[Se]C</td><td>The molecule is an organoselenium compound that is selenium substituted by a methyl group at position 2. It has a role as a metabolite. It derives from a selenium.</td></tr></table>
+
+Table 7: Five examples of SAPubChem-41  
+
+<table><tr><td>Info</td><td>SAPubChem-41</td><td>ChEBI-20</td><td>PCdes</td><td>DrugBank-23</td></tr><tr><td>Train</td><td>34,407</td><td>26,407</td><td>10,500</td><td>17,109</td></tr><tr><td>Validation</td><td>3,301</td><td>3,301</td><td>1,500</td><td>3,667</td></tr><tr><td>Test</td><td>3,300</td><td>3,300</td><td>3,000</td><td>3,666</td></tr><tr><td>LSMILES</td><td>69.96</td><td>81.56</td><td>56.47</td><td>54.11</td></tr><tr><td>LDescription</td><td>41.59</td><td>52.88</td><td>72.47</td><td>65.04</td></tr><tr><td>Data source</td><td colspan="3">PubChem</td><td>DrugBank</td></tr></table>
+
+Table 8: Details about the existing datasets and ours (SAPubChem-41).  $L_{\text{SMILES}}$  denotes the average length of SMILES while  $L_{\text{Description}}$  denotes the average word count per description.  
+
+<table><tr><td>Model</td><td>Suffix</td><td>Parameters (M)</td></tr><tr><td>Text+Chem T5</td><td>base</td><td>220</td></tr><tr><td>Ada-T5</td><td>-</td><td>220</td></tr><tr><td>MolT5</td><td>base</td><td>220</td></tr></table>
+
+Table 9: Model Size
+
+<table><tr><td>model</td><td>dataset</td><td>BLEU-2</td><td>BLEU-4</td><td>Rouge-1</td><td>Rouge-2</td><td>Rouge-L</td><td>Meteor</td></tr><tr><td rowspan="2">Text+Chem T5</td><td>ChEBI-20</td><td>0.580</td><td>0.490</td><td>0.647</td><td>0.498</td><td>0.586</td><td>0.604</td></tr><tr><td>SAPPubChem-41</td><td>0.584</td><td>0.496</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.614</td></tr><tr><td rowspan="2">Ada-T5</td><td>ChEBI-20</td><td>0.424</td><td>0.316</td><td>0.543</td><td>0.369</td><td>0.483</td><td>0.464</td></tr><tr><td>SAPPubChem-41</td><td>0.444</td><td>0.350</td><td>0.575</td><td>0.410</td><td>0.514</td><td>0.493</td></tr><tr><td rowspan="2">MolT5</td><td>ChEBI-20</td><td>0.540</td><td>0.457</td><td>0.634</td><td>0.485</td><td>0.578</td><td>0.569</td></tr><tr><td>SAPPubChem-41</td><td>0.551</td><td>0.464</td><td>0.638</td><td>0.484</td><td>0.585</td><td>0.572</td></tr></table>
+
+Table 10: The complete results of different models for molecular captioning on ChEBI-20, SAPubChem-41 datasets.  
+
+<table><tr><td>size</td><td>dataset</td><td>BL</td><td>Acc</td><td>Lev</td><td>MACCS</td><td>RDK</td><td>Morgan</td><td>Val</td></tr><tr><td rowspan="2">Text+Chem T5</td><td>ChEBI-20</td><td>0.750</td><td>0.212</td><td>27.39</td><td>0.874</td><td>0.767</td><td>0.697</td><td>0.792</td></tr><tr><td>SAPubChem-41</td><td>0.827</td><td>0.243</td><td>18.59</td><td>0.887</td><td>0.792</td><td>0.726</td><td>0.937</td></tr><tr><td rowspan="2">Ada-T5</td><td>ChEBI-20</td><td>0.699</td><td>0.182</td><td>27.48</td><td>0.869</td><td>0.753</td><td>0.672</td><td>0.886</td></tr><tr><td>SAPubChem-41</td><td>0.714</td><td>0.213</td><td>26.52</td><td>0.879</td><td>0.772</td><td>0.699</td><td>0.854</td></tr><tr><td rowspan="2">MolT5</td><td>ChEBI-20</td><td>0.769</td><td>0.081</td><td>24.49</td><td>0.721</td><td>0.588</td><td>0.529</td><td>0.772</td></tr><tr><td>SAPubChem-41</td><td>0.775</td><td>0.093</td><td>33.16</td><td>0.814</td><td>0.668</td><td>0.582</td><td>0.788</td></tr></table>
+
+Table 11: The complete results of different models for molecular generation on ChEBI-20, SAPubChem-41 datasets.  
+
+<table><tr><td>size</td><td>dataset</td><td>BLEU-2</td><td>BLEU-4</td><td>Rouge-1</td><td>Rouge-2</td><td>Rouge-L</td><td>Meteor</td></tr><tr><td rowspan="2">2k</td><td>random</td><td>0.534</td><td>0.447</td><td>0.641</td><td>0.489</td><td>0.579</td><td>0.575</td></tr><tr><td>refined</td><td>0.574</td><td>0.486</td><td>0.651</td><td>0.502</td><td>0.590</td><td>0.601</td></tr><tr><td rowspan="2">4k</td><td>random</td><td>0.537</td><td>0.451</td><td>0.641</td><td>0.489</td><td>0.579</td><td>0.577</td></tr><tr><td>refined</td><td>0.580</td><td>0.486</td><td>0.653</td><td>0.503</td><td>0.592</td><td>0.602</td></tr><tr><td rowspan="2">6k</td><td>random</td><td>0.532</td><td>0.446</td><td>0.640</td><td>0.489</td><td>0.577</td><td>0.573</td></tr><tr><td>refined</td><td>0.583</td><td>0.490</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.611</td></tr><tr><td rowspan="2">8k</td><td>random</td><td>0.527</td><td>0.443</td><td>0.640</td><td>0.489</td><td>0.579</td><td>0.571</td></tr><tr><td>refined</td><td>0.584</td><td>0.496</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.614</td></tr><tr><td rowspan="2">10k</td><td>random</td><td>0.526</td><td>0.442</td><td>0.639</td><td>0.489</td><td>0.579</td><td>0.571</td></tr><tr><td>refined</td><td>0.583</td><td>0.491</td><td>0.652</td><td>0.503</td><td>0.591</td><td>0.610</td></tr></table>
+
+Table 12: Full results of molecule captioning task using different sampling strategy.  
+
+<table><tr><td>size</td><td>dataset</td><td>BLEU</td><td>Accuracy</td><td>Levenshtein</td><td>MACCS FTS</td><td>RDK FTS</td><td>Morgan FTS</td><td>Validity</td></tr><tr><td rowspan="2">2k</td><td>random</td><td>0.715</td><td>0.215</td><td>23.82</td><td>0.873</td><td>0.766</td><td>0.703</td><td>0.868</td></tr><tr><td>refined</td><td>0.821</td><td>0.236</td><td>19.31</td><td>0.885</td><td>0.788</td><td>0.721</td><td>0.931</td></tr><tr><td rowspan="2">4k</td><td>random</td><td>0.720</td><td>0.219</td><td>23.79</td><td>0.876</td><td>0.768</td><td>0.706</td><td>0.862</td></tr><tr><td>refined</td><td>0.832</td><td>0.237</td><td>18.69</td><td>0.885</td><td>0.788</td><td>0.722</td><td>0.938</td></tr><tr><td rowspan="2">6k</td><td>random</td><td>0.709</td><td>0.220</td><td>24.38</td><td>0.876</td><td>0.767</td><td>0.706</td><td>0.858</td></tr><tr><td>refined</td><td>0.828</td><td>0.240</td><td>19.28</td><td>0.886</td><td>0.788</td><td>0.724</td><td>0.934</td></tr><tr><td rowspan="2">8k</td><td>random</td><td>0.715</td><td>0.221</td><td>24.27</td><td>0.875</td><td>0.763</td><td>0.701</td><td>0.855</td></tr><tr><td>refined</td><td>0.827</td><td>0.243</td><td>18.59</td><td>0.887</td><td>0.792</td><td>0.726</td><td>0.937</td></tr><tr><td rowspan="2">10k</td><td>random</td><td>0.715</td><td>0.207</td><td>24.37</td><td>0.870</td><td>0.759</td><td>0.696</td><td>0.864</td></tr><tr><td>refined</td><td>0.830</td><td>0.238</td><td>19.03</td><td>0.883</td><td>0.786</td><td>0.721</td><td>0.937</td></tr></table>
+
+Table 13: Full results of molecule generation task using different sampling strategy.
+
+<table><tr><td>model</td><td>dataset</td><td>BLEU-2</td><td>BLEU-4</td><td>Rouge-1</td><td>Rouge-2</td><td>Rouge-L</td><td>Meteor</td></tr><tr><td rowspan="3">Text+Chem T5</td><td>ChEBI-20</td><td>0.580</td><td>0.490</td><td>0.647</td><td>0.498</td><td>0.586</td><td>0.604</td></tr><tr><td>PCdes</td><td>0.352</td><td>0.266</td><td>0.439</td><td>0.274</td><td>0.373</td><td>0.382</td></tr><tr><td>SAPubChem-41</td><td>0.584</td><td>0.496</td><td>0.653</td><td>0.504</td><td>0.593</td><td>0.614</td></tr><tr><td rowspan="3">Ada-T5</td><td>ChEBI-20</td><td>0.424</td><td>0.316</td><td>0.543</td><td>0.369</td><td>0.483</td><td>0.464</td></tr><tr><td>PCdes</td><td>0.289</td><td>0.188</td><td>0.428</td><td>0.243</td><td>0.367</td><td>0.324</td></tr><tr><td>SAPubChem-41</td><td>0.444</td><td>0.350</td><td>0.575</td><td>0.410</td><td>0.514</td><td>0.493</td></tr><tr><td rowspan="3">MolT5</td><td>ChEBI-20</td><td>0.540</td><td>0.457</td><td>0.634</td><td>0.485</td><td>0.578</td><td>0.569</td></tr><tr><td>PCdes</td><td>0.165</td><td>0.078</td><td>0.290</td><td>0.118</td><td>0.233</td><td>0.204</td></tr><tr><td>SAPubChem-41</td><td>0.551</td><td>0.464</td><td>0.638</td><td>0.484</td><td>0.585</td><td>0.572</td></tr></table>
+
+Table 14: The complete results of different models for molecular captioning on SAPubChem-41 and other existing datasets. The best scores are in bold.  
+
+<table><tr><td>size</td><td>dataset</td><td>BL</td><td>Acc</td><td>Lev</td><td>MACCS</td><td>RDK</td><td>Morgan</td><td>Val</td></tr><tr><td rowspan="3">Text+Chem T5</td><td>ChEBI-20</td><td>0.750</td><td>0.212</td><td>27.39</td><td>0.874</td><td>0.767</td><td>0.697</td><td>0.792</td></tr><tr><td>PCdes</td><td>0.614</td><td>0.105</td><td>30.43</td><td>0.697</td><td>0.544</td><td>0.459</td><td>0.849</td></tr><tr><td>SAPPubChem-41</td><td>0.827</td><td>0.243</td><td>18.59</td><td>0.887</td><td>0.792</td><td>0.726</td><td>0.937</td></tr><tr><td rowspan="3">Ada-T5</td><td>ChEBI-20</td><td>0.699</td><td>0.182</td><td>27.48</td><td>0.869</td><td>0.753</td><td>0.672</td><td>0.886</td></tr><tr><td>PCdes</td><td>0.579</td><td>0.089</td><td>38.99</td><td>0.778</td><td>0.600</td><td>0.492</td><td>0.947</td></tr><tr><td>SAPPubChem-41</td><td>0.714</td><td>0.213</td><td>26.52</td><td>0.879</td><td>0.772</td><td>0.699</td><td>0.854</td></tr><tr><td rowspan="3">MolT5</td><td>ChEBI-20</td><td>0.769</td><td>0.081</td><td>24.49</td><td>0.721</td><td>0.588</td><td>0.529</td><td>0.772</td></tr><tr><td>PCdes</td><td>0.476</td><td>0.009</td><td>53.27</td><td>0.635</td><td>0.432</td><td>0.330</td><td>0.711</td></tr><tr><td>SAPPubChem-41</td><td>0.775</td><td>0.093</td><td>33.16</td><td>0.814</td><td>0.668</td><td>0.582</td><td>0.788</td></tr></table>
+
+Table 15: The complete results of different models for molecular generation on SAPubChem-41 and other existing datasets. The best scores are in bold.

2024/Enhancing Cross Text-Molecule Learning by Self-Augmentation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:863aa4befac6ec26d5ae08520512e65f23633796ea5fd201968edd1a23c6b720
+size 1397755

2024/Enhancing Distractor Generation for Multiple-Choice Questions with Retrieval Augmented Pretraining and Knowledge Graph Integration/e2e79252-24e8-4303-bd7f-47125cbcbdea_content_list.json

@@ -0,0 +1,1759 @@
+[
+    {
+        "type": "text",
+        "text": "Enhancing Distractor Generation for Multiple-Choice Questions with Retrieval Augmented Pretraining and Knowledge Graph Integration",
+        "text_level": 1,
+        "bbox": [
+            139,
+            80,
+            855,
+            122
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Han-Cheng Yu, Yu-An Shih, Kin-Man Law, Kai-Yu Hsieh, Yu-Chen Cheng, Hsin-Chih Ho, Zih-An Lin, Wen-Chuan Hsu, Yao-Chung Fan*",
+        "bbox": [
+            231,
+            130,
+            769,
+            180
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science and Engineering, National Chung Hsing University, Taiwan",
+        "bbox": [
+            292,
+            181,
+            707,
+            214
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "yfan@nchu.edu.tw",
+        "bbox": [
+            401,
+            216,
+            600,
+            231
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we tackle the task of distractor generation (DG) for multiple-choice questions. Our study introduces two key designs. First, we propose retrieval augmented pretraining, which involves refining the language model pretraining to align it more closely with the downstream task of DG. Second, we explore the integration of knowledge graphs to enhance the performance of DG. Through experiments with benchmarking datasets, we show that our models significantly outperform the state-of-the-art results. Our best-performing model advances the F1@3 score from 14.80 to 16.47 in MCQ dataset and from 15.92 to 16.50 in Sciq dataset.",
+        "bbox": [
+            141,
+            286,
+            460,
+            500
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            518,
+            260,
+            533
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Multiple-Choice Questions (MCQs) are widely used to evaluate a learner's knowledge. However, creating them manually requires a significant amount of time and effort from educators. Well-designed MCQs help in accurately assessing learners' abilities. The most challenging part of creating MCQs is designing appropriate incorrect options, commonly called distractors. Therefore, researchers have focused on the automatic distractor generation for MCQs in recent years. In this paper, we explore two directions for improving DG methods.",
+        "bbox": [
+            112,
+            546,
+            489,
+            739
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Task-Specific Pretraining Current SOTA DG methods, such as utilizing BERT for DG (Chiang et al., 2022) or harnessing T5 for Text2Text DG (Wang et al., 2023), are all based on pre-trained language models (LMs). Existing LMs undergo task-agnostic pertaining. However, recently, the idea of task-specific pretraining has gained prominence. The aim is to refine the pretraining process to closely align with the downstream task. Therefore, our first endeavor in this paper is to investigate Task-Specific Pretraining for the DG task.",
+        "bbox": [
+            112,
+            741,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In line with this goal, we propose Retrieval Augmented Pretraining (RAP) for task-specific pretraining. The main idea involves using MCQ answers to retrieve relevant sentences/passages from a large corpus, such as Wikipedia, to create pseudo questions, and then use the generated pseudo question for task-specific pertaining.",
+        "bbox": [
+            507,
+            253,
+            884,
+            363
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1 illustrates an example of a create pseudo question. Assume that we are given a set of MCQ options, the answer kidneys, and three distractors lungs, pancreas, liver. The idea is to use the answer option kidneys to retrieve a sentence, i.e., the kidneys are two reddish-brown bean-shaped organs and replace the answer with the [Mask] token to generate a pseudo question, i.e., the [Mask] are two reddish-brown bean-shaped organs. The pseudo question and the distractors are then served as training data for RAP.",
+        "bbox": [
+            507,
+            365,
+            884,
+            541
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Knowledge Augmented Generation Recent research trends show that leveraging knowledge graphs (KGs) can enhance the performance of LM-based text generation tasks, as discussed in (Yasunaga et al., 2022) and (Zhang et al., 2022). The perspective is that LMs and KGs should complement each other. KGs go beyond text by offering structural information, representing entities as nodes and their relationships as edges. This structured knowledge enhances the ability for multi-step reasoning. In line with this motivation, we propose knowledge-augmented generation (KAG), where we build upon the candidate augmentation strategy proposed by the state-of-the-art DG method (Wang et al., 2023). The idea is to retrieve knowledge triplets from a knowledge graph to serve as auxiliary information for the text2text DG model.",
+        "bbox": [
+            507,
+            542,
+            884,
+            815
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The contributions of this paper are",
+        "bbox": [
+            527,
+            816,
+            786,
+            832
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "- Our methods achieve a remarkable improvement in state-of-the-art DG results. Our best-performing model elevates the F1@3 score from 14.80 to 16.47 in MCQ and from 15.92 to 16.50 in Sciq, showcasing the effectiveness",
+        "bbox": [
+            531,
+            838,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "11019",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2024, pages 11019-11029 August 11-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            220,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e4ae3d58e4da5cd3034667e84b5abd236321a6261fb7749064bab0be4a9d9583.jpg",
+        "image_caption": [
+            "Figure 1: Retrieved Augmented Pretraining"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            84,
+            487,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "of our approach.",
+        "bbox": [
+            149,
+            294,
+            278,
+            310
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Extensive experimental evaluation with the benchmarking datasets are conducted and the insights of incorporating task-pretraining and knowledge triplet provision are discussed. Our study unveils promising directions for further development in DG by showcasing the efficacy of knowledge augmentation and task-specific pretraining.",
+        "bbox": [
+            136,
+            319,
+            490,
+            450
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The rest of this paper is organized as follows. Section 2 reviews the works of automatic distractor generation in the literature. In Section 3 we present the proposed methods. Section 4 reports the performance evaluation and Section 5 concludes this work and discuss the future work.",
+        "bbox": [
+            112,
+            458,
+            489,
+            555
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            567,
+            270,
+            582
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Distractor Generation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            592,
+            334,
+            607
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The current DG research can be divided into two directions:",
+        "bbox": [
+            112,
+            613,
+            485,
+            644
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Generating and Ranking (GR) framework The framework incorporates a general-purpose knowledge base to effectively create a small distractor candidate set, and a feature-rich learning-to-rank model to select distractors. Specifically, the GR architecture consists of two stages.",
+        "bbox": [
+            112,
+            645,
+            487,
+            741
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "First, it generates candidate distractors, and then it ranks these candidates based on semantic rules and linguistic features to select the final distractor. In the GR framework, there are two approaches for generating distractor candidates: using a knowledge base (Ren and Zhu, 2021) or using a language model (LM) (Chiang et al., 2022). These approaches have significantly improved the quality and diversity of distractors compared to traditional rule-based methods (Liang et al., 2017, 2018), making them the state-of-the-art in DG.",
+        "bbox": [
+            112,
+            741,
+            490,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Text2Text generation architecture The Text2Text generation architecture differs from the GR architecture by formulating distractor generation as a Text2Text task. Specifically, it concatenates the question stem with the answer and inputs it into a generative language model (e.g., T5 or GPT) to train the model to generate a set of distractors. Currently, research based on the Text2Text architecture (Wang et al., 2023) represents the state-of-the-art in distractor generation.",
+        "bbox": [
+            507,
+            84,
+            885,
+            247
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Task-specific Pretraining",
+        "text_level": 1,
+        "bbox": [
+            507,
+            262,
+            754,
+            279
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In recent related work, researchers have explored the integration of task-specific priors into BERT language model pretraining to enhance performance in low-resource finetuning tasks (Wang et al., 2020). Traditional pretraining methods often prioritize generic natural language knowledge, overlooking task-specific information, which can lead to overfitting in low-resource downstream tasks. To address this issue, (Wang et al., 2020) propose integrating task-specific label embeddings into the self-attention layers during pretraining. This integration enables the model to filter out task-irrelevant information and improve task-specific knowledge during fine-tuning, resulting in reduced overfitting and improved performance on downstream tasks with limited resources.",
+        "bbox": [
+            505,
+            287,
+            885,
+            544
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Another related work, (Zhang et al., 2020), adopts a task-specific pretraining approach for text summarization tasks. The authors replace a portion of the original sentences in the text with mask tokens, allowing the model to learn to generate masked sentences. Comparing this approach to traditional masked language modeling (MLM) for pretraining, they find that closer alignment between the pretraining task and the downstream task leads to improved performance. Additionally, they experiment with pretraining on various types of datasets and observe that utilizing a dataset similar to the downstream task yields better results. For example, when pretraining on a news dataset and fine-tuning on another news dataset, the performance surpasses that of fine-tuning on a web article dataset.",
+        "bbox": [
+            507,
+            546,
+            885,
+            803
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Building upon these findings, we draw inspiration from the aforementioned studies to design our proposed approach, known as RAP (Retrieval-Augmented Pretraining), aimed at enhancing the performance of distractor generation (DG). RAP leverages task-specific priors and incorporates them into the pretraining phase to capture and uti",
+        "bbox": [
+            507,
+            806,
+            885,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "11020",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "lize relevant information specific to the distractor generation task. By aligning the pretraining task with the downstream DG task and considering task-specific information, we aim to improve the quality and effectiveness of distractor generation.",
+        "bbox": [
+            112,
+            84,
+            489,
+            165
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Knowledge Augmented Generation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            177,
+            436,
+            192
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We have conducted a survey of research that leverages Knowledge Graphs (KGs) to enhance performance in various Natural Language Processing (NLP) tasks, including Question Answering, Entity Typing/Relation Classification, Query Answering, Question Generation, and Distractor Generation (Yasunaga et al., 2022, 2021; Zhang et al., 2022; Feng et al., 2020; He et al., 2019; Ren et al., 2020; Fei et al., 2022; Chen et al., 2023; Ren and Zhu, 2021).",
+        "bbox": [
+            112,
+            198,
+            489,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Notably, a significant amount of research has focused on using Knowledge Graphs to improve question-answering systems, particularly for tasks involving multihop reasoning. These studies utilize Knowledge Graph triplets for inferencing and have demonstrated the utility of structured background knowledge in strengthening NLP tasks.",
+        "bbox": [
+            112,
+            359,
+            487,
+            470
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For example, (Yasunaga et al., 2021) proposed an end-to-end question-answering model that combines language models and Knowledge Graphs to handle reasoning tasks. (Zhang et al., 2022) introduced a novel model that incorporates joint information exchange between language models and Knowledge Graphs, achieving superior performance across multiple domains.",
+        "bbox": [
+            112,
+            472,
+            489,
+            600
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "While Knowledge Graph integration has shown promising results in Question-Answering tasks, its application to Distractor Generation remains relatively limited (Ren and Zhu, 2021).",
+        "bbox": [
+            112,
+            601,
+            489,
+            665
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Therefore, our goal is to leverage KGs as auxiliary tools to enhance distractor generation, improving the relevance of generated options to the question. This approach not only creates more challenging options for multiple-choice questions but also enhances the model's efficiency in utilizing Knowledge Graphs, leading to more effective Knowledge Augmented Generation.",
+        "bbox": [
+            112,
+            665,
+            489,
+            794
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            807,
+            263,
+            824
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Retrieval Augmented Pretraining",
+        "text_level": 1,
+        "bbox": [
+            112,
+            833,
+            421,
+            848
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our approach involves generating pseudo questions to continuously train a pre-trained language model (e.g., T5 or BART) before fine-tuning it for downstream Distractor Generation (DG) tasks.",
+        "bbox": [
+            112,
+            854,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Here's how our idea works: Given an MCQ training instance consisting of a question stem  $(q)$ , an answer option  $(a)$ , and a set of distractors  $(D)$ , we utilize  $a$  to retrieve a single sentence or a short passage from a corpus (such as Wikipedia). Subsequently, we mask out  $a$  from the retrieved sentence/passage, replacing the answer position with a mask token ([Mask]). This process results in a modified sentence/passage known as a pseudo question  $(\\tilde{q}_{\\otimes [\\mathsf{Mask}]})$ .",
+        "bbox": [
+            507,
+            84,
+            884,
+            244
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With  $\\tilde{q}_{\\otimes[\\mathsf{Mask}]}$  our goal is to pretrain a LM as follows.",
+        "bbox": [
+            507,
+            246,
+            882,
+            275
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL (\\theta) = - \\sum_ {i = 1} ^ {| D |} t _ {i} \\log p \\left(\\hat {t} _ {i} \\mid \\hat {t} _ {<   i}, \\tilde {q} _ {\\otimes [ \\text {M a s k} ]}, a; \\theta\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            534,
+            285,
+            853,
+            330
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "-  $t_i$ : the token sequence given by the concatenation of the ground truth distractors  $d_1||d_2||d_3$ .",
+        "bbox": [
+            531,
+            340,
+            884,
+            372
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Alternatives for RAP Training Setting There are two alternatives for RAP training setting.",
+        "bbox": [
+            507,
+            382,
+            882,
+            414
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Data Augmentation RAP: We can treat RAP as a data augmentation mechanism by using the same dataset for RAP and DG.",
+            "- Cross-Domain RAP: We can use two different distractor generation datasets (SciQ and MCQ datasets in this study). One dataset is used for RAP and the other for DG fine-tuning."
+        ],
+        "bbox": [
+            531,
+            424,
+            880,
+            546
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "There are different impacts and implications for RAP training adoption. We provide experimental study and discussion for this part in Subsection 4.6.",
+        "bbox": [
+            507,
+            557,
+            882,
+            605
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Boosting RAP with ground-truth distractor The RAP idea can be extended to using groundtruth distractors to generate pseudo questions. That is, we can use ground-truth distractors to retrieve a sentence/passage and set the rest of options (i.e.,  $a$  and other distractors). That is,",
+        "bbox": [
+            507,
+            614,
+            882,
+            709
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL (\\theta) = - \\sum_ {i = 1} ^ {| D |} t _ {i} \\log p \\left(\\hat {t} _ {i} \\mid \\hat {t} _ {<   i}, \\tilde {q} _ {\\otimes [ \\text {M a s k} ]}, d _ {1}; \\theta\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            531,
+            721,
+            858,
+            765
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "-  $t_i$ : the token sequence given by the concatenation of the other options  $a||d_2||d_3$ .",
+        "bbox": [
+            531,
+            776,
+            884,
+            808
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Knowledge Augmented Generation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            818,
+            831,
+            833
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We extend the idea of the candidate augmentation strategy proposed by (Wang et al., 2023). We explore incorporating knowledge triplets for DG. Specifically, as shown in Figure 2, our KAG proceeds by (1) a knowledge triplet retrieval stage",
+        "bbox": [
+            507,
+            838,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "11021",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(Subsection 3.2.1) and (2) a triplet re-ranker stage (Subsection 3.2.2). The retriever stage is to retrieve triplets with respect to  $q$  and  $a$ , and the re-ranker stage aims to rank the triplets by their estimated relevancy with respect to  $q$  and  $a$",
+        "bbox": [
+            112,
+            84,
+            487,
+            165
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Retrieving Triplet from KG",
+        "text_level": 1,
+        "bbox": [
+            112,
+            173,
+            391,
+            187
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our process for obtaining knowledge triplets is as follows. First, for a given  $q$  and answer  $a$ , we employ an language model trained for distractor generation (e.g., (Chiang et al., 2022)) to generate a set of candidate distractors  $\\{\\hat{d}_1, \\dots, \\hat{d}_k\\}$ . In addition, we conduct keyword extraction over  $q$  and  $a$  to extract keywords from them. We then union the two sets of keywords (denoted by  $W$ ).",
+        "bbox": [
+            112,
+            192,
+            487,
+            319
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Figure 3, with  $W$  and a given knowledge graph  $G(V,E)$ , the triplet set is given by",
+        "bbox": [
+            112,
+            321,
+            485,
+            369
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nK = \\left\\{\\left(u, e _ {u, v}, v\\right) \\mid u, v \\in W \\text {a n d} e _ {u, v} \\in E \\right\\}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            134,
+            378,
+            465,
+            395
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Triplet Ranker",
+        "text_level": 1,
+        "bbox": [
+            112,
+            412,
+            292,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The input to triplet ranker is a triplet set  $T$  from KG Retrieval stage and let  $q||a$  be the concatenation of a question  $q$  and an answer  $a$ . We experiment with the following two alternatives for ranking triplets.",
+        "bbox": [
+            112,
+            431,
+            487,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "1. Unsupervised Sentence Embedding Distance: The ranking process begins by encoding  $q \\mid a$  and each knowledge triplet  $\\tau_{i}$  into their respective vector representations. Next, we compute a relevancy score for each triplet by comparing its vector representation with that of  $q \\mid a$ . The higher the relevancy score, the more likely the triplet is deemed relevant to  $q \\mid a$ . Specifically,",
+        "bbox": [
+            129,
+            504,
+            489,
+            650
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {s c o r e s} = \\left[ \\operatorname {C o s - S i m i l a r i t y} \\left(\\tau_ {i}, q \\| a\\right), \\forall \\tau_ {i} \\in K \\right]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            152,
+            658,
+            482,
+            676
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We select top- $k$  triplets for the later knowledge augmented generation.",
+        "bbox": [
+            149,
+            684,
+            485,
+            715
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2. Supervised Triplet Classification: As an alternative, we propose a binary classification formulation to estimate the relevancy of a knowledge triplet  $\\tau_{i}$  with respect to  $q||a$ . During the training phase, we label a triplet as relevant if it contains the answer or ground-truth distractors, and as irrelevant otherwise. Specifically, we concatenate  $q||a$  and  $\\tau_{i}$ , separated by a special token [SEP] as input and the relevancy as output. During the inference phase, we utilize the LM to re-sort the top- $k$  triplet set based on the confidence scores.",
+        "bbox": [
+            127,
+            726,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.3 KAG Training",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            684,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For a given training instance  $(q, a, D, K)$ , our goal is to train a generation model conditioned on  $q$ ,  $a$ , and  $K$  by minimizing the negative log-likelihood of the correct token  $t_i$  of  $D$  given the preceding tokens and the conditions.",
+        "bbox": [
+            507,
+            103,
+            882,
+            183
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "-  $K$ : the set of knowledge triplet  $\\{\\tau_1, \\dots, \\tau_k\\}$  retrieved from a KG.",
+        "bbox": [
+            529,
+            193,
+            880,
+            225
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The idea is to retrieve a set of knowledge triplet  $\\{\\tau_1,\\dots,\\tau_k\\}$  from a KG and concatenate the knowledge triplet with the original input text as an augmented text input for generation. Specifically, the loss function is",
+        "bbox": [
+            507,
+            237,
+            882,
+            316
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL (\\theta) = - \\sum_ {i = 1} ^ {| D |} t _ {i} \\log p (\\hat {t} _ {i} | \\hat {t} _ {<   i}, q, a, \\{\\tau_ {i} \\}; \\theta)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            539,
+            324,
+            850,
+            369
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The observation behind the knowledge triplet augmentation strategy is to inject more information for generation through the knowledge graph in hope to boost the DG performance.",
+        "bbox": [
+            507,
+            378,
+            882,
+            442
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            453,
+            648,
+            469
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Dataset",
+        "text_level": 1,
+        "bbox": [
+            507,
+            479,
+            615,
+            492
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We use SciQ (Welbl et al., 2017) and MCQ dataset (the dataset released by Ren and Zhu, 2021) for performance evaluation. For the details about the two benchmark datasets and our dataset setting.",
+        "bbox": [
+            507,
+            499,
+            880,
+            564
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Sciq dataset The Sciq dataset is a multi-domain multiple-choice question dataset consisting of 13,000 questions in the fields of physics, chemistry, biology, and other natural sciences. These questions are open-ended and require reading and understanding relevant scientific knowledge, followed by reasoning and answering the questions. The questions are presented in a multiple-choice format, with each question accompanied by an answer and three incorrect options. According to the Sciq dataset configuration, the 13,000 data points are split into train/dev/test datasets with a ratio of 11,700/1,000/1,000.",
+        "bbox": [
+            507,
+            571,
+            882,
+            780
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Since the Sciq question sentences do not contain **blanks**, in the candidate generation method, we utilize generative LM, specifically ChatGPT, to generate the candidate set.",
+        "bbox": [
+            507,
+            782,
+            880,
+            846
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "MCQ dataset MCQ dataset is a cross-domain cloze-style dataset, that includes the domains of science, vocabulary, common sense, and trivia. Each data is composed of a sentence containing",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "11022",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/64b9613a4bbefc0e0dec8344885017a53ff3ee2728f9a010f03f3c321823488e.jpg",
+        "image_caption": [
+            "Figure 2: Knowledge Augmented Generation"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            80,
+            885,
+            300
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/4cda7a2c14278d15006ff6f5eef58ec449972b25397464a70be2800fb8d87ced.jpg",
+        "image_caption": [
+            "Figure 3: Retrieve Triplet from KG: we extract keyword from a given question, answer and candidate set as a entity set  $W$  and retrieve relevant triplet set from KG with keyword entities"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            340,
+            887,
+            501
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/833f821b7401ab1a75523eba67ec6d7001efc8d61b50361d2e961d8c6a41d0ce.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"4\">MCQ</td><td colspan=\"4\">Sciq</td></tr><tr><td>Train</td><td>Dev</td><td>Test</td><td>All</td><td>Train</td><td>Dev</td><td>Test</td><td>All</td></tr><tr><td># of Questions</td><td>2088</td><td>233</td><td>258</td><td>2580</td><td>11700</td><td>1000</td><td>1000</td><td>13700</td></tr></table>",
+        "bbox": [
+            117,
+            558,
+            485,
+            593
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: The statistics of the training, development and test sets of MCQ, Sciq.",
+        "bbox": [
+            112,
+            602,
+            485,
+            632
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "**blank** of cloze stem, answer, and distractors. According to the setting reported by (Ren and Q. Zhu, 2021), MCQ contains 2880 questions and is randomly divided into train/dev/test with a ratio of 8:1:1.",
+        "bbox": [
+            112,
+            692,
+            487,
+            772
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We obtain the MCQ dataset from GitHub link shared by (Ren and Q. Zhu, 2021). However, we find there is a slight difference between the numbers in the shared dataset and reported in the paper. In the shared dataset, it only contains train and test data (with 2321/258). Thus, we use this data setting in our experiments. For dev data, we use 9:1 split from train as dev data.",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            562,
+            707,
+            577
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Automatic Metric Following the SOTA DG by (Wang et al., 2023), we evaluate the DG quality by F1 score  $(\\mathrm{F}1@3)$ , precision  $(\\mathrm{P}@1, \\mathrm{P}@3)$ , and recall  $(\\mathrm{R}@1, \\mathrm{R}@3)$ .  $\\mathrm{P}@k$  represents the ratio of correctly labeled top-k generated distractors, while  $\\mathrm{R}@k$  indicates the ratio of correctly predicted labels among the ground truth.  $\\mathrm{F}1@k$  is the harmonic mean of  $\\mathrm{P}@k$  and  $\\mathrm{R}@k$ . Notably, when the label size is 3,  $\\mathrm{P}@3$  and  $\\mathrm{R}@3$  will be the same, resulting in the same  $\\mathrm{F}1@3$  score. Since both the Sciq test data and MCQ test data contain 3 distractors, we report the scores of  $\\mathrm{P}@1$  and  $\\mathrm{F}1@3$  in the experiments.",
+        "bbox": [
+            507,
+            585,
+            882,
+            791
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Human Evaluation Metric In addition to Automatic Metrics, we also consider human evaluation to assess the performance of the model on various aspects. During the evaluation process, we randomly select 14 questions and generate distractor sets using various models, which are then assessed by five human labelers. For the evaluation of the",
+        "bbox": [
+            507,
+            806,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "11023",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Distractor Generation models, we consider the following criteria for human evaluation:",
+        "bbox": [
+            112,
+            84,
+            489,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Relevance: This metric assesses the relevance of the generated distractors to the question. This means that the distractors should be related to the theme, content, or context of the question.",
+            "2. Distractiveness: This metric evaluates the level of distraction that the generated distractors bring to the correct answer. The distractors should be able to divert the answer away from the correct answer, increasing the difficulty of the question.",
+            "3. Utility: This is an overall quality assessment of the generated distractors, considering aspects such as fluency, relevance, duplication, and distractiveness. This is a comprehensive metric that takes into account all facets of performance."
+        ],
+        "bbox": [
+            127,
+            131,
+            487,
+            429
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We employ a five-point rating system, where each score interval represents different levels of relevance, ranging from \"totally irrelevant\" (1 point) to \"completely relevant\" (5 points). These evaluation metrics will help us assess the quality and performance of the distractors generated by the models in a comprehensive manner.",
+        "bbox": [
+            112,
+            445,
+            487,
+            557
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            571,
+            344,
+            586
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "See Appendix.",
+        "bbox": [
+            112,
+            594,
+            226,
+            608
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Evaluation Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            621,
+            307,
+            636
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2 presents the results of the compared methods on the two benchmarking datasets. We have the following notes for the results.",
+        "bbox": [
+            112,
+            644,
+            487,
+            690
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "RAP indeed improve performance By comparing the performance difference between Text2Text (T5) and Text2Text (RAP-T5), we can see that the introduction of RAP elevates the performance of MCQ experiments from 11.45 to 14.80 (F1@3). In the case of Sciq, it increases from 15.92 to 16.06.",
+        "bbox": [
+            112,
+            693,
+            487,
+            788
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "RAP more effective in low resource settings Furthermore, continuing the above observations, we found that the improvement brought by RAP in the Sciq experiment is not as significant as in the MCQ experiment. We attribute this result to the much smaller size of the MCQ dataset compared to Sciq. When the training data is abundant, the gains from pretraining naturally become limited. Therefore, in",
+        "bbox": [
+            112,
+            791,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "the MCQ setting with only 2088 training examples, the introduction of RAP has a noticeable impact.",
+        "bbox": [
+            507,
+            84,
+            880,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "KG also boosts the performance By comparing T5 candidate augmentation with KAG (T5), we observe that the introduction of KG indeed brings a significant improvement. In Sciq, we observe an increase in F1@3 from 12.83 to 15.38. In MCQ, it increases from 14.80 to 16.47.",
+        "bbox": [
+            507,
+            120,
+            880,
+            216
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our methods outperform ChatGPT. We also compare the performance of ChatGPT (in a zero-shot manner) on both datasets. From the experimental results, it is evident that our best-performing models consistently outperform ChatGPT. In the comparison with Sciq, KAG (RAP-T5) achieves an F1@3 score of  $15+$  compared to ChatGPT's 10.61. An interesting observation is that while ChatGPT still lags behind the KAG+RAP-based approach in the MCQ comparison, the performance gap is smaller. We attribute this to the fact that some questions in Sciq require common-sense reasoning in addition to knowledge-based answering. Since ChatGPT generates responses in a zero-shot learning manner during testing, it does not fully capture the characteristics of the Sciq dataset. In the case of the MCQ dataset, due to the limited training data, the differences between various methods and ChatGPT are not as substantial as in the Sciq comparison.",
+        "bbox": [
+            507,
+            222,
+            882,
+            543
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Combining KAG with RAP did not yield additive effects We also observed that combining KAG with RAP did not yield additive effects. Taking the comparison of the T5 model on the Sciq dataset as an example, we found that the performance did not improve when using both KAG and RAP together (KAG(RAP-T5): 15.15) compared to using RAP alone (Text2Text(RAP-T5): 16.06) or KAG alone (KAG(T5): 15.38). Similar observations were made in other settings. We speculate that the additional noise introduced by KAG during the incorporation of knowledge triplets may be the reason behind this. Although we designed a reranker to select triplets, the results were not consistently stable. To validate this hypothesis, we attempted a variation of KAG (RAP-T5) (denote by with-only-answer triplet) where, instead of using a reranker, we selected triplets containing the answer to input into the Text2Text generator. We observed significant improvement with this approach in the Sciq dataset, but no improvement in the MCQ dataset. This suggests that there is considerable room for improvement in the design of the triplet ranker. Se",
+        "bbox": [
+            507,
+            549,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "11024",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/470fc20718f7ab42515aa084e7476be045244711b3d6c98c3349b81c5e8e5126.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Method</td><td>P@1</td><td>R@1</td><td>F1@3</td><td>MRR</td><td>NDCG@3</td></tr><tr><td rowspan=\"17\">MCQ</td><td>(Ren and Zhu, 2021)</td><td>10.58</td><td>-</td><td>9.19</td><td>17.51</td><td>-</td></tr><tr><td>(Chiang et al., 2022)</td><td>10.81</td><td>3.60</td><td>7.72</td><td>18.15</td><td>15.39</td></tr><tr><td>(Wang et al., 2023) (BART)</td><td>14.28</td><td>4.76</td><td>11.45</td><td>21.49</td><td>23.70</td></tr><tr><td>Text2text (RAP-BART)</td><td>18.14</td><td>6.04</td><td>12.35</td><td>24.06</td><td>25.78</td></tr><tr><td>(Wang et al., 2023) (BART w/ c.a.)</td><td>19.69</td><td>6.56</td><td>13.12</td><td>25.03</td><td>26.26</td></tr><tr><td>KAG (BART)</td><td>20.07</td><td>6.69</td><td>14.41</td><td>27.02</td><td>29.37</td></tr><tr><td>KAG (RAP-BART)</td><td>16.60</td><td>5.75</td><td>12.22</td><td>22.65</td><td>24.50</td></tr><tr><td>KAG (BART) (with only answer triplet)</td><td>14.28</td><td>4.76</td><td>12.74</td><td>21.55</td><td>24.13</td></tr><tr><td>KAG (RAP-BART)(with only answer triplet)</td><td>13.12</td><td>4.37</td><td>11.84</td><td>20.01</td><td>22.35</td></tr><tr><td>(Wang et al., 2023) (T5)</td><td>18.53</td><td>6.17</td><td>11.45</td><td>23.61</td><td>25.08</td></tr><tr><td>Text2text (RAP-T5)</td><td>22.39</td><td>7.46</td><td>14.80</td><td>29.02</td><td>30.72</td></tr><tr><td>(Wang et al., 2023) (T5 w/ c.a.)</td><td>16.60</td><td>5.53</td><td>14.80</td><td>24.90</td><td>27.61</td></tr><tr><td>KAG (T5)</td><td>20.07</td><td>6.69</td><td>16.47</td><td>28.57</td><td>30.99</td></tr><tr><td>KAG (RAP-T5)</td><td>20.07</td><td>6.69</td><td>14.92</td><td>26.44</td><td>28.23</td></tr><tr><td>KAG (T5)(with only answer triplet)</td><td>22.00</td><td>7.35</td><td>15.18</td><td>28.70</td><td>30.61</td></tr><tr><td>KAG (RAP-T5)(with only answer triplet)</td><td>21.62</td><td>7.20</td><td>14.41</td><td>27.22</td><td>28.60</td></tr><tr><td>ChatGPT</td><td>18.91</td><td>6.30</td><td>13.38</td><td>25.86</td><td>27.97</td></tr><tr><td rowspan=\"15\">Sciq</td><td>(Wang et al., 2023) (BART)</td><td>19.30</td><td>6.46</td><td>14.94</td><td>26.40</td><td>28.36</td></tr><tr><td>Text2text (RAP-BART)</td><td>20.10</td><td>6.73</td><td>16.32</td><td>27.56</td><td>29.68</td></tr><tr><td>(Wang et al., 2023) (BART w/ c.a.)</td><td>21.26</td><td>7.08</td><td>14.98</td><td>27.63</td><td>29.28</td></tr><tr><td>KAG (BART)</td><td>16.24</td><td>5.41</td><td>13.99</td><td>23.62</td><td>25.93</td></tr><tr><td>KAG (RAP-BART)</td><td>19.70</td><td>6.58</td><td>15.42</td><td>26.61</td><td>28.69</td></tr><tr><td>KAG (BART) (with only answer triplet)</td><td>17.15</td><td>5.57</td><td>14.87</td><td>25.00</td><td>27.27</td></tr><tr><td>KAG (RAP-BART)(with only answer triplet)</td><td>18.35</td><td>6.15</td><td>15.52</td><td>25.81</td><td>28.22</td></tr><tr><td>(Wang et al., 2023) Text2text (T5)</td><td>24.27</td><td>8.09</td><td>15.92</td><td>30.32</td><td>31.99</td></tr><tr><td>Text2text (RAP-T5)</td><td>24.30</td><td>8.09</td><td>16.06</td><td>29.98</td><td>31.49</td></tr><tr><td>(Wang et al., 2023) (T5 w/ c.a.)</td><td>22.06</td><td>7.35</td><td>12.83</td><td>25.54</td><td>26.47</td></tr><tr><td>KAG (T5)</td><td>22.50</td><td>7.51</td><td>15.38</td><td>29.10</td><td>31.04</td></tr><tr><td>KAG (RAP-T5)</td><td>18.50</td><td>6.18</td><td>15.15</td><td>26.00</td><td>28.40</td></tr><tr><td>KAG (T5)(with only answer triplet)</td><td>23.70</td><td>7.91</td><td>16.50</td><td>30.41</td><td>32.39</td></tr><tr><td>KAG (RAP-T5)(with only answer triplet)</td><td>23.00</td><td>7.68</td><td>16.46</td><td>29.51</td><td>31.41</td></tr><tr><td>ChatGPT</td><td>15.17</td><td>5.16</td><td>10.61</td><td>19.39</td><td>20.68</td></tr></table>",
+        "bbox": [
+            196,
+            80,
+            801,
+            492
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/ad40508409c380608395e3344a549be903e18e63a7e0cdf79db69de70a82f3c7.jpg",
+        "table_caption": [
+            "Table 2: DG Results on the Compared Datasets: In this set of experiment, we mainly compare our designs with the SOTA method (Wang et al., 2023). c.a. represents candidate augmentation strategy introduced by (Wang et al., 2023)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Method</td><td># of masked Sentences/Passages</td></tr><tr><td rowspan=\"3\">MCQ-train</td><td>RAP-S</td><td>17,225</td></tr><tr><td>RAP-P</td><td>17,362</td></tr><tr><td>RAP-P w/ GTD</td><td>69,356</td></tr><tr><td rowspan=\"3\">MCQ-all</td><td>RAP-S</td><td>10,894</td></tr><tr><td>RAP-P</td><td>10,995</td></tr><tr><td>RAP-P w/ GTD</td><td>44,294</td></tr><tr><td rowspan=\"3\">Sciq-train</td><td>RAP-S</td><td>85,463</td></tr><tr><td>RAP-P</td><td>86,269</td></tr><tr><td>RAP-P w/ GTD</td><td>307,915</td></tr><tr><td rowspan=\"3\">Sciq-all</td><td>RAP-S</td><td>91,798</td></tr><tr><td>RAP-P</td><td>92,731</td></tr><tr><td>RAP-P w/ GTD</td><td>327,546</td></tr></table>",
+        "bbox": [
+            119,
+            579,
+            482,
+            736
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: The statistics of the RAP training. Note that our RAP has the following variations. First, in RAP, we can choose to retrieve a single sentence or a passage, giving rise to different RAP strategies. Thus, we denote RAP-S as retrieving sentence for pseudo question generation and RAP-P for retrieving passage. Second, we have the option to use the answer or the distractor as the starting point for retrieving the sentence/passage. We denote the method using distractor as starting point as RAP-S w/GTD or RAP-S/RAP-P (as using answer as starting point)",
+        "bbox": [
+            112,
+            746,
+            487,
+            903
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "lecting relevant triplets for generation is an area that warrants further exploration in future research.",
+        "bbox": [
+            507,
+            571,
+            882,
+            602
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Human Evaluation Results Table 5 presents the results of various methods in generating distractor sets for the MCQ cloze tasks. Upon analysis, we observe that in terms of Relevance, Distractiveness, and Utility, both RAP and KAG scored relatively high and exceeded the previous soft method, T5 Candidate Augmentation.",
+        "bbox": [
+            507,
+            613,
+            882,
+            725
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5 presents the results of various methods in generating distractor sets for the MCQ cloze tasks. Upon analysis, we observe that in terms of Relevance, Distractiveness, and Utility, both RAP and KAG scored high and surpassed or comparable to the prior SOTA method.",
+        "bbox": [
+            507,
+            726,
+            882,
+            821
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "These results suggest that we have successfully guided our model to leverage additional knowledge to generate higher quality distractors. The reason why RAP outperforms KAG might be that the KG (ConceptNet) cannot encompass knowledge from all natural domains, whereas RAP, having access to",
+        "bbox": [
+            507,
+            822,
+            880,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "11025",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/47ee0b0cf7f96db46f076bfd9632d58c0a1690eb93ae1bd0f165cd8c0433449c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>DataSet for \nTraing and \nTesting DG model</td><td>Method</td><td>Pretrain Dataset</td><td>P@1</td><td>R@1</td><td>F1@3</td><td>MRR</td><td>NDCG@3</td></tr><tr><td rowspan=\"14\">MCQ</td><td>BART off-shelf</td><td>None</td><td>14.28</td><td>4.76</td><td>11.45</td><td>21.49</td><td>23.70</td></tr><tr><td>RAP-SBART</td><td>MCQ-Train</td><td>16.21</td><td>5.4</td><td>11.96</td><td>22.2</td><td>23.99</td></tr><tr><td>RAP-PBART</td><td>MCQ-Train</td><td>18.14</td><td>6.04</td><td>12.35</td><td>24.06</td><td>25.78</td></tr><tr><td>RAP-PBART /w GTD</td><td>MCQ-Train</td><td>15.05</td><td>5.01</td><td>11.58</td><td>20.65</td><td>22.33</td></tr><tr><td>RAP-SBART</td><td>Sciq-all</td><td>15.05</td><td>5.01</td><td>13.25</td><td>22.77</td><td>25.44</td></tr><tr><td>RAP-PBART</td><td>Sciq-all</td><td>20.84</td><td>6.94</td><td>15.57</td><td>28.57</td><td>30.85</td></tr><tr><td>RAP-PBART /w GTD</td><td>Sciq-all</td><td>22.77</td><td>7.59</td><td>17.88</td><td>31.33</td><td>33.97</td></tr><tr><td>T5 off-shelf</td><td>None</td><td>18.53</td><td>6.17</td><td>11.45</td><td>23.61</td><td>25.08</td></tr><tr><td>RAP-ST5</td><td>MCQ-Train</td><td>23.93</td><td>7.97</td><td>14.15</td><td>29.27</td><td>30.67</td></tr><tr><td>RAP-PT5</td><td>MCQ-Train</td><td>22.39</td><td>7.46</td><td>14.8</td><td>29.02</td><td>30.72</td></tr><tr><td>RAP-PT5 /w GTD</td><td>MCQ-Train</td><td>21.62</td><td>7.2</td><td>13.77</td><td>27.15</td><td>28.70</td></tr><tr><td>RAP-ST5</td><td>Sciq-all</td><td>25.86</td><td>8.62</td><td>15.57</td><td>31.46</td><td>33.12</td></tr><tr><td>RAP-PT5</td><td>Sciq-all</td><td>31.66</td><td>10.55</td><td>18.91</td><td>37.7</td><td>39.43</td></tr><tr><td>RAP-PT5 /w GTD</td><td>Sciq-all</td><td>27.79</td><td>9.26</td><td>18.66</td><td>34.42</td><td>36.31</td></tr><tr><td rowspan=\"14\">Sciq</td><td>BART off-shelf</td><td>None</td><td>19.3</td><td>6.46</td><td>14.94</td><td>26.40</td><td>28.36</td></tr><tr><td>RAP-SBART</td><td>Sciq-train</td><td>19.3</td><td>6.45</td><td>14.66</td><td>25.61</td><td>27.45</td></tr><tr><td>RAP-PBART</td><td>Sciq-train</td><td>20.1</td><td>6.73</td><td>16.32</td><td>27.56</td><td>29.68</td></tr><tr><td>RAP-PBART /w GTD</td><td>Sciq-train</td><td>22.4</td><td>7.48</td><td>16.39</td><td>28.9</td><td>30.81</td></tr><tr><td>RAP-SBART</td><td>MCQ-all</td><td>20.4</td><td>6.83</td><td>15.65</td><td>27.19</td><td>29.09</td></tr><tr><td>RAP-PBART</td><td>MCQ-all</td><td>20.2</td><td>6.75</td><td>15.58</td><td>27.41</td><td>29.5</td></tr><tr><td>RAP-PBART /w GTD</td><td>MCQ-all</td><td>20.7</td><td>6.91</td><td>16.06</td><td>28.16</td><td>30.32</td></tr><tr><td>T5 off-shelf</td><td>None</td><td>24.27</td><td>8.09</td><td>15.92</td><td>30.32</td><td>31.99</td></tr><tr><td>RAP-ST5</td><td>Sciq-train</td><td>23.6</td><td>7.86</td><td>15.62</td><td>29.13</td><td>30.69</td></tr><tr><td>RAP-PT5</td><td>Sciq-train</td><td>24.3</td><td>8.09</td><td>16.06</td><td>29.98</td><td>31.49</td></tr><tr><td>RAP-PT5 /w GTD</td><td>Sciq-train</td><td>25.0</td><td>8.34</td><td>16.46</td><td>30.18</td><td>31.54</td></tr><tr><td>RAP-ST5</td><td>MCQ-all</td><td>24.1</td><td>8.04</td><td>15.26</td><td>29.63</td><td>31.12</td></tr><tr><td>RAP-PT5</td><td>MCQ-all</td><td>23.1</td><td>7.71</td><td>16.31</td><td>29.56</td><td>31.49</td></tr><tr><td>RAP-PT5 /w GTD</td><td>MCQ-all</td><td>23.7</td><td>7.89</td><td>16.36</td><td>29.73</td><td>31.25</td></tr></table>",
+        "bbox": [
+            157,
+            80,
+            842,
+            480
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/4438d5465ced2fb5eb71e73bc9e845844cb3f5b8414247de4c07697c7f57ffb8.jpg",
+        "table_caption": [
+            "Table 4: RAP Performance Overview with In-Domain and Cross-Domain Study."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Relevance</td><td>Distractiveness</td><td>Utility</td></tr><tr><td>(Wang et al., 2023)</td><td>3.81</td><td>3.35</td><td>3.55</td></tr><tr><td>KAG(T5)</td><td>4.05</td><td>3.25</td><td>3.77</td></tr><tr><td>Text2text(RAP-T5)</td><td>4.45</td><td>3.87</td><td>4.07</td></tr><tr><td>ChatGPT</td><td>3.14</td><td>2.67</td><td>3</td></tr><tr><td>ground truth</td><td>3.52</td><td>3.02</td><td>3.4</td></tr></table>",
+        "bbox": [
+            119,
+            524,
+            482,
+            602
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/39e1b71a94f77c6226f8f14306620d0380d4bff83430beb961e09a65287a0098.jpg",
+        "table_caption": [
+            "Table 5: The human evaluation for MCQ dataset."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>F1@3</td></tr><tr><td>T5 KAG</td><td>16.47</td></tr><tr><td>w/o c.a</td><td>13.64</td></tr><tr><td>w/o c.a (Masked LM)</td><td>15.18</td></tr><tr><td>w/o c.a. (Generative LM)</td><td>13.89</td></tr><tr><td>w/o reranker</td><td>14.15</td></tr></table>",
+        "bbox": [
+            556,
+            524,
+            836,
+            626
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "an extensive external corpus (Wikipedia), includes a vast array of facts across various domains. This vastness could potentially explain why KAG scored slightly lower than RAP.",
+        "bbox": [
+            112,
+            655,
+            487,
+            718
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we noted that the ground truth quality for MCQ questions was not consistently high. The lowest performance was found in Distractiveness. This could likely be attributed to the fact that some of MCQ questions are generally designed for elementary and middle school grade levels, which could impact the complexity and misdirection quality of the distractors.",
+        "bbox": [
+            112,
+            721,
+            487,
+            848
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 Ablation Study on KAG",
+        "text_level": 1,
+        "bbox": [
+            112,
+            864,
+            347,
+            879
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conducted ablation experiments on the MCQ dataset to assess the impact of different compo",
+        "bbox": [
+            112,
+            887,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Ablation-study of our KAG model components using the MCQ Test Dataset",
+        "bbox": [
+            507,
+            636,
+            882,
+            665
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "nents in our system. The results are summarized in Table 6, providing insights into the importance and contributions of each component.",
+        "bbox": [
+            507,
+            684,
+            882,
+            733
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Candidate Generation Generating candidate sets using Masked LM or Generative LM is a key component in our system, enhancing KG retrieval. Removing these sets and relying solely on sentence and answer-based triplet retrieval significantly reduced the number of retrievable triplets, potentially introducing irrelevant noise and resulting in a performance drop (F1@3: 16.47 to 13.64). Using Masked LM or Generative LM alone also affected F1@3. This emphasizes the value of incorporating candidate sets, as they provide crucial information",
+        "bbox": [
+            507,
+            741,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "11026",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "for identifying important triplets in the knowledge graph.",
+        "bbox": [
+            112,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Reranker The Reranker, our second key component, sorts triplets to select those more relevant to distractors/answers. Without sorting by the Reranker, randomly selecting triplets for triplet augmentation led to a decrease in F1@3 (16.47 to 14.15). This highlights the Reranker's importance in identifying important triplets and improving downstream task performance.",
+        "bbox": [
+            112,
+            136,
+            489,
+            265
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4.6 Study on Cross-Domain for RAP",
+        "text_level": 1,
+        "bbox": [
+            112,
+            286,
+            418,
+            303
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we report our studies on cross-domain and in-domain for RAP. The experiment results are summarized in Table 4.",
+        "bbox": [
+            112,
+            313,
+            489,
+            361
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We have the following observations: Firstly, in terms of cross-domain effects (using the MCQ dataset for pretraining and fine-tuning the DG model with the Sciq dataset, or vice versa), we found that using Sciq-all for pretraining the language model resulted in a significant improvement in the downstream MCQ dataset for distractor generation. Specifically, using the T5 model, the RAP-based approach increased the F1@3 score from 11.45 to 18.91 (almost a 1.5-fold improvement). This notable improvement was expected, considering that Sciq is a larger dataset compared to MCQ. However, when we reversed the process and used MCQ for pretraining, followed by testing with the Sciq dataset, the improvement in performance was not as significant. Taking the BART model as an example, the direct use of BART off-shelf achieved a score of 14.94, while RAP-SBART) only improved it to 16.06. Although there was some improvement, it was not as substantial as the effects observed when using Sciq pretraining.",
+        "bbox": [
+            112,
+            365,
+            489,
+            703
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Regarding in-domain effects (using the same dataset for both pretraining and fine-tuning), we found consistent improvements in the generated results. For example, using the T5 model on the Sciq test dataset, the T5 off-shell score was 15.92, while  $\\mathrm{RAP - P_{T5}}$  improved it to 16.46. We also observed similar improvements in the MCQ dataset.",
+        "bbox": [
+            112,
+            707,
+            487,
+            819
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In conclusion, our proposed RAP approach has shown improved generation performance from both data augmentation and task-specific pretraining perspectives. The experiments demonstrate that incorporating RAP into the training process leads to significant enhancements in distractor generation.",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            640,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce the utilization of task-specific pretraining and knowledge base. Our experimental results highlight a significant performance improvement achieved through the integration. our work represents a novel contribution to the field of distractor generation for MCQs, showcasing the potential of combining retrieval augmented pretraining and KGs to achieve superior results.",
+        "bbox": [
+            507,
+            109,
+            884,
+            237
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            249,
+            645,
+            263
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We report the following limitations for the KAG method:",
+        "bbox": [
+            507,
+            274,
+            880,
+            306
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Combining KAG with RAP did not yield additive effects. We believe that the key factor lies in that the knowledge triplets may introduce additional noise that interferes with the Text2Text model's generation process. Thus, there is significant room for improvement in the current design of the triplet ranker. Selecting relevant triplets for generation is an aspect that requires further strengthening in this research.",
+            "- The current evaluation and training heavily rely on token scores, which can only reflect the similarity to the ground truth but cannot fully represent the quality of the generated output. The research conclusions thus far have also been established solely based on token scores.",
+            "- During the pretraining stage of RAP, when retrieving sentences/passages, if the answer is a rare or specialized term, the external corpus may fail to find matching sentences/passages. Therefore, when applying the RAP framework to other knowledge domains, having an abundant knowledge corpus becomes crucial."
+        ],
+        "bbox": [
+            531,
+            317,
+            882,
+            724
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            736,
+            672,
+            752
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by NSTC 112-2634-F-005 - 002-project Smart Sustainable New Agriculture Research Center (SMARTer), NSTC Taiwan Project under grant 112-2221-E-005 -075 -MY3, and Delta Research Center, Delta Electronics, Inc.",
+        "bbox": [
+            507,
+            760,
+            884,
+            841
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            509,
+            869,
+            608,
+            883
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Yu Chen, Lingfei Wu, and Mohammed J Zaki. 2023. Toward subgraph-guided knowledge graph question",
+        "bbox": [
+            509,
+            891,
+            882,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "11027",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "generation with graph neural networks. IEEE Transactions on Neural Networks and Learning Systems.",
+            "Shang-Hsuan Chiang, Ssu-Cheng Wang, and Yao-Chung Fan. 2022. Cdgp: Automatic cloze distractor generation based on pre-trained language model. In *Findings of the Association for Computational Linguistics: EMNLP* 2022.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.",
+            "Zichu Fei, Qi Zhang, Tao Gui, Di Liang, Sirui Wang, Wei Wu, and Xuan-Jing Huang. 2022. Cqg: A simple and effective controlled generation framework for multi-hop question generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6896-6906.",
+            "Yanlin Feng, Xinyue Chen, Bill Yuchen Lin, Peifeng Wang, Jun Yan, and Xiang Ren. 2020. Scalable multihop relational reasoning for knowledge-aware question answering. arXiv preprint arXiv:2005.00646.",
+            "Bin He, Di Zhou, Jinghui Xiao, Qun Liu, Nicholas Jing Yuan, Tong Xu, et al. 2019. Integrating graph contextualized knowledge into pre-trained language models. arXiv preprint arXiv:1912.00147.",
+            "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461.",
+            "Chen Liang, Xiao Yang, Neisarg Dave, Drew Wham, Bart Pursel, and C Lee Giles. 2018. Distractor generation for multiple choice questions using learning to rank. In Proceedings of the thirteenth workshop on innovative use of NLP for building educational applications, pages 284-290.",
+            "Chen Liang, Xiao Yang, Drew Wham, Bart Pursel, Rebecca Passonneaur, and C Lee Giles. 2017. Distractor generation with generative adversarial nets for automatically creating fill-in-the-blank questions. In Proceedings of the Knowledge Capture Conference, pages 1-4.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.",
+            "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084.",
+            "Hongyu Ren, Weihua Hu, and Jure Leskovec. 2020. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. arXiv preprint arXiv:2002.05969."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Siyu Ren and Kenny Q. Zhu. 2021. Knowledge-driven distractor generation for cloze-style multiple choice questions. Proceedings of the AAAI Conference on Artificial Intelligence, 35(5):4339-4347.",
+            "Siyu Ren and Kenny Q Zhu. 2021. Knowledge-driven distractor generation for cloze-style multiple choice questions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 4339-4347.",
+            "Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI conference on artificial intelligence, volume 31.",
+            "Hui-Juan Wang, Kai-Yu, Han-Cheng Yu, and Yao-Chung Fan. 2023. Distractor generation based on text2text language models with pseudo kullback-Leibler divergence regulation. In Proceedings of the Findings of ACL 2023.",
+            "Rui Wang, Shijing Si, Guoyin Wang, Lei Zhang, Lawrence Carin, and Ricardo Henao. 2020. Integrating task specific information into pretrained language models for low resource fine tuning. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3181-3186.",
+            "Johannes Welbl, Nelson F Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. arXiv preprint arXiv:1707.06209.",
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.",
+            "Michihiro Yasunaga, Antoine Bosselut, Hongyu Ren, Xikun Zhang, Christopher D Manning, Percy S Liang, and Jure Leskovec. 2022. Deep bidirectional language-knowledge graph pretraining. Advances in Neural Information Processing Systems, 35:37309-37323.",
+            "Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut, Percy Liang, and Jure Leskovec. 2021. Qagnn: Reasoning with language models and knowledge graphs for question answering. arXiv preprint arXiv:2104.06378.",
+            "Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter Liu. 2020. Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In International Conference on Machine Learning, pages 11328-11339. PMLR.",
+            "Xikun Zhang, Antoine Bosselut, Michihiro Yasunaga, Hongyu Ren, Percy Liang, Christopher D Manning, and Jure Leskovec. 2022. Greaselm: Graph reasoning enhanced language models for question answering. arXiv preprint arXiv:2201.08860."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            889
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "bbox": [
+            510,
+            903,
+            596,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "11028",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            356,
+            99
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Our models are implemented based on models from Hugging Face (Wolf et al., 2019). For generating candidate set, we use an language model trained for distractor generation from (Chiang et al., 2022) and gpt3.5 turbo through ChatGPT API with zero-shot prompting. For triplet reranker, we use SentenceBERT (Reimers and Gurevych, 2019) to estimate the initial sentence embedding distance. For triplet classification, we use Sentence-BERT and BERT (Devlin et al., 2018) as the default PLM. For the triplet augmented generator, we experiment with BART (Lewis et al., 2019) and T5 (Raffel et al., 2020) as base generation models.",
+        "bbox": [
+            112,
+            111,
+            489,
+            319
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "During training, we use AdamW as the optimizer and an initial learning rate of 2e-5 for BERT, BART and 1e-4 for T5 models. All experiments are conducted using two NVIDIA Titan RTX GPUs.",
+        "bbox": [
+            112,
+            321,
+            489,
+            385
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "BART-based generator With MCQ data, the maximum number of epochs is set to 40 with a batch size of on two NVIDIA TITAN RTX GPUs for the distraction generation with triplet augmentation with a batch size of 32 and a maximum number of triplets is set to 50. With Sciq data, the maximum number of epochs is set to 50 with a batch size of 32 on two NVIDIA TITAN RTX GPUs for the distraction generation with triplet augmentation with a batch size of 32 and a maximum number of triplets is set to 50. The average running time for BART-based generators is 30 minutes (1.5 hours) on MCQ (Sciq).",
+        "bbox": [
+            112,
+            397,
+            489,
+            606
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "T5-based generator With MCQ data, the maximum number of epochs is set to 40 with a batch size of on two NVIDIA TITAN RTX GPUs for the distraction generation with triplet augmentation with a batch size of 32 and a maximum number of triplets is set to 50. With Sciq data, the maximum number of epochs is set to 50 with a batch size of 32 on two NVIDIA TITAN RTX GPUs for the distraction generation with triplet augmentation with a batch size of 32 and a maximum number of triplets is set to 50. The average running time for BART-based generators is 40 minutes (3 hours) on MCQ (Sciq).",
+        "bbox": [
+            112,
+            618,
+            489,
+            827
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Knowledge Graph We utilize ConceptNet (Speer et al., 2017) as our KG, which is a graph that encompasses a wide range of knowledge domains. The KG consists of a total of 800k nodes and 2M edges. For each question's KG retrieval set, we pre",
+        "bbox": [
+            112,
+            839,
+            489,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "process the retrieval subgraph using the approach described in (Feng et al., 2020). Specifically, we process the textual corpus of each question, including the question text, answer options, and entities in the candidate set.",
+        "bbox": [
+            507,
+            84,
+            884,
+            164
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11029",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    }
+]