Add Batch acc772bc-3906-4564-b984-dbf69420805b data

.gitattributes

@@ -5785,3 +5785,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/DemoSG_[[:space:]]Demonstration-enhanced[[:space:]]Schema-guided[[:space:]]Generation[[:space:]]for[[:space:]]Low-resource[[:space:]]Event[[:space:]]Extraction/aeee44ed-001c-4372-8524-70dccc8b0406_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Democratizing[[:space:]]LLMs_[[:space:]]An[[:space:]]Exploration[[:space:]]of[[:space:]]Cost-Performance[[:space:]]Trade-offs[[:space:]]in[[:space:]]Self-Refined[[:space:]]Open-Source[[:space:]]Models/4770b9fd-de77-4b56-899d-c982de9ae291_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Demystifying[[:space:]]Prompts[[:space:]]in[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Perplexity[[:space:]]Estimation/375ed177-07a2-4de2-b0ce-c697ad94dc04_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dense[[:space:]]Retrieval[[:space:]]as[[:space:]]Indirect[[:space:]]Supervision[[:space:]]for[[:space:]]Large-space[[:space:]]Decision[[:space:]]Making/6485a745-f89c-470f-a30b-5390c646537c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Density-Aware[[:space:]]Prototypical[[:space:]]Network[[:space:]]for[[:space:]]Few-Shot[[:space:]]Relation[[:space:]]Classification/3cc00c50-e0b4-4c6e-9a21-2f4635fbe6e7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DepNeCTI_[[:space:]]Dependency-based[[:space:]]Nested[[:space:]]Compound[[:space:]]Type[[:space:]]Identification[[:space:]]for[[:space:]]Sanskrit/9cf22d4e-25cc-4736-8b7f-6f6bea41e671_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DepWiGNN_[[:space:]]A[[:space:]]Depth-wise[[:space:]]Graph[[:space:]]Neural[[:space:]]Network[[:space:]]for[[:space:]]Multi-hop[[:space:]]Spatial[[:space:]]Reasoning[[:space:]]in[[:space:]]Text/82978807-d92d-4dd8-83a8-263f59cfaa61_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Descriptive[[:space:]]Prompt[[:space:]]Paraphrasing[[:space:]]for[[:space:]]Target-Oriented[[:space:]]Multimodal[[:space:]]Sentiment[[:space:]]Classification/9ad0a505-4aa3-4e82-936d-bde5e9f488e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DetectLLM_[[:space:]]Leveraging[[:space:]]Log[[:space:]]Rank[[:space:]]Information[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Detection[[:space:]]of[[:space:]]Machine-Generated[[:space:]]Text/b6e08278-e610-46f1-92c0-6af873c348c2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Detecting[[:space:]]Erroneously[[:space:]]Recognized[[:space:]]Handwritten[[:space:]]Byzantine[[:space:]]Text/e23fb75f-6ceb-4c26-9972-d5c23aa51bcb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Detecting[[:space:]]Syntactic[[:space:]]Change[[:space:]]with[[:space:]]Pre-trained[[:space:]]Transformer[[:space:]]Models/5f1aa19b-edb6-4425-a7bb-8580ef6a41b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Detrimental[[:space:]]Contexts[[:space:]]in[[:space:]]Open-Domain[[:space:]]Question[[:space:]]Answering/1aae5c2b-0c07-4ccb-ad79-139133468d78_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DiFair_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Disentangled[[:space:]]Assessment[[:space:]]of[[:space:]]Gender[[:space:]]Knowledge[[:space:]]and[[:space:]]Bias/933134df-cd8a-4a72-9de3-e519ab59c119_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DiQAD_[[:space:]]A[[:space:]]Benchmark[[:space:]]Dataset[[:space:]]for[[:space:]]Open-domain[[:space:]]Dialogue[[:space:]]Quality[[:space:]]Assessment/1399a40f-bbbe-4231-8d56-69d6243a85d8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DialGuide_[[:space:]]Aligning[[:space:]]Dialogue[[:space:]]Model[[:space:]]Behavior[[:space:]]with[[:space:]]Developer[[:space:]]Guidelines/612d8c57-e446-486e-b87e-57f59d7a8f9b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dialect[[:space:]]Transfer[[:space:]]for[[:space:]]Swiss[[:space:]]German[[:space:]]Speech[[:space:]]Translation/597e5e5e-8508-4783-b929-57c76fc7e604_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dialect-to-Standard[[:space:]]Normalization_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Multilingual[[:space:]]Evaluation/4a1c12bb-c5a5-4d3f-b0e9-22019cf29318_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DialogQAE_[[:space:]]N-to-N[[:space:]]Question[[:space:]]Answer[[:space:]]Pair[[:space:]]Extraction[[:space:]]from[[:space:]]Customer[[:space:]]Service[[:space:]]Chatlog/22e5a4fb-fc61-41f3-81ea-a8dbdfc15a42_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dialogue[[:space:]]Act-Aided[[:space:]]Backchannel[[:space:]]Prediction[[:space:]]Using[[:space:]]Multi-Task[[:space:]]Learning/5d014702-ac80-4c48-a145-be9fc63baa8f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dialogue[[:space:]]Medical[[:space:]]Information[[:space:]]Extraction[[:space:]]with[[:space:]]Medical-Item[[:space:]]Graph[[:space:]]and[[:space:]]Dialogue-Status[[:space:]]Enriched[[:space:]]Representation/a22fbec1-6742-410d-847b-ac7f3fa21983_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Difference-Masking_[[:space:]]Choosing[[:space:]]What[[:space:]]to[[:space:]]Mask[[:space:]]in[[:space:]]Continued[[:space:]]Pretraining/2c3fe2b1-076b-4b1b-b362-ac94b7f2fd76_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DiffuSeq-v2_[[:space:]]Bridging[[:space:]]Discrete[[:space:]]and[[:space:]]Continuous[[:space:]]Text[[:space:]]Spaces[[:space:]]for[[:space:]]Accelerated[[:space:]]Seq2Seq[[:space:]]Diffusion[[:space:]]Models/5b9ffd86-8df0-428f-b1af-07797b37210d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DiffuVST_[[:space:]]Narrating[[:space:]]Fictional[[:space:]]Scenes[[:space:]]with[[:space:]]Global-History-Guided[[:space:]]Denoising[[:space:]]Models/7cd9450a-86e5-4b15-8890-ed82b603301d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Diffusion[[:space:]]Language[[:space:]]Model[[:space:]]with[[:space:]]Query-Document[[:space:]]Relevance[[:space:]]for[[:space:]]Query-Focused[[:space:]]Summarization/ef792535-71bf-454f-8642-63dbe13fb39b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DiffusionRet_[[:space:]]Diffusion-Enhanced[[:space:]]Generative[[:space:]]Retriever[[:space:]]using[[:space:]]Constrained[[:space:]]Decoding/bed838a2-75ec-4c0b-afad-166110b7bd92_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DiffusionSL_[[:space:]]Sequence[[:space:]]Labeling[[:space:]]via[[:space:]]Tag[[:space:]]Diffusion[[:space:]]Process/8b7817c2-0523-4fff-b2e6-8815f54b1cbe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dimensions[[:space:]]of[[:space:]]Online[[:space:]]Conflict_[[:space:]]Towards[[:space:]]Modeling[[:space:]]Agonism/056ddabc-8aa8-47a9-93b8-8f7b2e69353d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dior-CVAE_[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models[[:space:]]and[[:space:]]Diffusion[[:space:]]Priors[[:space:]]for[[:space:]]Variational[[:space:]]Dialog[[:space:]]Generation/a3821ddd-9397-40c0-86b2-705a60919607_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Discourse[[:space:]]Sense[[:space:]]Flows_[[:space:]]Modelling[[:space:]]the[[:space:]]Rhetorical[[:space:]]Style[[:space:]]of[[:space:]]Documents[[:space:]]across[[:space:]]Various[[:space:]]Domains/d897d0a8-37c1-42c0-8a25-96e6fbe21a3e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Discovering[[:space:]]Highly[[:space:]]Influential[[:space:]]Shortcut[[:space:]]Reasoning_[[:space:]]An[[:space:]]Automated[[:space:]]Template-Free[[:space:]]Approach/1563fae4-370d-46b9-8c0e-999b6163381f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Disentangling[[:space:]]Extraction[[:space:]]and[[:space:]]Reasoning[[:space:]]in[[:space:]]Multi-hop[[:space:]]Spatial[[:space:]]Reasoning/65578046-bebb-4564-a9a4-26371aaa5db0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Disentangling[[:space:]]Structure[[:space:]]and[[:space:]]Style_[[:space:]]Political[[:space:]]Bias[[:space:]]Detection[[:space:]]in[[:space:]]News[[:space:]]by[[:space:]]Inducing[[:space:]]Document[[:space:]]Hierarchy/ef86f313-91c3-4564-a0c5-bd98a83311e3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Disfluent[[:space:]]Cues[[:space:]]for[[:space:]]Enhanced[[:space:]]Speech[[:space:]]Understanding[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/34df72a7-369e-401e-9d76-bc58d87b574c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dissecting[[:space:]]In-Context[[:space:]]Learning[[:space:]]of[[:space:]]Translations[[:space:]]in[[:space:]]GPT-3/353a4934-cd3d-4c42-ae6b-fbb0fbd273cb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DistillCSE_[[:space:]]Distilled[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Sentence[[:space:]]Embeddings/c35c9e5b-4287-4b90-901c-3abb957654d9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Distilling[[:space:]]ChatGPT[[:space:]]for[[:space:]]Explainable[[:space:]]Automated[[:space:]]Student[[:space:]]Answer[[:space:]]Assessment/f13e993b-dc90-4229-b463-8136e35f4e22_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Diversifying[[:space:]]language[[:space:]]models[[:space:]]for[[:space:]]lesser-studied[[:space:]]languages[[:space:]]and[[:space:]]language-usage[[:space:]]contexts_[[:space:]]A[[:space:]]case[[:space:]]of[[:space:]]second[[:space:]]language[[:space:]]Korean/a46af694-2fd7-4772-beab-57f98ac6242c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Do[[:space:]]Stochastic[[:space:]]Parrots[[:space:]]have[[:space:]]Feelings[[:space:]]Too_[[:space:]]Improving[[:space:]]Neural[[:space:]]Detection[[:space:]]of[[:space:]]Synthetic[[:space:]]Text[[:space:]]via[[:space:]]Emotion[[:space:]]Recognition/6db6c252-f81f-402d-9e5f-f0b9487853ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Do[[:space:]]“English”[[:space:]]Named[[:space:]]Entity[[:space:]]Recognizers[[:space:]]Work[[:space:]]Well[[:space:]]on[[:space:]]Global[[:space:]]Englishes_/6886174a-78e0-41ed-8d92-ad5085b08003_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DocAsRef_[[:space:]]An[[:space:]]Empirical[[:space:]]Study[[:space:]]on[[:space:]]Repurposing[[:space:]]Reference-based[[:space:]]Summary[[:space:]]Quality[[:space:]]Metrics[[:space:]]as[[:space:]]Reference-free[[:space:]]Metrics/330c8365-dee6-4c39-9978-63eb62ae6c67_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DocSplit_[[:space:]]Simple[[:space:]]Contrastive[[:space:]]Pretraining[[:space:]]for[[:space:]]Large[[:space:]]Document[[:space:]]Embeddings/61b75f3f-ea1a-485c-b350-0949a8c98210_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DocTrack_[[:space:]]A[[:space:]]Visually-Rich[[:space:]]Document[[:space:]]Dataset[[:space:]]Really[[:space:]]Aligned[[:space:]]with[[:space:]]Human[[:space:]]Eye[[:space:]]Movement[[:space:]]for[[:space:]]Machine[[:space:]]Reading/9a8a902c-4029-426d-a9f4-e29aca0f06d2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Does[[:space:]]Listener[[:space:]]Gaze[[:space:]]in[[:space:]]Face-to-Face[[:space:]]Interaction[[:space:]]Follow[[:space:]]the[[:space:]]Entropy[[:space:]]Rate[[:space:]]Constancy[[:space:]]Principle_[[:space:]]An[[:space:]]Empirical[[:space:]]Study/51706ed8-173a-49d4-be08-38179f4f864b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dolphin_[[:space:]]A[[:space:]]Challenging[[:space:]]and[[:space:]]Diverse[[:space:]]Benchmark[[:space:]]for[[:space:]]Arabic[[:space:]]NLG/b0ee8c30-c5ca-491d-9b9d-c5aec88d1f56_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Domain[[:space:]]Adaptation[[:space:]]for[[:space:]]Conversational[[:space:]]Query[[:space:]]Production[[:space:]]with[[:space:]]the[[:space:]]RAG[[:space:]]Model[[:space:]]Feedback/8d54abbc-68dc-4e40-95aa-d992b85b7587_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Domain[[:space:]]Adaptation[[:space:]]for[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]Using[[:space:]]Robust[[:space:]]Internal[[:space:]]Representations/3cc7c212-88a3-420c-8615-4e9593c75e5c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Domain[[:space:]]Private[[:space:]]Transformers[[:space:]]for[[:space:]]Multi-Domain[[:space:]]Dialog[[:space:]]Systems/60a49d4f-3c64-4b58-8435-dad9146e351c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Don’t[[:space:]]Add,[[:space:]]don’t[[:space:]]Miss_[[:space:]]Effective[[:space:]]Content[[:space:]]Preserving[[:space:]]Generation[[:space:]]from[[:space:]]Pre-Selected[[:space:]]Text[[:space:]]Spans/6b92295f-064d-4669-b32c-0096d671b3e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Don’t[[:space:]]waste[[:space:]]a[[:space:]]single[[:space:]]annotation_[[:space:]]improving[[:space:]]single-label[[:space:]]classifiers[[:space:]]through[[:space:]]soft[[:space:]]labels/ef707eab-c929-48c8-a6a7-7a082d846c48_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Drilling[[:space:]]Down[[:space:]]into[[:space:]]the[[:space:]]Discourse[[:space:]]Structure[[:space:]]with[[:space:]]LLMs[[:space:]]for[[:space:]]Long[[:space:]]Document[[:space:]]Question[[:space:]]Answering/e6a179fa-0154-425a-899d-39a5515a260d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dual[[:space:]]Contrastive[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Incremental[[:space:]]Text[[:space:]]Classification/0c23ce14-067b-49e9-84ad-4f865c1c158a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dynamic[[:space:]]Low-rank[[:space:]]Estimation[[:space:]]for[[:space:]]Transformer-based[[:space:]]Language[[:space:]]Models/2d886487-9fbc-4061-b000-ac113eadac0a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dynamic[[:space:]]Open-book[[:space:]]Prompt[[:space:]]for[[:space:]]Conversational[[:space:]]Recommender[[:space:]]System/52051efe-c786-4f37-902c-51b2ae67eece_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dynamic[[:space:]]Stance_[[:space:]]Modeling[[:space:]]Discussions[[:space:]]by[[:space:]]Labeling[[:space:]]the[[:space:]]Interactions/8c80450d-ff06-4e44-9bcf-79db3276f612_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dynamic[[:space:]]Stashing[[:space:]]Quantization[[:space:]]for[[:space:]]Efficient[[:space:]]Transformer[[:space:]]Training/744ba489-38b0-471f-85a6-5c5b8dddb190_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Dynamic[[:space:]]Voting[[:space:]]for[[:space:]]Efficient[[:space:]]Reasoning[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/b997df11-b086-4887-ad48-44d42a49401b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/EARA_[[:space:]]Improving[[:space:]]Biomedical[[:space:]]Semantic[[:space:]]Textual[[:space:]]Similarity[[:space:]]with[[:space:]]Entity-Aligned[[:space:]]Attention[[:space:]]and[[:space:]]Retrieval[[:space:]]Augmentation/bda4dd44-0832-44b5-b59b-f09d846b7beb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ECHo_[[:space:]]A[[:space:]]Visio-Linguistic[[:space:]]Dataset[[:space:]]for[[:space:]]Event[[:space:]]Causality[[:space:]]Inference[[:space:]]via[[:space:]]Human-Centric[[:space:]]Reasoning/35f5853e-7600-40be-8b8b-554cac00acb4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/EMO-KNOW_[[:space:]]A[[:space:]]Large[[:space:]]Scale[[:space:]]Dataset[[:space:]]on[[:space:]]Emotion-Cause/e02f1c19-44c7-47de-8ac9-8ac7f0d72b47_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ESPVR_[[:space:]]Entity[[:space:]]Spans[[:space:]]Position[[:space:]]Visual[[:space:]]Regions[[:space:]]for[[:space:]]Multimodal[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition/e7b559e8-9bcf-42db-9917-580c643cef65_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Ecologically[[:space:]]Valid[[:space:]]Explanations[[:space:]]for[[:space:]]Label[[:space:]]Variation[[:space:]]in[[:space:]]NLI/c6c81cea-17af-4edb-942a-95ea7aad8d91_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/EconBERTa_[[:space:]]Towards[[:space:]]Robust[[:space:]]Extraction[[:space:]]of[[:space:]]Named[[:space:]]Entities[[:space:]]in[[:space:]]Economics/e4db376d-9a77-4172-8af6-625f56b777f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/EffEval_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Evaluation[[:space:]]of[[:space:]]Efficiency[[:space:]]for[[:space:]]MT[[:space:]]Evaluation[[:space:]]Metrics/b0753757-2eeb-416b-9a95-0d4aa597e75a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Effects[[:space:]]of[[:space:]]Human[[:space:]]Adversarial[[:space:]]and[[:space:]]Affable[[:space:]]Samples[[:space:]]on[[:space:]]BERT[[:space:]]Generalization/1f2b1b3d-2880-4d3d-9682-5dbb839efd4e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Efficient[[:space:]]Continue[[:space:]]Training[[:space:]]of[[:space:]]Temporal[[:space:]]Language[[:space:]]Model[[:space:]]with[[:space:]]Structural[[:space:]]Information/4e01312d-f9ea-48a2-ac0d-45b87a14409b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Efficient[[:space:]]Cross-Task[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Few-Shot[[:space:]]Conversational[[:space:]]Emotion[[:space:]]Recognition/d55c3273-2161-46ae-adc0-b3cf30e5ba4b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Efficient[[:space:]]Data[[:space:]]Learning[[:space:]]for[[:space:]]Open[[:space:]]Information[[:space:]]Extraction[[:space:]]with[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models/2930ad8d-b924-4379-920e-36da32fb8d1e_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Dense Retrieval as Indirect Supervision for Large-space Decision Making/6485a745-f89c-470f-a30b-5390c646537c_content_list.json

@@ -0,0 +1,1524 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/9fe764c8a2ca174d97908e9600f24354e6e2dbc1f477f18d882a5dd37836d839.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            260,
+            91,
+            292,
+            112
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dense Retrieval as Indirect Supervision for Large-space Decision Making",
+        "text_level": 1,
+        "bbox": [
+            292,
+            102,
+            736,
+            142
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Nan Xu Fei Wang Mingtao Dong Muhao Chen",
+        "bbox": [
+            248,
+            158,
+            749,
+            181
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Southern California University of California, Davis {nanx, fwang598, mingtaod}@usc.edu muhchen@ucdavis.edu",
+        "bbox": [
+            215,
+            183,
+            786,
+            223
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Many discriminative natural language understanding (NLU) tasks have large label spaces. Learning such a process of large-space decision making is particularly challenging due to the lack of training instances per label and the difficulty of selection among many fine-grained labels. Inspired by dense retrieval methods for passage finding in open-domain QA, we propose a reformulation of large-space discriminative NLU tasks as a learning-to-retrieve task, leading to a novel solution named Dense Decision Retrieval (DDR). Instead of predicting fine-grained decisions as logits, DDR adopts a dual-encoder architecture that learns to predict by retrieving from a decision thesaurus. This approach not only leverages rich indirect supervision signals from easy-to-consume learning resources for dense retrieval, it also leads to enhanced prediction generalizability with a semantically meaningful representation of the large decision space. When evaluated on tasks with decision spaces ranging from hundreds to hundred-thousand scales, DDR outperforms strong baselines greatly by  $27.54\\%$  in  $P@1$  on two extreme multi-label classification tasks,  $1.17\\%$  in F1 score ultra-fine entity typing, and  $1.26\\%$  in accuracy on three few-shot intent classification tasks on average.",
+        "bbox": [
+            141,
+            281,
+            460,
+            682
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            695,
+            258,
+            709
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Many discriminative natural language understanding (NLU) tasks require making fine-grained decisions from a large candidate decision space. For example, a task-oriented dialogue system, when responding to users' requests, needs to frequently detect their intents from hundreds of options (Zhang et al., 2020b; Ham et al., 2020). Description-based recommendation in e-commerce needs to search from millions of products in response to users' descriptions (Gupta et al., 2021; Xiong et al., 2022).",
+        "bbox": [
+            112,
+            721,
+            489,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Previous studies for such NLU tasks still train classifiers as the solution (Gupta et al., 2021; Yu et al., 2022; Lin et al., 2023a). However, we argue that this straightforward approach is less practical for large-space decision making for several reasons. First, more decision labels naturally lead to data scarcity, since collecting sufficient training data for all labels will need significantly more cost. This issue is also accompanied by the issue of rare labels in the long tail of highly skewed distributions suffering severely from the lack of sufficient training instances (Zhang et al., 2023), leading to overgeneralization bias where frequent labels are more likely to be predicted compared with rare ones (Xu et al., 2022). Second, the non-semantic logit-based representation for decision labels in a classifier also makes the model hard to generalize to rarely seen labels, and not adaptable to unseen ones in training. This issue also impairs the applicability of the decision-making model to real-world scenarios, such as recommendation (Xiong et al., 2022) and task-oriented parsing (Zhao et al., 2022a), where the decision space may expand rapidly and crucial labels may be absent in training.",
+        "bbox": [
+            507,
+            252,
+            884,
+            639
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "On the contrary, motivated by semantic similarity of examples annotated with identical labels, recent studies propose contrastive learning schemes that leverage Siamese encoding architectures to maximize similarity scores of representation for positive example pairs (Henderson et al., 2020; Zhang et al., 2020a; Dahiya et al., 2021a; Mehri and Eric, 2021; Zhang et al., 2021; Xiong et al., 2022). Meanwhile, inductive bias from NLU tasks such as masked language modeling (Mehri et al., 2020; Dai et al., 2021) and natural language inference (NLI; Li et al. 2022; Du et al. 2022) has shown beneficial for learning on rare and unseen labels via indirect supervision (Yin et al., 2023). However, when dealing with very large decision spaces, existing methods still face critical trade-offs between",
+        "bbox": [
+            507,
+            661,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ Code and resources are available at https://github.com/luka-group/DDR.",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "15021",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15021-15033 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            208,
+            945,
+            786,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "generalizability and efficiency of prediction.2",
+        "bbox": [
+            114,
+            84,
+            450,
+            99
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Inspired by the recent success of dense retrieval methods that learn to select answer-descriptive passages from millions of candidate documents for open-domain QA (Karpukhin et al., 2020; Lee et al., 2021; Zhan et al., 2021), we propose an indirectly supervised solution named Dense Decision Retrieval (DDR). DDR provides a general reformulation of large-space decision making as learning to retrieve from a semantically meaningful decision thesaurus constructed based on task-relevant resources. The model adopts the dual-encoder architecture from dense retrieval models to embed input texts and label descriptions, and learns to predict by retrieving from the informative decision thesaurus instead of predicting fine-grained decisions as logits. In this way, DDR not only leverages rich indirect supervision signals from other easy-to-consume learning resources for dense retrieval. It also leads to enhanced prediction performance and generalizability with a semantically meaningful representation of the large decision space. We evaluate DDR on large decision spaces ranging from hundreds for few-shot intent detection, ten-thousand for ultra-fine entity typing, to hundred-thousand scales for extreme multi-label classification. DDR obtains state-of-the-art performance on 6 benchmark datasets in multiple few-shot settings, improving the most competitive baselines by  $27.54\\%$  in  $\\mathbb{P}@\\mathbb{1}$  for extreme classification,  $1.17\\%$  in F1 for entity typing and  $1.26\\%$  in accuracy for intent detection on average. Ablation studies show that both the constructed informative label thesaurus and indirect supervision from dense retrieval contribute to performance gain.",
+        "bbox": [
+            115,
+            105,
+            490,
+            653
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The technical contributions of this work are three-fold. First, we present a novel and strong solution,  $DDR$ , for NLU tasks with large-space decision making that leverages indirect supervision from dense retrieval. Second, we provide semantically meaningful decision thesaurus construction that further improves the decision-making ability of  $DDR$ . Third, we comprehensively verify the effectiveness of  $DDR$  on tasks of fine-grained text classification, semantic typing and intent detection where the size of decision spaces range from hundreds to hundreds of thousands.",
+        "bbox": [
+            112,
+            659,
+            489,
+            851
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            665,
+            98
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Indirect Supervision Indirectly supervised methods (Roth, 2017; He et al., 2021; Yin et al., 2023) seek to transfer supervision signals from a more resource-rich task to enhance a specific more resource-limited task. A method of this kind often involves reformulation of the target task to the source task. Previous studies have investigated using source tasks such as NLI (Li et al., 2022; Yin et al., 2019; Lyu et al., 2021, inter alia), extractive QA (Wu et al., 2020; FitzGerald et al., 2018; Li et al., 2020, inter alia), abstractive QA (Zhao et al., 2022a; Du and Ji, 2022) and conditioned generation (Lu et al., 2022; Huang et al., 2022b; Hsu et al., 2022, inter alia) to enhance more expensive information extraction or semantic parsing tasks. Recent studies also transformed these technologies to specialized domains such as medicine (Xu et al., 2023) and software engineering (Zhao et al., 2022b) where model generalization and lack of annotations are more significant challenges. There has also been foundational work which studies the informativeness of supervision signals in such settings (He et al., 2021).",
+        "bbox": [
+            507,
+            129,
+            885,
+            500
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, the aforementioned studies are not designed for discriminative tasks with very large decision spaces, and do not apply directly due to issues such as high inference costs (NLI) and requiring decisions to be inclusive to the input (QA). We instead propose dense retriever that naturally serves as a proper and efficient form of indirect supervision for large-space decision making.",
+        "bbox": [
+            507,
+            506,
+            885,
+            636
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "NLU with Large Decision Spaces Many concrete NLU tasks deal with large decision spaces, including description-based recommendation in e-commerce (Gupta et al., 2021; Xiong et al., 2022) and Web search (Gupta et al., 2021), user intent detection in task-oriented dialog systems (Zhang et al., 2020b; Ham et al., 2020), and fine-grained semantic typing (Choi et al., 2018; Chen et al., 2020), etc. Previous studies for such tasks either train classifiers (Yu et al., 2022; Lin et al., 2023a) or rely on contrastive learning from scratch (Zhang et al., 2021; Xiong et al., 2022), which are generally impaired by insufficiency of training data and hard to generalize to rarely seen and unseen labels. These challenges motivate us to explore a practical solution with indirect supervision from a dense retriever.",
+        "bbox": [
+            507,
+            645,
+            885,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "For example, NLI-based inference leads to  $k$  times more inference cost where  $k$  being the size of the decision space, which is prohibitive for large-space decision making tasks.",
+        "bbox": [
+            112,
+            879,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "15022",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d21e79feccc28c3d1f6d52726dfd7f12363d46193d3f14c57384969389b2300.jpg",
+        "image_caption": [
+            "Figure 1: Overview of DDR that reformulates general large-space decision making tasks as dense retrieval. Label thesaurus is constructed with detailed descriptions from publicly available resources and the dual-encoder learns to maximize similarity between embeddings of input and label thesaurus entry with indirect supervision from a pre-trained dense retriever."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            84,
+            884,
+            284
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            376,
+            216,
+            390
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We first describe the reformulation of large-space decision making task as dense retrieval (§3.1). We then introduce several automatic ways to construct label thesaurus of high quality (§3.2). Lastly, we demonstrate the dual-encoder architecture of DDR (illustrated in Fig. 1) to retrieve decisions (§3.3).",
+        "bbox": [
+            112,
+            403,
+            487,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Preliminary",
+        "text_level": 1,
+        "bbox": [
+            112,
+            513,
+            253,
+            527
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The mechanism of dense retrieval for general decision-making in large spaces can be formalized as follows. Given a textual input, such as \"Illy Ground Espresso Classico Coffee: Classico, classic roast coffee has a lingering sweetness and delicate...\" for description-based product recommendation, a model learns to infer their corresponding labels by mapping input to textual descriptions in a large label thesaurus (such as names, brands, component, etc.), instead of simply predicting the label index.",
+        "bbox": [
+            110,
+            535,
+            487,
+            709
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Formally speaking, given a textual input  $x$ , a decision space covering  $L$  labels and a label thesaurus that contains corresponding entries (label descriptions)  $D = \\{d_{1}, d_{2}, \\ldots, d_{L}\\}$ . We first split every entry into text passages of a maximum length (e.g., 100 words) as the basic retrieval units and get  $M$  total passages  $P = \\{p_{1}, p_{2}, \\ldots, p_{M}\\}^{3}$ . A retriever  $R: (x, P) \\to P_{R}$  is a function that takes as input a text sequence  $x$  and the label thesaurus  $P$  and",
+        "bbox": [
+            112,
+            713,
+            487,
+            857
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "returns a ranking list of passages  $P_R$ , where labels represented by higher ranked passages are more likely to be relevant to the input. As demonstrated in Fig. 1, DDR also leverages indirect supervision from open-domain QA by initializing parameters of retriever  $R$  with those from DPR.",
+        "bbox": [
+            507,
+            376,
+            882,
+            472
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Thesaurus Construction",
+        "text_level": 1,
+        "bbox": [
+            507,
+            491,
+            746,
+            505
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Depending on the target task, the decision thesaurus can provide knowledge about the decision space from various perspectives. Given label names, we create informative descriptions for each label automatically without extra expenses. We can refer to publicly available resources for constructing thesaurus of high quality. To name a few:",
+        "bbox": [
+            507,
+            514,
+            882,
+            627
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1) Lexicographical knowledge bases: dictionaries such as WordNet (Miller, 1995) and Wiktionary<sup>4</sup> provide accurate definitions as well as plentiful use cases to help understand words and phrases.",
+            "2) Large Language Models: LLMs trained on instructions like ChatGPT and Vicuna-13B (Chiang et al., 2023) are able to generate comprehensive and reasonable descriptions of terms when they are prompted for explanation.",
+            "3) Training examples: directly adopting input text when labels with detailed information also appear as example input, or aggregating input texts from multiple examples assigned with the same label if they share very high-level semantic similarities."
+        ],
+        "bbox": [
+            507,
+            629,
+            882,
+            853
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We describe thesaurus construction in practice for each investigated downstream task in  $\\S 4$ .",
+        "bbox": [
+            507,
+            856,
+            880,
+            887
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>Since a label-descriptive entry, such as those in XMC tasks, is sometimes too long to encode with a Transformer encoder, we break it down into multiple passages to avoid the quadratic dependency w.r.t input length.",
+        "bbox": [
+            112,
+            868,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>4</sup>https://en.wikipedia.org/",
+        "bbox": [
+            536,
+            903,
+            742,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "15023",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3 Learning to Retrieve Decisions",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            401,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Encoders Since a key motivation of  $DDR$  is to leverage indirect supervision from retrieval tasks, we adopt the dual-encoder architecture similar to a dense retriever such as DPR (Karpukhin et al., 2020), where two sentence encoders with special tokens (BERT (Devlin et al., 2019) by default),  $E_{x}$  and  $E_{p}$ , are leveraged to represent input text  $x$  and passage  $p$  individually. The similarity between input and passage is computed using dot product of their vectors denoted by  $\\operatorname{sim}(x,p)$ .",
+        "bbox": [
+            112,
+            109,
+            487,
+            269
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Training The objective of the reformulated large-space decision making problem is to optimize the dual encoders such that relevant pairs of input text and label description (thesaurus entry) should have higher similarity than irrelevant pairs in the embedding space. Hence, the overall learning objective is a process of contrastive learning.",
+        "bbox": [
+            112,
+            274,
+            487,
+            385
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To tackle single-label classification where each instance contains one input  $x_{i}$ , one relevant (positive) passage  $x_{i}^{+}$ , along with  $n$  irrelevant (negative) passages  $p_{i,j}^{-}$ , we minimize the negative log likelihood of the positive label description",
+        "bbox": [
+            112,
+            387,
+            489,
+            467
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} L \\left(x _ {i}, p _ {i} ^ {+}, p _ {i, 1} ^ {-}, \\dots , p _ {i, n} ^ {-}\\right) \\tag {1} \\\\ = - \\log \\frac {e ^ {\\operatorname {s i m} \\left(x _ {i} , p _ {i} ^ {+}\\right)}}{e ^ {\\operatorname {s i m} \\left(x _ {i} , p _ {i} ^ {+}\\right)} + \\sum_ {j = 1} ^ {n} e ^ {\\operatorname {s i m} \\left(x _ {i} , p _ {i , j} ^ {-}\\right)}}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            147,
+            480,
+            487,
+            546
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We extend to multi-label classification tasks where each input has multiple positive passages and more than one of them participates in model update simultaneously. Accordingly, the cross entropy loss is minimized to encourage similarity between  $m$  positive pairs instead:",
+        "bbox": [
+            112,
+            556,
+            487,
+            651
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} L \\left(x _ {i}, p _ {i, 1} ^ {+}, \\dots , p _ {i, m} ^ {+}, p _ {i, 1} ^ {-}, \\dots , p _ {i, n} ^ {-}\\right) \\tag {2} \\\\ = - \\sum_ {k = 1} ^ {m} \\log \\frac {e ^ {\\mathrm {s i m} (x _ {i} , p _ {i , k} ^ {+})}}{\\sum_ {j = 1} ^ {m} e ^ {\\mathrm {s i m} (x _ {i} , p _ {i , j} ^ {+})} + \\sum_ {j = 1} ^ {n} e ^ {\\mathrm {s i m} (x _ {i} , p _ {i , j} ^ {-})}}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            114,
+            664,
+            500,
+            733
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Positive, negative and hard negative passages: the decision thesaurus entries that describe true labels in thesaurus for each input are deemed as positive. In-batch negatives are positive entries for other input in the same mini-batch, making the computation more efficient while achieving better performance compared with random passages or other negative entries (Gillick et al., 2019;",
+        "bbox": [
+            112,
+            743,
+            487,
+            871
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Karpukhin et al., 2020). If not otherwise specified, we train with positive and in-batch negatives to obtain a weak  $DDR$  model firstly; we then retrieve label thesaurus with the weak model for training examples and leverage wrongly predicted label description as hard negatives, which are augmented to training data to obtain a stronger  $DDR$ .",
+        "bbox": [
+            507,
+            84,
+            882,
+            197
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We note that, different from open-domain QA where one positive pair can be considered per time with Eq. 1, considering multiple positive pairs is very important for tasks where the decision space though large but follows highly skewed distributions, when in-batch negatives are adopted for model update. In the mini-batch, multiple inputs may share same labels that are popular in the head of long-tail label distribution. In this case, these input instances have multiple positive passages: one randomly sampled from label passages as normal and some from positives of other in-batch inputs.",
+        "bbox": [
+            507,
+            198,
+            882,
+            391
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inference After training is completed, we encode all passages in label thesaurus once with  $E_{p}$  and index them using FAISS offline (Johnson et al., 2019). Given a new input, we obtain its embedding with  $E_{x}$  and retrieve the top- $k$  passages with embeddings closest to that of input.",
+        "bbox": [
+            507,
+            394,
+            882,
+            491
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            504,
+            653,
+            520
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We evaluate DDR on three NLU tasks with large deiction spaces in two criteria, 1) a minority of label space is observed during training: Extreme Multi-label Classification (§4.1) and Entity Typing (§4.2), or 2) limited amounts of examples are available per label: Intent Classification (§4.3). We investigate decision spaces of up to hundreds of thousands. Dataset statistics are shown in Appx. Tab. 1.",
+        "bbox": [
+            507,
+            530,
+            882,
+            675
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Extreme Multi-label Classification",
+        "text_level": 1,
+        "bbox": [
+            507,
+            687,
+            823,
+            702
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Task Many real-world applications can be formulated as eXtreme Multi-label Classification (XMC) problems, where relevant labels from the set of an extremely large size are expected to be predicted given text input. Considering the ever-growing label set with newly added products or websites and the time-consuming example-label pairs collection, we follow the few-shot XMC setting (Gupta et al., 2021; Xiong et al., 2022) to choose relevant labels from all available seen and unseen labels. $^6$",
+        "bbox": [
+            507,
+            712,
+            882,
+            872
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>6</sup>By randomly sampling  $1\\%$  and  $5\\%$  labels from the large decision space, we only use their instances for model training and predict on the test set covering the whole label space.",
+        "bbox": [
+            507,
+            879,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "For example, in cases where a BERT model is used, the [CLS] token is used for representation of input text and the decision thesaurus entries",
+        "bbox": [
+            112,
+            879,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "15024",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/abc78cd58aa2bf73599e6b9ad822dc843aa509eed166241c302c42534fea7ebf.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Setting</td><td>Train</td><td>Dev</td><td>Test</td><td>Labels</td></tr><tr><td colspan=\"6\">Extreme Multi-label Classification</td></tr><tr><td rowspan=\"2\">LF-Amazon-131K</td><td>1%</td><td>6227</td><td>-</td><td rowspan=\"2\">134,835</td><td rowspan=\"2\">131,073</td></tr><tr><td>5%</td><td>29,719</td><td>-</td></tr><tr><td rowspan=\"2\">LF-WikiSeeAlso-320K</td><td>1%</td><td>13,920</td><td>-</td><td rowspan=\"2\">177,515</td><td rowspan=\"2\">312,330</td></tr><tr><td>5%</td><td>62,704</td><td>-</td></tr><tr><td colspan=\"6\">Ultra-fine Entity Typing</td></tr><tr><td>UFET</td><td>Crowd</td><td>1998</td><td>1998</td><td>1998</td><td>10,331</td></tr><tr><td colspan=\"6\">Intent Classification</td></tr><tr><td rowspan=\"2\">BANKING</td><td>5-shot</td><td>385</td><td rowspan=\"2\">1540</td><td rowspan=\"2\">3080</td><td rowspan=\"2\">77</td></tr><tr><td>10-shot</td><td>770</td></tr><tr><td rowspan=\"2\">HWU</td><td>5-shot</td><td>320</td><td rowspan=\"2\">1076</td><td rowspan=\"2\">1076</td><td rowspan=\"2\">64</td></tr><tr><td>10-shot</td><td>640</td></tr><tr><td rowspan=\"2\">CLINC</td><td>5-shot</td><td>750</td><td rowspan=\"2\">3000</td><td rowspan=\"2\">4500</td><td rowspan=\"2\">150</td></tr><tr><td>10-shot</td><td>1500</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            495,
+            331
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Statistics and experimental settings on datasets with label spaces ranging from 64 to  ${320}\\mathrm{\\;K}$  .",
+        "bbox": [
+            112,
+            338,
+            487,
+            369
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Metrics The public benchmark dataset LF-Amazon-131K collects pairs of relevant products with textual descriptions for commercial recommendation, while LF-WikiSeeAlso-320K<sup>7</sup> aims to link relevant Wikipedia passages for reference (Bhatia et al., 2016; Gupta et al., 2021).",
+        "bbox": [
+            112,
+            397,
+            489,
+            493
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use metrics widely adopted by XMC (Chien et al., 2023) and IR (Thakur et al., 2021) literature: precision@k for the proportion of the top-  $k$  predicted labels to be true labels, and recall@k for the proportion of true labels found in the top-  $k$  predictions.",
+        "bbox": [
+            112,
+            495,
+            489,
+            592
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baselines We compare DDR with strong XMC baselines in three categories: 1) Transformer-based models that learn semantically meaningful sentence embeddings with siamese or triplet network structures, including Sentence-BERT (Reimers and Gurevych, 2019) and MPNet (Song et al., 2020); 2) competitive methods originally proposed for scalable and accurate predictions over extremely large label spaces, including XR-Linear (Yu et al., 2022) that recursively learns to traverse an input from the root of a hierarchical label tree to a few leaf node clusters, Astec (Dahiya et al., 2021b) with four sub-tasks for varying trade-offs between accuracy",
+        "bbox": [
+            112,
+            596,
+            489,
+            806
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "and scalability, and SiameseXML (Dahiya et al., 2021a) based on a novel probabilistic model to meld Siamese architectures with high-capacity extreme classifiers; 3) methods specifically designed to improve performance with few-shot labels available, including ZestXML (Gupta et al., 2021) that learns to project a data point's features close to the features of its relevant labels through a highly sparsified linear transform, and MACLR (Xiong et al., 2022) that pre-trains Transformer-based encoders with a self-supervised contrastive loss.",
+        "bbox": [
+            507,
+            84,
+            884,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details The label space of LF-WikiSeeAlso-320K is composed of relevant passage titles, hence the corresponding full content should be available in Wikipedia and we use them as the label thesaurus. For relevant product advertisement and recommendation, concrete product description provides sufficient information as the label thesaurus for LF-Amazon-131K. We extract product textual descriptions mainly from Amazon data in XMC benchmark (Bhatia et al., 2016) and public Amazon product meta datasets (Ni et al., 2019), leading to  $95.67\\%$  of labels well documented. As introduced in  $\\S 3.2$ , we then use ChatGPT to obtain relevant information for remaining undocumented labels.",
+        "bbox": [
+            507,
+            266,
+            884,
+            507
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We perform label thesaurus construction on Intel(R) Xeon(R) Gold 5217 CPU @ 3.00GHz with 32 CPUs and 8 cores per CPU. It takes 59.12 minutes to construct label thesaurus for LF-WikiSeeAlso-320K, while 28.24 minutes for LF-Amazon-131K. We leverage the same infrastructure to prepare label thesaurus in following tasks unless otherwise specified.",
+        "bbox": [
+            507,
+            508,
+            885,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Due to a lack of explicit knowledge about the connection between the majority of labels and their examples for training, DDR only learns to distinguish positive example-label pairs from in-batch negatives without further negative mining. For inference, with the constructed label thesaurus, we first retrieve the large label space with a test example based on Sentence-BERT embeddings of the title and the first sentence in an unsupervised way. After training DDR with few-shot labels, we mimic the practice of DPR+BM25 for open-",
+        "bbox": [
+            507,
+            640,
+            885,
+            816
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "${}^{7}$  DPR was trained on the clean, text-partition of articles from the Wikipedia dump for question answering. Lists such as the \"See also\" section have been removed during preprocessing. Hence example-label pairs in LF-WikiSeeAlso-320K are unseen during DPR training and it is fair to fine-tune DPR on this task.",
+        "bbox": [
+            112,
+            820,
+            487,
+            891
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "8Following (Xiong et al., 2022), we use  $k = \\{1,3,5\\}$  for precision@k and  $k = \\{1,3,5,10,100\\}$  for recall@k.",
+        "bbox": [
+            112,
+            892,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>9</sup>Since the specific dump version is unclear, we look for different resources (e.g., latest dump, train and test passages, Wikipedia API search) and cover 99.95% of labels in the end.",
+        "bbox": [
+            507,
+            831,
+            882,
+            866
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{10}$ https://chat.openai.com/",
+        "bbox": [
+            532,
+            866,
+            729,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "11 Similarly, for following tasks, we obtain label information by prompting LLMs if the label definition or description is unavailable from existing resources.",
+        "bbox": [
+            507,
+            881,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "15025",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/e42e56058994447c9c6ae8a4d38009457238b24e3b987ce2997a25ed5c42fc19.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"8\">1% Labels</td><td colspan=\"7\">5% Labels</td><td></td></tr><tr><td>Precision@1</td><td>@3</td><td>@5</td><td>Recall@1</td><td>@3</td><td>@5</td><td>@10</td><td>@100</td><td>Precision@1</td><td>@3</td><td>@5</td><td>@1</td><td>@3</td><td>@5</td><td>@10</td><td>@100</td></tr><tr><td colspan=\"16\">LF-Amazon-131K</td><td></td></tr><tr><td>XR-Linear</td><td>1.53</td><td>0.57</td><td>0.36</td><td>0.67</td><td>0.75</td><td>0.78</td><td>0.81</td><td>0.92</td><td>5.09</td><td>2.09</td><td>1.32</td><td>2.36</td><td>2.86</td><td>3.02</td><td>3.18</td><td>3.74</td></tr><tr><td>Astec</td><td>0.94</td><td>0.44</td><td>0.29</td><td>0.55</td><td>0.78</td><td>0.84</td><td>0.91</td><td>1.13</td><td>3.94</td><td>1.92</td><td>1.26</td><td>2.31</td><td>2.34</td><td>3.66</td><td>4.00</td><td>4.96</td></tr><tr><td>SiameseXML</td><td>1.45</td><td>0.56</td><td>0.35</td><td>0.84</td><td>0.96</td><td>1.00</td><td>1.03</td><td>1.16</td><td>5.36</td><td>2.23</td><td>1.41</td><td>3.15</td><td>3.89</td><td>4.08</td><td>4.27</td><td>4.82</td></tr><tr><td>ZestXML</td><td>10.10</td><td>9.19</td><td>7.34</td><td>5.63</td><td>14.46</td><td>18.61</td><td>23.73</td><td>32.69</td><td>12.33</td><td>10.00</td><td>8.71</td><td>6.84</td><td>17.19</td><td>21.97</td><td>28.19</td><td>46.49</td></tr><tr><td>Sentence-BERT</td><td>12.64</td><td>9.82</td><td>7.80</td><td>6.97</td><td>15.34</td><td>19.74</td><td>25.33</td><td>43.53</td><td>15.47</td><td>12.24</td><td>9.64</td><td>8.63</td><td>19.23</td><td>24.40</td><td>30.82</td><td>49.22</td></tr><tr><td>MPNet</td><td>14.78</td><td>11.55</td><td>8.97</td><td>8.28</td><td>18.24</td><td>22.84</td><td>28.54</td><td>45.89</td><td>15.03</td><td>11.88</td><td>9.28</td><td>8.47</td><td>18.74</td><td>23.69</td><td>29.93</td><td>48.84</td></tr><tr><td>MACLR</td><td>18.74</td><td>16.07</td><td>12.52</td><td>10.73</td><td>25.44</td><td>31.89</td><td>39.17</td><td>57.55</td><td>19.56</td><td>16.19</td><td>12.64</td><td>11.15</td><td>25.65</td><td>32.18</td><td>39.63</td><td>58.45</td></tr><tr><td>DDR (w/o thesaurus)</td><td>15.39</td><td>10.92</td><td>7.87</td><td>8.25</td><td>19.15</td><td>24.10</td><td>31.51</td><td>54.32</td><td>21.21</td><td>13.77</td><td>10.23</td><td>12.02</td><td>24.18</td><td>30.29</td><td>38.19</td><td>58.75</td></tr><tr><td>DDR (scratch)</td><td>19.55</td><td>14.48</td><td>11.50</td><td>10.87</td><td>22.50</td><td>29.00</td><td>37.80</td><td>56.53</td><td>23.77</td><td>16.48</td><td>12.75</td><td>13.31</td><td>25.54</td><td>31.85</td><td>40.46</td><td>60.30</td></tr><tr><td>DDR (full)</td><td>22.39</td><td>16.38</td><td>12.40</td><td>12.57</td><td>25.64</td><td>31.44</td><td>38.44</td><td>57.72</td><td>25.22</td><td>17.96</td><td>13.63</td><td>14.18</td><td>27.94</td><td>34.40</td><td>41.72</td><td>61.17</td></tr><tr><td colspan=\"16\">LF-WikiSeeAlso-320K</td><td></td></tr><tr><td>XR-Linear</td><td>1.24</td><td>0.57</td><td>0.37</td><td>0.42</td><td>0.58</td><td>0.63</td><td>0.68</td><td>0.76</td><td>4.69</td><td>2.20</td><td>1.46</td><td>1.82</td><td>2.41</td><td>2.63</td><td>2.82</td><td>3.42</td></tr><tr><td>Astec</td><td>1.25</td><td>0.60</td><td>0.41</td><td>0.69</td><td>0.98</td><td>1.11</td><td>1.27</td><td>1.56</td><td>5.90</td><td>2.80</td><td>1.86</td><td>3.26</td><td>4.49</td><td>4.95</td><td>5.49</td><td>6.83</td></tr><tr><td>SiameseXML</td><td>1.81</td><td>0.75</td><td>0.48</td><td>1.03</td><td>1.26</td><td>1.33</td><td>1.41</td><td>1.67</td><td>6.83</td><td>3.15</td><td>2.06</td><td>3.88</td><td>5.15</td><td>5.56</td><td>6.02</td><td>7.09</td></tr><tr><td>ZestXML</td><td>8.74</td><td>6.78</td><td>5.41</td><td>4.68</td><td>9.70</td><td>12.21</td><td>15.73</td><td>24.98</td><td>10.06</td><td>8.11</td><td>6.60</td><td>5.33</td><td>11.49</td><td>14.74</td><td>19.57</td><td>40.46</td></tr><tr><td>Sentence-BERT</td><td>16.30</td><td>12.62</td><td>10.08</td><td>9.30</td><td>18.92</td><td>23.78</td><td>30.40</td><td>52.92</td><td>18.47</td><td>14.19</td><td>11.29</td><td>10.82</td><td>21.55</td><td>26.77</td><td>33.92</td><td>57.02</td></tr><tr><td>MPNet</td><td>17.14</td><td>12.64</td><td>9.96</td><td>9.98</td><td>18.98</td><td>23.45</td><td>29.67</td><td>50.75</td><td>18.59</td><td>13.99</td><td>11.08</td><td>10.89</td><td>21.12</td><td>26.10</td><td>32.82</td><td>54.70</td></tr><tr><td>MACLR</td><td>19.09</td><td>14.57</td><td>11.53</td><td>11.39</td><td>22.34</td><td>27.63</td><td>34.81</td><td>57.92</td><td>20.99</td><td>15.57</td><td>12.26</td><td>12.59</td><td>23.94</td><td>29.41</td><td>36.78</td><td>59.81</td></tr><tr><td>DDR (w/o thesaurus)</td><td>17.52</td><td>15.93</td><td>14.23</td><td>4.89</td><td>13.11</td><td>18.88</td><td>25.93</td><td>46.93</td><td>18.14</td><td>16.65</td><td>15.39</td><td>5.12</td><td>14.10</td><td>21.81</td><td>30.00</td><td>52.54</td></tr><tr><td>DDR (scratch)</td><td>21.26</td><td>15.81</td><td>13.08</td><td>11.16</td><td>22.52</td><td>29.55</td><td>40.14</td><td>64.94</td><td>25.04</td><td>18.95</td><td>15.05</td><td>12.46</td><td>25.70</td><td>32.18</td><td>40.24</td><td>66.22</td></tr><tr><td>DDR (full)</td><td>24.70</td><td>18.45</td><td>14.69</td><td>13.04</td><td>26.31</td><td>32.99</td><td>41.08</td><td>65.39</td><td>27.78</td><td>19.98</td><td>15.66</td><td>15.03</td><td>28.87</td><td>35.47</td><td>43.35</td><td>68.66</td></tr></table>",
+        "bbox": [
+            112,
+            80,
+            884,
+            422
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Results of few-shot XMC where the training subset covers  $1\\%$  (left) and  $5\\%$  (right) labels from the whole set. DDR outperforms the second best MACLR in both settings of two datasets. Indirect supervision from DPR boosts performance against training from scratch (the second row from the bottom), while label thesaurus construction improves accuracy over those using textual label names (the third row from the bottom).",
+        "bbox": [
+            112,
+            431,
+            882,
+            491
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "domain QA (Karpukhin et al., 2020) for reranking, where prediction scores on test set from  $DDR$  and Sentence-BERT are linearly combined as the final ranking score.",
+        "bbox": [
+            112,
+            514,
+            487,
+            579
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results We report performance of DDR with(out) indirect supervision from dense retrieval for open-domain QA in Tab. 2. On average, DDR boosts performance on LF-Amazon-131K by  $8.50\\%$  and LF-WikiSeeAlso-320K by  $21.73\\%$  compared with the second best baseline MACLR. When only  $1\\%$  labels are seen during training, we observe that DDR shows much better transferability than other pre-trained models, e.g., average performance improved by  $12.59\\%$  over MACLR and  $39.80\\%$  over MPNet. As more labels are available for fine-tuning the model, the performance gain from DDR over others becomes even more significant:  $17.64\\%$  over MACLR and  $43.38\\%$  over MPNet under  $5\\%$  labels setting.",
+        "bbox": [
+            112,
+            583,
+            489,
+            828
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Ultra-fine Entity Typing",
+        "text_level": 1,
+        "bbox": [
+            112,
+            843,
+            352,
+            859
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Task Entities can often be described by very fine grained-types (Choi et al., 2018) and the ultra-fine entity typing task aims at predicting one or more",
+        "bbox": [
+            112,
+            870,
+            487,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "fine-grained words or phrases that describe the type(s) of that specific mention (Xu et al., 2022). Consider the sentence \"He had blamed Israel for failing to implement its agreements.\" Besides person and male, the mention \"He\" has other very specific types that can be inferred from the context, such as leader or official for the \"blamed\" behavior and \"implement its agreements\" affair. Ultra-fine entity typing has a broad impact on various NLP tasks that depend on type understanding, including coreference resolution (Onoe and Durrett, 2020), event detection (Le and Nguyen, 2021) and relation extraction (Zhou and Chen, 2022).",
+        "bbox": [
+            507,
+            514,
+            884,
+            724
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Metrics We leverage the UFET dataset (Choi et al., 2018) to evaluate benefits of DDR with indirect supervision from dense retrieval. Among 10,331 entity types, accurately selecting finer labels (121 labels such as engineer) was more challenging to predict than coarse-grained labels (9 labels such as person), and this issue is exacerbated when dealing with ultra-fine types (10,201 labels such as flight engineer). Following recent entity typing literature (Li et al., 2022; Du et al., 2022), we train DDR on (originally provided) limited",
+        "bbox": [
+            507,
+            741,
+            884,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "15026",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/c5eee48e820b5eb42cee7a84f12c02290df1960b17be5d59dd684da70bd171ab.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td>BiLSTM</td><td>48.1</td><td>23.2</td><td>31.3</td></tr><tr><td>LabelGCN</td><td>50.3</td><td>29.2</td><td>36.9</td></tr><tr><td>LDET</td><td>51.5</td><td>33.0</td><td>40.1</td></tr><tr><td>Box4Types</td><td>52.8</td><td>38.8</td><td>44.8</td></tr><tr><td>LRN</td><td>54.5</td><td>38.9</td><td>45.4</td></tr><tr><td>UniST</td><td>50.2</td><td>49.6</td><td>49.9</td></tr><tr><td>MLMET</td><td>53.6</td><td>45.3</td><td>49.1</td></tr><tr><td>LITE</td><td>52.4</td><td>48.9</td><td>50.6</td></tr><tr><td>Context-TE</td><td>53.7</td><td>49.4</td><td>51.5</td></tr><tr><td>DDR (w/o thesaurus)</td><td>51.6</td><td>46.0</td><td>48.6</td></tr><tr><td>DDR (scratch)</td><td>53.0</td><td>50.5</td><td>51.7</td></tr><tr><td>DDR (full)</td><td>51.9</td><td>52.3</td><td>52.1</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            482,
+            259
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Results of UFET task. DDR outperforms competitive baselines with (upper 5 methods) or without (lower 3 methods) inductive bias from task pre-training.",
+        "bbox": [
+            112,
+            269,
+            487,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "crowd-sourced examples without relying on distant resources such as knowledge bases or head words from the Gigaword corpus. We follow prior studies (Choi et al., 2018) to evaluate macro-averaged precision, recall and F1.",
+        "bbox": [
+            112,
+            343,
+            487,
+            423
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Baselines We consider two categories of competitive entity typing models as baselines: 1) methods capturing the example-label and label-label relations, e.g., BiLSTM (Choi et al., 2018) that concatenates the context representation learned by a bidirectional LSTM and the mention representation learned by a CNN, LabelGCN (Xiong et al., 2019) that learns to encode global label co-occurrence statistics and their word-level similarities, LRN (Liu et al., 2021a) that models the coarse-to-fine label dependency as causal chains, Box4Types (Onoe et al., 2021) that captures hierarchies of types as topological relations of boxes, and UniST (Huang et al., 2022a) that conduct name-based label ranking; 2) methods leveraging inductive bias from pre-trained models for entity typing, e.g., MLMET (Dai et al., 2021) that utilizes the pretrained BERT to predict the most probable words for “[MASK]” earlier incorporated around the mention as type labels, LITE (Li et al., 2022) and Context-TE (Du et al., 2022) that both leverage indirect supervision from pre-trained natural language inference.",
+        "bbox": [
+            115,
+            430,
+            489,
+            800
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details The dataset UFET first asked crowd workers to annotate entity's types and then used WordNet (Miller, 1995) to expand these types automatically by generating all their synonyms and hypernyms based on the most common sense. Therefore, we automatically obtain label thesaurus entries from definitions and examples in",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "WordNet and Wiktionary, which covers  $99.99\\%$  of the whole label set. The time cost for prior label thesaurus construction is around 2.07 minutes.",
+        "bbox": [
+            507,
+            84,
+            882,
+            131
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "initialized with the original DPR checkpoint pretrained on open-domain QA datasets, DDR firstly optimizes the model given positive example-label pairs and in-batch negatives. With the fine-tuned model, we then perform dense retrieval on the label set for each training example, keeping label documents with high scores but not in the true label set as hard negatives. DDR further updates the model with these additional hard negatives. For inference of multi-label entity type classification, we adopt labels with retrieval scores higher than a threshold that leads to the best F1 score on the development set.",
+        "bbox": [
+            507,
+            133,
+            884,
+            341
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Results In Tab. 3, we show performance of DDR and other entity typing methods. DDR obtains the state-of-the-art F1 score over the best baseline training from scratch (LRN) by 6.7 and the best baseline with inductive bias from the language model (Context-TE) by 0.6.",
+        "bbox": [
+            507,
+            346,
+            882,
+            443
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Few-shot Intent Classification",
+        "text_level": 1,
+        "bbox": [
+            507,
+            455,
+            791,
+            469
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Task As a fundamental element in task-oriented dialog systems, intent detection is normally conducted in the NLU component for identifying a user's intent given an utterance (Ham et al., 2020). Recently, accurately identifying intents in the few-shot setting has attracted much attention due to data scarcity issues resulted from the cost of data collection as well as privacy and ethical concerns (Lin et al., 2023b). Following the few-shot intent detection benchmark (Zhang et al., 2022), we focus on the challenging 5-shot and 10-shot settings.",
+        "bbox": [
+            507,
+            480,
+            882,
+            656
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Metrics To evaluate the effectiveness of DDR for NLU with large decision spaces, we pick three challenging intent datasets with a relatively large number of semantically similar intent labels. Banking77 (Casanueva et al., 2020) is a single-domain dataset that provides very fine-grained 77 intents in a Banking domain. HWU64 (Liu et al., 2021b) is a multi-domain (21 domains) dataset recorded by a home assistant robot including 64 intents ranging from setting alarms, playing music, search, to movie recommendation. CLINC150 (Larson et al., 2019) prompts crowd workers to provide questions or commands in the manner they would interact with an artificially intelligent assistant covering 150 intent classes over 10 domains. We report accuracy",
+        "bbox": [
+            507,
+            661,
+            884,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "15027",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/c1cbe7cd96f4871c60d27fdcbffdba4297b9a14eaa995fb06ad3d0749871faed.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Banking77</td><td colspan=\"2\">HWU64</td><td colspan=\"2\">CLINC150</td></tr><tr><td>5-shot</td><td>10-shot</td><td>5-shot</td><td>10-shot</td><td>5-shot</td><td>10-shot</td></tr><tr><td colspan=\"7\">w/ Data Augmentation</td></tr><tr><td>ICDA-X S (+1x)</td><td>80.29</td><td>86.72</td><td>81.32</td><td>85.59</td><td>91.16</td><td>93.71</td></tr><tr><td>ICDA-S (+4x)</td><td>81.95</td><td>87.37</td><td>81.97</td><td>86.25</td><td>91.22</td><td>93.98</td></tr><tr><td>ICDA-M (+16x)</td><td>84.01</td><td>88.64</td><td>81.84</td><td>87.36</td><td>91.93</td><td>94.71</td></tr><tr><td>ICDA-L (+64x)</td><td>83.90</td><td>89.12</td><td>81.97</td><td>86.94</td><td>92.41</td><td>94.73</td></tr><tr><td>ICDA-X L (+128x)</td><td>83.90</td><td>89.79</td><td>82.45</td><td>87.41</td><td>92.62</td><td>94.84</td></tr><tr><td colspan=\"7\">w/o Data Augmentation</td></tr><tr><td>USE</td><td>76.29</td><td>84.23</td><td>77.79</td><td>83.75</td><td>87.82</td><td>90.85</td></tr><tr><td>ConveRT</td><td>75.32</td><td>83.32</td><td>76.95</td><td>82.65</td><td>89.22</td><td>92.62</td></tr><tr><td>USE+ConveRT</td><td>77.75</td><td>85.19</td><td>80.01</td><td>85.83</td><td>90.49</td><td>93.26</td></tr><tr><td>ConvBERT</td><td>-</td><td>83.63</td><td>-</td><td>83.77</td><td>-</td><td>92.10</td></tr><tr><td>+ MLM</td><td>-</td><td>83.99</td><td>-</td><td>84.52</td><td>-</td><td>92.75</td></tr><tr><td>+ MLM + Example</td><td>-</td><td>84.09</td><td>-</td><td>83.44</td><td>-</td><td>92.35</td></tr><tr><td>+ Combined</td><td>-</td><td>85.95</td><td>-</td><td>86.28</td><td>-</td><td>93.97</td></tr><tr><td>DNNC</td><td>80.40</td><td>86.71</td><td>80.46</td><td>84.72</td><td>91.02</td><td>93.76</td></tr><tr><td>CPFT</td><td>80.86</td><td>87.20</td><td>82.03</td><td>87.13</td><td>92.34</td><td>94.18</td></tr><tr><td>- Pretraining</td><td>76.75</td><td>84.83</td><td>76.02</td><td>82.96</td><td>88.19</td><td>91.55</td></tr><tr><td>Context-TE</td><td>80.76</td><td>85.53</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DDR (w/o thesaurus)</td><td>78.86</td><td>84.16</td><td>80.30</td><td>83.74</td><td>88.86</td><td>91.58</td></tr><tr><td>DDR (scratch)</td><td>82.11</td><td>85.97</td><td>80.76</td><td>84.71</td><td>88.58</td><td>91.16</td></tr><tr><td>DDR (full)</td><td>83.86 (XL)</td><td>88.25 (M)</td><td>84.29 (XL)</td><td>86.34 (S)</td><td>92.71 (XL)</td><td>94.58 (M)</td></tr></table>",
+        "bbox": [
+            166,
+            80,
+            828,
+            391
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Few-shot Intent Detection Accuracy on three benchmark datasets. DDR achieves much higher accuracy than existing strong learning baselines without data augmentation, while competitive with ICDA requiring a considerable amount of augmented data. ICDA prepares additional synthetic data with the scale ranging from the same amount (ICDA-  $XS$ ) to 128 times (ICDA-  $XL$ ) of the original few-shot train size. In the bottom row, we also mark performance of DDR with the most comparable ICDA variant. Results not available in original papers are marked as  $-$ .",
+        "bbox": [
+            112,
+            401,
+            882,
+            488
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "for this single-label classification task.",
+        "bbox": [
+            112,
+            512,
+            401,
+            527
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Baselines There are two families of intent learning algorithms to cope with limited amounts of example-label pairs: 1) Classifiers based on sentence embeddings from PLMs, including USE (Yang et al., 2020) embed sentences from 16 languages into a single semantic space using a multi-task trained dual-encoder, ConvBERT (Henderson et al., 2020) that pretrains a dual-encoder model on Reddit Data by maximizing the similarity score of embeddings of positive input-response pairs; ConvBERT (Mehri et al., 2020) that adopts a BERT-base model pre-trained on a large open-domain dialogue corpus; 2) methods leverage semantic similarity between utterances from different users, e.g. ConvBERT+Combined (Mehri and Eric, 2021) that extends the ConvBERT model with task-adaptive masked language modelling (MLM) and infers the intent of the utterance based on the similarity to the examples corresponding to each intent, DNNC (Zhang et al., 2020a) that leverages BERT-style pairwise encoding to train a binary classifier that estimates the best matched training example for a user input, CPFT (Zhang et al., 2021) that conducts contrastive pre-training on example sentences",
+        "bbox": [
+            112,
+            533,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "from six public intent datasets to discriminate semantically similar utterances.",
+        "bbox": [
+            507,
+            512,
+            884,
+            543
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Different from DDR and prior baselines, ICDA (Lin et al., 2023a), a most recent work tackles the challenge of limited data for intent detection by generating high-quality synthetic training data. It first fine-tunes a PLM on a small seed of training data for new data point synthesisation, and then employs pointwise V-information (PVI) based filtering to remove unhelpful ones.[12]",
+        "bbox": [
+            507,
+            546,
+            884,
+            675
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details Based on the observation that utterances labeled by identical intents share very similar sentence-level semantic meanings, we first utilize the whole set of unlabeled training examples to represent their corresponding pseudo labels predicted by Sentence-BERT (Reimers and Gurevych, 2019). We perform prediction on a single NVIDIA RTX A5000 GPU and it takes 4.17 minutes to complete label thesaurus construction for Banking77, 6.45 minutes for HWU64 and 7.61",
+        "bbox": [
+            507,
+            681,
+            884,
+            841
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "12 Although the learning and augmentation direction are orthogonal hence incommensurable, we display results with augmented training data as well to reflect the advantage of DDR over ICDA in terms of their required scale of augmented synthetic data to achieve similar good performance.",
+        "bbox": [
+            507,
+            857,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "15028",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/6c49bb824f9dadce8dc07a70685a3df96e4d0b259171e5f11316bed226fd6620.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Setting</td><td>Sentence-BERT (1st round)</td><td>DDR (2nd round)</td></tr><tr><td rowspan=\"2\">Banking77</td><td>5-shot</td><td>84.29</td><td>83.86</td></tr><tr><td>10-shot</td><td>88.02</td><td>88.25</td></tr><tr><td rowspan=\"2\">HWU64</td><td>5-shot</td><td>82.81</td><td>84.29</td></tr><tr><td>10-shot</td><td>85.69</td><td>86.34</td></tr><tr><td rowspan=\"2\">CLINC150</td><td>5-shot</td><td>91.82</td><td>92.71</td></tr><tr><td>10-shot</td><td>94.07</td><td>94.58</td></tr></table>",
+        "bbox": [
+            136,
+            80,
+            463,
+            206
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Impact on intent detection from label thesaurus constructed with predictions from Sentence-BERT and DDR for the unlabeled whole set of the training set.",
+        "bbox": [
+            112,
+            217,
+            487,
+            260
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "minutes for CLINC150. After training DDR with this first version of label thesaurus, we then update pseudo labels for each training example with predictions from DDR for the second-round training.",
+        "bbox": [
+            112,
+            288,
+            487,
+            351
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Regardless of the basis for label thesaurus construction, DDR is trained for two phases in each round: only few-shot positive example-label pairs and in-batch negatives are used in the first phase, while labels wrongly predicted for training examples from the prior phase are used as additional hard negatives in the second phase. After the second round of training, the latest DDR makes predictions on the whole set of training examples for the final label thesaurus construction, from which DDR retrieves the label with the highest score as the prediction.",
+        "bbox": [
+            115,
+            354,
+            489,
+            546
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Results Tab. 4 presents performance of different intent detection methods on 5- and 10-shot setting. Without extra training data synthesis, we observe significant performance gain from the proposed DDR over existing baselines. Although ICDA obtains enhanced accuracy by increasing data augmentation scale, DDR is able to achieve comparable performance without crafting 4x even 128x synthetic instances. In Tab. 5, we additionally study the quality of constructed label thesaurus in accordance with the prediction on the unlabeled whole training set from Sentence-BERT and DDR after the first round of training. We find that with a more accurate pseudo connection between example and label leads to a higher quality of label thesaurus construction, and finally benefits intent detection.",
+        "bbox": [
+            112,
+            552,
+            489,
+            810
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusions",
+        "text_level": 1,
+        "bbox": [
+            112,
+            826,
+            253,
+            841
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we focus on discriminative NLU tasks with large decision spaces. By reformulating these tasks as learning-to-retrieve tasks, we are able to leverage rich indirect supervision signals",
+        "bbox": [
+            112,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "from dense retrieval. Moreover, by representing decision spaces with thesaurus, we provide rich semantic knowledge of each label to improve the understanding of labels for dense decision retrievers. Experiments on 6 benchmark datasets show the effectiveness of our method on decision spaces scaling from hundreds of candidates to hundreds of thousands of candidates. Future work can extend our method to more large-space decision making tasks, especially in the low-resource setting.",
+        "bbox": [
+            507,
+            84,
+            884,
+            244
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            256,
+            672,
+            273
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We appreciate the reviewers for their insightful comments and suggestions. The logo for  $DDR$  used in this paper is sourced from the Wikipedia page $^{13}$  under CC BY-SA 3.0, for which we appreciate the contribution of the Wikipedia community.",
+        "bbox": [
+            507,
+            282,
+            882,
+            363
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Nan Xu is supported by the USC PhD Fellowship. Fei Wang is supported by the Annenberg Fellowship and the Amazon ML Fellowship. Mingtao Dong is supported by the Provost's Research Fellowship. Muhao Chen is supported by the NSF Grant IIS 2105329, the NSF Grant ITE 2333736, the DARPA MCS program under Contract No. N660011924033 with the United States Office Of Naval Research, a Cisco Research Award, two Amazon Research Awards, and a Keston Research Award. The computing of this work has been partly supported by a subaward of NSF Cloudbank 1925001 through UCSD.",
+        "bbox": [
+            507,
+            363,
+            884,
+            574
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            586,
+            613,
+            602
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The proposed DDR leverages the dual-encoder architecture with the inner dot product to compute embedding similarity, obtaining state-of-the-art performance on three investigated challenging large-space decision making tasks. More expressive models for embedding input and label thesaurus, such as joint encoding by a cross-encoder is not discussed. Moreover, other ways to model the connection between input and label thesaurus entry, such as using a sequence-to-sequence language model to generate the label name given input text and label thesaurus entry, is not explored yet. We believe adopting more advanced dense retrieval algorithms can further promote performance for large-space decision making. We leave this as an exciting future direction.",
+        "bbox": [
+            507,
+            612,
+            882,
+            869
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_footnote",
+        "text": "13https://en.wikipedia.org/wiki/File: Dance_Dance_Revolution_dance_pad Icon.png",
+        "bbox": [
+            507,
+            891,
+            823,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "15029",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "K. Bhatia, K. Dahiya, H. Jain, P. Kar, A. Mittal, Y. Prabhu, and M. Varma. 2016. The extreme classification repository: Multi-label datasets and code.",
+            "Inigo Casanueva, Tadas Temčinas, Daniela Gerz, Matthew Henderson, and Ivan Vulić. 2020. Efficient intent detection with dual sentence encoders. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 38-45, Online. Association for Computational Linguistics.",
+            "Muhao Chen, Hongming Zhang, Haoyu Wang, and Dan Roth. 2020. What are you trying to do? semantic typing of event processes. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 531-542, Online. Association for Computational Linguistics.",
+            "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.",
+            "Eli Chien, Jiong Zhang, Cho-Jui Hsieh, Jyun-Yu Jiang, Wei-Cheng Chang, Olgica Milenkovic, and Hsiang-Fu Yu. 2023. Pina: Leveraging side information in extreme multi-label classification via predicted instance neighborhood aggregation. arXiv preprint arXiv:2305.12349.",
+            "Eunsol Choi, Omer Levy, Yejin Choi, and Luke Zettlemoyer. 2018. Ultra-fine entity typing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 87-96, Melbourne, Australia. Association for Computational Linguistics.",
+            "Kunal Dahiya, Ananye Agarwal, Deepak Saini, K Gururaj, Jian Jiao, Amit Singh, Sumeet Agarwal, Purushottam Kar, and Manik Varma. 2021a. SiameseXML: Siamese networks meet extreme classifiers with 100m labels. In International Conference on Machine Learning, pages 2330-2340. PMLR.",
+            "Kunal Dahiya, Deepak Saini, Anshul Mittal, Ankush Shaw, Kushal Dave, Akshay Soni, Himanshu Jain, Sumeet Agarwal, and Manik Varma. 2021b. Deepxml: A deep extreme multi-label learning framework applied to short text documents. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining, pages 31-39.",
+            "Hongliang Dai, Yangqiu Song, and Haixun Wang. 2021. Ultra-fine entity typing with weak supervision from a masked language model. 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 1790-1799, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            108,
+            487,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "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.",
+            "Jiangshu Du, Wenpeng Yin, Congying Xia, and Philip S Yu. 2022. Learning to select from multiple options. arXiv preprint arXiv:2212.00301.",
+            "Xinya Du and Heng Ji. 2022. Retrieval-augmented generative question answering for event argument extraction. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4649-4666, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Nicholas FitzGerald, Julian Michael, Luheng He, and Luke Zettlemoyer. 2018. Large-scale QA-SRL parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2051-2060, Melbourne, Australia. Association for Computational Linguistics.",
+            "Daniel Gillick, Sayali Kulkarni, Larry Lansing, Alessandro Presta, Jason Baldridge, Eugene Ie, and Diego Garcia-Olano. 2019. Learning dense representations for entity retrieval. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 528-537, Hong Kong, China. Association for Computational Linguistics.",
+            "Nilesh Gupta, Sakina Bohra, Yashoteja Prabhu, Saurabh Purohit, and Manik Varma. 2021. Generalized zero-shot extreme multi-label learning. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 527-535.",
+            "Donghoon Ham, Jeong-Gwan Lee, Youngsoo Jang, and Kee-Eung Kim. 2020. End-to-end neural pipeline for goal-oriented dialogue systems using GPT-2. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 583-592, Online. Association for Computational Linguistics.",
+            "Hangfeng He, Mingyuan Zhang, Qiang Ning, and Dan Roth. 2021. Foreseeing the benefits of incidental supervision. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1782-1800, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Matthew Henderson, Iñigo Casanueva, Nikola Mrkšić, Pei-Hao Su, Tsung-Hsien Wen, and Ivan Vulić. 2020. ConveRT: Efficient and accurate conversational representations from transformers. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2161-2174, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "15030",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "I-Hung Hsu, Kuan-Hao Huang, Elizabeth Boschee, Scott Miller, Prem Natarajan, Kai-Wei Chang, and Nanyun Peng. 2022. DEGREE: A data-efficient generation-based event extraction model. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1890-1908, Seattle, United States. Association for Computational Linguistics.",
+            "James Y. Huang, Bangzheng Li, Jiashu Xu, and Muhao Chen. 2022a. Unified semantic typing with meaningful label inference. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2642-2654, Seattle, United States. Association for Computational Linguistics.",
+            "Kuan-Hao Huang, I-Hung Hsu, Prem Natarajan, Kai-Wei Chang, and Nanyun Peng. 2022b. Multilingual generative language models for zero-shot crosslingual event argument extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4633–4646, Dublin, Ireland. Association for Computational Linguistics.",
+            "Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547.",
+            "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online. Association for Computational Linguistics.",
+            "Stefan Larson, Anish Mahendran, Joseph J. Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K. Kummerfeld, Kevin Leach, Michael A. Laurenzano, Lingjia Tang, and Jason Mars. 2019. An evaluation dataset for intent classification and out-of-scope prediction. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1311-1316, Hong Kong, China. Association for Computational Linguistics.",
+            "Duong Le and Thien Huu Nguyen. 2021. Fine-grained event trigger detection. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2745-2752, Online. Association for Computational Linguistics.",
+            "Jinhyuk Lee, Mujeen Sung, Jaewoo Kang, and Danqi Chen. 2021. Learning dense representations of phrases at scale. 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"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Papers), pages 6634-6647, Online. Association for Computational Linguistics.",
+            "Bangzheng Li, Wenpeng Yin, and Muhao Chen. 2022. Ultra-fine entity typing with indirect supervision from natural language inference. Transactions of the Association for Computational Linguistics, 10:607-622.",
+            "Xiaoya Li, Jingrong Feng, Yuxian Meng, Qinghong Han, Fei Wu, and Jiwei Li. 2020. A unified MRC framework for named entity recognition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5849-5859, Online. Association for Computational Linguistics.",
+            "Yen-Ting Lin, Alexandros Papangelis, Seokhwan Kim, Sungjin Lee, Devamanyu Hazarika, Mahdi Namazifar, Di Jin, Yang Liu, and Dilek Hakkani-Tur. 2023a. Selective in-context data augmentation for intent detection using pointwise v-information. arXiv preprint arXiv:2302.05096.",
+            "Yen-Ting Lin, Alexandros Papangelis, Seokhwan Kim, Sungjin Lee, Devamanyu Hazarika, Mahdi Namazifar, Di Jin, Yang Liu, and Dilek Hakkani-Tur. 2023b. Selective in-context data augmentation for intent detection using pointwise V-information. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1463-1476, Dubrovnik, Croatia. Association for Computational Linguistics.",
+            "Qing Liu, Hongyu Lin, Xinyan Xiao, Xianpei Han, Le Sun, and Hua Wu. 2021a. Fine-grained entity typing via label reasoning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4611-4622, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Xingkun Liu, Arash Eshghi, Pawel Swietojanski, and Verena Rieser. 2021b. Benchmarking natural language understanding services for building conversational agents. In Increasing Naturalness and Flexibility in Spoken Dialogue Interaction: 10th International Workshop on Spoken Dialogue Systems, pages 165-183. Springer.",
+            "Keming Lu, I-Hung Hsu, Wenxuan Zhou, Mingyu Derek Ma, and Muhao Chen. 2022. Summarization as indirect supervision for relation extraction. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 6575-6594, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Qing Lyu, Hongming Zhang, Elior Sulem, and Dan Roth. 2021. Zero-shot event extraction via transfer learning: Challenges and insights. 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 322-332, Online. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "15031",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Shikib Mehri and Mihail Eric. 2021. Example-driven intent prediction with observers. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2979-2992, Online. Association for Computational Linguistics.",
+            "Shikib Mehri, Mihail Eric, and Dilek Hakkani-Tur. 2020. Dialogue: A natural language understanding benchmark for task-oriented dialogue. arXiv preprint arXiv:2009.13570.",
+            "George A Miller. 1995. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41.",
+            "Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 188-197, Hong Kong, China. Association for Computational Linguistics.",
+            "Yasumasa Onoe, Michael Boratko, Andrew McCallum, and Greg Durrett. 2021. Modeling fine-grained entity types with box embeddings. 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 2051-2064, Online. Association for Computational Linguistics.",
+            "Yasumasa Onoe and Greg Durrett. 2020. Interpretable entity representations through large-scale typing. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 612-624, Online. Association for Computational Linguistics.",
+            "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.",
+            "Dan Roth. 2017. Incidental supervision: Moving beyond supervised learning. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA, pages 4885-4890.",
+            "Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and TieYan Liu. 2020. Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems, 33:16857-16867.",
+            "Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).",
+            "Wei Wu, Fei Wang, Arianna Yuan, Fei Wu, and Jiwei Li. 2020. CorefQA: Coreference resolution as query-based span prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6953-6963, Online. Association for Computational Linguistics.",
+            "Wenhan Xiong, Jiawei Wu, Deren Lei, Mo Yu, Shiyu Chang, Xiaoxiao Guo, and William Yang Wang. 2019. Imposing label-relational inductive bias for extremely fine-grained entity typing. 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 773-784, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Yuanhao Xiong, Wei-Cheng Chang, Cho-Jui Hsieh, Hsiang-Fu Yu, and Inderjit Dhillon. 2022. Extreme Zero-Shot learning for extreme text classification. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5455-5468, Seattle, United States. Association for Computational Linguistics.",
+            "Jiashu Xu, Mingyu Derek Ma, and Muhao Chen. 2023. Can nli provide proper indirect supervision for low-resource biomedical relation extraction? In Proceedings of the 2023 Conference of the Association for Computational Linguistics: Human Language Technologies.",
+            "Nan Xu, Fei Wang, Bangzheng Li, Mingtao Dong, and Muhao Chen. 2022. Does your model classify entities reasonably? diagnosing and mitigating spurious correlations in entity typing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8642-8658, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Yinfei Yang, Daniel Cer, Amin Ahmad, Mandy Guo, Jax Law, Noah Constant, Gustavo Hernandez Abrego, Steve Yuan, Chris Tar, Yun-hsuan Sung, Brian Strope, and Ray Kurzweil. 2020. Multilingual universal sentence encoder for semantic retrieval. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 87-94, Online. Association for Computational Linguistics.",
+            "Wenpeng Yin, Muhao Chen, Ben Zhou, Qiang Ning, Kai-Wei Chang, and Dan Roth. 2023. Indirectly supervised natural language processing. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 6: Tutorial Abstracts), pages 32–40, Toronto, Canada. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "15032",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Wenpeng Yin, Jamaal Hay, and Dan Roth. 2019. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3914-3923, Hong Kong, China. Association for Computational Linguistics.",
+            "Hsiang-Fu Yu, Jiong Zhang, Wei-Cheng Chang, Jyun-Yu Jiang, Wei Li, and Cho-Jui Hsieh. 2022. Pecos: Prediction for enormous and correlated output spaces. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4848-4849.",
+            "Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1503-1512.",
+            "Jianguo Zhang, Trung Bui, Seunghyun Yoon, Xiang Chen, Zhiwei Liu, Congying Xia, Quan Hung Tran, Walter Chang, and Philip Yu. 2021. Few-shot intent detection via contrastive pre-training and fine-tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1906-1912, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Jianguo Zhang, Kazuma Hashimoto, Wenhao Liu, Chien-Sheng Wu, Yao Wan, Philip Yu, Richard Socher, and Caiming Xiong. 2020a. Discriminative nearest neighbor few-shot intent detection by transferring natural language inference. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5064-5082, Online. Association for Computational Linguistics.",
+            "Jianguo Zhang, Kazuma Hashimoto, Yao Wan, Zhiwei Liu, Ye Liu, Caiming Xiong, and Philip Yu. 2022. Are pre-trained transformers robust in intent classification? a missing ingredient in evaluation of out-of-scope intent detection. In Proceedings of the 4th Workshop on NLP for Conversational AI, pages 12-20, Dublin, Ireland. Association for Computational Linguistics.",
+            "Jianguo Zhang, Kazuma Hashimoto, Chien-Sheng Wu, Yao Wang, Philip Yu, Richard Socher, and Caiming Xiong. 2020b. Find or classify? dual strategy for slot-value predictions on multi-domain dialog state tracking. In Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics, pages 154-167, Barcelona, Spain (Online). Association for Computational Linguistics.",
+            "Ruohong Zhang, Yau-Shian Wang, Yiming Yang, Donghan Yu, Tom Vu, and Likun Lei. 2023. Long-tailed extreme multi-label text classification by the retrieval of generated pseudo label descriptions. In Findings of the Association for Computational Linguistics:"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "EACL 2023, pages 1092-1106, Dubrovnik, Croatia.  \nAssociation for Computational Linguistics.",
+            "Wenting Zhao, Konstantine Arkoudas, Weiqi Sun, and Claire Cardie. 2022a. Compositional task-oriented parsing as abstractive question answering. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4418-4427, Seattle, United States. Association for Computational Linguistics.",
+            "Zhongkai Zhao, Bonan Kou, Mohamed Yilmaz Ibrahim, Muhao Chen, and Tianyi Zhang. 2022b. Knowledge-based version incompatibility detection for deep learning. In Proceedings of the 19th International Conference on Mining Software Repositories, pages 247-251.",
+            "Wenxuan Zhou and Muhao Chen. 2022. An improved baseline for sentence-level relation extraction. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 161-168, Online only. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            430
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "15033",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Dense Retrieval as Indirect Supervision for Large-space Decision Making/6485a745-f89c-470f-a30b-5390c646537c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d168899f19d87314479373ae0ea2a66a6f6ad8bf388d56d189b37f048105d591
+size 756895

2023/Dense Retrieval as Indirect Supervision for Large-space Decision Making/full.md

@@ -0,0 +1,264 @@
+![](images/9fe764c8a2ca174d97908e9600f24354e6e2dbc1f477f18d882a5dd37836d839.jpg)
+
+# Dense Retrieval as Indirect Supervision for Large-space Decision Making
+
+Nan Xu Fei Wang Mingtao Dong Muhao Chen
+
+University of Southern California University of California, Davis {nanx, fwang598, mingtaod}@usc.edu muhchen@ucdavis.edu
+
+# Abstract
+
+Many discriminative natural language understanding (NLU) tasks have large label spaces. Learning such a process of large-space decision making is particularly challenging due to the lack of training instances per label and the difficulty of selection among many fine-grained labels. Inspired by dense retrieval methods for passage finding in open-domain QA, we propose a reformulation of large-space discriminative NLU tasks as a learning-to-retrieve task, leading to a novel solution named Dense Decision Retrieval (DDR). Instead of predicting fine-grained decisions as logits, DDR adopts a dual-encoder architecture that learns to predict by retrieving from a decision thesaurus. This approach not only leverages rich indirect supervision signals from easy-to-consume learning resources for dense retrieval, it also leads to enhanced prediction generalizability with a semantically meaningful representation of the large decision space. When evaluated on tasks with decision spaces ranging from hundreds to hundred-thousand scales, DDR outperforms strong baselines greatly by  $27.54\%$  in  $P@1$  on two extreme multi-label classification tasks,  $1.17\%$  in F1 score ultra-fine entity typing, and  $1.26\%$  in accuracy on three few-shot intent classification tasks on average.
+
+# 1 Introduction
+
+Many discriminative natural language understanding (NLU) tasks require making fine-grained decisions from a large candidate decision space. For example, a task-oriented dialogue system, when responding to users' requests, needs to frequently detect their intents from hundreds of options (Zhang et al., 2020b; Ham et al., 2020). Description-based recommendation in e-commerce needs to search from millions of products in response to users' descriptions (Gupta et al., 2021; Xiong et al., 2022).
+
+Previous studies for such NLU tasks still train classifiers as the solution (Gupta et al., 2021; Yu et al., 2022; Lin et al., 2023a). However, we argue that this straightforward approach is less practical for large-space decision making for several reasons. First, more decision labels naturally lead to data scarcity, since collecting sufficient training data for all labels will need significantly more cost. This issue is also accompanied by the issue of rare labels in the long tail of highly skewed distributions suffering severely from the lack of sufficient training instances (Zhang et al., 2023), leading to overgeneralization bias where frequent labels are more likely to be predicted compared with rare ones (Xu et al., 2022). Second, the non-semantic logit-based representation for decision labels in a classifier also makes the model hard to generalize to rarely seen labels, and not adaptable to unseen ones in training. This issue also impairs the applicability of the decision-making model to real-world scenarios, such as recommendation (Xiong et al., 2022) and task-oriented parsing (Zhao et al., 2022a), where the decision space may expand rapidly and crucial labels may be absent in training.
+
+On the contrary, motivated by semantic similarity of examples annotated with identical labels, recent studies propose contrastive learning schemes that leverage Siamese encoding architectures to maximize similarity scores of representation for positive example pairs (Henderson et al., 2020; Zhang et al., 2020a; Dahiya et al., 2021a; Mehri and Eric, 2021; Zhang et al., 2021; Xiong et al., 2022). Meanwhile, inductive bias from NLU tasks such as masked language modeling (Mehri et al., 2020; Dai et al., 2021) and natural language inference (NLI; Li et al. 2022; Du et al. 2022) has shown beneficial for learning on rare and unseen labels via indirect supervision (Yin et al., 2023). However, when dealing with very large decision spaces, existing methods still face critical trade-offs between
+
+generalizability and efficiency of prediction.2
+
+Inspired by the recent success of dense retrieval methods that learn to select answer-descriptive passages from millions of candidate documents for open-domain QA (Karpukhin et al., 2020; Lee et al., 2021; Zhan et al., 2021), we propose an indirectly supervised solution named Dense Decision Retrieval (DDR). DDR provides a general reformulation of large-space decision making as learning to retrieve from a semantically meaningful decision thesaurus constructed based on task-relevant resources. The model adopts the dual-encoder architecture from dense retrieval models to embed input texts and label descriptions, and learns to predict by retrieving from the informative decision thesaurus instead of predicting fine-grained decisions as logits. In this way, DDR not only leverages rich indirect supervision signals from other easy-to-consume learning resources for dense retrieval. It also leads to enhanced prediction performance and generalizability with a semantically meaningful representation of the large decision space. We evaluate DDR on large decision spaces ranging from hundreds for few-shot intent detection, ten-thousand for ultra-fine entity typing, to hundred-thousand scales for extreme multi-label classification. DDR obtains state-of-the-art performance on 6 benchmark datasets in multiple few-shot settings, improving the most competitive baselines by  $27.54\%$  in  $\mathbb{P}@\mathbb{1}$  for extreme classification,  $1.17\%$  in F1 for entity typing and  $1.26\%$  in accuracy for intent detection on average. Ablation studies show that both the constructed informative label thesaurus and indirect supervision from dense retrieval contribute to performance gain.
+
+The technical contributions of this work are three-fold. First, we present a novel and strong solution,  $DDR$ , for NLU tasks with large-space decision making that leverages indirect supervision from dense retrieval. Second, we provide semantically meaningful decision thesaurus construction that further improves the decision-making ability of  $DDR$ . Third, we comprehensively verify the effectiveness of  $DDR$  on tasks of fine-grained text classification, semantic typing and intent detection where the size of decision spaces range from hundreds to hundreds of thousands.
+
+# 2 Related Work
+
+Indirect Supervision Indirectly supervised methods (Roth, 2017; He et al., 2021; Yin et al., 2023) seek to transfer supervision signals from a more resource-rich task to enhance a specific more resource-limited task. A method of this kind often involves reformulation of the target task to the source task. Previous studies have investigated using source tasks such as NLI (Li et al., 2022; Yin et al., 2019; Lyu et al., 2021, inter alia), extractive QA (Wu et al., 2020; FitzGerald et al., 2018; Li et al., 2020, inter alia), abstractive QA (Zhao et al., 2022a; Du and Ji, 2022) and conditioned generation (Lu et al., 2022; Huang et al., 2022b; Hsu et al., 2022, inter alia) to enhance more expensive information extraction or semantic parsing tasks. Recent studies also transformed these technologies to specialized domains such as medicine (Xu et al., 2023) and software engineering (Zhao et al., 2022b) where model generalization and lack of annotations are more significant challenges. There has also been foundational work which studies the informativeness of supervision signals in such settings (He et al., 2021).
+
+However, the aforementioned studies are not designed for discriminative tasks with very large decision spaces, and do not apply directly due to issues such as high inference costs (NLI) and requiring decisions to be inclusive to the input (QA). We instead propose dense retriever that naturally serves as a proper and efficient form of indirect supervision for large-space decision making.
+
+NLU with Large Decision Spaces Many concrete NLU tasks deal with large decision spaces, including description-based recommendation in e-commerce (Gupta et al., 2021; Xiong et al., 2022) and Web search (Gupta et al., 2021), user intent detection in task-oriented dialog systems (Zhang et al., 2020b; Ham et al., 2020), and fine-grained semantic typing (Choi et al., 2018; Chen et al., 2020), etc. Previous studies for such tasks either train classifiers (Yu et al., 2022; Lin et al., 2023a) or rely on contrastive learning from scratch (Zhang et al., 2021; Xiong et al., 2022), which are generally impaired by insufficiency of training data and hard to generalize to rarely seen and unseen labels. These challenges motivate us to explore a practical solution with indirect supervision from a dense retriever.
+
+![](images/6d21e79feccc28c3d1f6d52726dfd7f12363d46193d3f14c57384969389b2300.jpg)  
+Figure 1: Overview of DDR that reformulates general large-space decision making tasks as dense retrieval. Label thesaurus is constructed with detailed descriptions from publicly available resources and the dual-encoder learns to maximize similarity between embeddings of input and label thesaurus entry with indirect supervision from a pre-trained dense retriever.
+
+# 3 Method
+
+We first describe the reformulation of large-space decision making task as dense retrieval (§3.1). We then introduce several automatic ways to construct label thesaurus of high quality (§3.2). Lastly, we demonstrate the dual-encoder architecture of DDR (illustrated in Fig. 1) to retrieve decisions (§3.3).
+
+# 3.1 Preliminary
+
+The mechanism of dense retrieval for general decision-making in large spaces can be formalized as follows. Given a textual input, such as "Illy Ground Espresso Classico Coffee: Classico, classic roast coffee has a lingering sweetness and delicate..." for description-based product recommendation, a model learns to infer their corresponding labels by mapping input to textual descriptions in a large label thesaurus (such as names, brands, component, etc.), instead of simply predicting the label index.
+
+Formally speaking, given a textual input  $x$ , a decision space covering  $L$  labels and a label thesaurus that contains corresponding entries (label descriptions)  $D = \{d_{1}, d_{2}, \ldots, d_{L}\}$ . We first split every entry into text passages of a maximum length (e.g., 100 words) as the basic retrieval units and get  $M$  total passages  $P = \{p_{1}, p_{2}, \ldots, p_{M}\}^{3}$ . A retriever  $R: (x, P) \to P_{R}$  is a function that takes as input a text sequence  $x$  and the label thesaurus  $P$  and
+
+returns a ranking list of passages  $P_R$ , where labels represented by higher ranked passages are more likely to be relevant to the input. As demonstrated in Fig. 1, DDR also leverages indirect supervision from open-domain QA by initializing parameters of retriever  $R$  with those from DPR.
+
+# 3.2 Thesaurus Construction
+
+Depending on the target task, the decision thesaurus can provide knowledge about the decision space from various perspectives. Given label names, we create informative descriptions for each label automatically without extra expenses. We can refer to publicly available resources for constructing thesaurus of high quality. To name a few:
+
+1) Lexicographical knowledge bases: dictionaries such as WordNet (Miller, 1995) and Wiktionary<sup>4</sup> provide accurate definitions as well as plentiful use cases to help understand words and phrases.  
+2) Large Language Models: LLMs trained on instructions like ChatGPT and Vicuna-13B (Chiang et al., 2023) are able to generate comprehensive and reasonable descriptions of terms when they are prompted for explanation.  
+3) Training examples: directly adopting input text when labels with detailed information also appear as example input, or aggregating input texts from multiple examples assigned with the same label if they share very high-level semantic similarities.
+
+We describe thesaurus construction in practice for each investigated downstream task in  $\S 4$ .
+
+# 3.3 Learning to Retrieve Decisions
+
+Encoders Since a key motivation of  $DDR$  is to leverage indirect supervision from retrieval tasks, we adopt the dual-encoder architecture similar to a dense retriever such as DPR (Karpukhin et al., 2020), where two sentence encoders with special tokens (BERT (Devlin et al., 2019) by default),  $E_{x}$  and  $E_{p}$ , are leveraged to represent input text  $x$  and passage  $p$  individually. The similarity between input and passage is computed using dot product of their vectors denoted by  $\operatorname{sim}(x,p)$ .
+
+Training The objective of the reformulated large-space decision making problem is to optimize the dual encoders such that relevant pairs of input text and label description (thesaurus entry) should have higher similarity than irrelevant pairs in the embedding space. Hence, the overall learning objective is a process of contrastive learning.
+
+To tackle single-label classification where each instance contains one input  $x_{i}$ , one relevant (positive) passage  $x_{i}^{+}$ , along with  $n$  irrelevant (negative) passages  $p_{i,j}^{-}$ , we minimize the negative log likelihood of the positive label description
+
+$$
+\begin{array}{l} L \left(x _ {i}, p _ {i} ^ {+}, p _ {i, 1} ^ {-}, \dots , p _ {i, n} ^ {-}\right) \tag {1} \\ = - \log \frac {e ^ {\operatorname {s i m} \left(x _ {i} , p _ {i} ^ {+}\right)}}{e ^ {\operatorname {s i m} \left(x _ {i} , p _ {i} ^ {+}\right)} + \sum_ {j = 1} ^ {n} e ^ {\operatorname {s i m} \left(x _ {i} , p _ {i , j} ^ {-}\right)}}. \\ \end{array}
+$$
+
+We extend to multi-label classification tasks where each input has multiple positive passages and more than one of them participates in model update simultaneously. Accordingly, the cross entropy loss is minimized to encourage similarity between  $m$  positive pairs instead:
+
+$$
+\begin{array}{l} L \left(x _ {i}, p _ {i, 1} ^ {+}, \dots , p _ {i, m} ^ {+}, p _ {i, 1} ^ {-}, \dots , p _ {i, n} ^ {-}\right) \tag {2} \\ = - \sum_ {k = 1} ^ {m} \log \frac {e ^ {\mathrm {s i m} (x _ {i} , p _ {i , k} ^ {+})}}{\sum_ {j = 1} ^ {m} e ^ {\mathrm {s i m} (x _ {i} , p _ {i , j} ^ {+})} + \sum_ {j = 1} ^ {n} e ^ {\mathrm {s i m} (x _ {i} , p _ {i , j} ^ {-})}}. \\ \end{array}
+$$
+
+Positive, negative and hard negative passages: the decision thesaurus entries that describe true labels in thesaurus for each input are deemed as positive. In-batch negatives are positive entries for other input in the same mini-batch, making the computation more efficient while achieving better performance compared with random passages or other negative entries (Gillick et al., 2019;
+
+Karpukhin et al., 2020). If not otherwise specified, we train with positive and in-batch negatives to obtain a weak  $DDR$  model firstly; we then retrieve label thesaurus with the weak model for training examples and leverage wrongly predicted label description as hard negatives, which are augmented to training data to obtain a stronger  $DDR$ .
+
+We note that, different from open-domain QA where one positive pair can be considered per time with Eq. 1, considering multiple positive pairs is very important for tasks where the decision space though large but follows highly skewed distributions, when in-batch negatives are adopted for model update. In the mini-batch, multiple inputs may share same labels that are popular in the head of long-tail label distribution. In this case, these input instances have multiple positive passages: one randomly sampled from label passages as normal and some from positives of other in-batch inputs.
+
+Inference After training is completed, we encode all passages in label thesaurus once with  $E_{p}$  and index them using FAISS offline (Johnson et al., 2019). Given a new input, we obtain its embedding with  $E_{x}$  and retrieve the top- $k$  passages with embeddings closest to that of input.
+
+# 4 Experiments
+
+We evaluate DDR on three NLU tasks with large deiction spaces in two criteria, 1) a minority of label space is observed during training: Extreme Multi-label Classification (§4.1) and Entity Typing (§4.2), or 2) limited amounts of examples are available per label: Intent Classification (§4.3). We investigate decision spaces of up to hundreds of thousands. Dataset statistics are shown in Appx. Tab. 1.
+
+# 4.1 Extreme Multi-label Classification
+
+Task Many real-world applications can be formulated as eXtreme Multi-label Classification (XMC) problems, where relevant labels from the set of an extremely large size are expected to be predicted given text input. Considering the ever-growing label set with newly added products or websites and the time-consuming example-label pairs collection, we follow the few-shot XMC setting (Gupta et al., 2021; Xiong et al., 2022) to choose relevant labels from all available seen and unseen labels. $^6$
+
+<table><tr><td>Dataset</td><td>Setting</td><td>Train</td><td>Dev</td><td>Test</td><td>Labels</td></tr><tr><td colspan="6">Extreme Multi-label Classification</td></tr><tr><td rowspan="2">LF-Amazon-131K</td><td>1%</td><td>6227</td><td>-</td><td rowspan="2">134,835</td><td rowspan="2">131,073</td></tr><tr><td>5%</td><td>29,719</td><td>-</td></tr><tr><td rowspan="2">LF-WikiSeeAlso-320K</td><td>1%</td><td>13,920</td><td>-</td><td rowspan="2">177,515</td><td rowspan="2">312,330</td></tr><tr><td>5%</td><td>62,704</td><td>-</td></tr><tr><td colspan="6">Ultra-fine Entity Typing</td></tr><tr><td>UFET</td><td>Crowd</td><td>1998</td><td>1998</td><td>1998</td><td>10,331</td></tr><tr><td colspan="6">Intent Classification</td></tr><tr><td rowspan="2">BANKING</td><td>5-shot</td><td>385</td><td rowspan="2">1540</td><td rowspan="2">3080</td><td rowspan="2">77</td></tr><tr><td>10-shot</td><td>770</td></tr><tr><td rowspan="2">HWU</td><td>5-shot</td><td>320</td><td rowspan="2">1076</td><td rowspan="2">1076</td><td rowspan="2">64</td></tr><tr><td>10-shot</td><td>640</td></tr><tr><td rowspan="2">CLINC</td><td>5-shot</td><td>750</td><td rowspan="2">3000</td><td rowspan="2">4500</td><td rowspan="2">150</td></tr><tr><td>10-shot</td><td>1500</td></tr></table>
+
+Table 1: Statistics and experimental settings on datasets with label spaces ranging from 64 to  ${320}\mathrm{\;K}$  .
+
+Datasets and Metrics The public benchmark dataset LF-Amazon-131K collects pairs of relevant products with textual descriptions for commercial recommendation, while LF-WikiSeeAlso-320K<sup>7</sup> aims to link relevant Wikipedia passages for reference (Bhatia et al., 2016; Gupta et al., 2021).
+
+We use metrics widely adopted by XMC (Chien et al., 2023) and IR (Thakur et al., 2021) literature: precision@k for the proportion of the top-  $k$  predicted labels to be true labels, and recall@k for the proportion of true labels found in the top-  $k$  predictions.
+
+Baselines We compare DDR with strong XMC baselines in three categories: 1) Transformer-based models that learn semantically meaningful sentence embeddings with siamese or triplet network structures, including Sentence-BERT (Reimers and Gurevych, 2019) and MPNet (Song et al., 2020); 2) competitive methods originally proposed for scalable and accurate predictions over extremely large label spaces, including XR-Linear (Yu et al., 2022) that recursively learns to traverse an input from the root of a hierarchical label tree to a few leaf node clusters, Astec (Dahiya et al., 2021b) with four sub-tasks for varying trade-offs between accuracy
+
+and scalability, and SiameseXML (Dahiya et al., 2021a) based on a novel probabilistic model to meld Siamese architectures with high-capacity extreme classifiers; 3) methods specifically designed to improve performance with few-shot labels available, including ZestXML (Gupta et al., 2021) that learns to project a data point's features close to the features of its relevant labels through a highly sparsified linear transform, and MACLR (Xiong et al., 2022) that pre-trains Transformer-based encoders with a self-supervised contrastive loss.
+
+Implementation Details The label space of LF-WikiSeeAlso-320K is composed of relevant passage titles, hence the corresponding full content should be available in Wikipedia and we use them as the label thesaurus. For relevant product advertisement and recommendation, concrete product description provides sufficient information as the label thesaurus for LF-Amazon-131K. We extract product textual descriptions mainly from Amazon data in XMC benchmark (Bhatia et al., 2016) and public Amazon product meta datasets (Ni et al., 2019), leading to  $95.67\%$  of labels well documented. As introduced in  $\S 3.2$ , we then use ChatGPT to obtain relevant information for remaining undocumented labels.
+
+We perform label thesaurus construction on Intel(R) Xeon(R) Gold 5217 CPU @ 3.00GHz with 32 CPUs and 8 cores per CPU. It takes 59.12 minutes to construct label thesaurus for LF-WikiSeeAlso-320K, while 28.24 minutes for LF-Amazon-131K. We leverage the same infrastructure to prepare label thesaurus in following tasks unless otherwise specified.
+
+Due to a lack of explicit knowledge about the connection between the majority of labels and their examples for training, DDR only learns to distinguish positive example-label pairs from in-batch negatives without further negative mining. For inference, with the constructed label thesaurus, we first retrieve the large label space with a test example based on Sentence-BERT embeddings of the title and the first sentence in an unsupervised way. After training DDR with few-shot labels, we mimic the practice of DPR+BM25 for open-
+
+<table><tr><td rowspan="2">Method</td><td colspan="8">1% Labels</td><td colspan="7">5% Labels</td><td></td></tr><tr><td>Precision@1</td><td>@3</td><td>@5</td><td>Recall@1</td><td>@3</td><td>@5</td><td>@10</td><td>@100</td><td>Precision@1</td><td>@3</td><td>@5</td><td>@1</td><td>@3</td><td>@5</td><td>@10</td><td>@100</td></tr><tr><td colspan="16">LF-Amazon-131K</td><td></td></tr><tr><td>XR-Linear</td><td>1.53</td><td>0.57</td><td>0.36</td><td>0.67</td><td>0.75</td><td>0.78</td><td>0.81</td><td>0.92</td><td>5.09</td><td>2.09</td><td>1.32</td><td>2.36</td><td>2.86</td><td>3.02</td><td>3.18</td><td>3.74</td></tr><tr><td>Astec</td><td>0.94</td><td>0.44</td><td>0.29</td><td>0.55</td><td>0.78</td><td>0.84</td><td>0.91</td><td>1.13</td><td>3.94</td><td>1.92</td><td>1.26</td><td>2.31</td><td>2.34</td><td>3.66</td><td>4.00</td><td>4.96</td></tr><tr><td>SiameseXML</td><td>1.45</td><td>0.56</td><td>0.35</td><td>0.84</td><td>0.96</td><td>1.00</td><td>1.03</td><td>1.16</td><td>5.36</td><td>2.23</td><td>1.41</td><td>3.15</td><td>3.89</td><td>4.08</td><td>4.27</td><td>4.82</td></tr><tr><td>ZestXML</td><td>10.10</td><td>9.19</td><td>7.34</td><td>5.63</td><td>14.46</td><td>18.61</td><td>23.73</td><td>32.69</td><td>12.33</td><td>10.00</td><td>8.71</td><td>6.84</td><td>17.19</td><td>21.97</td><td>28.19</td><td>46.49</td></tr><tr><td>Sentence-BERT</td><td>12.64</td><td>9.82</td><td>7.80</td><td>6.97</td><td>15.34</td><td>19.74</td><td>25.33</td><td>43.53</td><td>15.47</td><td>12.24</td><td>9.64</td><td>8.63</td><td>19.23</td><td>24.40</td><td>30.82</td><td>49.22</td></tr><tr><td>MPNet</td><td>14.78</td><td>11.55</td><td>8.97</td><td>8.28</td><td>18.24</td><td>22.84</td><td>28.54</td><td>45.89</td><td>15.03</td><td>11.88</td><td>9.28</td><td>8.47</td><td>18.74</td><td>23.69</td><td>29.93</td><td>48.84</td></tr><tr><td>MACLR</td><td>18.74</td><td>16.07</td><td>12.52</td><td>10.73</td><td>25.44</td><td>31.89</td><td>39.17</td><td>57.55</td><td>19.56</td><td>16.19</td><td>12.64</td><td>11.15</td><td>25.65</td><td>32.18</td><td>39.63</td><td>58.45</td></tr><tr><td>DDR (w/o thesaurus)</td><td>15.39</td><td>10.92</td><td>7.87</td><td>8.25</td><td>19.15</td><td>24.10</td><td>31.51</td><td>54.32</td><td>21.21</td><td>13.77</td><td>10.23</td><td>12.02</td><td>24.18</td><td>30.29</td><td>38.19</td><td>58.75</td></tr><tr><td>DDR (scratch)</td><td>19.55</td><td>14.48</td><td>11.50</td><td>10.87</td><td>22.50</td><td>29.00</td><td>37.80</td><td>56.53</td><td>23.77</td><td>16.48</td><td>12.75</td><td>13.31</td><td>25.54</td><td>31.85</td><td>40.46</td><td>60.30</td></tr><tr><td>DDR (full)</td><td>22.39</td><td>16.38</td><td>12.40</td><td>12.57</td><td>25.64</td><td>31.44</td><td>38.44</td><td>57.72</td><td>25.22</td><td>17.96</td><td>13.63</td><td>14.18</td><td>27.94</td><td>34.40</td><td>41.72</td><td>61.17</td></tr><tr><td colspan="16">LF-WikiSeeAlso-320K</td><td></td></tr><tr><td>XR-Linear</td><td>1.24</td><td>0.57</td><td>0.37</td><td>0.42</td><td>0.58</td><td>0.63</td><td>0.68</td><td>0.76</td><td>4.69</td><td>2.20</td><td>1.46</td><td>1.82</td><td>2.41</td><td>2.63</td><td>2.82</td><td>3.42</td></tr><tr><td>Astec</td><td>1.25</td><td>0.60</td><td>0.41</td><td>0.69</td><td>0.98</td><td>1.11</td><td>1.27</td><td>1.56</td><td>5.90</td><td>2.80</td><td>1.86</td><td>3.26</td><td>4.49</td><td>4.95</td><td>5.49</td><td>6.83</td></tr><tr><td>SiameseXML</td><td>1.81</td><td>0.75</td><td>0.48</td><td>1.03</td><td>1.26</td><td>1.33</td><td>1.41</td><td>1.67</td><td>6.83</td><td>3.15</td><td>2.06</td><td>3.88</td><td>5.15</td><td>5.56</td><td>6.02</td><td>7.09</td></tr><tr><td>ZestXML</td><td>8.74</td><td>6.78</td><td>5.41</td><td>4.68</td><td>9.70</td><td>12.21</td><td>15.73</td><td>24.98</td><td>10.06</td><td>8.11</td><td>6.60</td><td>5.33</td><td>11.49</td><td>14.74</td><td>19.57</td><td>40.46</td></tr><tr><td>Sentence-BERT</td><td>16.30</td><td>12.62</td><td>10.08</td><td>9.30</td><td>18.92</td><td>23.78</td><td>30.40</td><td>52.92</td><td>18.47</td><td>14.19</td><td>11.29</td><td>10.82</td><td>21.55</td><td>26.77</td><td>33.92</td><td>57.02</td></tr><tr><td>MPNet</td><td>17.14</td><td>12.64</td><td>9.96</td><td>9.98</td><td>18.98</td><td>23.45</td><td>29.67</td><td>50.75</td><td>18.59</td><td>13.99</td><td>11.08</td><td>10.89</td><td>21.12</td><td>26.10</td><td>32.82</td><td>54.70</td></tr><tr><td>MACLR</td><td>19.09</td><td>14.57</td><td>11.53</td><td>11.39</td><td>22.34</td><td>27.63</td><td>34.81</td><td>57.92</td><td>20.99</td><td>15.57</td><td>12.26</td><td>12.59</td><td>23.94</td><td>29.41</td><td>36.78</td><td>59.81</td></tr><tr><td>DDR (w/o thesaurus)</td><td>17.52</td><td>15.93</td><td>14.23</td><td>4.89</td><td>13.11</td><td>18.88</td><td>25.93</td><td>46.93</td><td>18.14</td><td>16.65</td><td>15.39</td><td>5.12</td><td>14.10</td><td>21.81</td><td>30.00</td><td>52.54</td></tr><tr><td>DDR (scratch)</td><td>21.26</td><td>15.81</td><td>13.08</td><td>11.16</td><td>22.52</td><td>29.55</td><td>40.14</td><td>64.94</td><td>25.04</td><td>18.95</td><td>15.05</td><td>12.46</td><td>25.70</td><td>32.18</td><td>40.24</td><td>66.22</td></tr><tr><td>DDR (full)</td><td>24.70</td><td>18.45</td><td>14.69</td><td>13.04</td><td>26.31</td><td>32.99</td><td>41.08</td><td>65.39</td><td>27.78</td><td>19.98</td><td>15.66</td><td>15.03</td><td>28.87</td><td>35.47</td><td>43.35</td><td>68.66</td></tr></table>
+
+Table 2: Results of few-shot XMC where the training subset covers  $1\%$  (left) and  $5\%$  (right) labels from the whole set. DDR outperforms the second best MACLR in both settings of two datasets. Indirect supervision from DPR boosts performance against training from scratch (the second row from the bottom), while label thesaurus construction improves accuracy over those using textual label names (the third row from the bottom).
+
+domain QA (Karpukhin et al., 2020) for reranking, where prediction scores on test set from  $DDR$  and Sentence-BERT are linearly combined as the final ranking score.
+
+Results We report performance of DDR with(out) indirect supervision from dense retrieval for open-domain QA in Tab. 2. On average, DDR boosts performance on LF-Amazon-131K by  $8.50\%$  and LF-WikiSeeAlso-320K by  $21.73\%$  compared with the second best baseline MACLR. When only  $1\%$  labels are seen during training, we observe that DDR shows much better transferability than other pre-trained models, e.g., average performance improved by  $12.59\%$  over MACLR and  $39.80\%$  over MPNet. As more labels are available for fine-tuning the model, the performance gain from DDR over others becomes even more significant:  $17.64\%$  over MACLR and  $43.38\%$  over MPNet under  $5\%$  labels setting.
+
+# 4.2 Ultra-fine Entity Typing
+
+Task Entities can often be described by very fine grained-types (Choi et al., 2018) and the ultra-fine entity typing task aims at predicting one or more
+
+fine-grained words or phrases that describe the type(s) of that specific mention (Xu et al., 2022). Consider the sentence "He had blamed Israel for failing to implement its agreements." Besides person and male, the mention "He" has other very specific types that can be inferred from the context, such as leader or official for the "blamed" behavior and "implement its agreements" affair. Ultra-fine entity typing has a broad impact on various NLP tasks that depend on type understanding, including coreference resolution (Onoe and Durrett, 2020), event detection (Le and Nguyen, 2021) and relation extraction (Zhou and Chen, 2022).
+
+Datasets and Metrics We leverage the UFET dataset (Choi et al., 2018) to evaluate benefits of DDR with indirect supervision from dense retrieval. Among 10,331 entity types, accurately selecting finer labels (121 labels such as engineer) was more challenging to predict than coarse-grained labels (9 labels such as person), and this issue is exacerbated when dealing with ultra-fine types (10,201 labels such as flight engineer). Following recent entity typing literature (Li et al., 2022; Du et al., 2022), we train DDR on (originally provided) limited
+
+<table><tr><td>Method</td><td>Precision</td><td>Recall</td><td>F1</td></tr><tr><td>BiLSTM</td><td>48.1</td><td>23.2</td><td>31.3</td></tr><tr><td>LabelGCN</td><td>50.3</td><td>29.2</td><td>36.9</td></tr><tr><td>LDET</td><td>51.5</td><td>33.0</td><td>40.1</td></tr><tr><td>Box4Types</td><td>52.8</td><td>38.8</td><td>44.8</td></tr><tr><td>LRN</td><td>54.5</td><td>38.9</td><td>45.4</td></tr><tr><td>UniST</td><td>50.2</td><td>49.6</td><td>49.9</td></tr><tr><td>MLMET</td><td>53.6</td><td>45.3</td><td>49.1</td></tr><tr><td>LITE</td><td>52.4</td><td>48.9</td><td>50.6</td></tr><tr><td>Context-TE</td><td>53.7</td><td>49.4</td><td>51.5</td></tr><tr><td>DDR (w/o thesaurus)</td><td>51.6</td><td>46.0</td><td>48.6</td></tr><tr><td>DDR (scratch)</td><td>53.0</td><td>50.5</td><td>51.7</td></tr><tr><td>DDR (full)</td><td>51.9</td><td>52.3</td><td>52.1</td></tr></table>
+
+Table 3: Results of UFET task. DDR outperforms competitive baselines with (upper 5 methods) or without (lower 3 methods) inductive bias from task pre-training.
+
+crowd-sourced examples without relying on distant resources such as knowledge bases or head words from the Gigaword corpus. We follow prior studies (Choi et al., 2018) to evaluate macro-averaged precision, recall and F1.
+
+Baselines We consider two categories of competitive entity typing models as baselines: 1) methods capturing the example-label and label-label relations, e.g., BiLSTM (Choi et al., 2018) that concatenates the context representation learned by a bidirectional LSTM and the mention representation learned by a CNN, LabelGCN (Xiong et al., 2019) that learns to encode global label co-occurrence statistics and their word-level similarities, LRN (Liu et al., 2021a) that models the coarse-to-fine label dependency as causal chains, Box4Types (Onoe et al., 2021) that captures hierarchies of types as topological relations of boxes, and UniST (Huang et al., 2022a) that conduct name-based label ranking; 2) methods leveraging inductive bias from pre-trained models for entity typing, e.g., MLMET (Dai et al., 2021) that utilizes the pretrained BERT to predict the most probable words for “[MASK]” earlier incorporated around the mention as type labels, LITE (Li et al., 2022) and Context-TE (Du et al., 2022) that both leverage indirect supervision from pre-trained natural language inference.
+
+Implementation Details The dataset UFET first asked crowd workers to annotate entity's types and then used WordNet (Miller, 1995) to expand these types automatically by generating all their synonyms and hypernyms based on the most common sense. Therefore, we automatically obtain label thesaurus entries from definitions and examples in
+
+WordNet and Wiktionary, which covers  $99.99\%$  of the whole label set. The time cost for prior label thesaurus construction is around 2.07 minutes.
+
+initialized with the original DPR checkpoint pretrained on open-domain QA datasets, DDR firstly optimizes the model given positive example-label pairs and in-batch negatives. With the fine-tuned model, we then perform dense retrieval on the label set for each training example, keeping label documents with high scores but not in the true label set as hard negatives. DDR further updates the model with these additional hard negatives. For inference of multi-label entity type classification, we adopt labels with retrieval scores higher than a threshold that leads to the best F1 score on the development set.
+
+Results In Tab. 3, we show performance of DDR and other entity typing methods. DDR obtains the state-of-the-art F1 score over the best baseline training from scratch (LRN) by 6.7 and the best baseline with inductive bias from the language model (Context-TE) by 0.6.
+
+# 4.3 Few-shot Intent Classification
+
+Task As a fundamental element in task-oriented dialog systems, intent detection is normally conducted in the NLU component for identifying a user's intent given an utterance (Ham et al., 2020). Recently, accurately identifying intents in the few-shot setting has attracted much attention due to data scarcity issues resulted from the cost of data collection as well as privacy and ethical concerns (Lin et al., 2023b). Following the few-shot intent detection benchmark (Zhang et al., 2022), we focus on the challenging 5-shot and 10-shot settings.
+
+Datasets and Metrics To evaluate the effectiveness of DDR for NLU with large decision spaces, we pick three challenging intent datasets with a relatively large number of semantically similar intent labels. Banking77 (Casanueva et al., 2020) is a single-domain dataset that provides very fine-grained 77 intents in a Banking domain. HWU64 (Liu et al., 2021b) is a multi-domain (21 domains) dataset recorded by a home assistant robot including 64 intents ranging from setting alarms, playing music, search, to movie recommendation. CLINC150 (Larson et al., 2019) prompts crowd workers to provide questions or commands in the manner they would interact with an artificially intelligent assistant covering 150 intent classes over 10 domains. We report accuracy
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Banking77</td><td colspan="2">HWU64</td><td colspan="2">CLINC150</td></tr><tr><td>5-shot</td><td>10-shot</td><td>5-shot</td><td>10-shot</td><td>5-shot</td><td>10-shot</td></tr><tr><td colspan="7">w/ Data Augmentation</td></tr><tr><td>ICDA-X S (+1x)</td><td>80.29</td><td>86.72</td><td>81.32</td><td>85.59</td><td>91.16</td><td>93.71</td></tr><tr><td>ICDA-S (+4x)</td><td>81.95</td><td>87.37</td><td>81.97</td><td>86.25</td><td>91.22</td><td>93.98</td></tr><tr><td>ICDA-M (+16x)</td><td>84.01</td><td>88.64</td><td>81.84</td><td>87.36</td><td>91.93</td><td>94.71</td></tr><tr><td>ICDA-L (+64x)</td><td>83.90</td><td>89.12</td><td>81.97</td><td>86.94</td><td>92.41</td><td>94.73</td></tr><tr><td>ICDA-X L (+128x)</td><td>83.90</td><td>89.79</td><td>82.45</td><td>87.41</td><td>92.62</td><td>94.84</td></tr><tr><td colspan="7">w/o Data Augmentation</td></tr><tr><td>USE</td><td>76.29</td><td>84.23</td><td>77.79</td><td>83.75</td><td>87.82</td><td>90.85</td></tr><tr><td>ConveRT</td><td>75.32</td><td>83.32</td><td>76.95</td><td>82.65</td><td>89.22</td><td>92.62</td></tr><tr><td>USE+ConveRT</td><td>77.75</td><td>85.19</td><td>80.01</td><td>85.83</td><td>90.49</td><td>93.26</td></tr><tr><td>ConvBERT</td><td>-</td><td>83.63</td><td>-</td><td>83.77</td><td>-</td><td>92.10</td></tr><tr><td>+ MLM</td><td>-</td><td>83.99</td><td>-</td><td>84.52</td><td>-</td><td>92.75</td></tr><tr><td>+ MLM + Example</td><td>-</td><td>84.09</td><td>-</td><td>83.44</td><td>-</td><td>92.35</td></tr><tr><td>+ Combined</td><td>-</td><td>85.95</td><td>-</td><td>86.28</td><td>-</td><td>93.97</td></tr><tr><td>DNNC</td><td>80.40</td><td>86.71</td><td>80.46</td><td>84.72</td><td>91.02</td><td>93.76</td></tr><tr><td>CPFT</td><td>80.86</td><td>87.20</td><td>82.03</td><td>87.13</td><td>92.34</td><td>94.18</td></tr><tr><td>- Pretraining</td><td>76.75</td><td>84.83</td><td>76.02</td><td>82.96</td><td>88.19</td><td>91.55</td></tr><tr><td>Context-TE</td><td>80.76</td><td>85.53</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DDR (w/o thesaurus)</td><td>78.86</td><td>84.16</td><td>80.30</td><td>83.74</td><td>88.86</td><td>91.58</td></tr><tr><td>DDR (scratch)</td><td>82.11</td><td>85.97</td><td>80.76</td><td>84.71</td><td>88.58</td><td>91.16</td></tr><tr><td>DDR (full)</td><td>83.86 (XL)</td><td>88.25 (M)</td><td>84.29 (XL)</td><td>86.34 (S)</td><td>92.71 (XL)</td><td>94.58 (M)</td></tr></table>
+
+Table 4: Few-shot Intent Detection Accuracy on three benchmark datasets. DDR achieves much higher accuracy than existing strong learning baselines without data augmentation, while competitive with ICDA requiring a considerable amount of augmented data. ICDA prepares additional synthetic data with the scale ranging from the same amount (ICDA-  $XS$ ) to 128 times (ICDA-  $XL$ ) of the original few-shot train size. In the bottom row, we also mark performance of DDR with the most comparable ICDA variant. Results not available in original papers are marked as  $-$ .
+
+for this single-label classification task.
+
+Baselines There are two families of intent learning algorithms to cope with limited amounts of example-label pairs: 1) Classifiers based on sentence embeddings from PLMs, including USE (Yang et al., 2020) embed sentences from 16 languages into a single semantic space using a multi-task trained dual-encoder, ConvBERT (Henderson et al., 2020) that pretrains a dual-encoder model on Reddit Data by maximizing the similarity score of embeddings of positive input-response pairs; ConvBERT (Mehri et al., 2020) that adopts a BERT-base model pre-trained on a large open-domain dialogue corpus; 2) methods leverage semantic similarity between utterances from different users, e.g. ConvBERT+Combined (Mehri and Eric, 2021) that extends the ConvBERT model with task-adaptive masked language modelling (MLM) and infers the intent of the utterance based on the similarity to the examples corresponding to each intent, DNNC (Zhang et al., 2020a) that leverages BERT-style pairwise encoding to train a binary classifier that estimates the best matched training example for a user input, CPFT (Zhang et al., 2021) that conducts contrastive pre-training on example sentences
+
+from six public intent datasets to discriminate semantically similar utterances.
+
+Different from DDR and prior baselines, ICDA (Lin et al., 2023a), a most recent work tackles the challenge of limited data for intent detection by generating high-quality synthetic training data. It first fine-tunes a PLM on a small seed of training data for new data point synthesisation, and then employs pointwise V-information (PVI) based filtering to remove unhelpful ones.[12]
+
+Implementation Details Based on the observation that utterances labeled by identical intents share very similar sentence-level semantic meanings, we first utilize the whole set of unlabeled training examples to represent their corresponding pseudo labels predicted by Sentence-BERT (Reimers and Gurevych, 2019). We perform prediction on a single NVIDIA RTX A5000 GPU and it takes 4.17 minutes to complete label thesaurus construction for Banking77, 6.45 minutes for HWU64 and 7.61
+
+<table><tr><td>Dataset</td><td>Setting</td><td>Sentence-BERT (1st round)</td><td>DDR (2nd round)</td></tr><tr><td rowspan="2">Banking77</td><td>5-shot</td><td>84.29</td><td>83.86</td></tr><tr><td>10-shot</td><td>88.02</td><td>88.25</td></tr><tr><td rowspan="2">HWU64</td><td>5-shot</td><td>82.81</td><td>84.29</td></tr><tr><td>10-shot</td><td>85.69</td><td>86.34</td></tr><tr><td rowspan="2">CLINC150</td><td>5-shot</td><td>91.82</td><td>92.71</td></tr><tr><td>10-shot</td><td>94.07</td><td>94.58</td></tr></table>
+
+Table 5: Impact on intent detection from label thesaurus constructed with predictions from Sentence-BERT and DDR for the unlabeled whole set of the training set.
+
+minutes for CLINC150. After training DDR with this first version of label thesaurus, we then update pseudo labels for each training example with predictions from DDR for the second-round training.
+
+Regardless of the basis for label thesaurus construction, DDR is trained for two phases in each round: only few-shot positive example-label pairs and in-batch negatives are used in the first phase, while labels wrongly predicted for training examples from the prior phase are used as additional hard negatives in the second phase. After the second round of training, the latest DDR makes predictions on the whole set of training examples for the final label thesaurus construction, from which DDR retrieves the label with the highest score as the prediction.
+
+Results Tab. 4 presents performance of different intent detection methods on 5- and 10-shot setting. Without extra training data synthesis, we observe significant performance gain from the proposed DDR over existing baselines. Although ICDA obtains enhanced accuracy by increasing data augmentation scale, DDR is able to achieve comparable performance without crafting 4x even 128x synthetic instances. In Tab. 5, we additionally study the quality of constructed label thesaurus in accordance with the prediction on the unlabeled whole training set from Sentence-BERT and DDR after the first round of training. We find that with a more accurate pseudo connection between example and label leads to a higher quality of label thesaurus construction, and finally benefits intent detection.
+
+# 5 Conclusions
+
+In this paper, we focus on discriminative NLU tasks with large decision spaces. By reformulating these tasks as learning-to-retrieve tasks, we are able to leverage rich indirect supervision signals
+
+from dense retrieval. Moreover, by representing decision spaces with thesaurus, we provide rich semantic knowledge of each label to improve the understanding of labels for dense decision retrievers. Experiments on 6 benchmark datasets show the effectiveness of our method on decision spaces scaling from hundreds of candidates to hundreds of thousands of candidates. Future work can extend our method to more large-space decision making tasks, especially in the low-resource setting.
+
+# Acknowledgement
+
+We appreciate the reviewers for their insightful comments and suggestions. The logo for  $DDR$  used in this paper is sourced from the Wikipedia page $^{13}$  under CC BY-SA 3.0, for which we appreciate the contribution of the Wikipedia community.
+
+Nan Xu is supported by the USC PhD Fellowship. Fei Wang is supported by the Annenberg Fellowship and the Amazon ML Fellowship. Mingtao Dong is supported by the Provost's Research Fellowship. Muhao Chen is supported by the NSF Grant IIS 2105329, the NSF Grant ITE 2333736, the DARPA MCS program under Contract No. N660011924033 with the United States Office Of Naval Research, a Cisco Research Award, two Amazon Research Awards, and a Keston Research Award. The computing of this work has been partly supported by a subaward of NSF Cloudbank 1925001 through UCSD.
+
+# Limitations
+
+The proposed DDR leverages the dual-encoder architecture with the inner dot product to compute embedding similarity, obtaining state-of-the-art performance on three investigated challenging large-space decision making tasks. More expressive models for embedding input and label thesaurus, such as joint encoding by a cross-encoder is not discussed. Moreover, other ways to model the connection between input and label thesaurus entry, such as using a sequence-to-sequence language model to generate the label name given input text and label thesaurus entry, is not explored yet. We believe adopting more advanced dense retrieval algorithms can further promote performance for large-space decision making. We leave this as an exciting future direction.
+
+# References
+
+K. Bhatia, K. Dahiya, H. Jain, P. Kar, A. Mittal, Y. Prabhu, and M. Varma. 2016. The extreme classification repository: Multi-label datasets and code.  
+Inigo Casanueva, Tadas Temčinas, Daniela Gerz, Matthew Henderson, and Ivan Vulić. 2020. Efficient intent detection with dual sentence encoders. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 38-45, Online. Association for Computational Linguistics.  
+Muhao Chen, Hongming Zhang, Haoyu Wang, and Dan Roth. 2020. What are you trying to do? semantic typing of event processes. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 531-542, Online. Association for Computational Linguistics.  
+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.  
+Eli Chien, Jiong Zhang, Cho-Jui Hsieh, Jyun-Yu Jiang, Wei-Cheng Chang, Olgica Milenkovic, and Hsiang-Fu Yu. 2023. Pina: Leveraging side information in extreme multi-label classification via predicted instance neighborhood aggregation. arXiv preprint arXiv:2305.12349.  
+Eunsol Choi, Omer Levy, Yejin Choi, and Luke Zettlemoyer. 2018. Ultra-fine entity typing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 87-96, Melbourne, Australia. Association for Computational Linguistics.  
+Kunal Dahiya, Ananye Agarwal, Deepak Saini, K Gururaj, Jian Jiao, Amit Singh, Sumeet Agarwal, Purushottam Kar, and Manik Varma. 2021a. SiameseXML: Siamese networks meet extreme classifiers with 100m labels. In International Conference on Machine Learning, pages 2330-2340. PMLR.  
+Kunal Dahiya, Deepak Saini, Anshul Mittal, Ankush Shaw, Kushal Dave, Akshay Soni, Himanshu Jain, Sumeet Agarwal, and Manik Varma. 2021b. Deepxml: A deep extreme multi-label learning framework applied to short text documents. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining, pages 31-39.  
+Hongliang Dai, Yangqiu Song, and Haixun Wang. 2021. Ultra-fine entity typing with weak supervision from a masked language model. 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 1790-1799, Online. Association for Computational Linguistics.
+
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Jiangshu Du, Wenpeng Yin, Congying Xia, and Philip S Yu. 2022. Learning to select from multiple options. arXiv preprint arXiv:2212.00301.  
+Xinya Du and Heng Ji. 2022. Retrieval-augmented generative question answering for event argument extraction. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4649-4666, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Nicholas FitzGerald, Julian Michael, Luheng He, and Luke Zettlemoyer. 2018. Large-scale QA-SRL parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2051-2060, Melbourne, Australia. Association for Computational Linguistics.  
+Daniel Gillick, Sayali Kulkarni, Larry Lansing, Alessandro Presta, Jason Baldridge, Eugene Ie, and Diego Garcia-Olano. 2019. Learning dense representations for entity retrieval. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 528-537, Hong Kong, China. Association for Computational Linguistics.  
+Nilesh Gupta, Sakina Bohra, Yashoteja Prabhu, Saurabh Purohit, and Manik Varma. 2021. Generalized zero-shot extreme multi-label learning. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 527-535.  
+Donghoon Ham, Jeong-Gwan Lee, Youngsoo Jang, and Kee-Eung Kim. 2020. End-to-end neural pipeline for goal-oriented dialogue systems using GPT-2. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 583-592, Online. Association for Computational Linguistics.  
+Hangfeng He, Mingyuan Zhang, Qiang Ning, and Dan Roth. 2021. Foreseeing the benefits of incidental supervision. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1782-1800, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Matthew Henderson, Iñigo Casanueva, Nikola Mrkšić, Pei-Hao Su, Tsung-Hsien Wen, and Ivan Vulić. 2020. ConveRT: Efficient and accurate conversational representations from transformers. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2161-2174, Online. Association for Computational Linguistics.
+
+I-Hung Hsu, Kuan-Hao Huang, Elizabeth Boschee, Scott Miller, Prem Natarajan, Kai-Wei Chang, and Nanyun Peng. 2022. DEGREE: A data-efficient generation-based event extraction model. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1890-1908, Seattle, United States. Association for Computational Linguistics.  
+James Y. Huang, Bangzheng Li, Jiashu Xu, and Muhao Chen. 2022a. Unified semantic typing with meaningful label inference. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2642-2654, Seattle, United States. Association for Computational Linguistics.  
+Kuan-Hao Huang, I-Hung Hsu, Prem Natarajan, Kai-Wei Chang, and Nanyun Peng. 2022b. Multilingual generative language models for zero-shot crosslingual event argument extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4633–4646, Dublin, Ireland. Association for Computational Linguistics.  
+Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547.  
+Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online. Association for Computational Linguistics.  
+Stefan Larson, Anish Mahendran, Joseph J. Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K. Kummerfeld, Kevin Leach, Michael A. Laurenzano, Lingjia Tang, and Jason Mars. 2019. An evaluation dataset for intent classification and out-of-scope prediction. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1311-1316, Hong Kong, China. Association for Computational Linguistics.  
+Duong Le and Thien Huu Nguyen. 2021. Fine-grained event trigger detection. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2745-2752, Online. Association for Computational Linguistics.  
+Jinhyuk Lee, Mujeen Sung, Jaewoo Kang, and Danqi Chen. 2021. Learning dense representations of phrases at scale. 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 6634-6647, Online. Association for Computational Linguistics.  
+Bangzheng Li, Wenpeng Yin, and Muhao Chen. 2022. Ultra-fine entity typing with indirect supervision from natural language inference. Transactions of the Association for Computational Linguistics, 10:607-622.  
+Xiaoya Li, Jingrong Feng, Yuxian Meng, Qinghong Han, Fei Wu, and Jiwei Li. 2020. A unified MRC framework for named entity recognition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5849-5859, Online. Association for Computational Linguistics.  
+Yen-Ting Lin, Alexandros Papangelis, Seokhwan Kim, Sungjin Lee, Devamanyu Hazarika, Mahdi Namazifar, Di Jin, Yang Liu, and Dilek Hakkani-Tur. 2023a. Selective in-context data augmentation for intent detection using pointwise v-information. arXiv preprint arXiv:2302.05096.  
+Yen-Ting Lin, Alexandros Papangelis, Seokhwan Kim, Sungjin Lee, Devamanyu Hazarika, Mahdi Namazifar, Di Jin, Yang Liu, and Dilek Hakkani-Tur. 2023b. Selective in-context data augmentation for intent detection using pointwise V-information. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1463-1476, Dubrovnik, Croatia. Association for Computational Linguistics.  
+Qing Liu, Hongyu Lin, Xinyan Xiao, Xianpei Han, Le Sun, and Hua Wu. 2021a. Fine-grained entity typing via label reasoning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4611-4622, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Xingkun Liu, Arash Eshghi, Pawel Swietojanski, and Verena Rieser. 2021b. Benchmarking natural language understanding services for building conversational agents. In Increasing Naturalness and Flexibility in Spoken Dialogue Interaction: 10th International Workshop on Spoken Dialogue Systems, pages 165-183. Springer.  
+Keming Lu, I-Hung Hsu, Wenxuan Zhou, Mingyu Derek Ma, and Muhao Chen. 2022. Summarization as indirect supervision for relation extraction. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 6575-6594, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Qing Lyu, Hongming Zhang, Elior Sulem, and Dan Roth. 2021. Zero-shot event extraction via transfer learning: Challenges and insights. 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 322-332, Online. Association for Computational Linguistics.
+
+Shikib Mehri and Mihail Eric. 2021. Example-driven intent prediction with observers. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2979-2992, Online. Association for Computational Linguistics.  
+Shikib Mehri, Mihail Eric, and Dilek Hakkani-Tur. 2020. Dialogue: A natural language understanding benchmark for task-oriented dialogue. arXiv preprint arXiv:2009.13570.  
+George A Miller. 1995. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41.  
+Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 188-197, Hong Kong, China. Association for Computational Linguistics.  
+Yasumasa Onoe, Michael Boratko, Andrew McCallum, and Greg Durrett. 2021. Modeling fine-grained entity types with box embeddings. 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 2051-2064, Online. Association for Computational Linguistics.  
+Yasumasa Onoe and Greg Durrett. 2020. Interpretable entity representations through large-scale typing. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 612-624, Online. Association for Computational Linguistics.  
+Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.  
+Dan Roth. 2017. Incidental supervision: Moving beyond supervised learning. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA, pages 4885-4890.  
+Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and TieYan Liu. 2020. Mpnet: Masked and permuted pretraining for language understanding. Advances in Neural Information Processing Systems, 33:16857-16867.  
+Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation
+
+of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).  
+Wei Wu, Fei Wang, Arianna Yuan, Fei Wu, and Jiwei Li. 2020. CorefQA: Coreference resolution as query-based span prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6953-6963, Online. Association for Computational Linguistics.  
+Wenhan Xiong, Jiawei Wu, Deren Lei, Mo Yu, Shiyu Chang, Xiaoxiao Guo, and William Yang Wang. 2019. Imposing label-relational inductive bias for extremely fine-grained entity typing. 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 773-784, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Yuanhao Xiong, Wei-Cheng Chang, Cho-Jui Hsieh, Hsiang-Fu Yu, and Inderjit Dhillon. 2022. Extreme Zero-Shot learning for extreme text classification. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5455-5468, Seattle, United States. Association for Computational Linguistics.  
+Jiashu Xu, Mingyu Derek Ma, and Muhao Chen. 2023. Can nli provide proper indirect supervision for low-resource biomedical relation extraction? In Proceedings of the 2023 Conference of the Association for Computational Linguistics: Human Language Technologies.  
+Nan Xu, Fei Wang, Bangzheng Li, Mingtao Dong, and Muhao Chen. 2022. Does your model classify entities reasonably? diagnosing and mitigating spurious correlations in entity typing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8642-8658, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Yinfei Yang, Daniel Cer, Amin Ahmad, Mandy Guo, Jax Law, Noah Constant, Gustavo Hernandez Abrego, Steve Yuan, Chris Tar, Yun-hsuan Sung, Brian Strope, and Ray Kurzweil. 2020. Multilingual universal sentence encoder for semantic retrieval. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 87-94, Online. Association for Computational Linguistics.  
+Wenpeng Yin, Muhao Chen, Ben Zhou, Qiang Ning, Kai-Wei Chang, and Dan Roth. 2023. Indirectly supervised natural language processing. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 6: Tutorial Abstracts), pages 32–40, Toronto, Canada. Association for Computational Linguistics.
+
+Wenpeng Yin, Jamaal Hay, and Dan Roth. 2019. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3914-3923, Hong Kong, China. Association for Computational Linguistics.  
+Hsiang-Fu Yu, Jiong Zhang, Wei-Cheng Chang, Jyun-Yu Jiang, Wei Li, and Cho-Jui Hsieh. 2022. Pecos: Prediction for enormous and correlated output spaces. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4848-4849.  
+Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1503-1512.  
+Jianguo Zhang, Trung Bui, Seunghyun Yoon, Xiang Chen, Zhiwei Liu, Congying Xia, Quan Hung Tran, Walter Chang, and Philip Yu. 2021. Few-shot intent detection via contrastive pre-training and fine-tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1906-1912, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Jianguo Zhang, Kazuma Hashimoto, Wenhao Liu, Chien-Sheng Wu, Yao Wan, Philip Yu, Richard Socher, and Caiming Xiong. 2020a. Discriminative nearest neighbor few-shot intent detection by transferring natural language inference. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5064-5082, Online. Association for Computational Linguistics.  
+Jianguo Zhang, Kazuma Hashimoto, Yao Wan, Zhiwei Liu, Ye Liu, Caiming Xiong, and Philip Yu. 2022. Are pre-trained transformers robust in intent classification? a missing ingredient in evaluation of out-of-scope intent detection. In Proceedings of the 4th Workshop on NLP for Conversational AI, pages 12-20, Dublin, Ireland. Association for Computational Linguistics.  
+Jianguo Zhang, Kazuma Hashimoto, Chien-Sheng Wu, Yao Wang, Philip Yu, Richard Socher, and Caiming Xiong. 2020b. Find or classify? dual strategy for slot-value predictions on multi-domain dialog state tracking. In Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics, pages 154-167, Barcelona, Spain (Online). Association for Computational Linguistics.  
+Ruohong Zhang, Yau-Shian Wang, Yiming Yang, Donghan Yu, Tom Vu, and Likun Lei. 2023. Long-tailed extreme multi-label text classification by the retrieval of generated pseudo label descriptions. In Findings of the Association for Computational Linguistics:
+
+EACL 2023, pages 1092-1106, Dubrovnik, Croatia.  
+Association for Computational Linguistics.  
+Wenting Zhao, Konstantine Arkoudas, Weiqi Sun, and Claire Cardie. 2022a. Compositional task-oriented parsing as abstractive question answering. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4418-4427, Seattle, United States. Association for Computational Linguistics.  
+Zhongkai Zhao, Bonan Kou, Mohamed Yilmaz Ibrahim, Muhao Chen, and Tianyi Zhang. 2022b. Knowledge-based version incompatibility detection for deep learning. In Proceedings of the 19th International Conference on Mining Software Repositories, pages 247-251.  
+Wenxuan Zhou and Muhao Chen. 2022. An improved baseline for sentence-level relation extraction. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 161-168, Online only. Association for Computational Linguistics.

2023/Dense Retrieval as Indirect Supervision for Large-space Decision Making/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1eb0ce5dd0075355407f138aa03e6e5e6cb8d1df65da5c02cc6fee8ec79b7cf7
+size 548794

2023/Density-Aware Prototypical Network for Few-Shot Relation Classification/3cc00c50-e0b4-4c6e-9a21-2f4635fbe6e7_content_list.json

@@ -0,0 +1,2216 @@
+[
+    {
+        "type": "text",
+        "text": "Density-Aware Prototypical Network for Few-Shot Relation Classification",
+        "text_level": 1,
+        "bbox": [
+            117,
+            89,
+            878,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jianfeng Wu $^{1}$  Mengting Hu $^{2*}$  Yike Wu $^{3}$  Bingzhe Wu $^{4}$  Yalan Xie $^{1}$  Mingming Liu $^{2}$  Renhong Cheng $^{1}$",
+        "bbox": [
+            242,
+            129,
+            757,
+            165
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup> College of Computer Science, Nankai University, <sup>2</sup> College of Software, Nankai University  \n<sup>3</sup> School of Journalism and Communication, Nankai University, <sup>4</sup> Tencent AI Lab  \nwjf@mail.nankai.edu.cn, mthu@nankai.edu.cn",
+        "bbox": [
+            124,
+            164,
+            875,
+            214
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In recent years, few-shot relation classification has evoked many research interests. Yet a more challenging problem, i.e. none-of-the-above (NOTA), is under-explored. Existing works mainly regard NOTA as an extra class and treat it the same as known relations. However, such a solution ignores the overall instance distribution, where NOTA instances are actually outliers and distributed unnaturally compared with known ones. In this paper, we propose a density-aware prototypical network (D-Proto) to treat various instances distinctly. Specifically, we design unique training objectives to separate known instances and isolate NOTA instances, respectively. This produces an ideal instance distribution, where known instances are dense yet NOTAs have a small density. Moreover, we propose a NOTA detection module to further enlarge the density of known samples, and discriminate NOTA and known samples accurately. Experimental results demonstrate that the proposed method outperforms strong baselines with robustness towards various NOTA rates. $^{1}$",
+        "bbox": [
+            141,
+            275,
+            460,
+            615
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            624,
+            258,
+            640
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Relation classification is a fundamental task in natural language processing field, aiming to recognize the relation between two entities in a given sentence. Recently, many works formulate this task in the few-shot learning scenario (Dong et al., 2020; Qu et al., 2020; Han et al., 2021; Brody et al., 2021). They focus on obtaining fast adaptation with a few samples by learning multiple episodes. Usually, each episode is composed of a support set and a query set, where the support set has  $N$ -way  $K$ -shot instances. As an example shown in Table 1, the support set provides two known relations, i.e. \"contains\" and \"founded by\". The first query instance has \"founded by\" relation, which can be",
+        "bbox": [
+            110,
+            649,
+            489,
+            873
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/4c03e3f1acfb2adbcb64c0b0a078f9ac72255fea658098bf6ed327bafefe0def.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Support Set</td></tr><tr><td>(1) contains</td><td>California is a state in the United States.</td></tr><tr><td>(2) founded by</td><td>Steve Jobs was one of the founders of Apple.</td></tr><tr><td colspan=\"2\">Query Set</td></tr><tr><td>(2)</td><td>Bill Gates founded Microsoft in 1975</td></tr><tr><td>NOTA</td><td>Einstein was born in 1879.</td></tr></table>",
+        "bbox": [
+            532,
+            249,
+            860,
+            404
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Table 1: An example for a 2-way 1-shot episode with  $50\\%$  NOTA rate, where 2 indicates the number of classes and each class has one instance. The head entity is marked in red, and the tail entity is marked in blue.",
+        "bbox": [
+            507,
+            414,
+            882,
+            470
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "recognized by reference to this support set. But for the second query instance, the relation between \"Einstein\" and \"1879\" is \"born in\". It cannot find a reference from the support set.",
+        "bbox": [
+            504,
+            498,
+            882,
+            562
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Gao et al. (2019b) define such relation as none-of-the-above (NOTA) and propose a new task setting. Its challenge relies on the accurate detection of both the known instances and NOTA ones. Previous works extend  $N$ -way to  $(N + 1)$ -way, regarding NOTA relation as an extra class. Then the problem is further solved by classification (Gao et al., 2019b), or computing distance with pre-defined learnable vectors (Sabo et al., 2021), or multiple choice (Liu et al., 2022). Despite their popularity and superiority, they only focus on detecting NOTA effectively rather than refining the distribution of NOTA instances, leaving the overall distribution of all instances ignored. However, NOTA is not a simple  $(N + 1)$ -th class, which tends to be outliers for the known  $N$  relations. As an illustration, MNAV may lead some NOTA instances to cluster around a particular vector. However, the clustering of these instances from different categories together is unnatural. Therefore, motivated by this, we study from the view of instance distribution.",
+        "bbox": [
+            507,
+            564,
+            884,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, our initial research question is what",
+        "bbox": [
+            526,
+            903,
+            880,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Mengting Hu is the corresponding author.",
+        "bbox": [
+            139,
+            879,
+            406,
+            892
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>The code is released on GitHub at https://github. com/Pisces-29/DProto.",
+        "bbox": [
+            115,
+            892,
+            487,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "2477",
+        "bbox": [
+            478,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2477-2489 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            779,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "instances distribution is ideal. For known samples, intra-class needs to be well clustered and interclass is clearly separated. And NOTA samples, without gathering, should be distributed away from all known instances. In light of these properties, we propose a density-aware prototypical network (D-Proto) that takes into account the density of different types of instances. Specifically, to guarantee the ideal instance distribution of known instances, we introduce neighborhood component analysis (NCA) (Laenen and Bertinetto, 2021), which involves all sample pairs out of the support and query sets. For NOTA instances, we propose a NOTA loss, which aims at pushing them away from all non-NOTA ones. By making NOTA instances isolated, both of them become more distinguishable in the embedding space.",
+        "bbox": [
+            112,
+            84,
+            489,
+            357
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Moreover, the ideal distribution will lead to distinct densities for NOTA and known samples. That is to say, known instances are distributed densely since the intra-class is clustered together. NOTA's density is small as it is isolated from others. Therefore, we propose a NOTA detection module by fully enhancing the density properties. Concretely, to promote the density of known instances, we design intra-class prototype enhancement. By adding prototypes and degenerative prototypes, their densities could be further enlarged. Based on this, we propose a prototype enhanced local outlier factor (PLOF) to calculate the outlier extent of an instance through distances with its neighbors. Finally, we propose a base score that uses the extremal distance ratio between query samples and  $N$  prototypes. By fusing PLOF and base score, the final NOTA score is obtained to accurately discriminate between NOTA instances and known ones.",
+        "bbox": [
+            115,
+            357,
+            489,
+            662
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, the main contributions of our work are as follows:",
+        "bbox": [
+            112,
+            664,
+            487,
+            694
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose D-Proto to address the more challenging few-shot relation classification task. To the best of our knowledge, this work is the first to model the overall instance distribution in an episode.",
+            "- We introduce the NCA objective and propose a NOTA loss to achieve ideal instance distribution. Based on its ideality, we further propose a prototype enhanced local outlier factor (PLOF) and a base score to fully leverage the density properties.",
+            "- Extensive experimental results demonstrated"
+        ],
+        "bbox": [
+            136,
+            705,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/ef8051553521c3146c5f387142b35ca933f9c7bd08475946c581640ac554e461.jpg",
+        "image_caption": [
+            "Figure 1: The framework overview of D-Proto. In this example, the support set is 2-way 3-shot. During training, two different query samples are for class 1 and NOTA, respectively. During testing, queries are first predicted via distances with prototypes and further revised with a NOTA detection module."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            514,
+            80,
+            882,
+            219
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the effectiveness and robustness of our approach on both fixed and random episode sampling with various NOTA rates.",
+        "bbox": [
+            544,
+            338,
+            884,
+            387
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Methodology",
+        "text_level": 1,
+        "bbox": [
+            507,
+            399,
+            658,
+            416
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Formulation and Method Overview",
+        "text_level": 1,
+        "bbox": [
+            507,
+            425,
+            835,
+            439
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given the training set  $\\mathcal{D}_{train}$ , we sample multiple episodes, where each of them is composed of  $\\{\\mathcal{S},\\mathcal{Q},\\mathcal{R}\\}$ . The relation set  $\\mathcal{R} = \\{r^1,r^2,\\dots,r^N\\}$  mentions  $N$  relations. The support set  $\\mathcal{S}$  follows  $N$ -way  $K$ -shot setting, which contains  $N$  relations, and each relation  $r^i$  has  $K$  instances.",
+        "bbox": [
+            507,
+            445,
+            884,
+            542
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {S} = \\left\\{\\left(s _ {j} ^ {i}, r _ {j} ^ {i}\\right) \\right\\}, 1 \\leq i \\leq N, 1 \\leq j \\leq K\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            547,
+            554,
+            843,
+            574
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $s_j^i$  denotes the  $j$ -th sample belonging to the relation  $r^i$ . The query set  $\\mathcal{Q}$  is composed of two kinds of instances: known instances, which belongs to the relation set  $\\mathcal{R}$  of the support set; NOTA instances, without references from the support set.",
+        "bbox": [
+            507,
+            583,
+            882,
+            668
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {Q} = \\{(q _ {1}, r _ {1}), \\dots , (q _ {n}, r _ {n}), (\\overline {{q}} _ {1}, \\overline {{r}}), \\dots , (\\overline {{q}} _ {m}, \\overline {{r}}) \\}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            521,
+            680,
+            870,
+            697
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $\\{(q_j,r_j)\\}_{j = 1}^n$  are the known instances and  $\\{(\\overline{q}_j,\\overline{r})\\}_{j = 1}^m$  are the NOTA instances. Here  $\\overline{r}$  denotes the NOTA relation. The query set  $\\mathcal{Q}$  is predicted with the help of support set  $S$ . Our goal is to recognize all samples in  $\\mathcal{Q}$  accurately.",
+        "bbox": [
+            507,
+            709,
+            884,
+            789
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The proposed approach is depicted in Figure 1. A sample is encoded with BERT  $E(\\cdot)$  (Devlin et al., 2019), yielding the vector representation of [CLS]. In the embedding space, we design two training objectives: 1) known sample separation  $\\mathcal{L}_{nca}$  aims to increase the separability of known samples by reducing intra-class distances and enlarging inter-class distances; 2) NOTA sample isolation  $\\mathcal{L}_{nota}$",
+        "bbox": [
+            505,
+            791,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "2478",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "push NOTA samples to be distinct from known ones, making them more easily identifiable as outliers. Ultimately, our model is optimized by the combination of two training objectives.",
+        "bbox": [
+            112,
+            84,
+            489,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {n c a} - \\alpha \\mathcal {L} _ {n o t a} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            223,
+            167,
+            487,
+            183
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha$  is a hyper-parameter to balance the contribution of NOTA loss.",
+        "bbox": [
+            112,
+            200,
+            489,
+            231
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the testing phase, we propose NOTA detection module, fusing PLOF and base score into NOTA score. Each sample's NOTA score then reveals its outlier extent. NOTA samples will be easily discriminated from known samples. Next, we will introduce each module in detail.",
+        "bbox": [
+            112,
+            234,
+            489,
+            330
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Known Sample Separation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            347,
+            371,
+            363
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Although considering the overall instance distribution is intuitive, its ideal is not easy to attain in the few-shot scenario. Due to the limitation of samples, a sample and its neighbors are all important for quantifying density. Therefore, we introduce neighborhood component analysis (NCA) loss (Laenen and Bertinetto, 2021). It considers the distances between all sample pairs from both the support set and the query set. In other words,  $K$  samples of some relation are pulled together interactively and away from all other  $K \\times (N - 1)$  samples.",
+        "bbox": [
+            112,
+            370,
+            489,
+            548
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Assume all instances in an episode are  $\\mathcal{B} = \\{(x_i,r_i)|(x_i,r_i)\\in \\{\\mathcal{S}\\cup \\mathcal{Q}\\} ,1\\leq i\\leq b\\}$  then all the known instances are defined as  $\\mathcal{B}_k = \\{(x_i,r_i)|(x_i,r_i)\\in \\mathcal{B},r_i\\neq \\overline{r}\\}$  where  $x_{i}$  is a instance,  $r_i$  is its relation,  $\\overline{r}$  represents the NOTA relation. NCA loss is formulated as below.",
+        "bbox": [
+            112,
+            550,
+            489,
+            646
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal{L}_{nca} = \\frac{-1}{|\\mathcal{B}_{k}|}\\sum_{\\substack{i\\in 1,\\ldots b\\\\ r_{i}\\neq \\overline{r}}}\\log \\\\ \\left( \\begin{array}{l} \\sum_ {j \\in 1,.., b} \\exp (- d ^ {2} (\\boldsymbol {x} _ {i}, \\boldsymbol {x} _ {j})) \\\\ \\frac {\\sum_ {k \\in 1 , . . . , b} \\exp (- d ^ {2} (\\boldsymbol {x} _ {i} , \\boldsymbol {x} _ {k}))}{r _ {k} \\neq \\bar {r}} \\end{array} \\right) \\tag {2} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            163,
+            658,
+            487,
+            808
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\boldsymbol{x}_i = E(\\boldsymbol{x}_i)$  is the vector representation for the instance  $x_i$ .  $d(\\cdot, \\cdot)$  represents the Euclidean distance.  $|\\mathcal{B}_k|$  indicates the number of samples in  $\\mathcal{B}_k$ . With  $\\mathcal{L}_{nca}$ , the intra-class is clustered together and the inter-class is away from each other, making a good distribution for the density of known samples.",
+        "bbox": [
+            112,
+            822,
+            490,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/ee4be0a89719bd16639b8a40bfe4df0c95628a94be7570f289031076d76c5fb8.jpg",
+        "image_caption": [
+            "Figure 2: Illustration of expanding instance sets for each relation in PLOF. We specifically augment with prototypes and degenerative prototypes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            531,
+            80,
+            867,
+            305
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 NOTA Sample Isolation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            384,
+            742,
+            400
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Since NOTA samples are outliers for  $N$  known relation clusters, they should be isolated from known ones. Meanwhile, if a query set has multiple NOTAs, they might belong to different relations like the second query instance of Table 1. Aggregating NOTA samples into the  $(N + 1)$ -th class may cause a negative impact on distinguishing relations. Therefore, we propose NOTA loss to push each NOTA instance away from all the known samples. The NOTA instances are denoted as  $\\mathcal{B}_{\\text{nota}} = \\{(x_i, r_i) | (x_i, r_i) \\in \\mathcal{B}, r_i = \\overline{r}\\}$ .  $\\mathcal{L}_{\\text{nota}}$  is defined as below.",
+        "bbox": [
+            507,
+            406,
+            884,
+            599
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal{L}_{nota} = \\frac{-1}{|\\mathcal{B}_{nota}|}\\sum_{\\substack{i\\in 1,\\ldots ,b\\\\ r_{i} = \\overline{r}}}\\log \\\\ \\left(\\sum_ {\\substack {j \\in 1, \\dots , b \\\\ r _ {j} \\neq \\bar {r}}} \\exp (- d ^ {2} (\\boldsymbol {x} _ {i}, \\boldsymbol {x} _ {j}))\\right) \\tag{3} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            559,
+            609,
+            882,
+            728
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this way, each NOTA sample is isolated from all others, which is beneficial for detecting known relations and NOTA via density.",
+        "bbox": [
+            507,
+            739,
+            882,
+            788
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.4 NOTA Detection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            800,
+            687,
+            814
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To further quantify densities accurately for few-shot samples, we promote the testing phase with PLOF and base score. They are aggregated into the final NOTA score, which indicates an instance's outlier degree through its nearest instances and prototypes. Next, we will introduce them individually.",
+        "bbox": [
+            507,
+            822,
+            884,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "2479",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/dbea161bd948359c2c1f58502fde5ed528b17a5b08513a22adcd67b509be60ec.jpg",
+        "image_caption": [
+            "Figure 3: A 3-neighborhood example. The radius of each circle represents 3-distance."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            189,
+            80,
+            415,
+            227
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Prototype Enhanced Local Outlier Factor (PLOF) Different from the naive prototypical network (Snell et al., 2017), we compute the sample pair distances as Eq. (2). However, during inference, we argue that the prototypes can be leveraged to enlarge the density. Therefore, as depicted in Figure 2, we first calculate the prototype for each relation as below and expand the relation's vector set with its prototype.",
+        "bbox": [
+            112,
+            291,
+            489,
+            436
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {c} ^ {\\boldsymbol {r} ^ {i}} = \\frac {1}{K} \\sum_ {j = 1} ^ {K} E \\left(s _ {j} ^ {i}\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            221,
+            445,
+            487,
+            491
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To further boost the density and improve the identification of NOTA samples, we propose a special type of prototype, referred to as a \"degenerative prototype\", which is computed with a degenerative set of instances.",
+        "bbox": [
+            112,
+            500,
+            487,
+            579
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol{c}_{m}^{r^{i}} = \\frac{1}{K - 1}\\sum_{\\substack{j = 1,\\ldots ,K\\\\ j\\neq m}}E(s_{j}^{i}),1\\leq m\\leq K \\quad (5)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            132,
+            587,
+            487,
+            636
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Intuitively, each degenerative prototype  $c_{m}^{r^i}$  is a vector obtained by averaging  $(K - 1)$  support embeddings belonging to its relation  $r^i$  after removing a support instance  $s_m^i$ . As shown in Figure 2, the augmentation of prototypes and degenerative prototypes significantly increases the density of known instances and the outlier status of NOTAs, thus improving the detection.",
+        "bbox": [
+            112,
+            646,
+            487,
+            778
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Before calculating the PLOF of each instance, the original instance set  $\\mathcal{B}$  is expanded to  $\\mathcal{B}_{plof} = \\mathcal{B} \\cup \\mathcal{P} \\cup \\mathcal{D}\\mathcal{P}$ .",
+        "bbox": [
+            112,
+            778,
+            487,
+            825
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {P} = \\left\\{\\boldsymbol {c} ^ {r ^ {i}} \\right\\}, 1 \\leq i \\leq N\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            834,
+            344,
+            853
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {D P} = \\{c _ {m} ^ {r ^ {i}} \\}, 1 \\leq i \\leq N, 1 \\leq m \\leq K\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            157,
+            858,
+            443,
+            877
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{P}$  is the prototype set and  $\\mathcal{D}\\mathcal{P}$  is the degenerative prototype set.",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/53949cb9b4b54bcb696b4afd9eb057005bab2bb14142fb8b18800c2ef0a58130.jpg",
+        "image_caption": [
+            "Figure 4: An example of a 2-way 3-shot scenario. For a known query, its base score is much smaller than that of a NOTA query."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            564,
+            80,
+            830,
+            231
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Then given a positive integer  $k$  ( $k < |\\mathcal{B}_{plof}|$ ), the  $k$ -distance  $d_{k}(\\boldsymbol{q})$  of an vector  $\\boldsymbol{q}$ , is defined as that for an instance  $\\boldsymbol{o} \\in \\mathcal{B}_{plof} \\backslash \\{\\boldsymbol{q}\\}$ , there are at least  $(k - 1)$  instances  $\\boldsymbol{v}$  of which holds that  $d(\\boldsymbol{q}, \\boldsymbol{v}) < d(\\boldsymbol{q}, \\boldsymbol{o})$ , where  $\\boldsymbol{v} \\in \\mathcal{B}_{plof} \\backslash \\{\\boldsymbol{q}, \\boldsymbol{o}\\}$ . Here the symbol  $\\backslash$  indicates except. After finding such instance  $\\boldsymbol{o}$ ,  $d_{k}(\\boldsymbol{q}) = d(\\boldsymbol{q}, \\boldsymbol{o})$ . The Euclidean distance between  $\\boldsymbol{q}$  and all the chosen instances are less than or equal to  $d_{k}(\\boldsymbol{q})$ . As shown in Figure 3, dashed circles surround the chosen instances for each target  $\\boldsymbol{q}$ . These instances are called the  $k$ -neighborhood set of  $\\boldsymbol{q}$ , denoted as  $N_{k}(\\boldsymbol{q}) = \\{\\boldsymbol{v} \\in \\mathcal{B}_{plof} \\backslash \\{\\boldsymbol{q}\\} | d(\\boldsymbol{q}, \\boldsymbol{v}) \\leq d_{k}(\\boldsymbol{q})\\}$ . Based on  $k$ -neighborhood, the local reachability distance  $lrd$  is defined as follows:",
+        "bbox": [
+            507,
+            306,
+            884,
+            546
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nl r d (\\boldsymbol {q}) = \\frac {\\left| N _ {k} (\\boldsymbol {q}) \\right|}{\\sum_ {\\boldsymbol {o} \\in N _ {k} (\\boldsymbol {q})} \\max  \\left\\{d _ {k} (\\boldsymbol {o}) , d (\\boldsymbol {q} , \\boldsymbol {o}) \\right\\}} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            552,
+            882,
+            590
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Obviously,  $lrd(\\pmb{q})$  is the inverse of the average reachability distance based on the  $k$ -neighborhood of  $\\pmb{q}$ . Based on it, we further define PLOF as the average of the ratio of the local reachability density of  $\\pmb{q}$  and those of  $\\pmb{q}$ 's  $k$ -neighborhood.",
+        "bbox": [
+            507,
+            599,
+            882,
+            678
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP L O F (\\boldsymbol {q}) = \\frac {\\sum_ {\\boldsymbol {v} \\in N _ {k} (\\boldsymbol {q})} \\frac {l r d (\\boldsymbol {v})}{l r d (\\boldsymbol {q})}}{| N _ {k} (\\boldsymbol {q}) |} \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            688,
+            882,
+            730
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "By considering the neighbors, PLOF better reflects the density. Meanwhile, the higher the PLOF value of  $\\pmb{q}$ , the more probable that  $\\pmb{q}$  is an outlier (e.g. a NOTA sample).",
+        "bbox": [
+            507,
+            737,
+            882,
+            802
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Base Score Query instances are recognized with the help of support prototypes. We observe an interesting property on the distances between the query instances and support prototypes. As depicted in Figure 4, the non-NOTA query is close to its own prototype but far away from the other prototypes. Yet the phenomenon is different for the",
+        "bbox": [
+            507,
+            806,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "2480",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "NOTA query, which is far away from all prototypes. This occurs because the proposed NOTA loss  $\\mathcal{L}_{\\text{nota}}$  pushes NOTA instances away from those known instances. Based upon this distance property, we propose a base score, which calculates the ratio of extremal distance to prototypes.",
+        "bbox": [
+            112,
+            84,
+            489,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS c o r e _ {b a s e} (\\boldsymbol {q}) = \\frac {\\underset {i = 1 , \\dots , N} {\\min} d ^ {2} (\\boldsymbol {q} , \\boldsymbol {c} ^ {\\boldsymbol {r} ^ {i}})}{\\underset {i = 1 , \\dots , N} {\\max} d ^ {2} (\\boldsymbol {q} , \\boldsymbol {c} ^ {\\boldsymbol {r} ^ {i}})} \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            186,
+            487,
+            240
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This definition could also help to distinguish non-NOTA and NOTA samples. The base score of a known query is smaller, while it is larger for a NOTA query.",
+        "bbox": [
+            112,
+            244,
+            489,
+            309
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "NOTA Score Finally, the NOTA score is the combination of PLOF and base score.",
+        "bbox": [
+            112,
+            313,
+            489,
+            346
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {S c o r e} _ {\\text {n o t a}} (\\boldsymbol {q}) = \\operatorname {P L O F} (\\boldsymbol {q}) \\cdot \\operatorname {S c o r e} _ {\\text {b a s e}} (\\boldsymbol {q}) \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            353,
+            487,
+            370
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During inference, a query instance can be detected by comparing its NOTA score with a threshold  $\\tau$ . Since the proposed NOTA score is well-designed, using this score can more effectively recognize instances. In the real application, the threshold can be estimated with k-fold validation. In our experiments, we choose a fixed threshold for simplification.",
+        "bbox": [
+            112,
+            376,
+            489,
+            505
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            516,
+            260,
+            532
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            539,
+            228,
+            554
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "FewRel 1.0 (Han et al., 2018) is a relation classification dataset containing 100 relations extracted from Wikipedia. Gao et al. (2019b) update it to FewRel 2.0 and propose the NOTA detection challenge. Our experiments follow the splits used in official FewRel benchmarks, which split the dataset into 64 classes for training, 16 for validation, and 20 for testing.",
+        "bbox": [
+            112,
+            561,
+            489,
+            689
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2 Episode Sampling",
+        "text_level": 1,
+        "bbox": [
+            112,
+            699,
+            302,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In §2.1 we introduce the  $N$ -way  $K$ -shot setting. Each episode sampling support set follows the  $N$ -way  $K$ -shot setting. But there are two ways to sample a query set.",
+        "bbox": [
+            112,
+            720,
+            489,
+            785
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- Fixed Sampling Gao et al. (2019b) propose fixing the NOTA rate of a query set to some specific value. For example, if a query set has 5 known queries and 5 NOTA ones, the NOTA rate of the query set would be calculated as  $5 / (5 + 5) \\times 100\\% = 50\\%$ . Then the total number of instances in that query set will be adjusted based on the NOTA rate.",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- Random Sampling Sabo et al. (2021) propose to fix the number of instances in the query set and assign a probability  $p$  to each instance for sampling. If  $p$  is less than a threshold  $\\tau_{na}$ , the instance is sampled as a known query. Otherwise, it is sampled as a NOTA. As such, the value of  $\\tau_{na}$  determines the NOTA rate. This sampling strategy results in a dynamic NOTA rate for each episode. For example, some episodes may have a NOTA rate of  $100\\%$ , while others may not have any NOTA instances, resulting in a more realistic simulation of real-world scenarios.",
+        "bbox": [
+            509,
+            84,
+            884,
+            275
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Baseline Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            299,
+            695,
+            313
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The following baseline methods are chosen for comparison. O-Proto (Tan et al., 2019) aims to solve few-shot out-of-domain detection. It is based on the prototypical network, which detects NOTA by cosine similarity. BERT-Pair (Gao et al., 2019b) adopts BERT to match a query instance and a support instance, which then yields a score indicating whether they share the same relation. MNAV (Sabo et al., 2021) is also based on the prototypical network. NOTA class is represented by some learnable vectors. A query instance is evaluated by its similarity with the NOTA vectors in the embedding space. MCMN (Liu et al., 2022) uses a pre-training and fine-tuning paradigm that converts all candidate relation descriptions into multiple-choice prompts.",
+        "bbox": [
+            507,
+            319,
+            884,
+            577
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Experimental Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            589,
+            724,
+            604
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4.1 Overall Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            609,
+            690,
+            624
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The results of two sampling strategies are presented in Table 2. It can be observed that D-Proto outperforms strong baselines and achieves the best average results on both sampling strategies. Especially, D-Proto significantly outperforms MCMN by  $+2.75\\%$  on accuracy and  $+2.66\\%$  on F1 under random sampling. The results indicate the effectiveness of D-Proto, with great stability and robustness. As mentioned in §3.2, random sampling better reflects practical applications. It can be seen that under fixed sampling, the proposed D-Proto achieves competitive results. And under random sampling, the improvements are more significant. This implies that D-Proto is more robust in real-world situations.",
+        "bbox": [
+            507,
+            629,
+            884,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Moreover, under random sampling, we see that the accuracy and F1 of MCMN decrease by  $13.36\\%$  and  $16.05\\%$ , respectively when the NOTA rate is",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "2481",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/8c11736ab147532d05df14488e7ab6b17dfccb358ce4168bd34f7842d6357ef5.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"9\">Random Sampling</td></tr><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">15%NOTA</td><td colspan=\"2\">30%NOTA</td><td colspan=\"2\">50%NOTA</td><td colspan=\"2\">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>O-Proto (Tan et al., 2019)</td><td>80.02</td><td>72.84</td><td>80.55</td><td>73.73</td><td>81.16</td><td>72.41</td><td>80.58</td><td>72.99</td></tr><tr><td>BERT-Pair (Gao et al., 2019b)</td><td>80.77</td><td>73.99</td><td>81.53</td><td>75.06</td><td>83.24</td><td>75.14</td><td>81.85</td><td>74.73</td></tr><tr><td>MNAV (Sabo et al., 2021)</td><td>85.02</td><td>78.81</td><td>83.67</td><td>77.78</td><td>81.59</td><td>73.94</td><td>83.42</td><td>76.84</td></tr><tr><td>MCMN (Liu et al., 2022)</td><td>88.24</td><td>83.54</td><td>82.50</td><td>76.39</td><td>74.88</td><td>67.49</td><td>81.87</td><td>75.80</td></tr><tr><td>ChatGPT</td><td>64.57</td><td>56.63</td><td>63.30</td><td>55.81</td><td>61.66</td><td>54.16</td><td>63.18</td><td>55.54</td></tr><tr><td>D-Proto (Ours)</td><td>85.76</td><td>79.79</td><td>84.80</td><td>79.27</td><td>83.30</td><td>76.32</td><td>84.62</td><td>78.46</td></tr><tr><td colspan=\"9\">Fixed Sampling</td></tr><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">15%NOTA</td><td colspan=\"2\">30%NOTA</td><td colspan=\"2\">50%NOTA</td><td colspan=\"2\">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>O-Proto (Tan et al., 2019)</td><td>81.38</td><td>80.45</td><td>81.42</td><td>81.04</td><td>81.38</td><td>79.39</td><td>81.38</td><td>80.29</td></tr><tr><td>BERT-Pair (Gao et al., 2019b)</td><td>82.83</td><td>81.98</td><td>83.76</td><td>83.00</td><td>85.49</td><td>82.62</td><td>84.02</td><td>82.53</td></tr><tr><td>MNAV (Sabo et al., 2021)</td><td>85.05</td><td>84.22</td><td>84.36</td><td>84.45</td><td>82.91</td><td>82.15</td><td>84.10</td><td>83.60</td></tr><tr><td>MCMN (Liu et al., 2022)</td><td>87.56</td><td>84.04</td><td>83.16</td><td>83.40</td><td>75.07</td><td>79.14</td><td>81.93</td><td>82.19</td></tr><tr><td>ChatGPT</td><td>63.59</td><td>56.26</td><td>61.83</td><td>54.46</td><td>58.41</td><td>51.14</td><td>61.28</td><td>53.95</td></tr><tr><td>D-Proto (Ours)</td><td>85.37</td><td>84.67</td><td>84.87</td><td>84.90</td><td>83.05</td><td>82.66</td><td>84.43</td><td>84.08</td></tr></table>",
+        "bbox": [
+            181,
+            80,
+            816,
+            351
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Evaluation results of baseline methods and the proposed D-Proto, in terms of accuracy  $(\\%)$  and F1  $(\\%)$ , on the FewRel dataset. The setting is 5-way 5-shot.",
+        "bbox": [
+            112,
+            360,
+            882,
+            390
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "set from  $15\\%$  to  $50\\%$ . On the contrary, D-Proto only sees decreases of  $2.46\\%$  and  $2.01\\%$ , respectively. A possible reason is that MCMN needs to meta-adapt with the support set. Yet a support set only has  $N$ -way  $K$ -shot samples, while NOTA exists in the query set. This tends to cause overfitting on the known instances but fails to detect NOTA. This further demonstrates the effectiveness and robustness of the proposed D-Proto model toward various NOTA rates.",
+        "bbox": [
+            110,
+            414,
+            489,
+            574
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we evaluate the performance of the ChatGPT $^2$  dialog system in the NOTA challenge. As demonstrated in Table 2, ChatGPT's performance on the NOTA task is far below other models. Therefore, we conclude that the model targeting the NOTA challenge in few-shot relation classification is still necessary. More details can be found in Appendix A.9.",
+        "bbox": [
+            112,
+            577,
+            489,
+            706
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.4.2 NOTA Score Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            718,
+            344,
+            734
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We further investigate the scores computed by DProto, which are depicted through box line diagrams in Figure 5. We first observe that the base scores alone are able to distinguish between known instances and NOTA ones, but the numerical values for the two classes are close. Specifically, the base scores of NOTA instances are generally greater than 0.4, while those of non-NOTA samples<sup>3</sup> are",
+        "bbox": [
+            112,
+            738,
+            489,
+            866
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "around 0.1. This leads to difficulty in choosing a good threshold since the range of proper thresholds is small. Thus, base scores alone may not be sufficient to distinguish between the two classes accurately.",
+        "bbox": [
+            507,
+            414,
+            884,
+            495
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Secondly, without adding prototypes and degenerative prototypes, we have the naive local outlier factor (LOF). It can be seen that both LOF and PLOF show numerical overlap for both non-NOTA and NOTA instances. However, in LOF, the scores for non-NOTA instances are centered around 0.9, while the scores for NOTA instances range from 1.0 to 1.2. The available range of threshold is limited. Contrarily, in PLOF, the scores for non-NOTA samples are in the range of 2 to 3, while the scores for NOTA instances are significantly larger than 4.",
+        "bbox": [
+            507,
+            497,
+            885,
+            673
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Finally, combining PLOF and base NOTA score forms our final NOTA score. It is found that the NOTA score can make two types of instances have a clear difference. The gap between medians becomes more obvious, and the NOTA score for the NOTA instance becomes larger. These all contribute to the detection of NOTA instances, showing the effectiveness of our proposed NOTA score.",
+        "bbox": [
+            507,
+            675,
+            885,
+            803
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.4.3 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            816,
+            690,
+            832
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To explore the impact of individual training objectives, we perform an ablation study. The results are reported in Table 3. It can be seen that removing  $\\mathcal{L}_{\\text {nota }}$  causes significant decreases in all evaluation metrics. This verifies that NOTA loss effectively",
+        "bbox": [
+            507,
+            839,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "2https://chat.openai.com",
+        "bbox": [
+            134,
+            879,
+            319,
+            891
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "3In the context of the section, non-NOTA samples can be considered equivalent to known ones.",
+        "bbox": [
+            115,
+            892,
+            485,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "2482",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/7dc9685c08e53eee1bd64182fb55849b421db44a43532ffc378a823ea522778a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">15%NOTA</td><td colspan=\"2\">30%NOTA</td><td colspan=\"2\">50%NOTA</td><td colspan=\"2\">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>D-Proto w/o Lnota</td><td>84.95</td><td>78.73</td><td>84.04</td><td>78.27</td><td>82.72</td><td>75.41</td><td>83.90</td><td>77.47</td></tr><tr><td>D-Proto w/o Scorebase</td><td>78.44</td><td>70.95</td><td>78.23</td><td>71.39</td><td>77.52</td><td>68.99</td><td>78.06</td><td>70.44</td></tr><tr><td>D-Proto w/o P</td><td>84.24</td><td>78.00</td><td>83.93</td><td>78.27</td><td>83.48</td><td>76.13</td><td>83.88</td><td>77.47</td></tr><tr><td>D-Proto w/o DP</td><td>85.40</td><td>79.20</td><td>83.39</td><td>77.55</td><td>80.32</td><td>73.01</td><td>83.03</td><td>76.58</td></tr><tr><td>D-Proto w/o {P ∪ DP}</td><td>83.27</td><td>76.83</td><td>83.04</td><td>77.10</td><td>82.37</td><td>74.73</td><td>82.89</td><td>76.22</td></tr><tr><td>D-Proto w. 20%DPK-2</td><td>84.04</td><td>77.76</td><td>83.80</td><td>78.10</td><td>83.46</td><td>76.10</td><td>83.77</td><td>77.32</td></tr><tr><td>D-Proto w. 40%DPK-2</td><td>83.60</td><td>77.29</td><td>83.49</td><td>77.47</td><td>83.75</td><td>76.36</td><td>83.61</td><td>77.04</td></tr><tr><td>D-Proto w. 60%DPK-2</td><td>83.35</td><td>76.99</td><td>83.63</td><td>77.88</td><td>83.88</td><td>76.49</td><td>83.62</td><td>77.12</td></tr><tr><td>D-Proto w. 80%DPK-2</td><td>83.19</td><td>76.81</td><td>83.58</td><td>77.83</td><td>84.01</td><td>76.62</td><td>83.59</td><td>77.08</td></tr><tr><td>D-Proto w. 100%DPK-2</td><td>83.70</td><td>77.37</td><td>83.74</td><td>78.02</td><td>83.72</td><td>76.37</td><td>83.72</td><td>77.25</td></tr><tr><td>D-Proto(Ours)</td><td>85.76</td><td>79.79</td><td>84.80</td><td>79.27</td><td>83.30</td><td>76.32</td><td>84.62</td><td>78.46</td></tr></table>",
+        "bbox": [
+            194,
+            80,
+            803,
+            266
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Ablation study results and degenerative prototypes study (%) of D-Proto on FewRel benchmark.",
+        "bbox": [
+            142,
+            275,
+            850,
+            291
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a7a87b5dcdbbb225a9964aa2f828eeadbebfc721f747299148b6c77388be56f7.jpg",
+        "image_caption": [
+            "Figure 5: Comparison of all scores presented in this paper under  $50\\%$  NOTA rate on the FewRel dataset. The dash-dotted line indicates the median of each score."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            144,
+            313,
+            458,
+            536
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "makes NOTA isolated, helping to recognize relations with densities.",
+        "bbox": [
+            112,
+            615,
+            487,
+            646
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Then removing the Base Score significantly impacts the model's performance. The Base Score is essential as it serves as the basis of PLOF. It can also be observed that the removal of either the prototype set  $\\mathcal{P}$  or the degenerative prototype set  $\\mathcal{D}\\mathcal{P}$  results in a significant decline in the evaluation metrics across different NOTA rates. Additionally, the removal of both  $\\mathcal{P}$  and  $\\mathcal{D}\\mathcal{P}$  results in an even more significant drop in the metrics. These findings suggest that both the prototype set and the degenerative prototype set are important components of the PLOF method and contribute significantly to its performance.",
+        "bbox": [
+            112,
+            649,
+            487,
+            857
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.4.4 Degenerative Prototypes Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            868,
+            438,
+            883
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We have a research question: does adding more degenerative prototypes also increase the perfor",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "mance of the model? To answer this question, we design the  $K - 2$  degenerative prototype set  $\\mathcal{D}\\mathcal{P}_{K - 2}$ .  $\\mathcal{D}\\mathcal{P}_{K - 2}$  consists of all  $K - 2$  degenerative prototypes, where each one is computed by averaging the embeddings of the  $K - 2$  support instances belonging to its relation  $r^i$  after excluding two support instances. It is worth noting that D-Proto only adds the degenerative prototypes  $\\mathcal{D}\\mathcal{P}_{K - 1}$ . Will including  $\\mathcal{D}\\mathcal{P}_{K - 2}$  bring further gains?",
+        "bbox": [
+            507,
+            316,
+            884,
+            476
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this experiment, we add various proportions of  $\\mathcal{D}\\mathcal{P}_{K - 2}$  into the support set. The results are displayed in the second part of Table 3. Our expectation is that  $\\mathcal{D}\\mathcal{P}_{K - 2}$  will increase the density of known samples and thus improve the performance. However, the results of this experiment show that while the performance of the D-Proto method does improve when  $\\mathcal{D}\\mathcal{P}_{K - 2}$  is added under a NOTA rate of  $50\\%$ , the performance decreases for NOTA rates of  $15\\%$  and  $30\\%$ . This suggests that while the use of  $\\mathcal{D}\\mathcal{P}_{K - 2}$  can indeed improve the detection of NOTA instances at higher densities, it also leads to the false detection of known instances as NOTA, particularly at lower NOTA rates.",
+        "bbox": [
+            507,
+            479,
+            884,
+            703
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.4.5 Visualization",
+        "text_level": 1,
+        "bbox": [
+            507,
+            719,
+            672,
+            734
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To further evaluate the representations learned by different models, we visualize the embeddings using the t-SNE algorithm (Van der Maaten and Hinton, 2008) in Figure 6. Under both the  $15\\%$  and  $30\\%$  NOTA rates, compared with O-Proto and MNAV, our method can make each relation cluster more densely. For instance, in Figure 6 (a), the relation 1 is scattered around in O-Proto and MNAV. However, in our method, this class is well aggregated. This shows that our method can learn more separable representations for known instances.",
+        "bbox": [
+            507,
+            741,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "2483",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/631389b77f0f4adeab6e8a51e746b43fe97ef04ea1a1ab1e7b85a3002337b166.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            126,
+            93,
+            297,
+            192
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d969ed3391fca5d0b59812f22c37bea2a3a282d2b913ad272971cd527c638c93.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            302,
+            93,
+            475,
+            192
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f0de52b7eef148a92fd7ad42ec45476ad51e9cfc006ed5fb5c918c7e04eed88.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            196,
+            295,
+            294
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/58973ce9ba740ea0bcf43d28eef834e00d5bbc4793fe9e9c07367c1677ad4927.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            305,
+            197,
+            475,
+            294
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/66b539df6d72c462de3454128976d19719f0dbf7d44ab0afb77de0b6dbe3a9c6.jpg",
+        "image_caption": [
+            "(a)  $15\\%$  NOTA rate"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            300,
+            295,
+            399
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/527207db514d81c6bf455acf3bc1810ee3c181e3e023fcbd131c87601ff03d81.jpg",
+        "image_caption": [
+            "(b)  $30\\%$  NOTA rate",
+            "Figure 6: Embedding distributions of different models in the same episode on the FewRel dataset. NOTA rates are  $15\\%$  and  $30\\%$ , respectively. The numbers 1 to 5 represent the relation. The red inverted triangles are NOTA instances."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            305,
+            300,
+            475,
+            399
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Moreover, in Figure 6 (a), we see that the NOTA instance is far away from known samples in our method. This observation is visible in the plot (b). This shows the validity of the proposed NOTA loss  $\\mathcal{L}_{\\text {nota }}$ , making NOTA instances more isolated. In MNAV, the two NOTA instances are close to a known sample. One possible reason is that MNAV makes multiple NOTA instances approach learnable vectors. Different classes may aggregate with each other. This tends to cause negative effects on distinguishing samples, as embeddings are not well separated. In summary, our method learns the best representations. With the well-clustered known samples and the isolated NOTA samples, D-Proto can both recognize them through the density-aware NOTA detection module accurately.",
+        "bbox": [
+            112,
+            525,
+            489,
+            782
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            796,
+            268,
+            810
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Previous works (Zelenko et al., 2003; Mooney and Bunescu, 2005; Zeng et al., 2014; Gormley et al., 2015; Zhou et al., 2016; Kumar, 2017; Zhang et al., 2018) for relation classification usually train models on a labeled dataset with a fixed number of classes, which cannot deal with unseen relations.",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Thus few-shot scenario has drawn much attention recently. By training on multiple episodes, models can learn transferrable knowledge for unseen testing episodes (Hu et al., 2021). Although many efforts (Hendrickx et al., 2010; Zhang et al., 2017; Obamuyide and Vlachos, 2019; Gao et al., 2019a; Baldini Soares et al., 2019; Ye and Ling, 2019; Dong et al., 2020) have been devoted to the few-shot relation classification. However, a real-world issue that has been proposed is the NOTA challenge (Gao et al., 2019b). This issue breaks the assumption of the traditional  $N$ -way  $K$ -shot setting. Therefore, detecting NOTA instances during inference needs to be explored.",
+        "bbox": [
+            507,
+            84,
+            884,
+            309
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Promising works have been proposed to detect NOTA instances in few-shot relation classification. One approach that has been widely adopted is the  $N + 1$  classification. This approach has been proposed by several works. One work is to use a sentence-pair model for classification, as proposed by Gao et al. (2019b) in their work on few-shot learning. Another work, Sabo et al. (2021) in their work on revisiting NOTA challenge, is to compute the distance between learnable vectors. The third one, proposed by Liu et al. (2022), adopts the pretraining fine-tuning paradigm and selects the correct relation from a set of options that includes NOTA. The above three methods, by considering NOTA instances as an additional class, are able to learn the boundary between the known relations and the NOTA instances, allowing the model to learn a more comprehensive representation of the data.",
+        "bbox": [
+            507,
+            309,
+            884,
+            613
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            636,
+            640,
+            651
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose D-Proto, a density-aware prototypical network for few-shot relation classification. We focus on the overall instance distribution and design special training objectives for distinct samples. For the known samples, intraclass becomes well-clustered and inter-class is well-separated. The NOTA ones are isolated as outliers. We further propose a NOTA detection module, which considers the local density and distance property, to distinguish both types of instances. And we use prototypes and degenerative prototypes to enhance the density of known instances to better identify NOTA. Experiments on a popular dataset demonstrate that D-Proto outperforms strong baselines. We also provide score analysis and visualization to verify the effectiveness of our method.",
+        "bbox": [
+            505,
+            661,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "2484",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we study the few-shot relation classification task, which contains NOTA instances. The proposed D-Proto method achieves some gains on this problem. Yet our work still has the following two limitations.",
+        "bbox": [
+            112,
+            110,
+            487,
+            187
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The first limitation is that NOTA instances are detected with a threshold  $\\tau$ . This threshold relies on human experiences or k-fold validation. However, using a threshold in the binary way methods, including O-Proto and ours, is not a big problem. Our viewpoint is that  $N + 1$  classification can efficiently accomplish the NOTA detection task by treating the NOTA as an additional class to classify with the known  $N$  classes. But the  $N + 1$  classification method ignores the distribution of the data. Therefore, despite that we need a threshold, the advantages of our method are also obvious.",
+        "bbox": [
+            115,
+            191,
+            489,
+            382
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The second limitation is that our experiment is not absolutely fair. The reason is that BERT-Pair (Gao et al., 2019b) requires high GPU memory. In the fixed sampling method of the episode, we used two NVIDIA RTX A6000 with 48G memory each to train BERT-Pair. The other baselines are trained with an NVIDIA TESLA V100-32G. Except for this difference, all other settings are the same.",
+        "bbox": [
+            112,
+            384,
+            487,
+            512
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgment",
+        "text_level": 1,
+        "bbox": [
+            114,
+            526,
+            270,
+            542
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We sincerely thank all the anonymous reviewers for providing valuable feedback. This work is supported by the National Natural Science Foundation of China (Grant No. 62302245), the Ministry of Education of the People's Republic of China Humanities and Social Sciences Youth Foundation (Grant No. 23YJCZH240), and the youth program of National Science Fund of Tianjin, China (Grant No. 22JCQNJC01340).",
+        "bbox": [
+            112,
+            552,
+            489,
+            697
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            725,
+            213,
+            739
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Livio Baldini Soares, Nicholas FitzGerald, Jeffrey Ling, and Tom Kwiatkowski. 2019. Matching the blanks: Distributional similarity for relation learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 2895-2905.",
+            "Sam Brody, Sichao Wu, and Adrian Benton. 2021. Towards realistic few-shot relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5338-5345, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            747,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "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 (NAACL-HLT), pages 4171-4186.",
+            "Bowen Dong, Yuan Yao, Ruobing Xie, Tianyu Gao, Xu Han, Zhiyuan Liu, Fen Lin, Leyu Lin, and Maosong Sun. 2020. Meta-information guided meta-learning for few-shot relation classification. In Proceedings of the 28th International Conference on Computational Linguistics (COLING), pages 1594-1605.",
+            "Tianyu Gao, Xu Han, Zhiyuan Liu, and Maosong Sun. 2019a. Hybrid attention-based prototypical networks for noisy few-shot relation classification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 33, pages 6407-6414.",
+            "Tianyu Gao, Xu Han, Hao Zhu, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. 2019b. Fewrel 2.0: Towards more challenging few-shot relation classification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6250-6255.",
+            "Matthew R. Gormley, Mo Yu, and Mark Dredze. 2015. Improved relation extraction with feature-rich compositional embedding models. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1774-1784.",
+            "Jiale Han, Bo Cheng, and Wei Lu. 2021. Exploring task difficulty for few-shot relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2605-2616, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Xu Han, Hao Zhu, Pengfei Yu, Ziyun Wang, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2018. Fewrel: A large-scale supervised few-shot relation classification dataset with state-of-the-art evaluation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4803-4809.",
+            "Iris Hendrickx, Su Nam Kim, Zornitsa Kozareva, Preslav Nakov, Diarmuid O Seaghdha, Sebastian Padó, Marco Pennacchiotti, Lorenza Romano, and Stan Szpakowicz. 2010. Semeval-2010 task 8: Multiway classification of semantic relations between pairs of nominals. In Proceedings of the 5th International Workshop on Semantic Evaluation, pages 33-38.",
+            "Mengting Hu, Shiwan Zhao, Honglei Guo, Chao Xue, Hang Gao, Tiegang Gao, Renhong Cheng, and Zhong Su. 2021. Multi-label few-shot learning for aspect category detection. In Proceedings of the 59th"
+        ],
+        "bbox": [
+            510,
+            85,
+            885,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "2485",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP), pages 6330-6340.",
+            "Shantanu Kumar. 2017. A survey of deep learning methods for relation extraction. arXiv preprint arXiv:1705.03645.",
+            "Steinar Laenen and Luca Bertinetto. 2021. On episodes, prototypical networks, and few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 34.",
+            "Fangchao Liu, Hongyu Lin, Xianpei Han, Boxi Cao, and Le Sun. 2022. Pre-training to match for unified low-shot relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5785-5795, Dublin, Ireland. Association for Computational Linguistics.",
+            "Raymond Mooney and Razvan Bunescu. 2005. Subsequence kernels for relation extraction. Advances in neural information processing systems (NeurIPS), 18.",
+            "Abiola Obamuyide and Andreas Vlachos. 2019. Model-agnostic meta-learning for relation classification with limited supervision. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5873-5879, Florence, Italy. Association for Computational Linguistics.",
+            "Meng Qu, Tianyu Gao, Louis-Pascal Xhonneux, and Jian Tang. 2020. Few-shot relation extraction via bayesian meta-learning on relation graphs. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 7867-7876.",
+            "Ofer Sabo, Yanai Elazar, Yoav Goldberg, and Ido Dagan. 2021. Revisiting few-shot relation classification: Evaluation data and classification schemes. Transactions of the Association for Computational Linguistics (TACL), 9:691-706.",
+            "Jake Snell, Kevin Swersky, and Richard Zemel. 2017. Prototypical networks for few-shot learning. Advances in neural information processing systems (NeurIPS), 30.",
+            "Ming Tan, Yang Yu, Haoyu Wang, Dakuo Wang, Saloni Potdar, Shiyu Chang, and Mo Yu. 2019. Out-of-domain detection for low-resource text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3566-3572.",
+            "Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research (JMLR), 9(11)."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zhi-Xiu Ye and Zhen-Hua Ling. 2019. Multi-level matching and aggregation network for few-shot relation classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 2872-2881.",
+            "Dmitry Zelenko, Chinatsu Aone, and Anthony Richardella. 2003. Kernel methods for relation extraction. Journal of machine learning research (JMLR), 3(Feb):1083-1106.",
+            "Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, and Jun Zhao. 2014. Relation classification via convolutional deep neural network. In Proceedings of the 28th International Conference on Computational Linguistics (COLING), pages 2335-2344.",
+            "Ningyu Zhang, Shumin Deng, Zhanling Sun, Xi Chen, Wei Zhang, and Huajun Chen. 2018. Attention-based capsule networks with dynamic routing for relation extraction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (ACL), pages 986-992.",
+            "Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor Angeli, and Christopher D Manning. 2017. Position-aware attention and supervised data improve slot filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 35-45.",
+            "Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen Li, Hongwei Hao, and Bo Xu. 2016. Attention-based bidirectional long short-term memory networks for relation classification. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL), pages 207-212."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            556
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "2486",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/39e78377a37d273e463e1cd2f44ec80b8428d2571b5f6f2149d5eef65dc3f23d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"9\">Random Sampling</td></tr><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">15%NOTA</td><td colspan=\"2\">30%NOTA</td><td colspan=\"2\">50%NOTA</td><td colspan=\"2\">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>O-Proto</td><td>80.02±0.65</td><td>72.84±0.88</td><td>80.55±0.97</td><td>73.73±1.18</td><td>81.16±1.04</td><td>72.41±1.28</td><td>80.58</td><td>72.99</td></tr><tr><td>BERT-Pair</td><td>80.77±2.35</td><td>73.99±2.67</td><td>81.53±1.60</td><td>75.06±1.92</td><td>83.24±0.85</td><td>75.14±1.38</td><td>81.85</td><td>74.73</td></tr><tr><td>MNAV</td><td>85.02±0.39</td><td>78.81±0.47</td><td>83.67±0.66</td><td>77.78±0.90</td><td>81.59±1.02</td><td>73.94±1.18</td><td>83.42</td><td>76.84</td></tr><tr><td>MCMN</td><td>88.24±1.83</td><td>83.54±2.45</td><td>82.50±3.40</td><td>76.39±4.34</td><td>74.88±5.83</td><td>67.49±6.64</td><td>81.87</td><td>75.80</td></tr><tr><td>ChatGPT</td><td>64.57±0.31</td><td>56.63±0.41</td><td>63.30±0.57</td><td>55.81±0.40</td><td>61.66±0.76</td><td>54.16±0.45</td><td>63.18</td><td>55.54</td></tr><tr><td>D-Proto</td><td>85.76±0.72</td><td>79.79±0.96</td><td>84.80±0.92</td><td>79.27±1.12</td><td>83.30±1.29</td><td>76.32±1.48</td><td>84.62</td><td>78.46</td></tr></table>",
+        "bbox": [
+            115,
+            82,
+            884,
+            215
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/142628a9d1ec856cdc35d9079d7233b3e777babd975ee3addbf4b8ef1c63656a.jpg",
+        "table_caption": [
+            "Fixed Sampling"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">15%NOTA</td><td colspan=\"2\">30%NOTA</td><td colspan=\"2\">50%NOTA</td><td colspan=\"2\">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>O-Proto</td><td>81.38±1.26</td><td>80.45±1.29</td><td>81.42±1.11</td><td>81.04±1.19</td><td>81.38±0.96</td><td>79.39±1.06</td><td>81.38</td><td>80.29</td></tr><tr><td>BERT-Pair</td><td>82.83±0.62</td><td>81.98±0.62</td><td>83.76±0.46</td><td>83.00±0.54</td><td>85.49±0.29</td><td>82.62±0.45</td><td>84.02</td><td>82.53</td></tr><tr><td>MNAV</td><td>85.05±0.97</td><td>84.22±1.03</td><td>84.36±1.34</td><td>84.45±1.15</td><td>82.91±2.61</td><td>82.15±1.85</td><td>84.10</td><td>83.60</td></tr><tr><td>MCMN</td><td>87.56±1.93</td><td>84.04±2.52</td><td>83.16±3.37</td><td>83.40±3.08</td><td>75.07±5.79</td><td>79.14±3.93</td><td>81.93</td><td>82.19</td></tr><tr><td>ChatGPT</td><td>63.59±0.50</td><td>56.26±0.27</td><td>61.83±0.41</td><td>54.46±0.32</td><td>58.41±0.51</td><td>51.14±0.43</td><td>61.28</td><td>53.95</td></tr><tr><td>D-Proto</td><td>85.37±0.61</td><td>84.67±0.61</td><td>84.87±0.71</td><td>84.90±0.66</td><td>83.05±1.00</td><td>82.66±1.10</td><td>84.43</td><td>84.08</td></tr></table>",
+        "bbox": [
+            115,
+            231,
+            882,
+            351
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Reproducibility",
+        "text_level": 1,
+        "bbox": [
+            112,
+            414,
+            290,
+            430
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.1 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            442,
+            349,
+            458
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Following Gao et al. (2019b), we exploit the same hyper-parameters. For a fair comparison, all baselines and our work use the same encoder. We choose BERT-base-uncased (Devlin et al., 2019) as the encoder. It is initialized by the pre-trained parameters and optimized during training. For the FewRel dataset, we set the batch size to 2, which means we feed two episodes into the model per batch. Each model is trained for 30 epochs with 1000 batches per epoch and tested with 10000 batches in the testing. An early stopping strategy is adopted, indicating the model will stop training when the performance on the validation set does not improve for 6 epochs. The best model on the validation set is saved for evaluation.",
+        "bbox": [
+            112,
+            464,
+            487,
+            705
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In Eq. (1), we set the hyper-parameters  $\\alpha$  to 1e-5 on the FewRel. For  $k$  of the local outlier factor in Eq. (7), we set it to be  $K$ . The reason is that in the  $N$ -way  $K$ -shot setting, a support instance has other  $(K - 1)$  neighbors with the same relation. And each relation has at least one prototype. This is consistent with the definition of  $k$ -neighborhood. All the reported results using the FewRel dataset are the average of five runs with fixed seeds [5, 10, 15, 20, 25].",
+        "bbox": [
+            112,
+            708,
+            487,
+            866
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "After the hyperparameter search, the threshold  $\\pmb{\\tau}$  is set to 0.9 in the experimental setting of FewRel.",
+        "bbox": [
+            112,
+            870,
+            487,
+            901
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "For MCMN, the sampling method in the pre",
+        "bbox": [
+            131,
+            903,
+            487,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/ee626e2f5875085deffd1d8fe95e50557059d14a6f708db992b22011486b4176.jpg",
+        "table_caption": [
+            "Table 4: Evaluation results of baseline methods and the proposed D-Proto, in terms of accuracy  $(\\%)$  and F1  $(\\%)$ , on the FewRel dataset. This table includes the standard deviation of baseline methods and D-Proto."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Number of parameters</td></tr><tr><td>O-Proto</td><td>109482240</td></tr><tr><td>MNAV</td><td>109497600</td></tr><tr><td>BERT-Pair</td><td>109483778</td></tr><tr><td>D-Proto</td><td>109482240</td></tr></table>",
+        "bbox": [
+            552,
+            412,
+            842,
+            511
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 5: The number of hyperparameters for each model.",
+        "bbox": [
+            507,
+            519,
+            882,
+            546
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "training follows Liu et al. (2022), while the sampling method in the testing is fixed sampling or random sampling.",
+        "bbox": [
+            507,
+            577,
+            882,
+            627
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.2 Computing Infrastructure",
+        "text_level": 1,
+        "bbox": [
+            507,
+            645,
+            766,
+            661
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Since BERT-Pair (Gao et al., 2019b) requires high GPU memory in the fixed sampling of 5-way 5-shot setting, we use 2 NVIDIA A6000s with 48G memory each to train the model. The rest of the baselines and our model are trained on a NVIDIA TESLA V100 with 32G of GPU memory.",
+        "bbox": [
+            507,
+            670,
+            882,
+            766
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.3 Number of Parameters per Model",
+        "text_level": 1,
+        "bbox": [
+            507,
+            785,
+            826,
+            802
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 5 shows the number of hyperparameters for each model.",
+        "bbox": [
+            507,
+            810,
+            882,
+            841
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.4 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            860,
+            710,
+            876
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We use two primary evaluation metrics, accuracy and macro F1. The basic implementation of accu",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "2487",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/311b5ed07bb5416c5b115290797dd48bebd4c59688f60d6e3ee53ca6fe611f2e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Number of instances</td></tr><tr><td>D-Proto w/o P</td><td>53</td></tr><tr><td>D-Proto w/o DP</td><td>33</td></tr><tr><td>D-Proto w/o {P ∪ DP}</td><td>28</td></tr><tr><td>D-Proto w. 20%DPK-2</td><td>68</td></tr><tr><td>D-Proto w. 40%DPK-2</td><td>78</td></tr><tr><td>D-Proto w. 60%DPK-2</td><td>88</td></tr><tr><td>D-Proto w. 80%DPK-2</td><td>98</td></tr><tr><td>D-Proto w. 100%DPK-2</td><td>108</td></tr><tr><td>D-Proto(Ours)</td><td>58</td></tr></table>",
+        "bbox": [
+            136,
+            80,
+            467,
+            231
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 6: The number of instances for each method in an episode. The experiment setting is 5-way 5-shot. The number of query instances is 3.",
+        "bbox": [
+            112,
+            240,
+            489,
+            285
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "racy is as follows:",
+        "bbox": [
+            112,
+            311,
+            253,
+            326
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nA c c u r a c y = \\frac {T P + T N}{T P + F P + T N + F N}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            154,
+            338,
+            443,
+            370
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "where  $TP$  are true positives,  $FP$  are false positives,  $TN$  are true negatives, and  $FN$  are false negatives. The definition of F1 is as follows:",
+        "bbox": [
+            112,
+            381,
+            489,
+            429
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP = \\frac {T P}{T P + F P}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            228,
+            439,
+            356,
+            470
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR = \\frac {T P}{T P + F N}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            231,
+            472,
+            357,
+            502
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF 1 = \\frac {2 \\times P \\times R}{P + R}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            231,
+            504,
+            369,
+            536
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Macro F1 calculates F1 for each class and finds their unweighted mean.",
+        "bbox": [
+            112,
+            548,
+            485,
+            580
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.5 Hyperparameters",
+        "text_level": 1,
+        "bbox": [
+            114,
+            593,
+            305,
+            609
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Following Gao et al. (2019b), the hyperparameters not mentioned in Section A.1 are as follows:",
+        "bbox": [
+            112,
+            615,
+            485,
+            646
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Learning rate: 2e-5.",
+            "Weight decay: 0.01.",
+            "Warmup steps: 300.",
+            "Gradient accumulation: 1.",
+            "Max length: 128.",
+            "- Hidden size: 768.",
+            "Test batch: 10000."
+        ],
+        "bbox": [
+            136,
+            659,
+            347,
+            850
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.6 Links to Dataset",
+        "text_level": 1,
+        "bbox": [
+            114,
+            865,
+            294,
+            879
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "FewRel (Han et al., 2018): https://thunlp.github.io/2/fewrel2nota.html.",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.7 Links to Toolkit",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            685,
+            98
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- BERT: https://huggingface.co/ transformers/v4.5.1/model_doc/bert.html#bertmodel",
+            "- BERT Tokenizer: https://huggingface.co/transformers/v4.5.1/model_doc/bert.html#berttokenizer",
+            "- Local Outlier Factor(LOF): https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.\n- LocalOutlierFactor.html",
+            "- Macro F1: https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html"
+        ],
+        "bbox": [
+            531,
+            108,
+            884,
+            366
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.8 Number of Instances per Method",
+        "text_level": 1,
+        "bbox": [
+            509,
+            386,
+            820,
+            401
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 6 shows the number of instances included in an episode for each method in ablation studying. The episode sampling method is random sampling.",
+        "bbox": [
+            507,
+            410,
+            882,
+            458
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.9 Prompt of ChatGPT",
+        "text_level": 1,
+        "bbox": [
+            509,
+            474,
+            721,
+            489
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate the prompt for using ChatGPT directly in the NOTA task, as follows:",
+        "bbox": [
+            507,
+            498,
+            880,
+            529
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question: The purpose of the relation classification task is to identify the relation between two entities (one is the head entity and the other is the tail entity) in a given sentence. For example, the following sentence \"California is a state in the United States\". Through the derivation of the whole sentence, the relation between the head entity \"California\" and the tail entity the \"United States\" is \"contains\". While the few-shot relation classification task consists of a support set and a query set. The support set generally contains  $N$  relations, and each relation contains  $K$  instances. This setup is called  $N$ -way  $K$ -shot. We need to predict the relation between two entities in each sentence given in the query set based on the support set. Now I need you to complete a few-shot relation classification task, which is set to 5-way 5-shot. The five categories in the support set are numbered from 0 to 4, respectively. Five sentences are given after",
+        "bbox": [
+            532,
+            550,
+            858,
+            904
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "2488",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "each category. The head entity of each sentence is surrounded by two special characters [head] and [/head], and the tail entity is surrounded by two special characters [tail] and [/tail]. You need to predict the relationship between two entities for each sentence in the query set from the support set. That is, if the relation of a query instance is similar to that of all instances of category 0 in the support set, then the category predicted by the query instance is category 0. Notice! If the relation between the two entities of a sentence in the query set cannot be referenced in the support set, the relation between the two entities of the sentence is called NOTA (None-of-the-above). When the NOTA relation is encountered, the category number of the predicted result is 5. Warning! You just need to return the class number for each instance predicted in the query set. For example, the relation of the first instance in the query set is relation 0, and the second instance is relation 5 (NOTA). Then your reply is: [0, 5]. If you get it, then I'll give you support sets and query sets next.",
+        "bbox": [
+            139,
+            91,
+            463,
+            508
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "ChatGPT: Yes, I understand the task. Please provide me with the support set and query set.",
+        "bbox": [
+            139,
+            510,
+            460,
+            558
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question: The following is the support set (5-way 5-shot. Instances in each category are separated by a special character [sep].):",
+        "bbox": [
+            139,
+            558,
+            460,
+            609
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "···",
+        "bbox": [
+            139,
+            611,
+            173,
+            621
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The following is the query set (The query set contains one instance.):",
+        "bbox": [
+            139,
+            623,
+            460,
+            653
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "··",
+        "bbox": [
+            141,
+            659,
+            173,
+            670
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "2489",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Density-Aware Prototypical Network for Few-Shot Relation Classification/3cc00c50-e0b4-4c6e-9a21-2f4635fbe6e7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae987a96340228ca57eda533605cb5677cb5492e22fe6286eb48fd2ca681d22d
+size 549193

2023/Density-Aware Prototypical Network for Few-Shot Relation Classification/full.md

@@ -0,0 +1,419 @@
+# Density-Aware Prototypical Network for Few-Shot Relation Classification
+
+Jianfeng Wu $^{1}$  Mengting Hu $^{2*}$  Yike Wu $^{3}$  Bingzhe Wu $^{4}$  Yalan Xie $^{1}$  Mingming Liu $^{2}$  Renhong Cheng $^{1}$
+
+<sup>1</sup> College of Computer Science, Nankai University, <sup>2</sup> College of Software, Nankai University  
+<sup>3</sup> School of Journalism and Communication, Nankai University, <sup>4</sup> Tencent AI Lab  
+wjf@mail.nankai.edu.cn, mthu@nankai.edu.cn
+
+# Abstract
+
+In recent years, few-shot relation classification has evoked many research interests. Yet a more challenging problem, i.e. none-of-the-above (NOTA), is under-explored. Existing works mainly regard NOTA as an extra class and treat it the same as known relations. However, such a solution ignores the overall instance distribution, where NOTA instances are actually outliers and distributed unnaturally compared with known ones. In this paper, we propose a density-aware prototypical network (D-Proto) to treat various instances distinctly. Specifically, we design unique training objectives to separate known instances and isolate NOTA instances, respectively. This produces an ideal instance distribution, where known instances are dense yet NOTAs have a small density. Moreover, we propose a NOTA detection module to further enlarge the density of known samples, and discriminate NOTA and known samples accurately. Experimental results demonstrate that the proposed method outperforms strong baselines with robustness towards various NOTA rates. $^{1}$
+
+# 1 Introduction
+
+Relation classification is a fundamental task in natural language processing field, aiming to recognize the relation between two entities in a given sentence. Recently, many works formulate this task in the few-shot learning scenario (Dong et al., 2020; Qu et al., 2020; Han et al., 2021; Brody et al., 2021). They focus on obtaining fast adaptation with a few samples by learning multiple episodes. Usually, each episode is composed of a support set and a query set, where the support set has  $N$ -way  $K$ -shot instances. As an example shown in Table 1, the support set provides two known relations, i.e. "contains" and "founded by". The first query instance has "founded by" relation, which can be
+
+<table><tr><td colspan="2">Support Set</td></tr><tr><td>(1) contains</td><td>California is a state in the United States.</td></tr><tr><td>(2) founded by</td><td>Steve Jobs was one of the founders of Apple.</td></tr><tr><td colspan="2">Query Set</td></tr><tr><td>(2)</td><td>Bill Gates founded Microsoft in 1975</td></tr><tr><td>NOTA</td><td>Einstein was born in 1879.</td></tr></table>
+
+Table 1: An example for a 2-way 1-shot episode with  $50\%$  NOTA rate, where 2 indicates the number of classes and each class has one instance. The head entity is marked in red, and the tail entity is marked in blue.
+
+recognized by reference to this support set. But for the second query instance, the relation between "Einstein" and "1879" is "born in". It cannot find a reference from the support set.
+
+Gao et al. (2019b) define such relation as none-of-the-above (NOTA) and propose a new task setting. Its challenge relies on the accurate detection of both the known instances and NOTA ones. Previous works extend  $N$ -way to  $(N + 1)$ -way, regarding NOTA relation as an extra class. Then the problem is further solved by classification (Gao et al., 2019b), or computing distance with pre-defined learnable vectors (Sabo et al., 2021), or multiple choice (Liu et al., 2022). Despite their popularity and superiority, they only focus on detecting NOTA effectively rather than refining the distribution of NOTA instances, leaving the overall distribution of all instances ignored. However, NOTA is not a simple  $(N + 1)$ -th class, which tends to be outliers for the known  $N$  relations. As an illustration, MNAV may lead some NOTA instances to cluster around a particular vector. However, the clustering of these instances from different categories together is unnatural. Therefore, motivated by this, we study from the view of instance distribution.
+
+In this paper, our initial research question is what
+
+instances distribution is ideal. For known samples, intra-class needs to be well clustered and interclass is clearly separated. And NOTA samples, without gathering, should be distributed away from all known instances. In light of these properties, we propose a density-aware prototypical network (D-Proto) that takes into account the density of different types of instances. Specifically, to guarantee the ideal instance distribution of known instances, we introduce neighborhood component analysis (NCA) (Laenen and Bertinetto, 2021), which involves all sample pairs out of the support and query sets. For NOTA instances, we propose a NOTA loss, which aims at pushing them away from all non-NOTA ones. By making NOTA instances isolated, both of them become more distinguishable in the embedding space.
+
+Moreover, the ideal distribution will lead to distinct densities for NOTA and known samples. That is to say, known instances are distributed densely since the intra-class is clustered together. NOTA's density is small as it is isolated from others. Therefore, we propose a NOTA detection module by fully enhancing the density properties. Concretely, to promote the density of known instances, we design intra-class prototype enhancement. By adding prototypes and degenerative prototypes, their densities could be further enlarged. Based on this, we propose a prototype enhanced local outlier factor (PLOF) to calculate the outlier extent of an instance through distances with its neighbors. Finally, we propose a base score that uses the extremal distance ratio between query samples and  $N$  prototypes. By fusing PLOF and base score, the final NOTA score is obtained to accurately discriminate between NOTA instances and known ones.
+
+In summary, the main contributions of our work are as follows:
+
+- We propose D-Proto to address the more challenging few-shot relation classification task. To the best of our knowledge, this work is the first to model the overall instance distribution in an episode.  
+- We introduce the NCA objective and propose a NOTA loss to achieve ideal instance distribution. Based on its ideality, we further propose a prototype enhanced local outlier factor (PLOF) and a base score to fully leverage the density properties.  
+- Extensive experimental results demonstrated
+
+![](images/ef8051553521c3146c5f387142b35ca933f9c7bd08475946c581640ac554e461.jpg)  
+Figure 1: The framework overview of D-Proto. In this example, the support set is 2-way 3-shot. During training, two different query samples are for class 1 and NOTA, respectively. During testing, queries are first predicted via distances with prototypes and further revised with a NOTA detection module.
+
+the effectiveness and robustness of our approach on both fixed and random episode sampling with various NOTA rates.
+
+# 2 Methodology
+
+# 2.1 Formulation and Method Overview
+
+Given the training set  $\mathcal{D}_{train}$ , we sample multiple episodes, where each of them is composed of  $\{\mathcal{S},\mathcal{Q},\mathcal{R}\}$ . The relation set  $\mathcal{R} = \{r^1,r^2,\dots,r^N\}$  mentions  $N$  relations. The support set  $\mathcal{S}$  follows  $N$ -way  $K$ -shot setting, which contains  $N$  relations, and each relation  $r^i$  has  $K$  instances.
+
+$$
+\mathcal {S} = \left\{\left(s _ {j} ^ {i}, r _ {j} ^ {i}\right) \right\}, 1 \leq i \leq N, 1 \leq j \leq K
+$$
+
+where  $s_j^i$  denotes the  $j$ -th sample belonging to the relation  $r^i$ . The query set  $\mathcal{Q}$  is composed of two kinds of instances: known instances, which belongs to the relation set  $\mathcal{R}$  of the support set; NOTA instances, without references from the support set.
+
+$$
+\mathcal {Q} = \{(q _ {1}, r _ {1}), \dots , (q _ {n}, r _ {n}), (\overline {{q}} _ {1}, \overline {{r}}), \dots , (\overline {{q}} _ {m}, \overline {{r}}) \}
+$$
+
+where  $\{(q_j,r_j)\}_{j = 1}^n$  are the known instances and  $\{(\overline{q}_j,\overline{r})\}_{j = 1}^m$  are the NOTA instances. Here  $\overline{r}$  denotes the NOTA relation. The query set  $\mathcal{Q}$  is predicted with the help of support set  $S$ . Our goal is to recognize all samples in  $\mathcal{Q}$  accurately.
+
+The proposed approach is depicted in Figure 1. A sample is encoded with BERT  $E(\cdot)$  (Devlin et al., 2019), yielding the vector representation of [CLS]. In the embedding space, we design two training objectives: 1) known sample separation  $\mathcal{L}_{nca}$  aims to increase the separability of known samples by reducing intra-class distances and enlarging inter-class distances; 2) NOTA sample isolation  $\mathcal{L}_{nota}$
+
+push NOTA samples to be distinct from known ones, making them more easily identifiable as outliers. Ultimately, our model is optimized by the combination of two training objectives.
+
+$$
+\mathcal {L} = \mathcal {L} _ {n c a} - \alpha \mathcal {L} _ {n o t a} \tag {1}
+$$
+
+where  $\alpha$  is a hyper-parameter to balance the contribution of NOTA loss.
+
+In the testing phase, we propose NOTA detection module, fusing PLOF and base score into NOTA score. Each sample's NOTA score then reveals its outlier extent. NOTA samples will be easily discriminated from known samples. Next, we will introduce each module in detail.
+
+# 2.2 Known Sample Separation
+
+Although considering the overall instance distribution is intuitive, its ideal is not easy to attain in the few-shot scenario. Due to the limitation of samples, a sample and its neighbors are all important for quantifying density. Therefore, we introduce neighborhood component analysis (NCA) loss (Laenen and Bertinetto, 2021). It considers the distances between all sample pairs from both the support set and the query set. In other words,  $K$  samples of some relation are pulled together interactively and away from all other  $K \times (N - 1)$  samples.
+
+Assume all instances in an episode are  $\mathcal{B} = \{(x_i,r_i)|(x_i,r_i)\in \{\mathcal{S}\cup \mathcal{Q}\} ,1\leq i\leq b\}$  then all the known instances are defined as  $\mathcal{B}_k = \{(x_i,r_i)|(x_i,r_i)\in \mathcal{B},r_i\neq \overline{r}\}$  where  $x_{i}$  is a instance,  $r_i$  is its relation,  $\overline{r}$  represents the NOTA relation. NCA loss is formulated as below.
+
+$$
+\begin{array}{l} \mathcal{L}_{nca} = \frac{-1}{|\mathcal{B}_{k}|}\sum_{\substack{i\in 1,\ldots b\\ r_{i}\neq \overline{r}}}\log \\ \left( \begin{array}{l} \sum_ {j \in 1,.., b} \exp (- d ^ {2} (\boldsymbol {x} _ {i}, \boldsymbol {x} _ {j})) \\ \frac {\sum_ {k \in 1 , . . . , b} \exp (- d ^ {2} (\boldsymbol {x} _ {i} , \boldsymbol {x} _ {k}))}{r _ {k} \neq \bar {r}} \end{array} \right) \tag {2} \\ \end{array}
+$$
+
+where  $\boldsymbol{x}_i = E(\boldsymbol{x}_i)$  is the vector representation for the instance  $x_i$ .  $d(\cdot, \cdot)$  represents the Euclidean distance.  $|\mathcal{B}_k|$  indicates the number of samples in  $\mathcal{B}_k$ . With  $\mathcal{L}_{nca}$ , the intra-class is clustered together and the inter-class is away from each other, making a good distribution for the density of known samples.
+
+![](images/ee4be0a89719bd16639b8a40bfe4df0c95628a94be7570f289031076d76c5fb8.jpg)  
+Figure 2: Illustration of expanding instance sets for each relation in PLOF. We specifically augment with prototypes and degenerative prototypes.
+
+# 2.3 NOTA Sample Isolation
+
+Since NOTA samples are outliers for  $N$  known relation clusters, they should be isolated from known ones. Meanwhile, if a query set has multiple NOTAs, they might belong to different relations like the second query instance of Table 1. Aggregating NOTA samples into the  $(N + 1)$ -th class may cause a negative impact on distinguishing relations. Therefore, we propose NOTA loss to push each NOTA instance away from all the known samples. The NOTA instances are denoted as  $\mathcal{B}_{\text{nota}} = \{(x_i, r_i) | (x_i, r_i) \in \mathcal{B}, r_i = \overline{r}\}$ .  $\mathcal{L}_{\text{nota}}$  is defined as below.
+
+$$
+\begin{array}{l} \mathcal{L}_{nota} = \frac{-1}{|\mathcal{B}_{nota}|}\sum_{\substack{i\in 1,\ldots ,b\\ r_{i} = \overline{r}}}\log \\ \left(\sum_ {\substack {j \in 1, \dots , b \\ r _ {j} \neq \bar {r}}} \exp (- d ^ {2} (\boldsymbol {x} _ {i}, \boldsymbol {x} _ {j}))\right) \tag{3} \\ \end{array}
+$$
+
+In this way, each NOTA sample is isolated from all others, which is beneficial for detecting known relations and NOTA via density.
+
+# 2.4 NOTA Detection
+
+To further quantify densities accurately for few-shot samples, we promote the testing phase with PLOF and base score. They are aggregated into the final NOTA score, which indicates an instance's outlier degree through its nearest instances and prototypes. Next, we will introduce them individually.
+
+![](images/dbea161bd948359c2c1f58502fde5ed528b17a5b08513a22adcd67b509be60ec.jpg)  
+Figure 3: A 3-neighborhood example. The radius of each circle represents 3-distance.
+
+Prototype Enhanced Local Outlier Factor (PLOF) Different from the naive prototypical network (Snell et al., 2017), we compute the sample pair distances as Eq. (2). However, during inference, we argue that the prototypes can be leveraged to enlarge the density. Therefore, as depicted in Figure 2, we first calculate the prototype for each relation as below and expand the relation's vector set with its prototype.
+
+$$
+\boldsymbol {c} ^ {\boldsymbol {r} ^ {i}} = \frac {1}{K} \sum_ {j = 1} ^ {K} E \left(s _ {j} ^ {i}\right) \tag {4}
+$$
+
+To further boost the density and improve the identification of NOTA samples, we propose a special type of prototype, referred to as a "degenerative prototype", which is computed with a degenerative set of instances.
+
+$$
+\boldsymbol{c}_{m}^{r^{i}} = \frac{1}{K - 1}\sum_{\substack{j = 1,\ldots ,K\\ j\neq m}}E(s_{j}^{i}),1\leq m\leq K \quad (5)
+$$
+
+Intuitively, each degenerative prototype  $c_{m}^{r^i}$  is a vector obtained by averaging  $(K - 1)$  support embeddings belonging to its relation  $r^i$  after removing a support instance  $s_m^i$ . As shown in Figure 2, the augmentation of prototypes and degenerative prototypes significantly increases the density of known instances and the outlier status of NOTAs, thus improving the detection.
+
+Before calculating the PLOF of each instance, the original instance set  $\mathcal{B}$  is expanded to  $\mathcal{B}_{plof} = \mathcal{B} \cup \mathcal{P} \cup \mathcal{D}\mathcal{P}$ .
+
+$$
+\mathcal {P} = \left\{\boldsymbol {c} ^ {r ^ {i}} \right\}, 1 \leq i \leq N
+$$
+
+$$
+\mathcal {D P} = \{c _ {m} ^ {r ^ {i}} \}, 1 \leq i \leq N, 1 \leq m \leq K
+$$
+
+where  $\mathcal{P}$  is the prototype set and  $\mathcal{D}\mathcal{P}$  is the degenerative prototype set.
+
+![](images/53949cb9b4b54bcb696b4afd9eb057005bab2bb14142fb8b18800c2ef0a58130.jpg)  
+Figure 4: An example of a 2-way 3-shot scenario. For a known query, its base score is much smaller than that of a NOTA query.
+
+Then given a positive integer  $k$  ( $k < |\mathcal{B}_{plof}|$ ), the  $k$ -distance  $d_{k}(\boldsymbol{q})$  of an vector  $\boldsymbol{q}$ , is defined as that for an instance  $\boldsymbol{o} \in \mathcal{B}_{plof} \backslash \{\boldsymbol{q}\}$ , there are at least  $(k - 1)$  instances  $\boldsymbol{v}$  of which holds that  $d(\boldsymbol{q}, \boldsymbol{v}) < d(\boldsymbol{q}, \boldsymbol{o})$ , where  $\boldsymbol{v} \in \mathcal{B}_{plof} \backslash \{\boldsymbol{q}, \boldsymbol{o}\}$ . Here the symbol  $\backslash$  indicates except. After finding such instance  $\boldsymbol{o}$ ,  $d_{k}(\boldsymbol{q}) = d(\boldsymbol{q}, \boldsymbol{o})$ . The Euclidean distance between  $\boldsymbol{q}$  and all the chosen instances are less than or equal to  $d_{k}(\boldsymbol{q})$ . As shown in Figure 3, dashed circles surround the chosen instances for each target  $\boldsymbol{q}$ . These instances are called the  $k$ -neighborhood set of  $\boldsymbol{q}$ , denoted as  $N_{k}(\boldsymbol{q}) = \{\boldsymbol{v} \in \mathcal{B}_{plof} \backslash \{\boldsymbol{q}\} | d(\boldsymbol{q}, \boldsymbol{v}) \leq d_{k}(\boldsymbol{q})\}$ . Based on  $k$ -neighborhood, the local reachability distance  $lrd$  is defined as follows:
+
+$$
+l r d (\boldsymbol {q}) = \frac {\left| N _ {k} (\boldsymbol {q}) \right|}{\sum_ {\boldsymbol {o} \in N _ {k} (\boldsymbol {q})} \max  \left\{d _ {k} (\boldsymbol {o}) , d (\boldsymbol {q} , \boldsymbol {o}) \right\}} \tag {6}
+$$
+
+Obviously,  $lrd(\pmb{q})$  is the inverse of the average reachability distance based on the  $k$ -neighborhood of  $\pmb{q}$ . Based on it, we further define PLOF as the average of the ratio of the local reachability density of  $\pmb{q}$  and those of  $\pmb{q}$ 's  $k$ -neighborhood.
+
+$$
+P L O F (\boldsymbol {q}) = \frac {\sum_ {\boldsymbol {v} \in N _ {k} (\boldsymbol {q})} \frac {l r d (\boldsymbol {v})}{l r d (\boldsymbol {q})}}{| N _ {k} (\boldsymbol {q}) |} \tag {7}
+$$
+
+By considering the neighbors, PLOF better reflects the density. Meanwhile, the higher the PLOF value of  $\pmb{q}$ , the more probable that  $\pmb{q}$  is an outlier (e.g. a NOTA sample).
+
+Base Score Query instances are recognized with the help of support prototypes. We observe an interesting property on the distances between the query instances and support prototypes. As depicted in Figure 4, the non-NOTA query is close to its own prototype but far away from the other prototypes. Yet the phenomenon is different for the
+
+NOTA query, which is far away from all prototypes. This occurs because the proposed NOTA loss  $\mathcal{L}_{\text{nota}}$  pushes NOTA instances away from those known instances. Based upon this distance property, we propose a base score, which calculates the ratio of extremal distance to prototypes.
+
+$$
+S c o r e _ {b a s e} (\boldsymbol {q}) = \frac {\underset {i = 1 , \dots , N} {\min} d ^ {2} (\boldsymbol {q} , \boldsymbol {c} ^ {\boldsymbol {r} ^ {i}})}{\underset {i = 1 , \dots , N} {\max} d ^ {2} (\boldsymbol {q} , \boldsymbol {c} ^ {\boldsymbol {r} ^ {i}})} \tag {8}
+$$
+
+This definition could also help to distinguish non-NOTA and NOTA samples. The base score of a known query is smaller, while it is larger for a NOTA query.
+
+NOTA Score Finally, the NOTA score is the combination of PLOF and base score.
+
+$$
+\operatorname {S c o r e} _ {\text {n o t a}} (\boldsymbol {q}) = \operatorname {P L O F} (\boldsymbol {q}) \cdot \operatorname {S c o r e} _ {\text {b a s e}} (\boldsymbol {q}) \tag {9}
+$$
+
+During inference, a query instance can be detected by comparing its NOTA score with a threshold  $\tau$ . Since the proposed NOTA score is well-designed, using this score can more effectively recognize instances. In the real application, the threshold can be estimated with k-fold validation. In our experiments, we choose a fixed threshold for simplification.
+
+# 3 Experiments
+
+# 3.1 Datasets
+
+FewRel 1.0 (Han et al., 2018) is a relation classification dataset containing 100 relations extracted from Wikipedia. Gao et al. (2019b) update it to FewRel 2.0 and propose the NOTA detection challenge. Our experiments follow the splits used in official FewRel benchmarks, which split the dataset into 64 classes for training, 16 for validation, and 20 for testing.
+
+# 3.2 Episode Sampling
+
+In §2.1 we introduce the  $N$ -way  $K$ -shot setting. Each episode sampling support set follows the  $N$ -way  $K$ -shot setting. But there are two ways to sample a query set.
+
+- Fixed Sampling Gao et al. (2019b) propose fixing the NOTA rate of a query set to some specific value. For example, if a query set has 5 known queries and 5 NOTA ones, the NOTA rate of the query set would be calculated as  $5 / (5 + 5) \times 100\% = 50\%$ . Then the total number of instances in that query set will be adjusted based on the NOTA rate.
+
+- Random Sampling Sabo et al. (2021) propose to fix the number of instances in the query set and assign a probability  $p$  to each instance for sampling. If  $p$  is less than a threshold  $\tau_{na}$ , the instance is sampled as a known query. Otherwise, it is sampled as a NOTA. As such, the value of  $\tau_{na}$  determines the NOTA rate. This sampling strategy results in a dynamic NOTA rate for each episode. For example, some episodes may have a NOTA rate of  $100\%$ , while others may not have any NOTA instances, resulting in a more realistic simulation of real-world scenarios.
+
+# 3.3 Baseline Methods
+
+The following baseline methods are chosen for comparison. O-Proto (Tan et al., 2019) aims to solve few-shot out-of-domain detection. It is based on the prototypical network, which detects NOTA by cosine similarity. BERT-Pair (Gao et al., 2019b) adopts BERT to match a query instance and a support instance, which then yields a score indicating whether they share the same relation. MNAV (Sabo et al., 2021) is also based on the prototypical network. NOTA class is represented by some learnable vectors. A query instance is evaluated by its similarity with the NOTA vectors in the embedding space. MCMN (Liu et al., 2022) uses a pre-training and fine-tuning paradigm that converts all candidate relation descriptions into multiple-choice prompts.
+
+# 3.4 Experimental Results
+
+# 3.4.1 Overall Results
+
+The results of two sampling strategies are presented in Table 2. It can be observed that D-Proto outperforms strong baselines and achieves the best average results on both sampling strategies. Especially, D-Proto significantly outperforms MCMN by  $+2.75\%$  on accuracy and  $+2.66\%$  on F1 under random sampling. The results indicate the effectiveness of D-Proto, with great stability and robustness. As mentioned in §3.2, random sampling better reflects practical applications. It can be seen that under fixed sampling, the proposed D-Proto achieves competitive results. And under random sampling, the improvements are more significant. This implies that D-Proto is more robust in real-world situations.
+
+Moreover, under random sampling, we see that the accuracy and F1 of MCMN decrease by  $13.36\%$  and  $16.05\%$ , respectively when the NOTA rate is
+
+<table><tr><td colspan="9">Random Sampling</td></tr><tr><td rowspan="2">Methods</td><td colspan="2">15%NOTA</td><td colspan="2">30%NOTA</td><td colspan="2">50%NOTA</td><td colspan="2">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>O-Proto (Tan et al., 2019)</td><td>80.02</td><td>72.84</td><td>80.55</td><td>73.73</td><td>81.16</td><td>72.41</td><td>80.58</td><td>72.99</td></tr><tr><td>BERT-Pair (Gao et al., 2019b)</td><td>80.77</td><td>73.99</td><td>81.53</td><td>75.06</td><td>83.24</td><td>75.14</td><td>81.85</td><td>74.73</td></tr><tr><td>MNAV (Sabo et al., 2021)</td><td>85.02</td><td>78.81</td><td>83.67</td><td>77.78</td><td>81.59</td><td>73.94</td><td>83.42</td><td>76.84</td></tr><tr><td>MCMN (Liu et al., 2022)</td><td>88.24</td><td>83.54</td><td>82.50</td><td>76.39</td><td>74.88</td><td>67.49</td><td>81.87</td><td>75.80</td></tr><tr><td>ChatGPT</td><td>64.57</td><td>56.63</td><td>63.30</td><td>55.81</td><td>61.66</td><td>54.16</td><td>63.18</td><td>55.54</td></tr><tr><td>D-Proto (Ours)</td><td>85.76</td><td>79.79</td><td>84.80</td><td>79.27</td><td>83.30</td><td>76.32</td><td>84.62</td><td>78.46</td></tr><tr><td colspan="9">Fixed Sampling</td></tr><tr><td rowspan="2">Methods</td><td colspan="2">15%NOTA</td><td colspan="2">30%NOTA</td><td colspan="2">50%NOTA</td><td colspan="2">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>O-Proto (Tan et al., 2019)</td><td>81.38</td><td>80.45</td><td>81.42</td><td>81.04</td><td>81.38</td><td>79.39</td><td>81.38</td><td>80.29</td></tr><tr><td>BERT-Pair (Gao et al., 2019b)</td><td>82.83</td><td>81.98</td><td>83.76</td><td>83.00</td><td>85.49</td><td>82.62</td><td>84.02</td><td>82.53</td></tr><tr><td>MNAV (Sabo et al., 2021)</td><td>85.05</td><td>84.22</td><td>84.36</td><td>84.45</td><td>82.91</td><td>82.15</td><td>84.10</td><td>83.60</td></tr><tr><td>MCMN (Liu et al., 2022)</td><td>87.56</td><td>84.04</td><td>83.16</td><td>83.40</td><td>75.07</td><td>79.14</td><td>81.93</td><td>82.19</td></tr><tr><td>ChatGPT</td><td>63.59</td><td>56.26</td><td>61.83</td><td>54.46</td><td>58.41</td><td>51.14</td><td>61.28</td><td>53.95</td></tr><tr><td>D-Proto (Ours)</td><td>85.37</td><td>84.67</td><td>84.87</td><td>84.90</td><td>83.05</td><td>82.66</td><td>84.43</td><td>84.08</td></tr></table>
+
+Table 2: Evaluation results of baseline methods and the proposed D-Proto, in terms of accuracy  $(\%)$  and F1  $(\%)$ , on the FewRel dataset. The setting is 5-way 5-shot.
+
+set from  $15\%$  to  $50\%$ . On the contrary, D-Proto only sees decreases of  $2.46\%$  and  $2.01\%$ , respectively. A possible reason is that MCMN needs to meta-adapt with the support set. Yet a support set only has  $N$ -way  $K$ -shot samples, while NOTA exists in the query set. This tends to cause overfitting on the known instances but fails to detect NOTA. This further demonstrates the effectiveness and robustness of the proposed D-Proto model toward various NOTA rates.
+
+Furthermore, we evaluate the performance of the ChatGPT $^2$  dialog system in the NOTA challenge. As demonstrated in Table 2, ChatGPT's performance on the NOTA task is far below other models. Therefore, we conclude that the model targeting the NOTA challenge in few-shot relation classification is still necessary. More details can be found in Appendix A.9.
+
+# 3.4.2 NOTA Score Analysis
+
+We further investigate the scores computed by DProto, which are depicted through box line diagrams in Figure 5. We first observe that the base scores alone are able to distinguish between known instances and NOTA ones, but the numerical values for the two classes are close. Specifically, the base scores of NOTA instances are generally greater than 0.4, while those of non-NOTA samples<sup>3</sup> are
+
+around 0.1. This leads to difficulty in choosing a good threshold since the range of proper thresholds is small. Thus, base scores alone may not be sufficient to distinguish between the two classes accurately.
+
+Secondly, without adding prototypes and degenerative prototypes, we have the naive local outlier factor (LOF). It can be seen that both LOF and PLOF show numerical overlap for both non-NOTA and NOTA instances. However, in LOF, the scores for non-NOTA instances are centered around 0.9, while the scores for NOTA instances range from 1.0 to 1.2. The available range of threshold is limited. Contrarily, in PLOF, the scores for non-NOTA samples are in the range of 2 to 3, while the scores for NOTA instances are significantly larger than 4.
+
+Finally, combining PLOF and base NOTA score forms our final NOTA score. It is found that the NOTA score can make two types of instances have a clear difference. The gap between medians becomes more obvious, and the NOTA score for the NOTA instance becomes larger. These all contribute to the detection of NOTA instances, showing the effectiveness of our proposed NOTA score.
+
+# 3.4.3 Ablation Study
+
+To explore the impact of individual training objectives, we perform an ablation study. The results are reported in Table 3. It can be seen that removing  $\mathcal{L}_{\text {nota }}$  causes significant decreases in all evaluation metrics. This verifies that NOTA loss effectively
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">15%NOTA</td><td colspan="2">30%NOTA</td><td colspan="2">50%NOTA</td><td colspan="2">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>D-Proto w/o Lnota</td><td>84.95</td><td>78.73</td><td>84.04</td><td>78.27</td><td>82.72</td><td>75.41</td><td>83.90</td><td>77.47</td></tr><tr><td>D-Proto w/o Scorebase</td><td>78.44</td><td>70.95</td><td>78.23</td><td>71.39</td><td>77.52</td><td>68.99</td><td>78.06</td><td>70.44</td></tr><tr><td>D-Proto w/o P</td><td>84.24</td><td>78.00</td><td>83.93</td><td>78.27</td><td>83.48</td><td>76.13</td><td>83.88</td><td>77.47</td></tr><tr><td>D-Proto w/o DP</td><td>85.40</td><td>79.20</td><td>83.39</td><td>77.55</td><td>80.32</td><td>73.01</td><td>83.03</td><td>76.58</td></tr><tr><td>D-Proto w/o {P ∪ DP}</td><td>83.27</td><td>76.83</td><td>83.04</td><td>77.10</td><td>82.37</td><td>74.73</td><td>82.89</td><td>76.22</td></tr><tr><td>D-Proto w. 20%DPK-2</td><td>84.04</td><td>77.76</td><td>83.80</td><td>78.10</td><td>83.46</td><td>76.10</td><td>83.77</td><td>77.32</td></tr><tr><td>D-Proto w. 40%DPK-2</td><td>83.60</td><td>77.29</td><td>83.49</td><td>77.47</td><td>83.75</td><td>76.36</td><td>83.61</td><td>77.04</td></tr><tr><td>D-Proto w. 60%DPK-2</td><td>83.35</td><td>76.99</td><td>83.63</td><td>77.88</td><td>83.88</td><td>76.49</td><td>83.62</td><td>77.12</td></tr><tr><td>D-Proto w. 80%DPK-2</td><td>83.19</td><td>76.81</td><td>83.58</td><td>77.83</td><td>84.01</td><td>76.62</td><td>83.59</td><td>77.08</td></tr><tr><td>D-Proto w. 100%DPK-2</td><td>83.70</td><td>77.37</td><td>83.74</td><td>78.02</td><td>83.72</td><td>76.37</td><td>83.72</td><td>77.25</td></tr><tr><td>D-Proto(Ours)</td><td>85.76</td><td>79.79</td><td>84.80</td><td>79.27</td><td>83.30</td><td>76.32</td><td>84.62</td><td>78.46</td></tr></table>
+
+Table 3: Ablation study results and degenerative prototypes study (%) of D-Proto on FewRel benchmark.
+
+![](images/a7a87b5dcdbbb225a9964aa2f828eeadbebfc721f747299148b6c77388be56f7.jpg)  
+Figure 5: Comparison of all scores presented in this paper under  $50\%$  NOTA rate on the FewRel dataset. The dash-dotted line indicates the median of each score.
+
+makes NOTA isolated, helping to recognize relations with densities.
+
+Then removing the Base Score significantly impacts the model's performance. The Base Score is essential as it serves as the basis of PLOF. It can also be observed that the removal of either the prototype set  $\mathcal{P}$  or the degenerative prototype set  $\mathcal{D}\mathcal{P}$  results in a significant decline in the evaluation metrics across different NOTA rates. Additionally, the removal of both  $\mathcal{P}$  and  $\mathcal{D}\mathcal{P}$  results in an even more significant drop in the metrics. These findings suggest that both the prototype set and the degenerative prototype set are important components of the PLOF method and contribute significantly to its performance.
+
+# 3.4.4 Degenerative Prototypes Analysis
+
+We have a research question: does adding more degenerative prototypes also increase the perfor
+
+mance of the model? To answer this question, we design the  $K - 2$  degenerative prototype set  $\mathcal{D}\mathcal{P}_{K - 2}$ .  $\mathcal{D}\mathcal{P}_{K - 2}$  consists of all  $K - 2$  degenerative prototypes, where each one is computed by averaging the embeddings of the  $K - 2$  support instances belonging to its relation  $r^i$  after excluding two support instances. It is worth noting that D-Proto only adds the degenerative prototypes  $\mathcal{D}\mathcal{P}_{K - 1}$ . Will including  $\mathcal{D}\mathcal{P}_{K - 2}$  bring further gains?
+
+In this experiment, we add various proportions of  $\mathcal{D}\mathcal{P}_{K - 2}$  into the support set. The results are displayed in the second part of Table 3. Our expectation is that  $\mathcal{D}\mathcal{P}_{K - 2}$  will increase the density of known samples and thus improve the performance. However, the results of this experiment show that while the performance of the D-Proto method does improve when  $\mathcal{D}\mathcal{P}_{K - 2}$  is added under a NOTA rate of  $50\%$ , the performance decreases for NOTA rates of  $15\%$  and  $30\%$ . This suggests that while the use of  $\mathcal{D}\mathcal{P}_{K - 2}$  can indeed improve the detection of NOTA instances at higher densities, it also leads to the false detection of known instances as NOTA, particularly at lower NOTA rates.
+
+# 3.4.5 Visualization
+
+To further evaluate the representations learned by different models, we visualize the embeddings using the t-SNE algorithm (Van der Maaten and Hinton, 2008) in Figure 6. Under both the  $15\%$  and  $30\%$  NOTA rates, compared with O-Proto and MNAV, our method can make each relation cluster more densely. For instance, in Figure 6 (a), the relation 1 is scattered around in O-Proto and MNAV. However, in our method, this class is well aggregated. This shows that our method can learn more separable representations for known instances.
+
+![](images/631389b77f0f4adeab6e8a51e746b43fe97ef04ea1a1ab1e7b85a3002337b166.jpg)
+
+![](images/d969ed3391fca5d0b59812f22c37bea2a3a282d2b913ad272971cd527c638c93.jpg)
+
+![](images/4f0de52b7eef148a92fd7ad42ec45476ad51e9cfc006ed5fb5c918c7e04eed88.jpg)
+
+![](images/58973ce9ba740ea0bcf43d28eef834e00d5bbc4793fe9e9c07367c1677ad4927.jpg)
+
+![](images/66b539df6d72c462de3454128976d19719f0dbf7d44ab0afb77de0b6dbe3a9c6.jpg)  
+(a)  $15\%$  NOTA rate
+
+![](images/527207db514d81c6bf455acf3bc1810ee3c181e3e023fcbd131c87601ff03d81.jpg)  
+(b)  $30\%$  NOTA rate  
+Figure 6: Embedding distributions of different models in the same episode on the FewRel dataset. NOTA rates are  $15\%$  and  $30\%$ , respectively. The numbers 1 to 5 represent the relation. The red inverted triangles are NOTA instances.
+
+Moreover, in Figure 6 (a), we see that the NOTA instance is far away from known samples in our method. This observation is visible in the plot (b). This shows the validity of the proposed NOTA loss  $\mathcal{L}_{\text {nota }}$ , making NOTA instances more isolated. In MNAV, the two NOTA instances are close to a known sample. One possible reason is that MNAV makes multiple NOTA instances approach learnable vectors. Different classes may aggregate with each other. This tends to cause negative effects on distinguishing samples, as embeddings are not well separated. In summary, our method learns the best representations. With the well-clustered known samples and the isolated NOTA samples, D-Proto can both recognize them through the density-aware NOTA detection module accurately.
+
+# 4 Related Work
+
+Previous works (Zelenko et al., 2003; Mooney and Bunescu, 2005; Zeng et al., 2014; Gormley et al., 2015; Zhou et al., 2016; Kumar, 2017; Zhang et al., 2018) for relation classification usually train models on a labeled dataset with a fixed number of classes, which cannot deal with unseen relations.
+
+Thus few-shot scenario has drawn much attention recently. By training on multiple episodes, models can learn transferrable knowledge for unseen testing episodes (Hu et al., 2021). Although many efforts (Hendrickx et al., 2010; Zhang et al., 2017; Obamuyide and Vlachos, 2019; Gao et al., 2019a; Baldini Soares et al., 2019; Ye and Ling, 2019; Dong et al., 2020) have been devoted to the few-shot relation classification. However, a real-world issue that has been proposed is the NOTA challenge (Gao et al., 2019b). This issue breaks the assumption of the traditional  $N$ -way  $K$ -shot setting. Therefore, detecting NOTA instances during inference needs to be explored.
+
+Promising works have been proposed to detect NOTA instances in few-shot relation classification. One approach that has been widely adopted is the  $N + 1$  classification. This approach has been proposed by several works. One work is to use a sentence-pair model for classification, as proposed by Gao et al. (2019b) in their work on few-shot learning. Another work, Sabo et al. (2021) in their work on revisiting NOTA challenge, is to compute the distance between learnable vectors. The third one, proposed by Liu et al. (2022), adopts the pretraining fine-tuning paradigm and selects the correct relation from a set of options that includes NOTA. The above three methods, by considering NOTA instances as an additional class, are able to learn the boundary between the known relations and the NOTA instances, allowing the model to learn a more comprehensive representation of the data.
+
+# 5 Conclusion
+
+In this paper, we propose D-Proto, a density-aware prototypical network for few-shot relation classification. We focus on the overall instance distribution and design special training objectives for distinct samples. For the known samples, intraclass becomes well-clustered and inter-class is well-separated. The NOTA ones are isolated as outliers. We further propose a NOTA detection module, which considers the local density and distance property, to distinguish both types of instances. And we use prototypes and degenerative prototypes to enhance the density of known instances to better identify NOTA. Experiments on a popular dataset demonstrate that D-Proto outperforms strong baselines. We also provide score analysis and visualization to verify the effectiveness of our method.
+
+# Limitations
+
+In this work, we study the few-shot relation classification task, which contains NOTA instances. The proposed D-Proto method achieves some gains on this problem. Yet our work still has the following two limitations.
+
+The first limitation is that NOTA instances are detected with a threshold  $\tau$ . This threshold relies on human experiences or k-fold validation. However, using a threshold in the binary way methods, including O-Proto and ours, is not a big problem. Our viewpoint is that  $N + 1$  classification can efficiently accomplish the NOTA detection task by treating the NOTA as an additional class to classify with the known  $N$  classes. But the  $N + 1$  classification method ignores the distribution of the data. Therefore, despite that we need a threshold, the advantages of our method are also obvious.
+
+The second limitation is that our experiment is not absolutely fair. The reason is that BERT-Pair (Gao et al., 2019b) requires high GPU memory. In the fixed sampling method of the episode, we used two NVIDIA RTX A6000 with 48G memory each to train BERT-Pair. The other baselines are trained with an NVIDIA TESLA V100-32G. Except for this difference, all other settings are the same.
+
+# Acknowledgment
+
+We sincerely thank all the anonymous reviewers for providing valuable feedback. This work is supported by the National Natural Science Foundation of China (Grant No. 62302245), the Ministry of Education of the People's Republic of China Humanities and Social Sciences Youth Foundation (Grant No. 23YJCZH240), and the youth program of National Science Fund of Tianjin, China (Grant No. 22JCQNJC01340).
+
+# References
+
+Livio Baldini Soares, Nicholas FitzGerald, Jeffrey Ling, and Tom Kwiatkowski. 2019. Matching the blanks: Distributional similarity for relation learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 2895-2905.  
+Sam Brody, Sichao Wu, and Adrian Benton. 2021. Towards realistic few-shot relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5338-5345, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
+
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (NAACL-HLT), pages 4171-4186.  
+Bowen Dong, Yuan Yao, Ruobing Xie, Tianyu Gao, Xu Han, Zhiyuan Liu, Fen Lin, Leyu Lin, and Maosong Sun. 2020. Meta-information guided meta-learning for few-shot relation classification. In Proceedings of the 28th International Conference on Computational Linguistics (COLING), pages 1594-1605.  
+Tianyu Gao, Xu Han, Zhiyuan Liu, and Maosong Sun. 2019a. Hybrid attention-based prototypical networks for noisy few-shot relation classification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 33, pages 6407-6414.  
+Tianyu Gao, Xu Han, Hao Zhu, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. 2019b. Fewrel 2.0: Towards more challenging few-shot relation classification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6250-6255.  
+Matthew R. Gormley, Mo Yu, and Mark Dredze. 2015. Improved relation extraction with feature-rich compositional embedding models. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1774-1784.  
+Jiale Han, Bo Cheng, and Wei Lu. 2021. Exploring task difficulty for few-shot relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2605-2616, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Xu Han, Hao Zhu, Pengfei Yu, Ziyun Wang, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2018. Fewrel: A large-scale supervised few-shot relation classification dataset with state-of-the-art evaluation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4803-4809.  
+Iris Hendrickx, Su Nam Kim, Zornitsa Kozareva, Preslav Nakov, Diarmuid O Seaghdha, Sebastian Padó, Marco Pennacchiotti, Lorenza Romano, and Stan Szpakowicz. 2010. Semeval-2010 task 8: Multiway classification of semantic relations between pairs of nominals. In Proceedings of the 5th International Workshop on Semantic Evaluation, pages 33-38.  
+Mengting Hu, Shiwan Zhao, Honglei Guo, Chao Xue, Hang Gao, Tiegang Gao, Renhong Cheng, and Zhong Su. 2021. Multi-label few-shot learning for aspect category detection. In Proceedings of the 59th
+
+Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP), pages 6330-6340.  
+Shantanu Kumar. 2017. A survey of deep learning methods for relation extraction. arXiv preprint arXiv:1705.03645.  
+Steinar Laenen and Luca Bertinetto. 2021. On episodes, prototypical networks, and few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 34.  
+Fangchao Liu, Hongyu Lin, Xianpei Han, Boxi Cao, and Le Sun. 2022. Pre-training to match for unified low-shot relation extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5785-5795, Dublin, Ireland. Association for Computational Linguistics.  
+Raymond Mooney and Razvan Bunescu. 2005. Subsequence kernels for relation extraction. Advances in neural information processing systems (NeurIPS), 18.  
+Abiola Obamuyide and Andreas Vlachos. 2019. Model-agnostic meta-learning for relation classification with limited supervision. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5873-5879, Florence, Italy. Association for Computational Linguistics.  
+Meng Qu, Tianyu Gao, Louis-Pascal Xhonneux, and Jian Tang. 2020. Few-shot relation extraction via bayesian meta-learning on relation graphs. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 7867-7876.  
+Ofer Sabo, Yanai Elazar, Yoav Goldberg, and Ido Dagan. 2021. Revisiting few-shot relation classification: Evaluation data and classification schemes. Transactions of the Association for Computational Linguistics (TACL), 9:691-706.  
+Jake Snell, Kevin Swersky, and Richard Zemel. 2017. Prototypical networks for few-shot learning. Advances in neural information processing systems (NeurIPS), 30.  
+Ming Tan, Yang Yu, Haoyu Wang, Dakuo Wang, Saloni Potdar, Shiyu Chang, and Mo Yu. 2019. Out-of-domain detection for low-resource text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3566-3572.  
+Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research (JMLR), 9(11).
+
+Zhi-Xiu Ye and Zhen-Hua Ling. 2019. Multi-level matching and aggregation network for few-shot relation classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 2872-2881.  
+Dmitry Zelenko, Chinatsu Aone, and Anthony Richardella. 2003. Kernel methods for relation extraction. Journal of machine learning research (JMLR), 3(Feb):1083-1106.  
+Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, and Jun Zhao. 2014. Relation classification via convolutional deep neural network. In Proceedings of the 28th International Conference on Computational Linguistics (COLING), pages 2335-2344.  
+Ningyu Zhang, Shumin Deng, Zhanling Sun, Xi Chen, Wei Zhang, and Huajun Chen. 2018. Attention-based capsule networks with dynamic routing for relation extraction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (ACL), pages 986-992.  
+Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor Angeli, and Christopher D Manning. 2017. Position-aware attention and supervised data improve slot filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 35-45.  
+Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen Li, Hongwei Hao, and Bo Xu. 2016. Attention-based bidirectional long short-term memory networks for relation classification. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL), pages 207-212.
+
+<table><tr><td colspan="9">Random Sampling</td></tr><tr><td rowspan="2">Methods</td><td colspan="2">15%NOTA</td><td colspan="2">30%NOTA</td><td colspan="2">50%NOTA</td><td colspan="2">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>O-Proto</td><td>80.02±0.65</td><td>72.84±0.88</td><td>80.55±0.97</td><td>73.73±1.18</td><td>81.16±1.04</td><td>72.41±1.28</td><td>80.58</td><td>72.99</td></tr><tr><td>BERT-Pair</td><td>80.77±2.35</td><td>73.99±2.67</td><td>81.53±1.60</td><td>75.06±1.92</td><td>83.24±0.85</td><td>75.14±1.38</td><td>81.85</td><td>74.73</td></tr><tr><td>MNAV</td><td>85.02±0.39</td><td>78.81±0.47</td><td>83.67±0.66</td><td>77.78±0.90</td><td>81.59±1.02</td><td>73.94±1.18</td><td>83.42</td><td>76.84</td></tr><tr><td>MCMN</td><td>88.24±1.83</td><td>83.54±2.45</td><td>82.50±3.40</td><td>76.39±4.34</td><td>74.88±5.83</td><td>67.49±6.64</td><td>81.87</td><td>75.80</td></tr><tr><td>ChatGPT</td><td>64.57±0.31</td><td>56.63±0.41</td><td>63.30±0.57</td><td>55.81±0.40</td><td>61.66±0.76</td><td>54.16±0.45</td><td>63.18</td><td>55.54</td></tr><tr><td>D-Proto</td><td>85.76±0.72</td><td>79.79±0.96</td><td>84.80±0.92</td><td>79.27±1.12</td><td>83.30±1.29</td><td>76.32±1.48</td><td>84.62</td><td>78.46</td></tr></table>
+
+Fixed Sampling  
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">15%NOTA</td><td colspan="2">30%NOTA</td><td colspan="2">50%NOTA</td><td colspan="2">Average</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>O-Proto</td><td>81.38±1.26</td><td>80.45±1.29</td><td>81.42±1.11</td><td>81.04±1.19</td><td>81.38±0.96</td><td>79.39±1.06</td><td>81.38</td><td>80.29</td></tr><tr><td>BERT-Pair</td><td>82.83±0.62</td><td>81.98±0.62</td><td>83.76±0.46</td><td>83.00±0.54</td><td>85.49±0.29</td><td>82.62±0.45</td><td>84.02</td><td>82.53</td></tr><tr><td>MNAV</td><td>85.05±0.97</td><td>84.22±1.03</td><td>84.36±1.34</td><td>84.45±1.15</td><td>82.91±2.61</td><td>82.15±1.85</td><td>84.10</td><td>83.60</td></tr><tr><td>MCMN</td><td>87.56±1.93</td><td>84.04±2.52</td><td>83.16±3.37</td><td>83.40±3.08</td><td>75.07±5.79</td><td>79.14±3.93</td><td>81.93</td><td>82.19</td></tr><tr><td>ChatGPT</td><td>63.59±0.50</td><td>56.26±0.27</td><td>61.83±0.41</td><td>54.46±0.32</td><td>58.41±0.51</td><td>51.14±0.43</td><td>61.28</td><td>53.95</td></tr><tr><td>D-Proto</td><td>85.37±0.61</td><td>84.67±0.61</td><td>84.87±0.71</td><td>84.90±0.66</td><td>83.05±1.00</td><td>82.66±1.10</td><td>84.43</td><td>84.08</td></tr></table>
+
+# A Reproducibility
+
+# A.1 Implementation Details
+
+Following Gao et al. (2019b), we exploit the same hyper-parameters. For a fair comparison, all baselines and our work use the same encoder. We choose BERT-base-uncased (Devlin et al., 2019) as the encoder. It is initialized by the pre-trained parameters and optimized during training. For the FewRel dataset, we set the batch size to 2, which means we feed two episodes into the model per batch. Each model is trained for 30 epochs with 1000 batches per epoch and tested with 10000 batches in the testing. An early stopping strategy is adopted, indicating the model will stop training when the performance on the validation set does not improve for 6 epochs. The best model on the validation set is saved for evaluation.
+
+In Eq. (1), we set the hyper-parameters  $\alpha$  to 1e-5 on the FewRel. For  $k$  of the local outlier factor in Eq. (7), we set it to be  $K$ . The reason is that in the  $N$ -way  $K$ -shot setting, a support instance has other  $(K - 1)$  neighbors with the same relation. And each relation has at least one prototype. This is consistent with the definition of  $k$ -neighborhood. All the reported results using the FewRel dataset are the average of five runs with fixed seeds [5, 10, 15, 20, 25].
+
+After the hyperparameter search, the threshold  $\pmb{\tau}$  is set to 0.9 in the experimental setting of FewRel.
+
+For MCMN, the sampling method in the pre
+
+Table 4: Evaluation results of baseline methods and the proposed D-Proto, in terms of accuracy  $(\%)$  and F1  $(\%)$ , on the FewRel dataset. This table includes the standard deviation of baseline methods and D-Proto.  
+
+<table><tr><td>Model</td><td>Number of parameters</td></tr><tr><td>O-Proto</td><td>109482240</td></tr><tr><td>MNAV</td><td>109497600</td></tr><tr><td>BERT-Pair</td><td>109483778</td></tr><tr><td>D-Proto</td><td>109482240</td></tr></table>
+
+Table 5: The number of hyperparameters for each model.
+
+training follows Liu et al. (2022), while the sampling method in the testing is fixed sampling or random sampling.
+
+# A.2 Computing Infrastructure
+
+Since BERT-Pair (Gao et al., 2019b) requires high GPU memory in the fixed sampling of 5-way 5-shot setting, we use 2 NVIDIA A6000s with 48G memory each to train the model. The rest of the baselines and our model are trained on a NVIDIA TESLA V100 with 32G of GPU memory.
+
+# A.3 Number of Parameters per Model
+
+Table 5 shows the number of hyperparameters for each model.
+
+# A.4 Evaluation Metrics
+
+We use two primary evaluation metrics, accuracy and macro F1. The basic implementation of accu
+
+<table><tr><td>Method</td><td>Number of instances</td></tr><tr><td>D-Proto w/o P</td><td>53</td></tr><tr><td>D-Proto w/o DP</td><td>33</td></tr><tr><td>D-Proto w/o {P ∪ DP}</td><td>28</td></tr><tr><td>D-Proto w. 20%DPK-2</td><td>68</td></tr><tr><td>D-Proto w. 40%DPK-2</td><td>78</td></tr><tr><td>D-Proto w. 60%DPK-2</td><td>88</td></tr><tr><td>D-Proto w. 80%DPK-2</td><td>98</td></tr><tr><td>D-Proto w. 100%DPK-2</td><td>108</td></tr><tr><td>D-Proto(Ours)</td><td>58</td></tr></table>
+
+Table 6: The number of instances for each method in an episode. The experiment setting is 5-way 5-shot. The number of query instances is 3.
+
+racy is as follows:
+
+$$
+A c c u r a c y = \frac {T P + T N}{T P + F P + T N + F N}
+$$
+
+where  $TP$  are true positives,  $FP$  are false positives,  $TN$  are true negatives, and  $FN$  are false negatives. The definition of F1 is as follows:
+
+$$
+P = \frac {T P}{T P + F P}
+$$
+
+$$
+R = \frac {T P}{T P + F N}
+$$
+
+$$
+F 1 = \frac {2 \times P \times R}{P + R}
+$$
+
+Macro F1 calculates F1 for each class and finds their unweighted mean.
+
+# A.5 Hyperparameters
+
+Following Gao et al. (2019b), the hyperparameters not mentioned in Section A.1 are as follows:
+
+- Learning rate: 2e-5.  
+Weight decay: 0.01.  
+Warmup steps: 300.  
+Gradient accumulation: 1.  
+Max length: 128.  
+- Hidden size: 768.  
+Test batch: 10000.
+
+# A.6 Links to Dataset
+
+FewRel (Han et al., 2018): https://thunlp.github.io/2/fewrel2nota.html.
+
+# A.7 Links to Toolkit
+
+- BERT: https://huggingface.co/ transformers/v4.5.1/model_doc/bert.html#bertmodel  
+- BERT Tokenizer: https://huggingface.co/transformers/v4.5.1/model_doc/bert.html#berttokenizer  
+- Local Outlier Factor(LOF): https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.
+- LocalOutlierFactor.html  
+- Macro F1: https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html
+
+# A.8 Number of Instances per Method
+
+Table 6 shows the number of instances included in an episode for each method in ablation studying. The episode sampling method is random sampling.
+
+# A.9 Prompt of ChatGPT
+
+We demonstrate the prompt for using ChatGPT directly in the NOTA task, as follows:
+
+Question: The purpose of the relation classification task is to identify the relation between two entities (one is the head entity and the other is the tail entity) in a given sentence. For example, the following sentence "California is a state in the United States". Through the derivation of the whole sentence, the relation between the head entity "California" and the tail entity the "United States" is "contains". While the few-shot relation classification task consists of a support set and a query set. The support set generally contains  $N$  relations, and each relation contains  $K$  instances. This setup is called  $N$ -way  $K$ -shot. We need to predict the relation between two entities in each sentence given in the query set based on the support set. Now I need you to complete a few-shot relation classification task, which is set to 5-way 5-shot. The five categories in the support set are numbered from 0 to 4, respectively. Five sentences are given after
+
+each category. The head entity of each sentence is surrounded by two special characters [head] and [/head], and the tail entity is surrounded by two special characters [tail] and [/tail]. You need to predict the relationship between two entities for each sentence in the query set from the support set. That is, if the relation of a query instance is similar to that of all instances of category 0 in the support set, then the category predicted by the query instance is category 0. Notice! If the relation between the two entities of a sentence in the query set cannot be referenced in the support set, the relation between the two entities of the sentence is called NOTA (None-of-the-above). When the NOTA relation is encountered, the category number of the predicted result is 5. Warning! You just need to return the class number for each instance predicted in the query set. For example, the relation of the first instance in the query set is relation 0, and the second instance is relation 5 (NOTA). Then your reply is: [0, 5]. If you get it, then I'll give you support sets and query sets next.
+
+ChatGPT: Yes, I understand the task. Please provide me with the support set and query set.
+
+Question: The following is the support set (5-way 5-shot. Instances in each category are separated by a special character [sep].):
+
+···
+
+The following is the query set (The query set contains one instance.):
+
+··

2023/Density-Aware Prototypical Network for Few-Shot Relation Classification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95c23a8bc77588b195c5f14ea9bafdc3c23b9cea0f34f0c5ece095da5297bb7f
+size 661941

2023/DepNeCTI_ Dependency-based Nested Compound Type Identification for Sanskrit/9cf22d4e-25cc-4736-8b7f-6f6bea41e671_content_list.json

@@ -0,0 +1,1669 @@
+[
+    {
+        "type": "text",
+        "text": "DepNeCTI: Dependency-based Nested Compound Type Identification for Sanskrit",
+        "text_level": 1,
+        "bbox": [
+            119,
+            89,
+            878,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jivnesh Sandhan†,* Yaswanth Narsupalli‡, Sreevatsa Muppirala‡, Sriram Krishnan§, Pavankumar Satuluri∥, Amba Kulkarni§ and Pawan Goyal‡",
+        "bbox": [
+            159,
+            143,
+            840,
+            183
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "†UC Berkeley, ‡IIT Kharagpur, §University of Hyderabad and ¶IIT Roorkee",
+        "bbox": [
+            193,
+            185,
+            805,
+            202
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "jivneshsandhan@gmail.com, yasshu.yaswanth@gmail.com, pawang@cse.iitkgp.ac.in",
+        "bbox": [
+            115,
+            205,
+            884,
+            223
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Multi-component compounding is a prevalent phenomenon in Sanskrit, and understanding the implicit structure of a compound's components is crucial for deciphering its meaning. Earlier approaches in Sanskrit have focused on binary compounds and neglected the multi-component compound setting. This work introduces the novel task of nested compound type identification (NeCTI), which aims to identify nested spans of a multi-component compound and decode the implicit semantic relations between them. To the best of our knowledge, this is the first attempt in the field of lexical semantics to propose this task.",
+        "bbox": [
+            141,
+            281,
+            460,
+            480
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We present 2 newly annotated datasets including an out-of-domain dataset for this task. We also benchmark these datasets by exploring the efficacy of the standard problem formulations such as nested named entity recognition, constituency parsing and seq2seq, etc. We present a novel framework named DepNeCTI: Dependency-based Nested Compound Type Identifier that surpasses the performance of the best baseline with an average absolute improvement of 13.1 points F1-score in terms of Labeled Span Score (LSS) and a 5-fold enhancement in inference efficiency. In line with the previous findings in the binary Sanskrit compound identification task, context provides benefits for the NeCTI task. The codebase and datasets are publicly available at: https://github.com/yaswanth-iitkgp/DepNeCTI",
+        "bbox": [
+            141,
+            486,
+            460,
+            741
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            757,
+            258,
+            772
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A compound is defined as a group of entities functioning as a single meaningful entity. The process of identifying the implied semantic relationship between the components of a compound in Sanskrit is known as Sanskrit Compound Type Identification (SaCTI) (Sandhan et al., 2022a) or Noun Compound Interpretation (NCI) (Ponkiya et al., 2021,",
+        "bbox": [
+            112,
+            783,
+            490,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2020). Within the literature, the NCI problem has been approached in two ways, namely, classification (Dima and Hinrichs, 2015; Fares et al., 2018; Ponkiya et al., 2021) and paraphrasing (Lapata and Keller, 2004; Ponkiya et al., 2018, 2020).",
+        "bbox": [
+            507,
+            252,
+            884,
+            332
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In Sanskrit literature, particularly in poetry, the use of multi-component compounds is ubiquitous (Kumar, 2012). According to the Digital Corpus of Sanskrit, more than  $41\\%$  of compounds contain three or more components (Krishna et al., 2016). However, earlier approaches focus solely on binary compounds and fail to address the complexities inherent in multi-component compounds adequately. Thus, we propose a new task, the Nested Compound Type Identification (NeCTI) Task, which focuses on identifying nested spans within a multicomponent compound and interpreting their implicit semantic relationships. Figure 1 illustrates an example of the NeCTI task, highlighting nested spans and their associated semantic relations using distinct colors.",
+        "bbox": [
+            507,
+            335,
+            884,
+            592
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8fce5a9d9993137775f22da5532a49bcdcb89d81850c7ea5d2f698e76259dd72.jpg",
+        "image_caption": [
+            "Figure 1: Illustration of the NeCTI task for the multi-component compound sumitra-ananda-vardhanah (Translation: Sumitra-delight-enhancer), highlighting nested spans and their associated semantic relations using distinct colors."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            551,
+            607,
+            840,
+            700
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The NeCTI task presents multiple challenges: (1) The number of potential solutions for a multicomponent compound grows exponentially as the number of components increases. (2) It often relies on contextual or world knowledge about the entities involved (Krishna et al., 2016). Even if a multicomponent compound shares the same components",
+        "bbox": [
+            507,
+            806,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* denotes the first two authors contributed equally.",
+        "bbox": [
+            139,
+            904,
+            450,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "13679",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13679-13692 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            208,
+            945,
+            786,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "and final form, the implicit relationship between the spans can only be deciphered with the aid of available contextual information (Kulkarni and Kumar, 2013; Krishna et al., 2016). For instance, as de",
+        "bbox": [
+            112,
+            84,
+            490,
+            149
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/265320e2dba0b3084032e52044623b15bc76e00e3533e531149e7c7a9ec77023.jpg",
+        "image_caption": [
+            "Figure 2: Illustration of the multi-component compound sumitra-ananda-vardhanah (Translation: Sumitra-delight-enhancer) with two valid parses depicted in green and red. The two parses correspond to two potential meanings. The green solution represents the correct interpretation within the provided context."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            156,
+            487,
+            225
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "picted in Figure 2, the multi-component compound sumitra-ananda-vardhanah (Translation: Sumitradelight-enhancer) can have two valid but distinct solutions, leading to different meanings. Resolving ambiguity to select the correct solution requires reliance on the provided context. Consequently, downstream Natural Language Processing (NLP) applications for Sanskrit, such as question answering (Terdalkar and Bhattacharya, 2019) and machine translation (Aralikatte et al., 2021), often exhibit sub-optimal performance when encountering compounds. Hence, the NeCTI task serves as a preliminary requirement for developing robust NLP technology for Sanskrit. Moreover, this dependency on contextual information eliminates the possibility of storing and conducting a lookup to identify the semantic types of nested spans.",
+        "bbox": [
+            115,
+            332,
+            489,
+            606
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Previous approaches (Kulkarni and Kumar, 2013; Krishna et al., 2016; Sandhan et al., 2019) addressing SaCTI have predominantly focused on binary compounds, neglecting the consideration of multi-component compounds. In multi-component compounds, the components exhibit semantic relationships akin to dependency relations, represented as directed labels within the dependency structure, which also facilitate the identification of the compound's headword through the labels directed towards it. Consequently, dependency formulation enables the simultaneous identification of both the structure or constituency span and the compound types. Thus, we propose a novel framework ( $\\S 4$ ) named DepNeCTI: Dependency-based Nested Compound type Identifier ( $\\S 3$ ). In summary, our contributions can be outlined as follows:",
+        "bbox": [
+            112,
+            607,
+            489,
+            879
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We introduce a novel task called Nested Compound Type Identification (§ 2).",
+        "bbox": [
+            114,
+            887,
+            489,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We present 2 newly annotated datasets and provide benchmarking by exploring the efficacy of various standard formulations for NeCTI (§ 5).",
+            "- We propose a novel framework DepNeCTI: Dependency-based Nested Compound type Identifier (§ 4), which reports an average 13.1 points F1-score in terms of LSS absolute gain and 5-fold enhancement in inference efficiency (§ 6) over the best baseline.",
+            "- We publicly release the codebase of DepNeCTI and benchmarked baselines, along with newly annotated datasets for the NeCTI task."
+        ],
+        "bbox": [
+            509,
+            84,
+            884,
+            297
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            509,
+            307,
+            732,
+            324
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The objective of the NeCTI task is to detect nested spans within a multi-component compound and decipher the implicit semantic relations among them. Our study focuses exclusively on this task and does not address the compound segmentation problem. It is assumed that the segmented components of the multi-component compound are already available. To obtain the segmentation of a compound, we rely on established resources such as the rule-based shallow parser (Goyal and Huet, 2016) or existing data-driven segmentation systems designed explicitly for Sanskrit (Hellwig and Nehrdich, 2018; Sandhan et al., 2022b).",
+        "bbox": [
+            507,
+            332,
+            884,
+            542
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/5f67455afe96c082a3588685841c0ceb83beea3e04841f5ca78b657aebe04ee0.jpg",
+        "image_caption": [
+            "Figure 3: Illustration of a number of ways (as per Catalan number in log scale) in which a multi-component compound can be parsed. Our dataset contains compounds that have a maximum of 16 components."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            547,
+            570,
+            843,
+            707
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Complexity of NeCTI Task: As the number of components in a multi-component compound increases, the number of possible parses grows exponentially. Our dataset encompasses multicomponent compounds ranging from 2 to 16 components. Figure 3 visually depicts the exponential growth of possible parses with increasing component count. The parsing of a compound word with",
+        "bbox": [
+            507,
+            790,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "13680",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "$n + 1$  components can be likened to the problem of fully parenthesizing  $n + 1$  factors in all possible ways (Kulkarni and Kumar, 2011). Thus, the total number of parse-options for a multi-component compound word with  $n + 1$  components corresponds to the Catalan Number  $C_n$ , where  $n \\geq 0$  (Huet, 2009).",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC _ {n} = \\frac {1}{n + 1} \\binom {2 n} {n} = \\frac {(2 n) !}{(n + 1) ! n !}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            173,
+            208,
+            428,
+            242
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $C_n$  represents the  $n^{th}$  Catalan number,  $\\binom{2n}{n}$  is the binomial coefficient, and! denotes factorial. Finally, the compatibility rules derived from Paninian grammar (Panini, 500 BCE) and contextual information are needed to disambiguate multiple possibilities.",
+        "bbox": [
+            112,
+            249,
+            489,
+            344
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Formally, in a given sentence  $X = \\{x_{1}, x_{2}, \\ldots, x_{M}\\}$  with  $M$  tokens, let  $x_{p}$  (1 ≤  $p \\leq M$ ) denote an  $N$ -component compound. It is worth noting that  $X$  may contain multiple instances of multi-component compounds. A valid solution corresponds to a full paraphesization of these  $N$  components; let  $P_{N}$  encompass all valid solutions for fully parenthesizing  $N$  factors, satisfying the cardinality  $|P_{N}| = C_{N-1}$ , where  $C_{N-1}$  represents the Catalan number. A valid solution for an  $N$ -component compound consists of  $N-1$  nested spans. The NeCTI system produces an output represented as a list of  $N-1$  tuples for  $x_{p}$ , given by  $Y_{p} = \\{[I_{1}^{H}, I_{1}^{T}, T_{1}], \\ldots, [I_{N-1}^{H}, I_{N-1}^{T}, T_{N-1}]\\}$ , such that  $Y_{p} \\in P_{N}$ .  $I_{i}^{H}$  and  $I_{i}^{T}$  denote the head and the tail indices, respectively, of the  $i^{th}$  span, and  $T_{i}$  corresponds to the label assigned to the respective span.",
+        "bbox": [
+            115,
+            347,
+            489,
+            652
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "How different is NeCTI compared to the Nested Named Entity Recognition (NNER) task?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            661,
+            487,
+            693
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "NNER is a component of information extraction that aims to identify and classify nested named entities within unstructured text, considering their hierarchical structure. In contrast, the NeCTI task focuses on identifying nested spans within a multi-component compound and decoding their implicit semantic relations. These tasks have several key differences: (1) NeCTI operates at the intra-word level, whereas NNER operates at the inter-word level, considering entities across a phrase. (2) In NeCTI, a multi-component compound can have multiple possible parses, requiring disambiguation through contextual cues and incorporating insights from Paninian grammar to address incompatibil",
+        "bbox": [
+            112,
+            694,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ities. Conversely, we could not find discussions related to these aspects in NNER literature. (3) NeCTI benefits from prior knowledge of the compound's location and segmented components. In contrast, the NNER task involves the additional challenge of identifying the location of entities within the text. Consequently, leveraging existing NNER frameworks for NeCTI is not straightforward due to their inability to provide explicit support for providing the location of compounds. Therefore, NeCTI presents unique characteristics and challenges that differentiate it from the NNER task, requiring specialized approaches tailored to its specific requirements.",
+        "bbox": [
+            507,
+            84,
+            884,
+            309
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "How different is NeCTI compared to the Multi-Word Expressions (MWE)? MWEs encompass various categories such as idioms (e.g., kick the bucket), named entities (e.g., World Health Organization), and compounds (e.g., telephone box), etc. Sanskrit compounds exhibit similarities to multi-word expressions, particularly multi-word compounds (nominal, noun, and verb), with respect to characteristics like collocation of components based on semantic relations between them and strict preference for ordering of the components.",
+        "bbox": [
+            507,
+            318,
+            884,
+            494
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Multi-word compounds in various languages involve adjacent lexemes juxtaposed with potential semantic relations. In contrast, Sanskrit compounds are intuitively constructed based on the semantic compatibility of their components. Additionally, in Sanskrit, compounds always appear as a single word, requiring mandatory sandhi (euphonic transformations) between their components. Conversely, certain categories of MWEs (idioms, complex function words, verb-particles, and light-verbals) have relatively fixed structures, predominantly with components separated by spaces. Furthermore, nesting within multi-word expressions is considered syntactic overlaps, while the nested structures of Sanskrit compounds result from successive combinations of components based on semantic relations, thereby clearly distinguishing Sanskrit compounds from MWEs.",
+        "bbox": [
+            507,
+            495,
+            884,
+            784
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Why NeCTI as a Dependency Parsing Task?",
+        "text_level": 1,
+        "bbox": [
+            507,
+            796,
+            870,
+            829
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The decision to formulate NeCTI as a dependency parsing task is driven by several considerations. Compounds with more than two components are typically formed through successive binary combinations, following specific semantic relations. This",
+        "bbox": [
+            507,
+            839,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "13681",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a6896d9f66fbadcc2c5aeba9e82b4b253e6f8bb33a856175c4bae5c4aac4d13a.jpg",
+        "image_caption": [
+            "Figure 4: Illustration of DepNeCTI with an example “sumitra-ananda-vardhanah laksmanah rāmam anujagāma” (Translation: “Lakṣmana, the one who enhanced the delight of Sumitra, followed Rāma”) where ‘sumitra-ananda-vardhanah’ is a multi-component compound word. We assume prior knowledge of compound segmentation and treat the individual components of multi-component compounds as separate words. That means the compound and its components are known apriori. However, the associations of the components, i.e. spans, are not known apriori. We propose two variants of DepNeCTI depending on the choice of encoder: DepNeCTI-LSTM and DepNeCTI-XLMR. To inform DepNeCTI-LSTM about compound (highlighted with color) and non-compound (highlighted with color) tokens, we employ span encoding. The span encoding uses two randomly initialized vectors (compound or non-compound) to inform the model whether the corresponding instance is a compound or non-compound word. On the other hand, DepNeCTI-XLMR is informed about compound's location in the input string using bracketing (for example, <sumitra-ananda-vardhanah>) and it lacks span encoding component. Next, to transform the compound-level parsing task into standard dependency parsing, we introduce (1) an additional token called “Global” (color) as a global head for all words in the sentence. (2) The compound head and non-compound words are connected to the Global token using the “Compound Root” and “Global Relation” relations, respectively. The hidden representations of the tokens are obtained using a Bi-LSTM or XLMR encoder. Finally, a Bi-affine (Dozat and Manning, 2017) dependency module is applied on top of the hidden representations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            231,
+            80,
+            769,
+            370
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "process creates a nested structure of binary compounds, except for a few exceptions, where the structure represents a constituency tree. However, treating this as a constituency parsing task poses challenges. The nested structure of compounds does not adhere to a syntactic structure but instead follows a semantic structure based on component relations. If we substitute the intermediate nodes with semantic relations, the same spans can be represented as dependency structures by annotating the types as directed relations. Moreover, the headwords in constituency spans are not explicitly marked but can be identified through their corresponding types. In contrast, dependency structures allow the determination of the headword based on labels directed towards it within the compound. Notably, dependency structures faithfully capture",
+        "bbox": [
+            112,
+            639,
+            489,
+            912
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "constituency information and can be mutually converted with their corresponding spans (Goyal and Kulkarni, 2014).",
+        "bbox": [
+            507,
+            639,
+            882,
+            684
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Summarily, the semantic relations among compound components resemble dependency relations, which can be represented as directed labels within the dependency parse structure. This approach succinctly represents the semantic relations without introducing intermediary nodes. Lastly, it enables the simultaneous identification of the structure or constituency span alongside the identification of compound types. We encourage readers to refer to Appendix § A for a more detailed illustration.",
+        "bbox": [
+            507,
+            688,
+            882,
+            848
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 DepNeCTI: The Proposed Framework",
+        "text_level": 1,
+        "bbox": [
+            507,
+            860,
+            875,
+            878
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 illustrates the proposed framework with an example \"sumitra-ananda-vardhanah laksmanah",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "13682",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ramanujagama\" (Translation: \"Laksmana, the one who enhanced the delight of Sumitra, followed Rama\") where 'sumitra-ananda-vardhanah' is a multi-component compound word. We assume prior knowledge of compound segmentation and treat the individual components of multicomponent compounds as separate words. We propose two variants of DepNeCTI depending on the choice of encoder: DepNeCTI-LSTM and DepNeCTI-XLMR. In DepNeCTI-LSTM, to differentiate between compound (highlighted with color) and non-compound (highlighted with color) tokens, we employ span encoding. The span encoding uses two randomly initialized vectors (compound or non-compound) to inform the model whether the corresponding instance is a compound or non-compound word. On the other hand, DepNeCTI-XLMR is informed about compound's location in the input string using bracketing (for example, <sumitra-ananda-vardhana>) and it lacks span encoding component. In order to convert the compound-level dependency parsing task into standard dependency parsing, we introduce two modifications. First, we introduce an additional token called \"Global\" (color) which serves as the global head for all words in the sentence. Second, we establish connections between the compound head and non-compound words to the Global token using the \"Compound Root\" and \"Global Relation\" relations, respectively.",
+        "bbox": [
+            112,
+            84,
+            492,
+            567
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Formally, in a given sentence  $X = \\{x_{1}, x_{2}, \\ldots, x_{M}\\}$  with  $M$  tokens, let  $x_{p}$  (1 ≤  $p \\leq M$ ) denote an  $N$ -component compound. Notably,  $X$  may contain multiple occurrences of multi-component compounds. The  $N$ -component compound  $(x_{p})$  is further split into its components  $(x_{p} = \\{x_{p}^{1}, x_{p}^{2}, \\ldots, x_{p}^{N}\\})$ . Next, we pass the overall sequence  $(X = \\{x_{1}, x_{2}, \\ldots, x_{p}^{1}, x_{p}^{2}, \\ldots, x_{p}^{N}, \\ldots, x_{M}\\})$  to the encoder to obtain hidden representations. The LSTM encoder concatenates a token's word, character and span embedding to obtain its representation and the XLMR encoder uses word-pieces. Finally, a Bi-affine (Dozat and Manning, 2017) dependency module is applied on top of it.",
+        "bbox": [
+            112,
+            568,
+            489,
+            810
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            822,
+            260,
+            838
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            848,
+            228,
+            862
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1 presents data on the total number of multi-component compounds (Figure 5) and their statistics. Our primary focus is on compounds with",
+        "bbox": [
+            112,
+            870,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "more than two components ( $n > 2$ ), while also considering binary compounds if they occur in the context. These datasets comprise segmented compound components, nested spans, context, and semantic relations among the nested spans. We offer two levels of annotations for these datasets: coarse (4 broad types) and fine-grained (86 sub-types). There are 4 broad semantic types of compounds: Avyayibhava (Indeclinable), Bahuvrihi (Exocentric), Tatpurusa (Endocentric) and Dvandva (Copulative). Again, each broader class is divided into multiple subclasses, leading to 86 fine-grained types. $^{1}$  Figure 6 shows class-wise label frequency in NecTIS fine-grained.",
+        "bbox": [
+            507,
+            84,
+            884,
+            311
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/d042aadf1ce3670041ca114a42ecb2fbdfe04c294976f0d8a66b9153c8d3a4fa.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Datasets</td><td>#Nested</td><td>#Train</td><td>#Test</td><td>#Dev</td><td>#Types</td></tr><tr><td>NeCTIS</td><td>17656</td><td>12431</td><td>3493</td><td>2405</td><td>4 (86)</td></tr><tr><td>NeCTIS OOD</td><td>1189</td><td>-</td><td>1189</td><td>-</td><td>4 (86)</td></tr></table>",
+        "bbox": [
+            522,
+            321,
+            870,
+            361
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Data statistics for NeCTIS and NeCTIS-OOD",
+        "bbox": [
+            509,
+            370,
+            880,
+            384
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/c3da73854968cd8851f14fbaeded14e077e1f3a601c48c7634b28e9c6c18ceb8.jpg",
+        "image_caption": [
+            "Figure 5: Frequency of  $n$ -component compounds."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            544,
+            414,
+            848,
+            569
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/41bad8dd6130d2436dd5d2a286774e27e3bdf4ae89273dcf8054357792745a7d.jpg",
+        "image_caption": [
+            "Figure 6: Frequency in NeCTIS fine-grained."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            542,
+            621,
+            848,
+            739
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We introduce two context-sensitive datasets: NeCTIS and NeCTIS OOD. The purpose of the additional dataset (NeCTIS OOD) is to create an out-of-domain testbed. The multi-component compound instances are extracted from various books",
+        "bbox": [
+            507,
+            780,
+            884,
+            860
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "The list of fine-grained labels and the corresponding examples can be found at: https://sanskrit.uohyd.ac.in/scl/GOLD_DATA/Tagging_Guidelines/samaasa_tagging16mar12-modified.pdf",
+        "bbox": [
+            507,
+            869,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "13683",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "categorized into 4 types based on subject content: philosophical, pauranic (Translation: Epic is a genre of ancient Indian literature encompassing historical stories, traditions, and legends.), literary, and ayurveda. The NeCTIS dataset encompasses compounds from books falling under the Philosophical, Literary, and ayurveda categories. The multicomponent compound instances extracted from the puranic category are included in the NeCTIS out-of-domain (NeCTIS OOD) dataset. Furthermore, the multi-component instances in NeCTIS belong to the prose domain, while NeCTIS-OOD pertains to the poetry domain. Poetry commonly uses multi-component compounding extensively (more exocentric compounds) to adhere to metrical constraints and convey complex concepts. Conversely, prose uses compounds in a more direct and less condensed manner. Furthermore, poets in the realm of poetry often enjoy the freedom to form novel compounds or employ unconventional ones to conform to meter requirements, rendering these compounds infrequent in regular usage.",
+        "bbox": [
+            110,
+            84,
+            492,
+            441
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dataset Annotation Process: We established a data creation process to address the unavailability of annotated context-sensitive multi-component compound data in Sanskrit. We employ a sufficient annotation budget sponsored by DeitY, 2009-2012 for the Sanskrit Hindi Machine Translation project to employ 6 institutes, each consisting of approximately 10 team members. Each team was organized in a hierarchical manner. There were 3 levels in the hierarchy: Junior linguist (Masters degree in Sanskrit), Senior linguist (Ph.D. in Sanskrit) and professional linguist (Professor in Sanskrit). The annotations from lower expertise were further checked as per the above-mentioned hierarchy. Subsequently, the annotated data underwent an exchange process with another team for correctness verification. Any ambiguities encountered during the annotation process were resolved through collective discussions conducted by the correctness-checking team. The available books were distributed among these teams, and each team was responsible for annotating their allocated books. The annotation guidelines<sup>2</sup> are essentially based on Sanskrit grammar which provides the syntactic and semantic criteria for annotation.",
+        "bbox": [
+            110,
+            450,
+            490,
+            851
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Elaborate commentaries accompany the majority",
+        "bbox": [
+            131,
+            854,
+            487,
+            871
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "of the texts, that discuss the semantics associated with the compounds, which are typically studied by students as a part of their coursework. Given these considerations, it is very unlikely for professional linguists, often professors instructing these texts, to make mistakes. The dataset was curated around 12 years ago, primarily with the aim of producing error-free gold-standard data. Consequently, the errors made by junior annotators were not recorded or measured, aligning with our focus on achieving error-free quality. The benchmark for determining correctness was based on the Pānīnīan grammar.",
+        "bbox": [
+            507,
+            84,
+            885,
+            280
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Baselines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            288,
+            628,
+            304
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We investigate the efficacy of various standard formulations (originally proposed for nested named entity recognition for English) for the proposed task. We adapt these systems to the NeCTI task by providing the location of the compounds to ensure a fair comparison with DepNeCTI. Since these baselines are leveraged from the nested named entity recognition task, they do not have explicit channels to provide the location of a compound word. Therefore, we provide this information in the input string itself with the help of brackets (for example, <sumitra-ānanda-vardhanah>):",
+        "bbox": [
+            505,
+            309,
+            885,
+            502
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Constituency Parsing (CP): Following Fu et al. (2020), we formulate NeCTI as constituency parsing with partially-observed trees, with all labeled compound spans as observed nodes in a constituency tree and non-compound spans as latent nodes. We leverage TreeCRF to model the observed and the latent nodes jointly.",
+            "- Bottom-up Constituency Parsing (BotCP): Following Yang and Tu (2021), we formulate NeCTI as a bottom-up constituency parsing, where a pointer network is leveraged for post-order traversal within a constituency tree to enhance parsing efficiency, enabling linear order parsing.",
+            "- Span Classifier (SpanCL): Following Yuan et al. (2022), we formulate NeCTI as a span classification problem, where triaffine mechanism is leveraged to learn a better span representation by integrating factors such as inside tokens, boundaries, labels, and related spans.",
+            "- Lexicalized Constituency Parsing (LexCP): Following Lou et al. (2022), we formulate NeCTI as lexicalized constituency parsing, which embeds a constituency and a dependency trees"
+        ],
+        "bbox": [
+            507,
+            514,
+            885,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "2The guidelines are available at: https://sanskrit.uohyd.ac.in/scl/GOLD_DATA/Tagging_Guidelines/samaasa_tagging16mar12-modified.pdf",
+        "bbox": [
+            112,
+            879,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "13684",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/6749b74aed49bb8f1500637b7642f1a77150fc7b1d427f741879cd2ff219b3cc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"7\">Coarse</td><td colspan=\"6\">Fine Grain</td></tr><tr><td colspan=\"3\">w/o context</td><td colspan=\"3\">w/ context</td><td colspan=\"3\">w/o context</td><td colspan=\"3\">w/ context</td><td></td></tr><tr><td></td><td>Models</td><td>USS</td><td>LSS</td><td>EM</td><td>USS</td><td>LSS</td><td>EM</td><td>USS</td><td>LSS</td><td>EM</td><td>USS</td><td>LSS</td><td>EM</td></tr><tr><td rowspan=\"7\">NeCTIS</td><td>BotCP</td><td>72.90</td><td>58.78</td><td>32.26</td><td>76.22</td><td>63.97</td><td>35.10</td><td>74.28</td><td>33.50</td><td>18.58</td><td>75.72</td><td>41.80</td><td>23.05</td></tr><tr><td>CP</td><td>76.83</td><td>61.71</td><td>39.97</td><td>64.73</td><td>46.27</td><td>30.14</td><td>77.38</td><td>41.86</td><td>27.22</td><td>70.56</td><td>32.26</td><td>21.11</td></tr><tr><td>LexCP</td><td>93.39</td><td>84.74</td><td>72.88</td><td>93.39</td><td>85.16</td><td>74.41</td><td>88.70</td><td>45.86</td><td>14.72</td><td>87.86</td><td>48.87</td><td>19.67</td></tr><tr><td>Seq2seq</td><td>92.54</td><td>84.11</td><td>59.90</td><td>91.18</td><td>80.45</td><td>52.89</td><td>92.67</td><td>65.63</td><td>30.65</td><td>92.94</td><td>68.19</td><td>34.35</td></tr><tr><td>SpanCL</td><td>92.84</td><td>84.80</td><td>69.12</td><td>93.13</td><td>84.74</td><td>69.67</td><td>93.12</td><td>69.38</td><td>52.17</td><td>92.69</td><td>68.05</td><td>50.82</td></tr><tr><td>DepNeCTI-LSTM</td><td>95.46</td><td>89.06</td><td>76.82</td><td>97.42</td><td>89.24</td><td>77.00</td><td>95.38</td><td>79.49</td><td>57.46</td><td>97.49</td><td>79.72</td><td>56.83</td></tr><tr><td>DepNeCTI-XLMR</td><td>96.21</td><td>90.83</td><td>79.85</td><td>96.16</td><td>90.67</td><td>79.45</td><td>96.35</td><td>83.36</td><td>63.92</td><td>96.34</td><td>83.19</td><td>63.30</td></tr><tr><td rowspan=\"7\">NeCTIS OOD</td><td>BotCP</td><td>72.81</td><td>48.08</td><td>21.63</td><td>73.89</td><td>51.18</td><td>22.57</td><td>72.30</td><td>20.52</td><td>8.64</td><td>73.87</td><td>30.10</td><td>13.32</td></tr><tr><td>CP</td><td>71.43</td><td>52.65</td><td>34.57</td><td>68.17</td><td>38.63</td><td>25.15</td><td>76.12</td><td>29.68</td><td>19.77</td><td>69.57</td><td>24.10</td><td>15.43</td></tr><tr><td>LexCP</td><td>89.90</td><td>71.44</td><td>50.00</td><td>91.57</td><td>72.60</td><td>52.73</td><td>83.45</td><td>32.85</td><td>7.53</td><td>83.38</td><td>34.18</td><td>8.31</td></tr><tr><td>Seq2Seq</td><td>84.26</td><td>71.71</td><td>45.33</td><td>90.13</td><td>71.41</td><td>44.00</td><td>92.16</td><td>53.55</td><td>24.51</td><td>92.87</td><td>53.61</td><td>24.96</td></tr><tr><td>SpanCL</td><td>91.51</td><td>72.69</td><td>56.30</td><td>90.24</td><td>71.89</td><td>54.46</td><td>89.19</td><td>50.21</td><td>32.66</td><td>90.88</td><td>50.50</td><td>34.00</td></tr><tr><td>DepNeCTI-LSTM</td><td>93.32</td><td>78.94</td><td>57.40</td><td>95.67</td><td>79.26</td><td>57.90</td><td>93.88</td><td>67.96</td><td>36.60</td><td>95.88</td><td>67.26</td><td>36.26</td></tr><tr><td>DepNeCTI-XLMR</td><td>95.56</td><td>84.24</td><td>65.50</td><td>95.56</td><td>84.45</td><td>65.00</td><td>95.56</td><td>74.26</td><td>45.70</td><td>95.45</td><td>73.54</td><td>44.37</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            884,
+            328
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Evaluation on the NeCTIS datasets, considering 2 levels of annotations (coarse and fine-grained) and in 2 settings (with and without context). The best-performing results are in bold, while the second-best results are underlined. The results are averaged over 3 runs. To assess the significance between the proposed system and the best baselines for each setting, a significance test in Accuracy metrics was conducted:  $p < 0.01$  (as per t-test).",
+        "bbox": [
+            112,
+            337,
+            882,
+            395
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "together. This formulation leverages the constituents' heads into the architecture which is crucial for the NeCTI task.",
+        "bbox": [
+            127,
+            420,
+            487,
+            467
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Seq2Seq: Following Yan et al. (2021), we formulate NeCTI as an entity span sequence generation task using the pretrained seq2seq framework. This generative framework can also identify discontinuous spans; however, NeCTI does not have such instances.",
+            "- DepNeCTI: We propose two variants of our system (§4) depending on the choice of encoders: DepNeCTI-LSTM and DepNeCTI-XLMR (Nguyen et al., 2021). The compound's location is provided in DepNeCTI-LSTM using span encoding; however, DepNeCTI-XLMR lacks span encoding and leverages this information from the input similar to other baselines."
+        ],
+        "bbox": [
+            114,
+            479,
+            489,
+            715
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics: We evaluate the performance using the Labeled/Unlabeled Span Score (LSS/USS) in terms of micro-averaged F1-score. We define LSS as a micro-averaged F1-score applied on tuples of predicted spans including their labels. We exclude labels of the spans while calculating USS. Additionally, we report the exact match (EM) which indicates the percentage of the compounds for which the predictions of all spans and their semantic relations are correctly identified. Refer to Appendix B for hyper-parameters and details of the computing infrastructure used.",
+        "bbox": [
+            112,
+            726,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Results",
+        "text_level": 1,
+        "bbox": [
+            509,
+            419,
+            613,
+            434
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2 presents the performance of the top-performing configurations among all baselines on the NeCTIS benchmark datasets' test set. The evaluation includes 2 levels of annotations (coarse and fine-grained) and in 2 settings (with and without context). While all baseline systems demonstrate competitiveness, no single baseline consistently outperforms the others across all settings. Consequently, we underline the best-performing numbers within each specific setting.",
+        "bbox": [
+            507,
+            448,
+            884,
+            609
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Our proposed system DepNeCTI surpasses all competing systems across all evaluation metrics, demonstrating an absolute average gain of 13.1 points (LSS) and 11.3 points (EM) compared to the best-performing baseline in each setting. Notably, our proposed system exhibits substantial performance superiority over the best baseline in fine-grained settings, particularly in low-resourced scenarios. This validates the effectiveness of our proposed system in low-resourced settings with fine-grained labels. The significant performance gap between our proposed system and the best baselines highlights the efficacy of employing a simple yet effective architecture inspired by the dependency parsing paradigm. These results establish new state-of-the-art benchmarks by integrating the contextual component into our novel framework. While most baselines (except BotCP) do not benefit from contextual information, DepNeCTI-LSTM",
+        "bbox": [
+            507,
+            613,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "13685",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "demonstrates slight improvements and DepNeCTI-XLMR shows on par improvements when leveraging contextual information. Furthermore, as the number of components grows, the number of potential solutions increases exponentially, leading to poor performance by the systems in such scenarios. Due to this exponential possibility, contextual information provides limited improvements compared to binary compound identification settings (Sandhan et al., 2022a). In other words, unlike the context-free setting, the introduction of context information does not warrant an expectation for the system to precisely generate the correct solution from the exponential candidate space. Figure 3 provides a visual representation that elucidates the concept of this exponential candidate space. A similar performance trend is observed for the NeCTIS-OOD dataset.",
+        "bbox": [
+            112,
+            84,
+            492,
+            374
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            386,
+            225,
+            401
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Here, we examine the proposed system, focusing on a comprehensive analysis and its applicability. For this purpose, we utilize the NeCTIS coarse dataset under the w/ context configuration. We report LSS in terms of macro-average F1-scores.",
+        "bbox": [
+            112,
+            412,
+            489,
+            494
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(1) Ablation analysis: In this study, we analyze the impact of different system components on the overall enhancements of DepNeCTI-LSTM. Ablations, documented in Table 3, present the evaluation metrics when a specific component is deactivated within DepNeCTI-LSTM. For instance, the absence of the span encoding component is denoted as “- Span Encoding”. The results indicate that removing any component leads to a decline in performance. Notably, Table 3 highlights the significance of the “Span Encoding” component in driving the improvements.",
+        "bbox": [
+            112,
+            502,
+            489,
+            695
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/db281849403599f19f2e4cfcdc94c2decd1d238f1ccc6abe387a6d585f4b9372.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>System</td><td>P</td><td>R</td><td>F1</td><td>EM</td></tr><tr><td>DepNeCTI-LSTM</td><td>89.24</td><td>89.24</td><td>89.24</td><td>77.00</td></tr><tr><td>- FastText (FT)</td><td>88.84</td><td>88.84</td><td>88.84</td><td>76.50</td></tr><tr><td>- Span Encoding (SE)</td><td>86.18</td><td>85.53</td><td>86.85</td><td>70.54</td></tr><tr><td>- FT - SE</td><td>84.25</td><td>84.23</td><td>84.24</td><td>67.86</td></tr></table>",
+        "bbox": [
+            127,
+            706,
+            475,
+            778
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Each ablation involved the removal of a singular component from DepNeCTI-LSTM. The ablation denoted as “-Span Encoding” entailed eliminating the span encoding component from the proposed system.",
+        "bbox": [
+            112,
+            787,
+            489,
+            846
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(2) How robust is the system when the number of components of a compound increases? We analyze the relationship between the F1-score and",
+        "bbox": [
+            112,
+            871,
+            489,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the number of components in compounds. For compounds with a small number of components, all systems demonstrate high performance, but our proposed systems consistently outperforms other baselines. However, as the number of components increases, the number of examples in each category decreases. Additionally, the number of potential solutions grows exponentially, following the Catalan number. Consequently, all systems experience a decline in performance.",
+        "bbox": [
+            507,
+            84,
+            885,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5ff8473746bb9318bf3a0541b7d8f58d627dc7872f6d44aa8d9c26fae00d945a.jpg",
+        "image_caption": [
+            "Figure 7: F1-Score against the number of components. The compounds with components  $N > 10$  are excluded."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            544,
+            256,
+            850,
+            385
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(3) Error analysis: We investigate whether all the systems are able to identify the location of a multi-component compound correctly. The motivation behind this experiment is to evaluate the capability of the baselines and the proposed architecture to leverage the information about the compound's location effectively. We define the span of text that corresponds to a compound as a global span which we know apriori. Figure 8 illustrates the effectiveness of each system in correctly identifying the global span of multi-component compounds. In",
+        "bbox": [
+            507,
+            449,
+            885,
+            627
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cd62abd30e89577376aaf5b03b1a6f1bc475b8e081afdc231261c5500e1295ad.jpg",
+        "image_caption": [
+            "Figure 8: Performance of the systems on identifying the global span of a compound."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            564,
+            642,
+            843,
+            760
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "DepNeCTI-LSTM, our span encoding effectively captures this information, resulting in a perfect  $100\\%$  score. However, even after providing the baselines with this information, they fail to use it due to limitations in their architectures. Interestingly, DepNeCTI-XLMR does not contain a span",
+        "bbox": [
+            507,
+            822,
+            885,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "13686",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "encoding component and leverages the compound's location information from the input as provided for the baselines. Still, DepNeCTI-XLMR reports the best performance due to its powerful word representation ability. It is worth noting that the NeC-TIS dataset exhibits an inherent bias towards left branching, as indicated by the nested tree structure. Consequently, all systems display a bias towards left branching as well. Therefore, due to the dominance of left-branching instances and increased variance due to less number of instances, a spike is observed in the results.",
+        "bbox": [
+            112,
+            84,
+            492,
+            277
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "(4) Efficiency of our proposed system: Figure 9 present the computational efficiency of our system measured in terms of the number of sentences processed per second. We compare the inference speed of different baselines on the NeCTI task. Our systems, DepNeCTI-LSTM and DepNeCTI-XLMR are leveraging a simple architecture and utilizing dependency parsing as an appropriate problem formulation, exhibits a 5-fold/3-fold improvement over the most efficient baseline, BotCP.",
+        "bbox": [
+            112,
+            285,
+            490,
+            447
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/90fe607bd03136c8e1c07b23fb267912538017d39f1d3f11633d60ab90874443.jpg",
+        "image_caption": [
+            "Figure 9: Inference speed of the competing systems."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            456,
+            455,
+            600
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Related Works",
+        "text_level": 1,
+        "bbox": [
+            112,
+            653,
+            278,
+            668
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Lexical semantics is a dedicated field focused on word meaning. Various tasks such as word-sense disambiguation (Bevilacqua and Navigli, 2020; Barba et al., 2021; Maurya et al., 2023; Maurya and Bahadur, 2022), relationship extraction (Tian et al., 2021; Nadgeri et al., 2021; Hu et al., 2022; Wang et al., 2022), and semantic role labeling (He et al., 2018; Kulkarni and Sharma, 2019; Zhang et al., 2022) play essential roles in determining word meaning. Moreover, when dealing with complex word structures such as compounds, named entities, and multi-word expressions, relying solely on basic word senses and relationships is inadequate. While efforts have been made in Noun Compound Identification (Ziering and van der Plas,",
+        "bbox": [
+            112,
+            677,
+            490,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "2015; Dima and Hinrichs, 2015; Fares et al., 2018; Shwartz and Waterson, 2018; Ponkiya et al., 2018, 2020, 2021), multi-word expression (MWE) (Constant et al., 2017; Gharbieh et al., 2017; Gooding et al., 2020; Premasiri and Ranasinghe, 2022) and named entity recognition (Fu et al., 2020; Yang and Tu, 2021; Lou et al., 2022; Yuan et al., 2022), the nested compounds remains unexplored.",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Sanskrit Compound Type Identification has attracted significant attention over the past decade. Decoding the meaning of a Sanskrit compound requires determining its constituents (Huet, 2009; Mittal, 2010; Hellwig and Nehrdich, 2018), understanding how these constituents are grouped (Kulkarni and Kumar, 2011), identifying the semantic relationship between them (Kumar, 2012), and ultimately generating a paraphrase of the compound (Kumar et al., 2009). Previous studies proposed rule-based approaches (Satuluri and Kulkarni, 2013; Kulkarni and Kumar, 2013), a data-driven approach (Sandhan et al., 2019) and a hybrid approach (Krishna et al., 2016) for SaCTI. Sandhan et al. (2022a) proposed a context-sensitive architecture for binary compounds.",
+        "bbox": [
+            507,
+            221,
+            885,
+            479
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Earlier works in Sanskrit solely focused on binary compounds, neglecting the identification of multi-component compound types; however, our proposed framework fills this research gap.",
+        "bbox": [
+            507,
+            480,
+            882,
+            544
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            556,
+            640,
+            570
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we focused on multi-component compounding in Sanskrit, which helps to decode the implicit structure of a compound to decipher its meaning. While previous approaches have primarily focused on binary compounds, we introduced a novel task called nested compound type identification (NeCTI). This task aims to identify nested spans within multi-component compounds and decode the implicit semantic relations between them, filling a gap in the field of lexical semantics. To facilitate research in this area, we created 2 newly annotated datasets, designed explicitly for the NeCTI task. These datasets were utilized to benchmark various problem formulations. Our novel framework DepNeCTI outperformed the best baseline system by achieving a stupendous absolute gain of 13.1 points F1-score in terms of LSS. Similar to the previous findings on binary Sanskrit compound identification, we discovered that our proposed system exhibits substantial performance superiority over the best baseline in low-resourced scenarios.",
+        "bbox": [
+            507,
+            581,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "13687",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            98
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We could not extend our framework to other languages exhibiting multi-component compounding phenomena due to the lack of availability of annotated datasets. It would be interesting to measure the effectiveness of rules from Pānīnīan grammar to discard incompatible semantic relations (Kulkarni, 2019, 2021).",
+        "bbox": [
+            112,
+            111,
+            489,
+            223
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            237,
+            265,
+            253
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "This work introduces a new task, along with annotated datasets and a framework, to address the nested compounding phenomenon in Sanskrit. The proposed resources aim to enhance the understanding of multi-component compounds and contribute to the improvement of machine translation systems. Regarding potential effects, we anticipate no harm to any community resulting from the use of our datasets and framework. However, we advise users to exercise caution, as our system is not flawless and may generate mispredictions. To ensure transparency and future research, we have publicly released all our annotated NeCTIS datasets and source codes. We confirm that our data collection adheres to the terms of use of the sources and respects the intellectual property and privacy rights of the original authors. Our annotation team consisted of qualified individuals, including Master's and Ph.D. degree holders, some of whom are Sanskrit professors. Annotators were compensated appropriately and provided with detailed instructions to ensure consistency in the annotation process. We remain committed to addressing ethical implications as we refine our systems and welcome feedback from the community to enhance our ethical practices.",
+        "bbox": [
+            112,
+            265,
+            490,
+            684
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            699,
+            285,
+            714
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We thank all the annotators from different institutes for helping us with NeCTIS data annotation. We would like to thank Zheng Yuan, Tsinghua University (Alibaba Group), for helping us with the hyperparameter section and adapting his system for the NeCTI task. We appreciate and thank all the anonymous reviewers for their constructive feedback towards improving this work. The work was supported in part by the National Language Translation Mission (NLTM): Bhashini project by Government of India.",
+        "bbox": [
+            112,
+            726,
+            490,
+            902
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            84,
+            608,
+            98
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Rahul Aralikatte, Miryam de Lhoneux, Anoop Kunchukuttan, and Anders Søgaard. 2021. Itihasa: A large-scale corpus for Sanskrit to English translation. In Proceedings of the 8th Workshop on Asian Translation (WAT2021), pages 191-197, Online. Association for Computational Linguistics.",
+            "Edoardo Barba, Luigi Procopio, and Roberto Navigli. 2021. ConSeC: Word sense disambiguation as continuous sense comprehension. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1492-1503, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Michele Bevilacqua and Roberto Navigli. 2020. Breaking through the  $80\\%$  glass ceiling: Raising the state of the art in word sense disambiguation by incorporating knowledge graph information. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2854-2864, Online. Association for Computational Linguistics.",
+            "Mathieu Constant, Gülşen Eryiğit, Johanna Monti, Lonneke van der Plas, Carlos Ramisch, Michael Rosner, and Amalia Todirascu. 2017. Survey: Multiword expression processing: A Survey. Computational Linguistics, 43(4):837-892.",
+            "Corina Dima and Erhard Hinrichs. 2015. Automatic noun compound interpretation using deep neural networks and word embeddings. In Proceedings of the 11th International Conference on Computational Semantics, pages 173-183, London, UK. Association for Computational Linguistics.",
+            "Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency parsing.",
+            "Murhaf Fares, Stephan Oepen, and Erik Velldal. 2018. Transfer and multi-task learning for noun-noun compound interpretation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1488-1498, Brussels, Belgium. Association for Computational Linguistics.",
+            "Yao Fu, Chuanqi Tan, Mosha Chen, Songfang Huang, and Fei Huang. 2020. Nested named entity recognition with partially-observed treecrfs.",
+            "Waseem Gharbieh, Virendrakumar Bhavsar, and Paul Cook. 2017. Deep learning models for multiword expression identification. In Proceedings of the 6th Joint Conference on Lexical and Computational Semantics (*SEM* 2017), pages 54-64, Vancouver, Canada. Association for Computational Linguistics.",
+            "Sian Gooding, Shiva Taslimipoor, and Ekaterina Kochmar. 2020. Incorporating multiword expressions in phrase complexity estimation. In Proceedings of the 1st Workshop on Tools and Resources to Empower People with REAding Difficulties (READI), pages 14-19, Marseille, France. European Language Resources Association."
+        ],
+        "bbox": [
+            510,
+            105,
+            885,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "13688",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Pawan Goyal and Gerard Huet. 2016. Design and analysis of a lean interface for sanskrit corpus annotation. Journal of Language Modelling, 4:145.",
+            "Pawan Goyal and Amba Kulkarni. 2014. Converting phrase structures to dependency structures in Sanskrit. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1834-1843, Dublin, Ireland. Dublin City University and Association for Computational Linguistics.",
+            "Luheng He, Kenton Lee, Omer Levy, and Luke Zettlemoyer. 2018. Jointly predicting predicates and arguments in neural semantic role labeling. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 364-369, Melbourne, Australia. Association for Computational Linguistics.",
+            "Oliver Hellwig and Sebastian Nehrdich. 2018. Sanskrit word segmentation using character-level recurrent and convolutional neural networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2754-2763, Brussels, Belgium. Association for Computational Linguistics.",
+            "Chengwei Hu, Deqing Yang, Haoliang Jin, Zhen Chen, and Yanghua Xiao. 2022. Improving continual relation extraction through prototypical contrastive learning. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1885-1895, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.",
+            "Gérard Huet. 2009. Sanskrit segmentation. In Proceedings of the South Asian Languages Analysis Roundtable XXVIII, Denton, Ohio.",
+            "Amrith Krishna, Pavankumar Satuluri, Shubham Sharma, Apurv Kumar, and Pawan Goyal. 2016. Compound type identification in Sanskrit: What roles do the corpus and grammar play? In Proceedings of the 6th Workshop on South and Southeast Asian Natural Language Processing (WSSANLP2016), pages 1-10, Osaka, Japan. The COLING 2016 Organizing Committee.",
+            "Amba Kulkarni. 2019. Sanskrit Parsing based on the Theories of Shabdabodha. D. K. PrintWorld and Indian Institute of Advanced Study.",
+            "Amba Kulkarni. 2021. Sanskrit parsing following indian theories of verbal cognition. In ACM Transactions on Asian and Low-Resource Language Information Processing, pages 1-38.",
+            "Amba Kulkarni and Anil Kumar. 2011. Statistical constituency parser for sanskrit compounds. In Proceedings of ICON-2011: 9th International Conference on Natural Language Processing, International Institute of Information Technology, Hyderabad, India. Macmillan Publishers."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Amba Kulkarni and Anil Kumar. 2013. Clues from astadhyay for compound type identification. In Proceedings of the 5th International Sanskrit Computational Linguistics Symposium, IIT Bombay, India. D. K. PrintWorld and Indian Institute of Advanced Study.",
+            "Amba Kulkarni and Dipti Sharma. 2019. Panjinian syntactico-semantic relation labels. In Proceedings of the Fifth International Conference on Dependency Linguistics (Depling, SyntaxFest 2019), pages 198-208, Paris, France. Association for Computational Linguistics.",
+            "Anil Kumar. 2012. An automatic Sanskrit compound processing. Ph.D. thesis, University of Hyderabad.",
+            "Anil Kumar, V Sheeba, and Amba Kulkarni. 2009. Sanskrit compound paraphrase generator. In Proceedings of the ICON 2009: 7th International Conference on Natural Language Processing, University of Hyderabad, Hyderabad. Macmillan Publishers.",
+            "Mirella Lapata and Frank Keller. 2004. The web as a baseline: Evaluating the performance of unsupervised web-based models for a range of NLP tasks. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pages 121-128, Boston, Massachusetts, USA. Association for Computational Linguistics.",
+            "Chao Lou, Songlin Yang, and Kewei Tu. 2022. Nested named entity recognition as latent lexicalized constituency parsing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6183-6198, Dublin, Ireland. Association for Computational Linguistics.",
+            "Xuezhe Ma, Zeong Hu, Jingzhou Liu, Nanyun Peng, Graham Neubig, and Eduard Hovy. 2018. Stackpointer networks for dependency parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1403-1414, Melbourne, Australia. Association for Computational Linguistics.",
+            "Archana Maurya and Promila Bahadur. 2022. *A Detailed Analysis of Word Sense Disambiguation Algorithms and Approaches for Indian Languages*, pages 693-710.",
+            "Archana Sachindeo Maurya, Promila Bahadur, and Srishti Garg. 2023. Approach toward word sense disambiguation for the english-to-sanskrit language using naive bayesian classification. In Proceedings of Third Doctoral Symposium on Computational Intelligence, pages 477-491, Singapore. Springer Nature Singapore.",
+            "Vipul Mittal. 2010. Automatic Sanskrit segmentizer using finite state transducers. In Proceedings of the"
+        ],
+        "bbox": [
+            512,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "13689",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ACL 2010 Student Research Workshop, pages 85-90, Uppsala, Sweden. Association for Computational Linguistics.",
+            "Abhishek Nadgeri, Anson Bastos, Kuldeep Singh, Isaiah Onando Mulang', Johannes Hoffart, Saeedeh Shekarpour, and Vijay Saraswat. 2021. KGPool: Dynamic knowledge graph context selection for relation extraction. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 535-548, Online. Association for Computational Linguistics.",
+            "Minh Van Nguyen, Viet Duc Lai, Amir Pouran Ben Vey-seh, and Thien Huu Nguyen. 2021. Trankit: A lightweight transformer-based toolkit for multilingual natural language processing. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 80-90, Online. Association for Computational Linguistics.",
+            "Panini. 500 BCE. Aştadhyayi.",
+            "Girishkumar Ponkiya, Diptesh Kanojia, Pushpak Bhattacharyya, and Girish Palshikar. 2021. FrameNet-assisted noun compound interpretation. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2901-2911, Online. Association for Computational Linguistics.",
+            "Girishkumar Ponkiya, Rudra Murthy, Pushpak Bhattacharyya, and Girish Palshikar. 2020. Looking inside noun compounds: Unsupervised prepositional and free paraphrasing. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4313-4323, Online. Association for Computational Linguistics.",
+            "Girishkumar Ponkiya, Kevin Patel, Pushpak Bhattacharyya, and Girish Palshikar. 2018. Treat us like the sequences we are: Prepositional paraphrasing of noun compounds using LSTM. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1827-1836, Santa Fe, New Mexico, USA. Association for Computational Linguistics.",
+            "Damith Premasiri and Tharindu Ranasinghe. 2022. Bert(s) to detect multiword expressions.",
+            "Jivnesh Sandhan, Ashish Gupta, Hrishikesh Terdalkar, Tushar Sandhan, Suvendu Samanta, Laxmidhar Behera, and Pawan Goyal. 2022a. A novel multi-task learning approach for context-sensitive compound type identification in Sanskrit. In Proceedings of the 29th International Conference on Computational Linguistics, pages 4071-4083, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.",
+            "Jivnesh Sandhan, Amrith Krishna, Pawan Goyal, and Laxmidhar Behera. 2019. Revisiting the role of feature engineering for compound type identification in Sanskrit. In Proceedings of the 6th International Sanskrit Computational Linguistics Symposium, pages"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "28-44, IIT Kharagpur, India. Association for Computational Linguistics.",
+            "Jivnesh Sandhan, Rathin Singha, Narein Rao, Suvendu Samanta, Laxmidhar Behera, and Pawan Goyal. 2022b. TransLIST: A transformer-based linguistically informed Sanskrit tokenizer. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 6902-6912, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Pavankumar Satuluri and Amba Kulkarni. 2013. Generation of sanskrit compounds. In Proceedings of ICON-2013: 10th International Conference on Natural Language Processing, CDAC Noida, India.",
+            "Vered Shwartz and Chris Waterson. 2018. Olive oil is made of olives, baby oil is made for babies: Interpreting noun compounds using paraphrases in a neural model. 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 218-224, New Orleans, Louisiana. Association for Computational Linguistics.",
+            "Hrishikesh Terdalkar and Arnab Bhattacharya. 2019. Framework for question-answering in Sanskrit through automated construction of knowledge graphs. In Proceedings of the 6th International Sanskrit Computational Linguistics Symposium, pages 97-116, IIT Kharagpur, India. Association for Computational Linguistics.",
+            "Yuanhe Tian, Guimin Chen, Yan Song, and Xiang Wan. 2021. Dependency-driven relation extraction with attentive graph convolutional networks. 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 4458-4471, Online. Association for Computational Linguistics.",
+            "Peiyi Wang, Yifan Song, Tianyu Liu, Rundong Gao, Binghuai Lin, Yunbo Cao, and Zhifang Sui. 2022. Less is more: Rethinking state-of-the-art continual relation extraction models with a frustratingly easy but effective approach.",
+            "Hang Yan, Tao Gui, Junqi Dai, Qipeng Guo, Zheng Zhang, and Xipeng Qiu. 2021. A unified generative framework for various ner subtasks.",
+            "Songlin Yang and Kewei Tu. 2021. Bottom-up constituency parsing and nested named entity recognition with pointer networks.",
+            "Zheng Yuan, Chuanqi Tan, Songfang Huang, and Fei Huang. 2022. Fusing heterogeneous factors with triaffine mechanism for nested named entity recognition.",
+            "Li Zhang, Ishan Jindal, and Yunyao Li. 2022. Label definitions improve semantic role labeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "13690",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Linguistics: Human Language Technologies, pages 5613-5620, Seattle, United States. Association for Computational Linguistics.",
+        "bbox": [
+            132,
+            85,
+            487,
+            124
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Patrick Ziering and Lonneke van der Plas. 2015. From a distance: Using cross-lingual word alignments for noun compound bracketing. In Proceedings of the 11th International Conference on Computational Semantics, pages 82–87, London, UK. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            135,
+            489,
+            214
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A Why NeCTI as a Dependency Parsing Task?",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            873,
+            115
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Compounds in language are semantic constructions. While a limited number of rules derived from Sanskrit Grammar aid in determining the syntactic structure of a compound, they offer limited assistance in uncovering its meaning. The meaning of a compound primarily arises from the semantic relationship between its components. This semantic relationship, known as \"sāmarthya,\" is expressed through various types of semantic compounds. These compound types also facilitate the identification of the headword within a compound. The headword can be one of the constituents or an entirely distinct word modified by the resulting compound. Compounds consisting of more than two components are typically formed through successive binary combinations of the components, following specific relations. As a result, a nested structure of binary compounds is created, with a few exceptions. Therefore, the semantic compound types play a vital role in determining the correct nested structure amidst the various possible structures of a compound. Each of these structures represents a constituency span, and a parser equipped with compound type identification assists in accurately identifying the appropriate constituency span.",
+        "bbox": [
+            507,
+            128,
+            884,
+            546
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Treating this as a constituency parsing task presents several challenges. Figure 10 illustrates the potential constituency spans for a 3-component compound  $a - b - c$ , with  $ab$  and  $bc$  representing the intermediate compounds. First, the nested structure of compounds does not conform to a syntactic structure; instead, it follows a semantic structure based on the relations between the components rather than their position or relative co-occurrence. Second, the intermediate compounds within the structure are not categories but stem forms. Each intermediate compound serves as an entity that modifies the meanings of one of its constituents, subsequently combining with another component to form a larger compound.",
+        "bbox": [
+            507,
+            548,
+            882,
+            788
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Kumar (2012) formulated this as constituency parsing, where the identification of semantic compatibility was performed based on the relative co-occurrence and relative position of the components. As shown in Figure 11, the intermediate compounds, which lacked additional information, were substituted with their respective semantic compound types in the type identification stage.",
+        "bbox": [
+            507,
+            790,
+            882,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "13691",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/14e82fef0833ca1f709f59be7d00163ebf433fb6e7e98af7148eb59c6227e1fb.jpg",
+        "image_caption": [
+            "Figure 10: Possible constituency spans for a three component compound a-b-c"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            90,
+            272,
+            225
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/e502b72898c158a349061a2bcc1720ef5361e89c3c12c3fe80521a95350d3874.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            327,
+            90,
+            453,
+            225
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/1df216ff9dac45c83cb764f96742822e5d44217f5e4110f5932a60083062abb9.jpg",
+        "image_caption": [
+            "Figure 11: Possible constituency spans for a three component compound a-b-c with semantic types"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            292,
+            289,
+            428
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/17add7d243851a830a55f970ac61f99249d78a125aa26f9e94a6e0c8a336c68e.jpg",
+        "image_caption": [
+            "Figure 12: Dependency Structure for the three component compound a-b-c with semantic types as labels"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            309,
+            292,
+            473,
+            428
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "The semantic types serve as semantic constructions for compounds, rather than being syntactic categories and play a crucial role in determining the meaning of the compound. For example the compound vidyālayaghanṭa has the nested structure «vidyā-alaya>T6-ghanṭa>T6. The type T6 indicatesṣaṭṭi tatpurusa compound (6<sup>th</sup> case determinative compound) inferring a possessive relationship. Vidyā (knowledge) and alaya (place) combine to form the intermediate compound vidyālaya (the place of knowledge, viz. school), which combines with ghantā (bell) to form the whole compound indicating school-bell. The possessive relation expressed by the compound type (T6) is akin to a dependency relation, and this holds true for other compound types as well.",
+        "bbox": [
+            112,
+            500,
+            489,
+            758
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Consequently, when represented with compound types, the same constituency spans can be depicted as dependency structures by annotating the types as directed relations and removing the extra nodes indicating the types. The dependency structures for the previous example  $(a - b - c)$  are shown in Figure 12. Furthermore, the head words are not explicitly marked in constituency spans and can only be identified through their corresponding types. However, with dependency structures, the",
+        "bbox": [
+            112,
+            758,
+            489,
+            919
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "head word can be determined by the labels directed towards it within the compound. Notably, these dependency structures faithfully capture the constituency information and are mutually convertible with their corresponding constituency spans.",
+        "bbox": [
+            507,
+            84,
+            884,
+            165
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/f34917bff55776dfcbb154e24bed17fdc8c1bfca679902f6ce608ff56da5175e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            566,
+            174,
+            823,
+            227
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/e5b123b04460721ef6649a19a35d1f2f6ca86990d7f76dab5d7a84d1f559bf49.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            566,
+            237,
+            823,
+            318
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "There are several considerations behind the decision to treat NeCTI as a dependency parsing task. First, the semantic relations among compound components resemble dependency relations and can be represented as directed labels within the dependency parse structure. Second, this approach concisely represents the semantic relations between compound components without introducing extra intermediary nodes. Lastly, it enables the simultaneous identification of the structure or constituency span alongside the identification of compound types.",
+        "bbox": [
+            507,
+            378,
+            884,
+            571
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B Experiment Details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            581,
+            714,
+            598
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Hyper-parameters: For our proposed system, we build on the top of the codebase from BiAFF, as developed by Ma et al. (2018). We configure the hyperparameters as follows: a batch size of 16, 100 training iterations, a dropout rate of 0.33, 2 stacked Bi-LSTM layers, a learning rate of 0.002, and the remaining parameters set identically to those used in the work of Ma et al. (2018). We use manual tuning for the hyper-parameter selection and F1-score criteria on dev set's performance. Our codebase is publicly available and released under a creative-common license. We use FastText word embeddings for the proposed framework.",
+        "bbox": [
+            505,
+            606,
+            882,
+            815
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Computing Infrastructure Used: We perform our experiments using a single GPU equipped with an NVIDIA A40, 48 GB GPU memory, and 10752 CUDA cores. We employ a single GPU with an NVIDIA Quadro RTX 4000, 8 GB GPU memory, and 2304 CUDA cores for our proposed system.",
+        "bbox": [
+            507,
+            822,
+            882,
+            919
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "13692",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/DepNeCTI_ Dependency-based Nested Compound Type Identification for Sanskrit/9cf22d4e-25cc-4736-8b7f-6f6bea41e671_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f406a57dd3fc2f016b2f4546b69f25ed29b878263807bff0905390717e8aa7d
+size 629022

2023/DepNeCTI_ Dependency-based Nested Compound Type Identification for Sanskrit/full.md

@@ -0,0 +1,303 @@
+# DepNeCTI: Dependency-based Nested Compound Type Identification for Sanskrit
+
+Jivnesh Sandhan†,* Yaswanth Narsupalli‡, Sreevatsa Muppirala‡, Sriram Krishnan§, Pavankumar Satuluri∥, Amba Kulkarni§ and Pawan Goyal‡
+
+†UC Berkeley, ‡IIT Kharagpur, §University of Hyderabad and ¶IIT Roorkee
+
+jivneshsandhan@gmail.com, yasshu.yaswanth@gmail.com, pawang@cse.iitkgp.ac.in
+
+# Abstract
+
+Multi-component compounding is a prevalent phenomenon in Sanskrit, and understanding the implicit structure of a compound's components is crucial for deciphering its meaning. Earlier approaches in Sanskrit have focused on binary compounds and neglected the multi-component compound setting. This work introduces the novel task of nested compound type identification (NeCTI), which aims to identify nested spans of a multi-component compound and decode the implicit semantic relations between them. To the best of our knowledge, this is the first attempt in the field of lexical semantics to propose this task.
+
+We present 2 newly annotated datasets including an out-of-domain dataset for this task. We also benchmark these datasets by exploring the efficacy of the standard problem formulations such as nested named entity recognition, constituency parsing and seq2seq, etc. We present a novel framework named DepNeCTI: Dependency-based Nested Compound Type Identifier that surpasses the performance of the best baseline with an average absolute improvement of 13.1 points F1-score in terms of Labeled Span Score (LSS) and a 5-fold enhancement in inference efficiency. In line with the previous findings in the binary Sanskrit compound identification task, context provides benefits for the NeCTI task. The codebase and datasets are publicly available at: https://github.com/yaswanth-iitkgp/DepNeCTI
+
+# 1 Introduction
+
+A compound is defined as a group of entities functioning as a single meaningful entity. The process of identifying the implied semantic relationship between the components of a compound in Sanskrit is known as Sanskrit Compound Type Identification (SaCTI) (Sandhan et al., 2022a) or Noun Compound Interpretation (NCI) (Ponkiya et al., 2021,
+
+2020). Within the literature, the NCI problem has been approached in two ways, namely, classification (Dima and Hinrichs, 2015; Fares et al., 2018; Ponkiya et al., 2021) and paraphrasing (Lapata and Keller, 2004; Ponkiya et al., 2018, 2020).
+
+In Sanskrit literature, particularly in poetry, the use of multi-component compounds is ubiquitous (Kumar, 2012). According to the Digital Corpus of Sanskrit, more than  $41\%$  of compounds contain three or more components (Krishna et al., 2016). However, earlier approaches focus solely on binary compounds and fail to address the complexities inherent in multi-component compounds adequately. Thus, we propose a new task, the Nested Compound Type Identification (NeCTI) Task, which focuses on identifying nested spans within a multicomponent compound and interpreting their implicit semantic relationships. Figure 1 illustrates an example of the NeCTI task, highlighting nested spans and their associated semantic relations using distinct colors.
+
+![](images/8fce5a9d9993137775f22da5532a49bcdcb89d81850c7ea5d2f698e76259dd72.jpg)  
+Figure 1: Illustration of the NeCTI task for the multi-component compound sumitra-ananda-vardhanah (Translation: Sumitra-delight-enhancer), highlighting nested spans and their associated semantic relations using distinct colors.
+
+The NeCTI task presents multiple challenges: (1) The number of potential solutions for a multicomponent compound grows exponentially as the number of components increases. (2) It often relies on contextual or world knowledge about the entities involved (Krishna et al., 2016). Even if a multicomponent compound shares the same components
+
+and final form, the implicit relationship between the spans can only be deciphered with the aid of available contextual information (Kulkarni and Kumar, 2013; Krishna et al., 2016). For instance, as de
+
+![](images/265320e2dba0b3084032e52044623b15bc76e00e3533e531149e7c7a9ec77023.jpg)  
+Figure 2: Illustration of the multi-component compound sumitra-ananda-vardhanah (Translation: Sumitra-delight-enhancer) with two valid parses depicted in green and red. The two parses correspond to two potential meanings. The green solution represents the correct interpretation within the provided context.
+
+picted in Figure 2, the multi-component compound sumitra-ananda-vardhanah (Translation: Sumitradelight-enhancer) can have two valid but distinct solutions, leading to different meanings. Resolving ambiguity to select the correct solution requires reliance on the provided context. Consequently, downstream Natural Language Processing (NLP) applications for Sanskrit, such as question answering (Terdalkar and Bhattacharya, 2019) and machine translation (Aralikatte et al., 2021), often exhibit sub-optimal performance when encountering compounds. Hence, the NeCTI task serves as a preliminary requirement for developing robust NLP technology for Sanskrit. Moreover, this dependency on contextual information eliminates the possibility of storing and conducting a lookup to identify the semantic types of nested spans.
+
+Previous approaches (Kulkarni and Kumar, 2013; Krishna et al., 2016; Sandhan et al., 2019) addressing SaCTI have predominantly focused on binary compounds, neglecting the consideration of multi-component compounds. In multi-component compounds, the components exhibit semantic relationships akin to dependency relations, represented as directed labels within the dependency structure, which also facilitate the identification of the compound's headword through the labels directed towards it. Consequently, dependency formulation enables the simultaneous identification of both the structure or constituency span and the compound types. Thus, we propose a novel framework ( $\S 4$ ) named DepNeCTI: Dependency-based Nested Compound type Identifier ( $\S 3$ ). In summary, our contributions can be outlined as follows:
+
+- We introduce a novel task called Nested Compound Type Identification (§ 2).
+
+- We present 2 newly annotated datasets and provide benchmarking by exploring the efficacy of various standard formulations for NeCTI (§ 5).  
+- We propose a novel framework DepNeCTI: Dependency-based Nested Compound type Identifier (§ 4), which reports an average 13.1 points F1-score in terms of LSS absolute gain and 5-fold enhancement in inference efficiency (§ 6) over the best baseline.  
+- We publicly release the codebase of DepNeCTI and benchmarked baselines, along with newly annotated datasets for the NeCTI task.
+
+# 2 Problem Formulation
+
+The objective of the NeCTI task is to detect nested spans within a multi-component compound and decipher the implicit semantic relations among them. Our study focuses exclusively on this task and does not address the compound segmentation problem. It is assumed that the segmented components of the multi-component compound are already available. To obtain the segmentation of a compound, we rely on established resources such as the rule-based shallow parser (Goyal and Huet, 2016) or existing data-driven segmentation systems designed explicitly for Sanskrit (Hellwig and Nehrdich, 2018; Sandhan et al., 2022b).
+
+![](images/5f67455afe96c082a3588685841c0ceb83beea3e04841f5ca78b657aebe04ee0.jpg)  
+Figure 3: Illustration of a number of ways (as per Catalan number in log scale) in which a multi-component compound can be parsed. Our dataset contains compounds that have a maximum of 16 components.
+
+Complexity of NeCTI Task: As the number of components in a multi-component compound increases, the number of possible parses grows exponentially. Our dataset encompasses multicomponent compounds ranging from 2 to 16 components. Figure 3 visually depicts the exponential growth of possible parses with increasing component count. The parsing of a compound word with
+
+$n + 1$  components can be likened to the problem of fully parenthesizing  $n + 1$  factors in all possible ways (Kulkarni and Kumar, 2011). Thus, the total number of parse-options for a multi-component compound word with  $n + 1$  components corresponds to the Catalan Number  $C_n$ , where  $n \geq 0$  (Huet, 2009).
+
+$$
+C _ {n} = \frac {1}{n + 1} \binom {2 n} {n} = \frac {(2 n) !}{(n + 1) ! n !}
+$$
+
+where  $C_n$  represents the  $n^{th}$  Catalan number,  $\binom{2n}{n}$  is the binomial coefficient, and! denotes factorial. Finally, the compatibility rules derived from Paninian grammar (Panini, 500 BCE) and contextual information are needed to disambiguate multiple possibilities.
+
+Formally, in a given sentence  $X = \{x_{1}, x_{2}, \ldots, x_{M}\}$  with  $M$  tokens, let  $x_{p}$  (1 ≤  $p \leq M$ ) denote an  $N$ -component compound. It is worth noting that  $X$  may contain multiple instances of multi-component compounds. A valid solution corresponds to a full paraphesization of these  $N$  components; let  $P_{N}$  encompass all valid solutions for fully parenthesizing  $N$  factors, satisfying the cardinality  $|P_{N}| = C_{N-1}$ , where  $C_{N-1}$  represents the Catalan number. A valid solution for an  $N$ -component compound consists of  $N-1$  nested spans. The NeCTI system produces an output represented as a list of  $N-1$  tuples for  $x_{p}$ , given by  $Y_{p} = \{[I_{1}^{H}, I_{1}^{T}, T_{1}], \ldots, [I_{N-1}^{H}, I_{N-1}^{T}, T_{N-1}]\}$ , such that  $Y_{p} \in P_{N}$ .  $I_{i}^{H}$  and  $I_{i}^{T}$  denote the head and the tail indices, respectively, of the  $i^{th}$  span, and  $T_{i}$  corresponds to the label assigned to the respective span.
+
+# How different is NeCTI compared to the Nested Named Entity Recognition (NNER) task?
+
+NNER is a component of information extraction that aims to identify and classify nested named entities within unstructured text, considering their hierarchical structure. In contrast, the NeCTI task focuses on identifying nested spans within a multi-component compound and decoding their implicit semantic relations. These tasks have several key differences: (1) NeCTI operates at the intra-word level, whereas NNER operates at the inter-word level, considering entities across a phrase. (2) In NeCTI, a multi-component compound can have multiple possible parses, requiring disambiguation through contextual cues and incorporating insights from Paninian grammar to address incompatibil
+
+ities. Conversely, we could not find discussions related to these aspects in NNER literature. (3) NeCTI benefits from prior knowledge of the compound's location and segmented components. In contrast, the NNER task involves the additional challenge of identifying the location of entities within the text. Consequently, leveraging existing NNER frameworks for NeCTI is not straightforward due to their inability to provide explicit support for providing the location of compounds. Therefore, NeCTI presents unique characteristics and challenges that differentiate it from the NNER task, requiring specialized approaches tailored to its specific requirements.
+
+How different is NeCTI compared to the Multi-Word Expressions (MWE)? MWEs encompass various categories such as idioms (e.g., kick the bucket), named entities (e.g., World Health Organization), and compounds (e.g., telephone box), etc. Sanskrit compounds exhibit similarities to multi-word expressions, particularly multi-word compounds (nominal, noun, and verb), with respect to characteristics like collocation of components based on semantic relations between them and strict preference for ordering of the components.
+
+Multi-word compounds in various languages involve adjacent lexemes juxtaposed with potential semantic relations. In contrast, Sanskrit compounds are intuitively constructed based on the semantic compatibility of their components. Additionally, in Sanskrit, compounds always appear as a single word, requiring mandatory sandhi (euphonic transformations) between their components. Conversely, certain categories of MWEs (idioms, complex function words, verb-particles, and light-verbals) have relatively fixed structures, predominantly with components separated by spaces. Furthermore, nesting within multi-word expressions is considered syntactic overlaps, while the nested structures of Sanskrit compounds result from successive combinations of components based on semantic relations, thereby clearly distinguishing Sanskrit compounds from MWEs.
+
+# 3 Why NeCTI as a Dependency Parsing Task?
+
+The decision to formulate NeCTI as a dependency parsing task is driven by several considerations. Compounds with more than two components are typically formed through successive binary combinations, following specific semantic relations. This
+
+![](images/a6896d9f66fbadcc2c5aeba9e82b4b253e6f8bb33a856175c4bae5c4aac4d13a.jpg)  
+Figure 4: Illustration of DepNeCTI with an example “sumitra-ananda-vardhanah laksmanah rāmam anujagāma” (Translation: “Lakṣmana, the one who enhanced the delight of Sumitra, followed Rāma”) where ‘sumitra-ananda-vardhanah’ is a multi-component compound word. We assume prior knowledge of compound segmentation and treat the individual components of multi-component compounds as separate words. That means the compound and its components are known apriori. However, the associations of the components, i.e. spans, are not known apriori. We propose two variants of DepNeCTI depending on the choice of encoder: DepNeCTI-LSTM and DepNeCTI-XLMR. To inform DepNeCTI-LSTM about compound (highlighted with color) and non-compound (highlighted with color) tokens, we employ span encoding. The span encoding uses two randomly initialized vectors (compound or non-compound) to inform the model whether the corresponding instance is a compound or non-compound word. On the other hand, DepNeCTI-XLMR is informed about compound's location in the input string using bracketing (for example, <sumitra-ananda-vardhanah>) and it lacks span encoding component. Next, to transform the compound-level parsing task into standard dependency parsing, we introduce (1) an additional token called “Global” (color) as a global head for all words in the sentence. (2) The compound head and non-compound words are connected to the Global token using the “Compound Root” and “Global Relation” relations, respectively. The hidden representations of the tokens are obtained using a Bi-LSTM or XLMR encoder. Finally, a Bi-affine (Dozat and Manning, 2017) dependency module is applied on top of the hidden representations.
+
+process creates a nested structure of binary compounds, except for a few exceptions, where the structure represents a constituency tree. However, treating this as a constituency parsing task poses challenges. The nested structure of compounds does not adhere to a syntactic structure but instead follows a semantic structure based on component relations. If we substitute the intermediate nodes with semantic relations, the same spans can be represented as dependency structures by annotating the types as directed relations. Moreover, the headwords in constituency spans are not explicitly marked but can be identified through their corresponding types. In contrast, dependency structures allow the determination of the headword based on labels directed towards it within the compound. Notably, dependency structures faithfully capture
+
+constituency information and can be mutually converted with their corresponding spans (Goyal and Kulkarni, 2014).
+
+Summarily, the semantic relations among compound components resemble dependency relations, which can be represented as directed labels within the dependency parse structure. This approach succinctly represents the semantic relations without introducing intermediary nodes. Lastly, it enables the simultaneous identification of the structure or constituency span alongside the identification of compound types. We encourage readers to refer to Appendix § A for a more detailed illustration.
+
+# 4 DepNeCTI: The Proposed Framework
+
+Figure 4 illustrates the proposed framework with an example "sumitra-ananda-vardhanah laksmanah
+
+ramanujagama" (Translation: "Laksmana, the one who enhanced the delight of Sumitra, followed Rama") where 'sumitra-ananda-vardhanah' is a multi-component compound word. We assume prior knowledge of compound segmentation and treat the individual components of multicomponent compounds as separate words. We propose two variants of DepNeCTI depending on the choice of encoder: DepNeCTI-LSTM and DepNeCTI-XLMR. In DepNeCTI-LSTM, to differentiate between compound (highlighted with color) and non-compound (highlighted with color) tokens, we employ span encoding. The span encoding uses two randomly initialized vectors (compound or non-compound) to inform the model whether the corresponding instance is a compound or non-compound word. On the other hand, DepNeCTI-XLMR is informed about compound's location in the input string using bracketing (for example, <sumitra-ananda-vardhana>) and it lacks span encoding component. In order to convert the compound-level dependency parsing task into standard dependency parsing, we introduce two modifications. First, we introduce an additional token called "Global" (color) which serves as the global head for all words in the sentence. Second, we establish connections between the compound head and non-compound words to the Global token using the "Compound Root" and "Global Relation" relations, respectively.
+
+Formally, in a given sentence  $X = \{x_{1}, x_{2}, \ldots, x_{M}\}$  with  $M$  tokens, let  $x_{p}$  (1 ≤  $p \leq M$ ) denote an  $N$ -component compound. Notably,  $X$  may contain multiple occurrences of multi-component compounds. The  $N$ -component compound  $(x_{p})$  is further split into its components  $(x_{p} = \{x_{p}^{1}, x_{p}^{2}, \ldots, x_{p}^{N}\})$ . Next, we pass the overall sequence  $(X = \{x_{1}, x_{2}, \ldots, x_{p}^{1}, x_{p}^{2}, \ldots, x_{p}^{N}, \ldots, x_{M}\})$  to the encoder to obtain hidden representations. The LSTM encoder concatenates a token's word, character and span embedding to obtain its representation and the XLMR encoder uses word-pieces. Finally, a Bi-affine (Dozat and Manning, 2017) dependency module is applied on top of it.
+
+# 5 Experiments
+
+# 5.1 Datasets
+
+Table 1 presents data on the total number of multi-component compounds (Figure 5) and their statistics. Our primary focus is on compounds with
+
+more than two components ( $n > 2$ ), while also considering binary compounds if they occur in the context. These datasets comprise segmented compound components, nested spans, context, and semantic relations among the nested spans. We offer two levels of annotations for these datasets: coarse (4 broad types) and fine-grained (86 sub-types). There are 4 broad semantic types of compounds: Avyayibhava (Indeclinable), Bahuvrihi (Exocentric), Tatpurusa (Endocentric) and Dvandva (Copulative). Again, each broader class is divided into multiple subclasses, leading to 86 fine-grained types. $^{1}$  Figure 6 shows class-wise label frequency in NecTIS fine-grained.
+
+<table><tr><td>Datasets</td><td>#Nested</td><td>#Train</td><td>#Test</td><td>#Dev</td><td>#Types</td></tr><tr><td>NeCTIS</td><td>17656</td><td>12431</td><td>3493</td><td>2405</td><td>4 (86)</td></tr><tr><td>NeCTIS OOD</td><td>1189</td><td>-</td><td>1189</td><td>-</td><td>4 (86)</td></tr></table>
+
+Table 1: Data statistics for NeCTIS and NeCTIS-OOD
+
+![](images/c3da73854968cd8851f14fbaeded14e077e1f3a601c48c7634b28e9c6c18ceb8.jpg)  
+Figure 5: Frequency of  $n$ -component compounds.
+
+![](images/41bad8dd6130d2436dd5d2a286774e27e3bdf4ae89273dcf8054357792745a7d.jpg)  
+Figure 6: Frequency in NeCTIS fine-grained.
+
+We introduce two context-sensitive datasets: NeCTIS and NeCTIS OOD. The purpose of the additional dataset (NeCTIS OOD) is to create an out-of-domain testbed. The multi-component compound instances are extracted from various books
+
+categorized into 4 types based on subject content: philosophical, pauranic (Translation: Epic is a genre of ancient Indian literature encompassing historical stories, traditions, and legends.), literary, and ayurveda. The NeCTIS dataset encompasses compounds from books falling under the Philosophical, Literary, and ayurveda categories. The multicomponent compound instances extracted from the puranic category are included in the NeCTIS out-of-domain (NeCTIS OOD) dataset. Furthermore, the multi-component instances in NeCTIS belong to the prose domain, while NeCTIS-OOD pertains to the poetry domain. Poetry commonly uses multi-component compounding extensively (more exocentric compounds) to adhere to metrical constraints and convey complex concepts. Conversely, prose uses compounds in a more direct and less condensed manner. Furthermore, poets in the realm of poetry often enjoy the freedom to form novel compounds or employ unconventional ones to conform to meter requirements, rendering these compounds infrequent in regular usage.
+
+Dataset Annotation Process: We established a data creation process to address the unavailability of annotated context-sensitive multi-component compound data in Sanskrit. We employ a sufficient annotation budget sponsored by DeitY, 2009-2012 for the Sanskrit Hindi Machine Translation project to employ 6 institutes, each consisting of approximately 10 team members. Each team was organized in a hierarchical manner. There were 3 levels in the hierarchy: Junior linguist (Masters degree in Sanskrit), Senior linguist (Ph.D. in Sanskrit) and professional linguist (Professor in Sanskrit). The annotations from lower expertise were further checked as per the above-mentioned hierarchy. Subsequently, the annotated data underwent an exchange process with another team for correctness verification. Any ambiguities encountered during the annotation process were resolved through collective discussions conducted by the correctness-checking team. The available books were distributed among these teams, and each team was responsible for annotating their allocated books. The annotation guidelines<sup>2</sup> are essentially based on Sanskrit grammar which provides the syntactic and semantic criteria for annotation.
+
+Elaborate commentaries accompany the majority
+
+of the texts, that discuss the semantics associated with the compounds, which are typically studied by students as a part of their coursework. Given these considerations, it is very unlikely for professional linguists, often professors instructing these texts, to make mistakes. The dataset was curated around 12 years ago, primarily with the aim of producing error-free gold-standard data. Consequently, the errors made by junior annotators were not recorded or measured, aligning with our focus on achieving error-free quality. The benchmark for determining correctness was based on the Pānīnīan grammar.
+
+# 5.2 Baselines
+
+We investigate the efficacy of various standard formulations (originally proposed for nested named entity recognition for English) for the proposed task. We adapt these systems to the NeCTI task by providing the location of the compounds to ensure a fair comparison with DepNeCTI. Since these baselines are leveraged from the nested named entity recognition task, they do not have explicit channels to provide the location of a compound word. Therefore, we provide this information in the input string itself with the help of brackets (for example, <sumitra-ānanda-vardhanah>):
+
+- Constituency Parsing (CP): Following Fu et al. (2020), we formulate NeCTI as constituency parsing with partially-observed trees, with all labeled compound spans as observed nodes in a constituency tree and non-compound spans as latent nodes. We leverage TreeCRF to model the observed and the latent nodes jointly.  
+- Bottom-up Constituency Parsing (BotCP): Following Yang and Tu (2021), we formulate NeCTI as a bottom-up constituency parsing, where a pointer network is leveraged for post-order traversal within a constituency tree to enhance parsing efficiency, enabling linear order parsing.  
+- Span Classifier (SpanCL): Following Yuan et al. (2022), we formulate NeCTI as a span classification problem, where triaffine mechanism is leveraged to learn a better span representation by integrating factors such as inside tokens, boundaries, labels, and related spans.  
+- Lexicalized Constituency Parsing (LexCP): Following Lou et al. (2022), we formulate NeCTI as lexicalized constituency parsing, which embeds a constituency and a dependency trees
+
+<table><tr><td rowspan="2"></td><td colspan="7">Coarse</td><td colspan="6">Fine Grain</td></tr><tr><td colspan="3">w/o context</td><td colspan="3">w/ context</td><td colspan="3">w/o context</td><td colspan="3">w/ context</td><td></td></tr><tr><td></td><td>Models</td><td>USS</td><td>LSS</td><td>EM</td><td>USS</td><td>LSS</td><td>EM</td><td>USS</td><td>LSS</td><td>EM</td><td>USS</td><td>LSS</td><td>EM</td></tr><tr><td rowspan="7">NeCTIS</td><td>BotCP</td><td>72.90</td><td>58.78</td><td>32.26</td><td>76.22</td><td>63.97</td><td>35.10</td><td>74.28</td><td>33.50</td><td>18.58</td><td>75.72</td><td>41.80</td><td>23.05</td></tr><tr><td>CP</td><td>76.83</td><td>61.71</td><td>39.97</td><td>64.73</td><td>46.27</td><td>30.14</td><td>77.38</td><td>41.86</td><td>27.22</td><td>70.56</td><td>32.26</td><td>21.11</td></tr><tr><td>LexCP</td><td>93.39</td><td>84.74</td><td>72.88</td><td>93.39</td><td>85.16</td><td>74.41</td><td>88.70</td><td>45.86</td><td>14.72</td><td>87.86</td><td>48.87</td><td>19.67</td></tr><tr><td>Seq2seq</td><td>92.54</td><td>84.11</td><td>59.90</td><td>91.18</td><td>80.45</td><td>52.89</td><td>92.67</td><td>65.63</td><td>30.65</td><td>92.94</td><td>68.19</td><td>34.35</td></tr><tr><td>SpanCL</td><td>92.84</td><td>84.80</td><td>69.12</td><td>93.13</td><td>84.74</td><td>69.67</td><td>93.12</td><td>69.38</td><td>52.17</td><td>92.69</td><td>68.05</td><td>50.82</td></tr><tr><td>DepNeCTI-LSTM</td><td>95.46</td><td>89.06</td><td>76.82</td><td>97.42</td><td>89.24</td><td>77.00</td><td>95.38</td><td>79.49</td><td>57.46</td><td>97.49</td><td>79.72</td><td>56.83</td></tr><tr><td>DepNeCTI-XLMR</td><td>96.21</td><td>90.83</td><td>79.85</td><td>96.16</td><td>90.67</td><td>79.45</td><td>96.35</td><td>83.36</td><td>63.92</td><td>96.34</td><td>83.19</td><td>63.30</td></tr><tr><td rowspan="7">NeCTIS OOD</td><td>BotCP</td><td>72.81</td><td>48.08</td><td>21.63</td><td>73.89</td><td>51.18</td><td>22.57</td><td>72.30</td><td>20.52</td><td>8.64</td><td>73.87</td><td>30.10</td><td>13.32</td></tr><tr><td>CP</td><td>71.43</td><td>52.65</td><td>34.57</td><td>68.17</td><td>38.63</td><td>25.15</td><td>76.12</td><td>29.68</td><td>19.77</td><td>69.57</td><td>24.10</td><td>15.43</td></tr><tr><td>LexCP</td><td>89.90</td><td>71.44</td><td>50.00</td><td>91.57</td><td>72.60</td><td>52.73</td><td>83.45</td><td>32.85</td><td>7.53</td><td>83.38</td><td>34.18</td><td>8.31</td></tr><tr><td>Seq2Seq</td><td>84.26</td><td>71.71</td><td>45.33</td><td>90.13</td><td>71.41</td><td>44.00</td><td>92.16</td><td>53.55</td><td>24.51</td><td>92.87</td><td>53.61</td><td>24.96</td></tr><tr><td>SpanCL</td><td>91.51</td><td>72.69</td><td>56.30</td><td>90.24</td><td>71.89</td><td>54.46</td><td>89.19</td><td>50.21</td><td>32.66</td><td>90.88</td><td>50.50</td><td>34.00</td></tr><tr><td>DepNeCTI-LSTM</td><td>93.32</td><td>78.94</td><td>57.40</td><td>95.67</td><td>79.26</td><td>57.90</td><td>93.88</td><td>67.96</td><td>36.60</td><td>95.88</td><td>67.26</td><td>36.26</td></tr><tr><td>DepNeCTI-XLMR</td><td>95.56</td><td>84.24</td><td>65.50</td><td>95.56</td><td>84.45</td><td>65.00</td><td>95.56</td><td>74.26</td><td>45.70</td><td>95.45</td><td>73.54</td><td>44.37</td></tr></table>
+
+Table 2: Evaluation on the NeCTIS datasets, considering 2 levels of annotations (coarse and fine-grained) and in 2 settings (with and without context). The best-performing results are in bold, while the second-best results are underlined. The results are averaged over 3 runs. To assess the significance between the proposed system and the best baselines for each setting, a significance test in Accuracy metrics was conducted:  $p < 0.01$  (as per t-test).
+
+together. This formulation leverages the constituents' heads into the architecture which is crucial for the NeCTI task.
+
+- Seq2Seq: Following Yan et al. (2021), we formulate NeCTI as an entity span sequence generation task using the pretrained seq2seq framework. This generative framework can also identify discontinuous spans; however, NeCTI does not have such instances.  
+- DepNeCTI: We propose two variants of our system (§4) depending on the choice of encoders: DepNeCTI-LSTM and DepNeCTI-XLMR (Nguyen et al., 2021). The compound's location is provided in DepNeCTI-LSTM using span encoding; however, DepNeCTI-XLMR lacks span encoding and leverages this information from the input similar to other baselines.
+
+Evaluation Metrics: We evaluate the performance using the Labeled/Unlabeled Span Score (LSS/USS) in terms of micro-averaged F1-score. We define LSS as a micro-averaged F1-score applied on tuples of predicted spans including their labels. We exclude labels of the spans while calculating USS. Additionally, we report the exact match (EM) which indicates the percentage of the compounds for which the predictions of all spans and their semantic relations are correctly identified. Refer to Appendix B for hyper-parameters and details of the computing infrastructure used.
+
+# 5.3 Results
+
+Table 2 presents the performance of the top-performing configurations among all baselines on the NeCTIS benchmark datasets' test set. The evaluation includes 2 levels of annotations (coarse and fine-grained) and in 2 settings (with and without context). While all baseline systems demonstrate competitiveness, no single baseline consistently outperforms the others across all settings. Consequently, we underline the best-performing numbers within each specific setting.
+
+Our proposed system DepNeCTI surpasses all competing systems across all evaluation metrics, demonstrating an absolute average gain of 13.1 points (LSS) and 11.3 points (EM) compared to the best-performing baseline in each setting. Notably, our proposed system exhibits substantial performance superiority over the best baseline in fine-grained settings, particularly in low-resourced scenarios. This validates the effectiveness of our proposed system in low-resourced settings with fine-grained labels. The significant performance gap between our proposed system and the best baselines highlights the efficacy of employing a simple yet effective architecture inspired by the dependency parsing paradigm. These results establish new state-of-the-art benchmarks by integrating the contextual component into our novel framework. While most baselines (except BotCP) do not benefit from contextual information, DepNeCTI-LSTM
+
+demonstrates slight improvements and DepNeCTI-XLMR shows on par improvements when leveraging contextual information. Furthermore, as the number of components grows, the number of potential solutions increases exponentially, leading to poor performance by the systems in such scenarios. Due to this exponential possibility, contextual information provides limited improvements compared to binary compound identification settings (Sandhan et al., 2022a). In other words, unlike the context-free setting, the introduction of context information does not warrant an expectation for the system to precisely generate the correct solution from the exponential candidate space. Figure 3 provides a visual representation that elucidates the concept of this exponential candidate space. A similar performance trend is observed for the NeCTIS-OOD dataset.
+
+# 6 Analysis
+
+Here, we examine the proposed system, focusing on a comprehensive analysis and its applicability. For this purpose, we utilize the NeCTIS coarse dataset under the w/ context configuration. We report LSS in terms of macro-average F1-scores.
+
+(1) Ablation analysis: In this study, we analyze the impact of different system components on the overall enhancements of DepNeCTI-LSTM. Ablations, documented in Table 3, present the evaluation metrics when a specific component is deactivated within DepNeCTI-LSTM. For instance, the absence of the span encoding component is denoted as “- Span Encoding”. The results indicate that removing any component leads to a decline in performance. Notably, Table 3 highlights the significance of the “Span Encoding” component in driving the improvements.
+
+<table><tr><td>System</td><td>P</td><td>R</td><td>F1</td><td>EM</td></tr><tr><td>DepNeCTI-LSTM</td><td>89.24</td><td>89.24</td><td>89.24</td><td>77.00</td></tr><tr><td>- FastText (FT)</td><td>88.84</td><td>88.84</td><td>88.84</td><td>76.50</td></tr><tr><td>- Span Encoding (SE)</td><td>86.18</td><td>85.53</td><td>86.85</td><td>70.54</td></tr><tr><td>- FT - SE</td><td>84.25</td><td>84.23</td><td>84.24</td><td>67.86</td></tr></table>
+
+Table 3: Each ablation involved the removal of a singular component from DepNeCTI-LSTM. The ablation denoted as “-Span Encoding” entailed eliminating the span encoding component from the proposed system.
+
+(2) How robust is the system when the number of components of a compound increases? We analyze the relationship between the F1-score and
+
+the number of components in compounds. For compounds with a small number of components, all systems demonstrate high performance, but our proposed systems consistently outperforms other baselines. However, as the number of components increases, the number of examples in each category decreases. Additionally, the number of potential solutions grows exponentially, following the Catalan number. Consequently, all systems experience a decline in performance.
+
+![](images/5ff8473746bb9318bf3a0541b7d8f58d627dc7872f6d44aa8d9c26fae00d945a.jpg)  
+Figure 7: F1-Score against the number of components. The compounds with components  $N > 10$  are excluded.
+
+(3) Error analysis: We investigate whether all the systems are able to identify the location of a multi-component compound correctly. The motivation behind this experiment is to evaluate the capability of the baselines and the proposed architecture to leverage the information about the compound's location effectively. We define the span of text that corresponds to a compound as a global span which we know apriori. Figure 8 illustrates the effectiveness of each system in correctly identifying the global span of multi-component compounds. In
+
+![](images/cd62abd30e89577376aaf5b03b1a6f1bc475b8e081afdc231261c5500e1295ad.jpg)  
+Figure 8: Performance of the systems on identifying the global span of a compound.
+
+DepNeCTI-LSTM, our span encoding effectively captures this information, resulting in a perfect  $100\%$  score. However, even after providing the baselines with this information, they fail to use it due to limitations in their architectures. Interestingly, DepNeCTI-XLMR does not contain a span
+
+encoding component and leverages the compound's location information from the input as provided for the baselines. Still, DepNeCTI-XLMR reports the best performance due to its powerful word representation ability. It is worth noting that the NeC-TIS dataset exhibits an inherent bias towards left branching, as indicated by the nested tree structure. Consequently, all systems display a bias towards left branching as well. Therefore, due to the dominance of left-branching instances and increased variance due to less number of instances, a spike is observed in the results.
+
+(4) Efficiency of our proposed system: Figure 9 present the computational efficiency of our system measured in terms of the number of sentences processed per second. We compare the inference speed of different baselines on the NeCTI task. Our systems, DepNeCTI-LSTM and DepNeCTI-XLMR are leveraging a simple architecture and utilizing dependency parsing as an appropriate problem formulation, exhibits a 5-fold/3-fold improvement over the most efficient baseline, BotCP.
+
+![](images/90fe607bd03136c8e1c07b23fb267912538017d39f1d3f11633d60ab90874443.jpg)  
+Figure 9: Inference speed of the competing systems.
+
+# 7 Related Works
+
+Lexical semantics is a dedicated field focused on word meaning. Various tasks such as word-sense disambiguation (Bevilacqua and Navigli, 2020; Barba et al., 2021; Maurya et al., 2023; Maurya and Bahadur, 2022), relationship extraction (Tian et al., 2021; Nadgeri et al., 2021; Hu et al., 2022; Wang et al., 2022), and semantic role labeling (He et al., 2018; Kulkarni and Sharma, 2019; Zhang et al., 2022) play essential roles in determining word meaning. Moreover, when dealing with complex word structures such as compounds, named entities, and multi-word expressions, relying solely on basic word senses and relationships is inadequate. While efforts have been made in Noun Compound Identification (Ziering and van der Plas,
+
+2015; Dima and Hinrichs, 2015; Fares et al., 2018; Shwartz and Waterson, 2018; Ponkiya et al., 2018, 2020, 2021), multi-word expression (MWE) (Constant et al., 2017; Gharbieh et al., 2017; Gooding et al., 2020; Premasiri and Ranasinghe, 2022) and named entity recognition (Fu et al., 2020; Yang and Tu, 2021; Lou et al., 2022; Yuan et al., 2022), the nested compounds remains unexplored.
+
+Sanskrit Compound Type Identification has attracted significant attention over the past decade. Decoding the meaning of a Sanskrit compound requires determining its constituents (Huet, 2009; Mittal, 2010; Hellwig and Nehrdich, 2018), understanding how these constituents are grouped (Kulkarni and Kumar, 2011), identifying the semantic relationship between them (Kumar, 2012), and ultimately generating a paraphrase of the compound (Kumar et al., 2009). Previous studies proposed rule-based approaches (Satuluri and Kulkarni, 2013; Kulkarni and Kumar, 2013), a data-driven approach (Sandhan et al., 2019) and a hybrid approach (Krishna et al., 2016) for SaCTI. Sandhan et al. (2022a) proposed a context-sensitive architecture for binary compounds.
+
+Earlier works in Sanskrit solely focused on binary compounds, neglecting the identification of multi-component compound types; however, our proposed framework fills this research gap.
+
+# 8 Conclusion
+
+In this work, we focused on multi-component compounding in Sanskrit, which helps to decode the implicit structure of a compound to decipher its meaning. While previous approaches have primarily focused on binary compounds, we introduced a novel task called nested compound type identification (NeCTI). This task aims to identify nested spans within multi-component compounds and decode the implicit semantic relations between them, filling a gap in the field of lexical semantics. To facilitate research in this area, we created 2 newly annotated datasets, designed explicitly for the NeCTI task. These datasets were utilized to benchmark various problem formulations. Our novel framework DepNeCTI outperformed the best baseline system by achieving a stupendous absolute gain of 13.1 points F1-score in terms of LSS. Similar to the previous findings on binary Sanskrit compound identification, we discovered that our proposed system exhibits substantial performance superiority over the best baseline in low-resourced scenarios.
+
+# Limitations
+
+We could not extend our framework to other languages exhibiting multi-component compounding phenomena due to the lack of availability of annotated datasets. It would be interesting to measure the effectiveness of rules from Pānīnīan grammar to discard incompatible semantic relations (Kulkarni, 2019, 2021).
+
+# Ethics Statement
+
+This work introduces a new task, along with annotated datasets and a framework, to address the nested compounding phenomenon in Sanskrit. The proposed resources aim to enhance the understanding of multi-component compounds and contribute to the improvement of machine translation systems. Regarding potential effects, we anticipate no harm to any community resulting from the use of our datasets and framework. However, we advise users to exercise caution, as our system is not flawless and may generate mispredictions. To ensure transparency and future research, we have publicly released all our annotated NeCTIS datasets and source codes. We confirm that our data collection adheres to the terms of use of the sources and respects the intellectual property and privacy rights of the original authors. Our annotation team consisted of qualified individuals, including Master's and Ph.D. degree holders, some of whom are Sanskrit professors. Annotators were compensated appropriately and provided with detailed instructions to ensure consistency in the annotation process. We remain committed to addressing ethical implications as we refine our systems and welcome feedback from the community to enhance our ethical practices.
+
+# Acknowledgements
+
+We thank all the annotators from different institutes for helping us with NeCTIS data annotation. We would like to thank Zheng Yuan, Tsinghua University (Alibaba Group), for helping us with the hyperparameter section and adapting his system for the NeCTI task. We appreciate and thank all the anonymous reviewers for their constructive feedback towards improving this work. The work was supported in part by the National Language Translation Mission (NLTM): Bhashini project by Government of India.
+
+# References
+
+Rahul Aralikatte, Miryam de Lhoneux, Anoop Kunchukuttan, and Anders Søgaard. 2021. Itihasa: A large-scale corpus for Sanskrit to English translation. In Proceedings of the 8th Workshop on Asian Translation (WAT2021), pages 191-197, Online. Association for Computational Linguistics.  
+Edoardo Barba, Luigi Procopio, and Roberto Navigli. 2021. ConSeC: Word sense disambiguation as continuous sense comprehension. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1492-1503, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Michele Bevilacqua and Roberto Navigli. 2020. Breaking through the  $80\%$  glass ceiling: Raising the state of the art in word sense disambiguation by incorporating knowledge graph information. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2854-2864, Online. Association for Computational Linguistics.  
+Mathieu Constant, Gülşen Eryiğit, Johanna Monti, Lonneke van der Plas, Carlos Ramisch, Michael Rosner, and Amalia Todirascu. 2017. Survey: Multiword expression processing: A Survey. Computational Linguistics, 43(4):837-892.  
+Corina Dima and Erhard Hinrichs. 2015. Automatic noun compound interpretation using deep neural networks and word embeddings. In Proceedings of the 11th International Conference on Computational Semantics, pages 173-183, London, UK. Association for Computational Linguistics.  
+Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency parsing.  
+Murhaf Fares, Stephan Oepen, and Erik Velldal. 2018. Transfer and multi-task learning for noun-noun compound interpretation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1488-1498, Brussels, Belgium. Association for Computational Linguistics.  
+Yao Fu, Chuanqi Tan, Mosha Chen, Songfang Huang, and Fei Huang. 2020. Nested named entity recognition with partially-observed treecrfs.  
+Waseem Gharbieh, Virendrakumar Bhavsar, and Paul Cook. 2017. Deep learning models for multiword expression identification. In Proceedings of the 6th Joint Conference on Lexical and Computational Semantics (*SEM* 2017), pages 54-64, Vancouver, Canada. Association for Computational Linguistics.  
+Sian Gooding, Shiva Taslimipoor, and Ekaterina Kochmar. 2020. Incorporating multiword expressions in phrase complexity estimation. In Proceedings of the 1st Workshop on Tools and Resources to Empower People with REAding Difficulties (READI), pages 14-19, Marseille, France. European Language Resources Association.
+
+Pawan Goyal and Gerard Huet. 2016. Design and analysis of a lean interface for sanskrit corpus annotation. Journal of Language Modelling, 4:145.  
+Pawan Goyal and Amba Kulkarni. 2014. Converting phrase structures to dependency structures in Sanskrit. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1834-1843, Dublin, Ireland. Dublin City University and Association for Computational Linguistics.  
+Luheng He, Kenton Lee, Omer Levy, and Luke Zettlemoyer. 2018. Jointly predicting predicates and arguments in neural semantic role labeling. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 364-369, Melbourne, Australia. Association for Computational Linguistics.  
+Oliver Hellwig and Sebastian Nehrdich. 2018. Sanskrit word segmentation using character-level recurrent and convolutional neural networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2754-2763, Brussels, Belgium. Association for Computational Linguistics.  
+Chengwei Hu, Deqing Yang, Haoliang Jin, Zhen Chen, and Yanghua Xiao. 2022. Improving continual relation extraction through prototypical contrastive learning. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1885-1895, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Gérard Huet. 2009. Sanskrit segmentation. In Proceedings of the South Asian Languages Analysis Roundtable XXVIII, Denton, Ohio.  
+Amrith Krishna, Pavankumar Satuluri, Shubham Sharma, Apurv Kumar, and Pawan Goyal. 2016. Compound type identification in Sanskrit: What roles do the corpus and grammar play? In Proceedings of the 6th Workshop on South and Southeast Asian Natural Language Processing (WSSANLP2016), pages 1-10, Osaka, Japan. The COLING 2016 Organizing Committee.  
+Amba Kulkarni. 2019. Sanskrit Parsing based on the Theories of Shabdabodha. D. K. PrintWorld and Indian Institute of Advanced Study.  
+Amba Kulkarni. 2021. Sanskrit parsing following indian theories of verbal cognition. In ACM Transactions on Asian and Low-Resource Language Information Processing, pages 1-38.  
+Amba Kulkarni and Anil Kumar. 2011. Statistical constituency parser for sanskrit compounds. In Proceedings of ICON-2011: 9th International Conference on Natural Language Processing, International Institute of Information Technology, Hyderabad, India. Macmillan Publishers.
+
+Amba Kulkarni and Anil Kumar. 2013. Clues from astadhyay for compound type identification. In Proceedings of the 5th International Sanskrit Computational Linguistics Symposium, IIT Bombay, India. D. K. PrintWorld and Indian Institute of Advanced Study.  
+Amba Kulkarni and Dipti Sharma. 2019. Panjinian syntactico-semantic relation labels. In Proceedings of the Fifth International Conference on Dependency Linguistics (Depling, SyntaxFest 2019), pages 198-208, Paris, France. Association for Computational Linguistics.  
+Anil Kumar. 2012. An automatic Sanskrit compound processing. Ph.D. thesis, University of Hyderabad.  
+Anil Kumar, V Sheeba, and Amba Kulkarni. 2009. Sanskrit compound paraphrase generator. In Proceedings of the ICON 2009: 7th International Conference on Natural Language Processing, University of Hyderabad, Hyderabad. Macmillan Publishers.  
+Mirella Lapata and Frank Keller. 2004. The web as a baseline: Evaluating the performance of unsupervised web-based models for a range of NLP tasks. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pages 121-128, Boston, Massachusetts, USA. Association for Computational Linguistics.  
+Chao Lou, Songlin Yang, and Kewei Tu. 2022. Nested named entity recognition as latent lexicalized constituency parsing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6183-6198, Dublin, Ireland. Association for Computational Linguistics.  
+Xuezhe Ma, Zeong Hu, Jingzhou Liu, Nanyun Peng, Graham Neubig, and Eduard Hovy. 2018. Stackpointer networks for dependency parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1403-1414, Melbourne, Australia. Association for Computational Linguistics.  
+Archana Maurya and Promila Bahadur. 2022. *A Detailed Analysis of Word Sense Disambiguation Algorithms and Approaches for Indian Languages*, pages 693-710.  
+Archana Sachindeo Maurya, Promila Bahadur, and Srishti Garg. 2023. Approach toward word sense disambiguation for the english-to-sanskrit language using naive bayesian classification. In Proceedings of Third Doctoral Symposium on Computational Intelligence, pages 477-491, Singapore. Springer Nature Singapore.  
+Vipul Mittal. 2010. Automatic Sanskrit segmentizer using finite state transducers. In Proceedings of the
+
+ACL 2010 Student Research Workshop, pages 85-90, Uppsala, Sweden. Association for Computational Linguistics.  
+Abhishek Nadgeri, Anson Bastos, Kuldeep Singh, Isaiah Onando Mulang', Johannes Hoffart, Saeedeh Shekarpour, and Vijay Saraswat. 2021. KGPool: Dynamic knowledge graph context selection for relation extraction. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 535-548, Online. Association for Computational Linguistics.  
+Minh Van Nguyen, Viet Duc Lai, Amir Pouran Ben Vey-seh, and Thien Huu Nguyen. 2021. Trankit: A lightweight transformer-based toolkit for multilingual natural language processing. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 80-90, Online. Association for Computational Linguistics.  
+Panini. 500 BCE. Aştadhyayi.  
+Girishkumar Ponkiya, Diptesh Kanojia, Pushpak Bhattacharyya, and Girish Palshikar. 2021. FrameNet-assisted noun compound interpretation. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2901-2911, Online. Association for Computational Linguistics.  
+Girishkumar Ponkiya, Rudra Murthy, Pushpak Bhattacharyya, and Girish Palshikar. 2020. Looking inside noun compounds: Unsupervised prepositional and free paraphrasing. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4313-4323, Online. Association for Computational Linguistics.  
+Girishkumar Ponkiya, Kevin Patel, Pushpak Bhattacharyya, and Girish Palshikar. 2018. Treat us like the sequences we are: Prepositional paraphrasing of noun compounds using LSTM. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1827-1836, Santa Fe, New Mexico, USA. Association for Computational Linguistics.  
+Damith Premasiri and Tharindu Ranasinghe. 2022. Bert(s) to detect multiword expressions.  
+Jivnesh Sandhan, Ashish Gupta, Hrishikesh Terdalkar, Tushar Sandhan, Suvendu Samanta, Laxmidhar Behera, and Pawan Goyal. 2022a. A novel multi-task learning approach for context-sensitive compound type identification in Sanskrit. In Proceedings of the 29th International Conference on Computational Linguistics, pages 4071-4083, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Jivnesh Sandhan, Amrith Krishna, Pawan Goyal, and Laxmidhar Behera. 2019. Revisiting the role of feature engineering for compound type identification in Sanskrit. In Proceedings of the 6th International Sanskrit Computational Linguistics Symposium, pages
+
+28-44, IIT Kharagpur, India. Association for Computational Linguistics.  
+Jivnesh Sandhan, Rathin Singha, Narein Rao, Suvendu Samanta, Laxmidhar Behera, and Pawan Goyal. 2022b. TransLIST: A transformer-based linguistically informed Sanskrit tokenizer. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 6902-6912, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Pavankumar Satuluri and Amba Kulkarni. 2013. Generation of sanskrit compounds. In Proceedings of ICON-2013: 10th International Conference on Natural Language Processing, CDAC Noida, India.  
+Vered Shwartz and Chris Waterson. 2018. Olive oil is made of olives, baby oil is made for babies: Interpreting noun compounds using paraphrases in a neural model. 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 218-224, New Orleans, Louisiana. Association for Computational Linguistics.  
+Hrishikesh Terdalkar and Arnab Bhattacharya. 2019. Framework for question-answering in Sanskrit through automated construction of knowledge graphs. In Proceedings of the 6th International Sanskrit Computational Linguistics Symposium, pages 97-116, IIT Kharagpur, India. Association for Computational Linguistics.  
+Yuanhe Tian, Guimin Chen, Yan Song, and Xiang Wan. 2021. Dependency-driven relation extraction with attentive graph convolutional networks. 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 4458-4471, Online. Association for Computational Linguistics.  
+Peiyi Wang, Yifan Song, Tianyu Liu, Rundong Gao, Binghuai Lin, Yunbo Cao, and Zhifang Sui. 2022. Less is more: Rethinking state-of-the-art continual relation extraction models with a frustratingly easy but effective approach.  
+Hang Yan, Tao Gui, Junqi Dai, Qipeng Guo, Zheng Zhang, and Xipeng Qiu. 2021. A unified generative framework for various ner subtasks.  
+Songlin Yang and Kewei Tu. 2021. Bottom-up constituency parsing and nested named entity recognition with pointer networks.  
+Zheng Yuan, Chuanqi Tan, Songfang Huang, and Fei Huang. 2022. Fusing heterogeneous factors with triaffine mechanism for nested named entity recognition.  
+Li Zhang, Ishan Jindal, and Yunyao Li. 2022. Label definitions improve semantic role labeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational
+
+Linguistics: Human Language Technologies, pages 5613-5620, Seattle, United States. Association for Computational Linguistics.
+
+Patrick Ziering and Lonneke van der Plas. 2015. From a distance: Using cross-lingual word alignments for noun compound bracketing. In Proceedings of the 11th International Conference on Computational Semantics, pages 82–87, London, UK. Association for Computational Linguistics.
+
+# A Why NeCTI as a Dependency Parsing Task?
+
+Compounds in language are semantic constructions. While a limited number of rules derived from Sanskrit Grammar aid in determining the syntactic structure of a compound, they offer limited assistance in uncovering its meaning. The meaning of a compound primarily arises from the semantic relationship between its components. This semantic relationship, known as "sāmarthya," is expressed through various types of semantic compounds. These compound types also facilitate the identification of the headword within a compound. The headword can be one of the constituents or an entirely distinct word modified by the resulting compound. Compounds consisting of more than two components are typically formed through successive binary combinations of the components, following specific relations. As a result, a nested structure of binary compounds is created, with a few exceptions. Therefore, the semantic compound types play a vital role in determining the correct nested structure amidst the various possible structures of a compound. Each of these structures represents a constituency span, and a parser equipped with compound type identification assists in accurately identifying the appropriate constituency span.
+
+Treating this as a constituency parsing task presents several challenges. Figure 10 illustrates the potential constituency spans for a 3-component compound  $a - b - c$ , with  $ab$  and  $bc$  representing the intermediate compounds. First, the nested structure of compounds does not conform to a syntactic structure; instead, it follows a semantic structure based on the relations between the components rather than their position or relative co-occurrence. Second, the intermediate compounds within the structure are not categories but stem forms. Each intermediate compound serves as an entity that modifies the meanings of one of its constituents, subsequently combining with another component to form a larger compound.
+
+Kumar (2012) formulated this as constituency parsing, where the identification of semantic compatibility was performed based on the relative co-occurrence and relative position of the components. As shown in Figure 11, the intermediate compounds, which lacked additional information, were substituted with their respective semantic compound types in the type identification stage.
+
+![](images/14e82fef0833ca1f709f59be7d00163ebf433fb6e7e98af7148eb59c6227e1fb.jpg)  
+Figure 10: Possible constituency spans for a three component compound a-b-c
+
+![](images/e502b72898c158a349061a2bcc1720ef5361e89c3c12c3fe80521a95350d3874.jpg)
+
+![](images/1df216ff9dac45c83cb764f96742822e5d44217f5e4110f5932a60083062abb9.jpg)  
+Figure 11: Possible constituency spans for a three component compound a-b-c with semantic types
+
+![](images/17add7d243851a830a55f970ac61f99249d78a125aa26f9e94a6e0c8a336c68e.jpg)  
+Figure 12: Dependency Structure for the three component compound a-b-c with semantic types as labels
+
+The semantic types serve as semantic constructions for compounds, rather than being syntactic categories and play a crucial role in determining the meaning of the compound. For example the compound vidyālayaghanṭa has the nested structure «vidyā-alaya>T6-ghanṭa>T6. The type T6 indicatesṣaṭṭi tatpurusa compound (6<sup>th</sup> case determinative compound) inferring a possessive relationship. Vidyā (knowledge) and alaya (place) combine to form the intermediate compound vidyālaya (the place of knowledge, viz. school), which combines with ghantā (bell) to form the whole compound indicating school-bell. The possessive relation expressed by the compound type (T6) is akin to a dependency relation, and this holds true for other compound types as well.
+
+Consequently, when represented with compound types, the same constituency spans can be depicted as dependency structures by annotating the types as directed relations and removing the extra nodes indicating the types. The dependency structures for the previous example  $(a - b - c)$  are shown in Figure 12. Furthermore, the head words are not explicitly marked in constituency spans and can only be identified through their corresponding types. However, with dependency structures, the
+
+head word can be determined by the labels directed towards it within the compound. Notably, these dependency structures faithfully capture the constituency information and are mutually convertible with their corresponding constituency spans.
+
+![](images/f34917bff55776dfcbb154e24bed17fdc8c1bfca679902f6ce608ff56da5175e.jpg)
+
+![](images/e5b123b04460721ef6649a19a35d1f2f6ca86990d7f76dab5d7a84d1f559bf49.jpg)
+
+There are several considerations behind the decision to treat NeCTI as a dependency parsing task. First, the semantic relations among compound components resemble dependency relations and can be represented as directed labels within the dependency parse structure. Second, this approach concisely represents the semantic relations between compound components without introducing extra intermediary nodes. Lastly, it enables the simultaneous identification of the structure or constituency span alongside the identification of compound types.
+
+# B Experiment Details
+
+Hyper-parameters: For our proposed system, we build on the top of the codebase from BiAFF, as developed by Ma et al. (2018). We configure the hyperparameters as follows: a batch size of 16, 100 training iterations, a dropout rate of 0.33, 2 stacked Bi-LSTM layers, a learning rate of 0.002, and the remaining parameters set identically to those used in the work of Ma et al. (2018). We use manual tuning for the hyper-parameter selection and F1-score criteria on dev set's performance. Our codebase is publicly available and released under a creative-common license. We use FastText word embeddings for the proposed framework.
+
+Computing Infrastructure Used: We perform our experiments using a single GPU equipped with an NVIDIA A40, 48 GB GPU memory, and 10752 CUDA cores. We employ a single GPU with an NVIDIA Quadro RTX 4000, 8 GB GPU memory, and 2304 CUDA cores for our proposed system.

2023/DepNeCTI_ Dependency-based Nested Compound Type Identification for Sanskrit/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:367b7a4605ebee450179325e6c63f506c8d71aa3ad58af1242b3c27ec7b444fb
+size 453767

2023/DepWiGNN_ A Depth-wise Graph Neural Network for Multi-hop Spatial Reasoning in Text/82978807-d92d-4dd8-83a8-263f59cfaa61_content_list.json

@@ -0,0 +1,1753 @@
+[
+    {
+        "type": "text",
+        "text": "DepWiGNN: A Depth-wise Graph Neural Network for Multi-hop Spatial Reasoning in Text *",
+        "text_level": 1,
+        "bbox": [
+            121,
+            89,
+            875,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Shuaiyi Li $^{1}$ , Yang Deng $^{2,\\dagger}$ , Wai Lam $^{1}$",
+        "bbox": [
+            337,
+            154,
+            658,
+            171
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup> The Chinese University of Hong Kong, <sup>2</sup> National University of Singapore li.shuaiyi@link.cuhk.edu.hk, ydeng@nus.edu.sg, wlam@se.cuhk.edu.hk",
+        "bbox": [
+            174,
+            171,
+            825,
+            204
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Spatial reasoning in text plays a crucial role in various real-world applications. Existing approaches for spatial reasoning typically infer spatial relations from pure text, which overlook the gap between natural language and symbolic structures. Graph neural networks (GNNs) have showcased exceptional proficiency in inducing and aggregating symbolic structures. However, classical GNNs face challenges in handling multi-hop spatial reasoning due to the over-smoothing issue, i.e., the performance decreases substantially as the number of graph layers increases. To cope with these challenges, we propose a novel Depth-Wise Graph Neural Network (DepWiGNN). Specifically, we design a novel node memory scheme and aggregate the information over the depth dimension instead of the breadth dimension of the graph, which empowers the ability to collect long dependencies without stacking multiple layers. Experimental results on two challenging multi-hop spatial reasoning datasets show that DepWiGNN outperforms existing spatial reasoning methods. The comparisons with the other three GNNs further demonstrate its superiority in capturing long dependency in the graph.",
+        "bbox": [
+            141,
+            278,
+            460,
+            649
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            659,
+            258,
+            673
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Spatial reasoning in text is crucial and indispensable in many areas, e.g., medical domain (Datta et al., 2020; Massa et al., 2015), navigations (Zhang et al., 2021; Zhang and Kordjamshidi, 2022; Chen et al., 2019) and robotics (Luo et al., 2023; Venkatesh et al., 2021). It has been demonstrated to be a challenging problem for both modern pretrained language models (PLMs) (Mirzaee et al., 2021; Deng et al., 2023a) and large language models (LLMs) like ChatGPT (Bang et al., 2023). However, early textual spatial reasoning datasets, e.g.,",
+        "bbox": [
+            112,
+            684,
+            489,
+            860
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/93f34138c5657851645495012225cbbb2be40164e2ca36b03e02f56c190b8c8c.jpg",
+        "image_caption": [
+            "Figure 1: An example of multihop spatial reasoning in text from the StepGame dataset (Shi et al., 2022)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            260,
+            884,
+            395
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "bAbI (Weston et al., 2016), suffer from the issue of over-simplicity and, therefore, is not qualified for revealing real textual spatial reasoning scenario. Recently, researchers propose several new benchmarks (Shi et al., 2022; Mirzaee and Kordjamshidi, 2022; Mirzaee et al., 2021) with an increased level of complexity, which involve more required reasoning steps, enhanced variety of spatial relation expression, and more. As shown in Figure 1, 4 steps of reasoning are required to answer the question, and spatial relation descriptions and categories are diverse.",
+        "bbox": [
+            507,
+            451,
+            884,
+            643
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To tackle the problem of multi-hop spatial reasoning, Shi et al. (2022) propose a recurrent memory network based on Tensor Product Representation (TPR) (Schlag and Schmidhuber, 2018a), which mimics the step-by-step reasoning by iteratively updating and removing episodes from memory. Specifically, TPR encodes symbolic knowledge hidden in natural language into distributed vector space to be used for deductive reasoning. Despite the effectiveness of applying TPR memory, the performance of this model is overwhelmed by modern PLMs (Mirzaee and Kordjamshidi, 2022). Moreover, these works typically overlook the gap between natural language and symbolic relations.",
+        "bbox": [
+            507,
+            645,
+            884,
+            869
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Graph Neural Neural Networks (GNNs) have been considerably used in multi-hop reasoning (Xu et al., 2021b; Chen et al., 2020b; Qiu et al., 2019).",
+        "bbox": [
+            507,
+            870,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*This work is substantially supported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Code: 14200719)",
+        "bbox": [
+            112,
+            866,
+            487,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author.",
+        "bbox": [
+            142,
+            904,
+            290,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "6459",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6459-6471 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "These methods often treat a single graph convolutional layer of node information propagation (node to its immediate neighbors) in GNN as one step of reasoning and expand it to multi-hop reasoning by stacking multiple layers. However, increasing the number of graph convolutional layers in deep neural structures can have a detrimental effect on model performance (Li et al., 2018). This phenomenon, known as the over-smoothing problem, occurs because each layer of graph convolutions causes adjacent nodes to become more similar to each other. This paradox poses a challenge for multi-hop reasoning: although multiple layers are needed to capture multi-hop dependencies, implementing them can fail to capture these dependencies due to the over-smoothing problem. Furthermore, many chain-finding problems, e.g., multihop spatial reasoning, only require specific depth path information to solve a single question and do not demand full breadth aggregation for all neighbors (Figure 1). Nevertheless, existing methods (Palm et al., 2018; Xu et al., 2021b; Chen et al., 2020b; Deng et al., 2022) for solving this kind of problem usually lie in the propagation conducted by iterations of breadth aggregation, which brings superfluous and irrelevant information that may distract the model from the key information.",
+        "bbox": [
+            115,
+            84,
+            490,
+            517
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In light of these challenges, we propose a novel graph-based method, named Depth-Wise Graph Neural Network (DepWiGNN), which operates over the depth instead of the breadth dimension of the graph to tackle the multi-hop spatial reasoning problem. It introduces a novel node memory implementation that only stores depth path information between nodes by applying the TPR technique. Specifically, it first initializes the node memory by filling the atomic information (spatial relation) between each pair of directly connected nodes, and then collects the relation between two indirectly connected nodes via depth-wisely retrieving and aggregating all atomic spatial relations reserved in the memory of each node in the path. The collected long-dependency information is further stored in the source node memory in the path and can be retrieved freely if the target node is given. Unlike typical existing GNNs (Morris et al., 2019; Velickovic et al., 2017; Hamilton et al., 2017; Kipf and Welling, 2017), DepWiGNN does not need to be stacked to gain long relationships between two distant nodes and, hence, is immune to the over-smoothing problem. Moreover, instead of",
+        "bbox": [
+            115,
+            533,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "aimlessly performing breadth aggregation on all immediate neighbors, it selectively prioritizes the key path information.",
+        "bbox": [
+            512,
+            84,
+            880,
+            131
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Experiments on two challenging multi-hop spatial reasoning datasets show that DepWiGNN not only outperforms existing spatial reasoning methods, but also enhances the spatial reasoning capability of PLMs. The comparisons with three GNNs verify that DepWiGNN surpasses classical graph convolutional layers in capturing long dependencies by a noticeable margin without harming the performance of short dependency collection.",
+        "bbox": [
+            512,
+            133,
+            882,
+            277
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Overall, our contributions are threefold:",
+        "bbox": [
+            529,
+            278,
+            823,
+            292
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a novel graph-based method, DepWiGNN, to perform propagation over the depth dimension of a graph, which can capture long dependency without the need of stacking layers and avoid the issue of over-smoothing.",
+            "- We implement a novel node memory scheme, which takes advantage of TPR mechanism, enabling convenient memory updating and retrieval operations through simple arithmetic operations instead of neural layers.",
+            "- DepWiGNN excels in multi-hop spatial reasoning tasks, surpassing existing methods in experimental evaluations on two challenging datasets. Besides, comparisons with three other GNNs highlight its superior ability to capture long dependencies within the graph. Our code will be released via https://github.com/Syon-Li/DepWiGNN."
+        ],
+        "bbox": [
+            509,
+            302,
+            882,
+            604
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related works",
+        "text_level": 1,
+        "bbox": [
+            512,
+            619,
+            667,
+            634
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Spatial Reasoning In Text has experienced a thriving development in recent years, supported by several benchmark datasets. Weston et al. (2016) proposes the bAbI project, which contains 20 QA tasks, including one focusing on textual spatial reasoning. However, some issues exist in bAbI, such as data leakage, overly short reasoning steps, and monotony of spatial relation categories and descriptions, which makes it fails to reflect the intricacy of spatial reasoning in natural language (Shi et al., 2022). Targeting these shortages, StepGame (Shi et al., 2022) expands the spatial relation types, the diversity of relation descriptions, and the required reasoning steps. Besides, SPARTQA (Mirzaee et al., 2021) augments the number of question types from one to four: find relation (FR), find blocks (FB), choose objects (CO), and yes/no (YN), while",
+        "bbox": [
+            510,
+            645,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "6460",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "SPARTUN (Mirzaee and Kordjamshidi, 2022) only has two question types (FR, YN) built with broader coverage of spatial relation types.",
+        "bbox": [
+            112,
+            84,
+            487,
+            134
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Tensor Product Representation (TPR) (Schlag and Schmidhuber, 2018a) is a mechanism to encode symbolic knowledge into a vector space, which can be applied to various natural language reasoning tasks (Huang et al., 2018; Chen et al., 2020a). For example, Schlag and Schmidhuber (2018a) perform reasoning by deconstructing the language into combinatorial symbolic representations and binding them using Third-order TPR, which can be further combined with RNN to improve the model capability of making sequential inference (Schlag et al., 2021). Shi et al. (2022) used a paragraph-level, TPR memory-augmented way to implement complex multi-hop spatial reasoning. However, existing methods typically apply TPR to pure text, which neglects the gap between natural language and symbolic structures.",
+        "bbox": [
+            112,
+            141,
+            489,
+            416
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Graph Neural Networks (GNNs) (Morris et al., 2019; Velickovic et al., 2017; Hamilton et al., 2017; Kipf and Welling, 2017) have been demonstrated to be effective in inducing and aggregating symbolic structures on other multi-hop question answering problems (Cao et al., 2019; Fang et al., 2020; Huang and Yang, 2021; Heo et al., 2022; Xu et al., 2021a; Deng et al., 2023b). In practice, the required number of graph layers grows with the multi-hop dependency between two distant nodes (Wang et al., 2021; Hong et al., 2022), which inevitably encounters the problem of oversmoothing (Li et al., 2018). Some researchers have conducted studies on relieving this problem (Wu et al., 2023; Min et al., 2020; Huang and Li, 2023; Yang et al., 2023; Liu et al., 2023; Koishekenov, 2023; Song et al., 2023). However, these methods are all breadth-aggregation-based, i.e., they only posed adjustments on breadth aggregation like improving the aggregation filters, scattering the aggregation target using the probabilistic tool, etc., but never jumping out it. In this work, we investigate a depth-wise aggregation approach that captures long-range dependencies across any distance without the need to increase the model depth.",
+        "bbox": [
+            112,
+            425,
+            490,
+            829
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            840,
+            218,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Problem Definition",
+        "text_level": 1,
+        "bbox": [
+            112,
+            866,
+            310,
+            881
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following previous studies on spatial reasoning in text (Mirzaee et al., 2021; Shi et al., 2022; Mirzaee",
+        "bbox": [
+            112,
+            887,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "and Kordjamshidi, 2022), we define the problem as follows: Given a story description  $S$  consisting of multiple sentences, the system aims to answer a question  $Q$  based on the story  $S$  by selecting a correct answer from the fixed set of given candidate answers regarding the spatial relations.",
+        "bbox": [
+            507,
+            84,
+            882,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 The DepWiNet",
+        "text_level": 1,
+        "bbox": [
+            507,
+            195,
+            672,
+            210
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As presented in Figure 2, the overall framework named DepWiNet consists of three modules: the entity representation extraction module, the DepWiGNN reasoning module, and the prediction module. The entity representation extraction module provides comprehensive entity embedding that is used in the reasoning module. A graph with recognized entities as nodes is constructed after obtaining entity representations, which later will be fed into DepWiGNN. The DepWiGNN reasons over the constructed graph and updates the node embedding correspondingly. The final prediction module adds the entity embedding from DepWiGNN to the embeddings from the first extraction module and applies a single step of attention (Vaswani et al., 2017) to generate the final result.",
+        "bbox": [
+            507,
+            218,
+            884,
+            475
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Entity Representation Extraction Module We leverage PLMs to extract entity representations<sup>1</sup>. The model takes the concatenation of the story  $S$  and the question  $Q$  as the input and output embeddings of each token. The output embeddings are further projected using a single linear layer.",
+        "bbox": [
+            507,
+            488,
+            884,
+            586
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {S}}, \\hat {\\mathbf {Q}} = \\mathbf {P L M} (S, Q) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            616,
+            599,
+            882,
+            617
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {S}} _ {\\alpha} = \\hat {\\mathbf {S}} \\mathbf {W} _ {\\alpha} ^ {T} \\quad \\hat {\\mathbf {Q}} _ {\\alpha} = \\hat {\\mathbf {Q}} \\mathbf {W} _ {\\alpha} ^ {T} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            578,
+            621,
+            882,
+            642
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{PLM}(\\cdot)$  denotes the PLM-based encoder, e.g., BERT (Devlin et al., 2019).  $\\hat{\\mathbf{S}}_{\\alpha}\\in \\mathbb{R}^{r_S\\times d_h}$ ,  $\\hat{\\mathbf{Q}}_{\\alpha}\\in \\mathbb{R}^{r_Q\\times d_h}$  denotes the list of projected tokens of size  $d_h$  in the story and question, and  $\\mathbf{W}_{\\alpha}\\in \\mathbb{R}^{d_h\\times d_h}$  is the shared projection matrix. The entity representation is just the mean pooling of all token embeddings belonging to that entity.",
+        "bbox": [
+            507,
+            657,
+            882,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Graph Construction The entities are first recognized from the input by employing rule-based entity recognition. Specifically, in StepGame (Shi et al., 2022), the entities are represented by a single capitalized letter, so we only need to locate all single capitalized letters. For SPARTUN",
+        "bbox": [
+            507,
+            783,
+            884,
+            879
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>Entity in this paper refers to the spatial roles defined in(Kordjamshidi et al., 2010).",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "6461",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/b5117fa614d2e7653a82fb581fdd79f1966f76c00f16390fee945063b2b92bdf.jpg",
+        "image_caption": [
+            "Figure 2: The DepWiNet framework. The entity representations are first extracted from the entity representation extraction module (left), and then a homogeneous graph is constructed based on the entity embeddings and fed into the DepWiNet reasoning module. The DepWiNet depth-wisely aggregates information for all indirectly connected node pairs, and stores it in node memories. The updated node embeddings are then passed to the prediction module."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            194,
+            85,
+            803,
+            300
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "and ReSQ, we use nthk RegexpParser $^2$  with self-designed grammars $^3$  to recognize entities. Entities and their embeddings are treated as nodes of the graph, and an edge between two entities exists if and only if the two entities co-occur in the same sentence. We also add an edge feature for each edge.",
+        "bbox": [
+            112,
+            395,
+            487,
+            507
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {i j} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f i = = j ,} \\\\ \\mathbf {e} _ {[ C L S ]} & \\text {o r t h e r w i s e .} \\end{array} \\right. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            519,
+            487,
+            560
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "If two entities are the same (self-loop), the edge feature is just a zero tensor with a size  $d_h$ ; otherwise, it is the sequence's last layer hidden state of the [CLS] token. The motivation behind treating [CLS] token as an edge feature is that it can help the model better understand the atomic relation  $(k = 1)$  between two nodes (entities) so as to facilitate later depth aggregation. A straightforward justification can be found in Table 1, where all three PLMs achieve very high accuracy at  $k = 1$  cases by using the [CLS] token, which demonstrates that the [CLS] token favors the understanding of the atomic relation.",
+        "bbox": [
+            112,
+            564,
+            489,
+            772
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "DepWiGNN Module The DepWiGNN module (middle part in Figure 2) is responsible for multi-hop reasoning and can collect arbitrarily distant dependency without layer stacking.",
+        "bbox": [
+            112,
+            781,
+            489,
+            846
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {V} = \\operatorname {D e p} \\operatorname {W i G N N} (\\mathcal {G}; V, E) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            186,
+            854,
+            487,
+            872
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $V \\in \\mathbb{R}^{|V| \\times d_h}$  and  $E \\in \\mathbb{R}^{|E| \\times d_h}$  are the nodes and edges of the graph. It comprises three components: node memory initialization, long dependency collection, and spatial relation retrieval. Details of these three components will be discussed in Section 3.3.",
+        "bbox": [
+            507,
+            392,
+            884,
+            491
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Prediction Module The node embedding updated in DepWiGNN is correspondingly added to entity embedding extracted from PLM.",
+        "bbox": [
+            507,
+            501,
+            882,
+            550
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {Z} _ {i} = \\left\\{ \\begin{array}{l l} Z _ {i} + \\hat {V} _ {i d x (i)} & \\text {i f i - t h t o k e n} \\in \\hat {V} \\\\ Z _ {i} & \\text {o r t h e r w i s e .} \\end{array} \\right. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            537,
+            560,
+            882,
+            602
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $Z_{i} = [\\hat{S};\\hat{Q}]$  is the ith token embedding from PLM and  $\\hat{V}$  denotes the updated entity representation set.  $idx(i)$  is the index of  $i$ -th token in the graph nodes.",
+        "bbox": [
+            507,
+            615,
+            884,
+            680
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Then, the sequence of token embeddings in the question and story are extracted separately to perform attention mechanism (Vaswani et al., 2017) with the query to be the sequence of question tokens embeddings  $\\hat{Z}_{\\hat{\\mathbf{Q}}}$ , key and value to be the sequence of story token embeddings  $\\hat{Z}_{\\hat{\\mathbf{S}}}$ .",
+        "bbox": [
+            507,
+            681,
+            885,
+            783
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR = \\sum_ {i = 0} ^ {r _ {Q}} \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\hat {Z} _ {\\hat {\\mathbf {Q}}} \\hat {Z} _ {\\hat {\\mathbf {S}}} ^ {T}}{\\sqrt {d _ {h}}}\\right) \\hat {Z} _ {\\hat {\\mathbf {S}}}\\right) _ {i} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            568,
+            794,
+            882,
+            837
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC = F F N (\\operatorname {l a y e r n m} (R)) \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            840,
+            880,
+            858
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where layernm means layer normalization and  $C$  is the final logits. The result embeddings are summed over the first dimension, layernormed, and",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2https://www.nltk.org/api/nltk.chunk.regexp.html",
+        "bbox": [
+            134,
+            879,
+            438,
+            892
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "3r\"\"\"\"NP:<ADJ|NOUN>+<NOUN|NUM>\"\" and",
+        "bbox": [
+            134,
+            892,
+            485,
+            904
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "r\"\"\"\"NP:ADJINOUN  $>$  +<VERB  $>$  +<NOUNNUM>",
+        "bbox": [
+            115,
+            904,
+            453,
+            916
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "6462",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/9282c7cc833d54129685fc4ca0316422997a1261d856cb969918d36c9b338705.jpg",
+        "image_caption": [
+            "Figure 3: The illustration of DepWiGNN."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            80,
+            487,
+            206
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "fed into a 3-layer feedforward neural network to acquire the final result. The overall framework is trained in an end-to-end manner to minimize the cross-entropy loss between the predicted candidate answer probabilities and the ground-truth answer.",
+        "bbox": [
+            112,
+            256,
+            487,
+            338
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Depth-wise Graph Neural Network",
+        "text_level": 1,
+        "bbox": [
+            112,
+            349,
+            436,
+            365
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Figure 3, we introduce the proposed graph neural network, called DepWiGNN, i.e., the operation  $\\hat{V} = \\mathsf{DepWiGNN}(\\mathcal{G};V,E)$ . Unlike existing GNNs (e.g. (Morris et al., 2019; Velickovic et al., 2017; Hamilton et al., 2017)), which counts the one-dimension node embedding itself as its memory to achieve the function of information reservation, updating, and retrieval, DepWiGNN employs a novel two-dimension node memory implementation approach that takes the advantage of TPR mechanism allowing the updating and retrieval operations of the memory to be conveniently realized by simple arithmetic operations like plus, minus and outer product. This essentially determines that the information propagation between any pair of nodes with any distance in the graph can be accomplished without the demand to iteratively apply neural layers.",
+        "bbox": [
+            112,
+            370,
+            489,
+            659
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Node Memory Initialization At this stage, the node memories of all nodes are initialized with the relations to their immediate neighbors. In the multi-hop spatial reasoning cases, the relations will be the atomic spatial orientation relations (which only needs one hop) of the destination node relative to the source node. For example, \"X is to the left of K and is on the same horizontal plane.\" We follow the TPR mechanism (Smolensky, 1990), which uses outer product operation to bind roles and fillers<sup>4</sup>. In this work, the calculated spatial vectors are considered to be the fillers. They will be bound with corresponding node embeddings and stored in the two-dimensional node memory. Explicitly, we first",
+        "bbox": [
+            112,
+            670,
+            489,
+            894
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "acquire spatial orientation filler  $f_{ij} \\in \\mathbb{R}^{d_h}$  by using a feedforward network, the input to FFN is concatenation in the form  $[V_i, E_{ij}, V_j]$ .  $V_i$  represents  $i$ -th node embedding.",
+        "bbox": [
+            507,
+            83,
+            884,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf _ {i j} = F F N \\left(\\left[ V _ {i}, E _ {i j}, V _ {j} \\right]\\right) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            596,
+            160,
+            882,
+            178
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM _ {i} = \\sum_ {V _ {j} \\in N (V _ {i})} f _ {i j} \\otimes V _ {j} \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            181,
+            882,
+            208
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The filler is bound together with the corresponding destination node  $V_{j}$  using the outer product. The initial node memory  $M_{i}$  for node  $V_{i}$  is just the summation of all outer product results of the fillers and corresponding neighbors (left part in Figure 3).",
+        "bbox": [
+            507,
+            216,
+            882,
+            298
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Long Dependency Collection We discuss how the model collects long dependencies in this section. Since the atomic relations are bound by corresponding destinations and have already been contained in the node memories, we can easily unbind all the atomic relation fillers in a path using the corresponding atomic destination node embedding (middle part in Figure 3). For each pair of indirectly connected nodes, we first find the shortest path between them using breadth-first search (BFS). Then, all the existing atomic relation fillers along the path are unbound using the embedding of each node in the path (Eq.10).",
+        "bbox": [
+            507,
+            305,
+            884,
+            514
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {f} _ {p _ {i} \\left(p _ {i + 1}\\right)} = M _ {p _ {i}} V _ {p _ {i + 1}} ^ {T} \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            613,
+            524,
+            882,
+            545
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {F}} = \\operatorname {A g g r e g a t o r} (\\mathbf {F}) \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            608,
+            546,
+            880,
+            565
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf _ {p _ {0} p _ {n}} = \\operatorname {l a y e r n m} (F F N (\\hat {\\mathbf {F}}) + \\hat {\\mathbf {F}}) \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            547,
+            569,
+            882,
+            588
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM _ {p _ {0}} = M _ {p _ {0}} + f _ {p _ {0} p _ {n}} \\otimes V _ {p _ {n}} ^ {T} \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            593,
+            590,
+            880,
+            611
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $p_i$  denotes the  $i$ -th element in the path and  $\\mathbf{F} = [\\hat{f}_{p_0p_1};\\dots;\\hat{f}_{p_{n-1}p_n}]$  is the retrieved filler set along the path. The collected relation fillers are aggregated using a selected depth aggregator like LSTM (Hochreiter and Schmidhuber, 1997), etc, and passed to a feedforward neural network to reason the relation filler between the source and destination node in the path (Eq.12). The result spatial filler is then bound with the target node embedding and added to the source node memory (Eq.13). In this way, each node memory will finally contain the spatial orientation information to every other connected node in the graph.",
+        "bbox": [
+            507,
+            620,
+            882,
+            829
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Spatial Relation Retrieval After the collection process is completed, every node memory contains spatial relation information to all other connected nodes in the graph. Therefore, we can conveniently retrieve the spatial information from a source node",
+        "bbox": [
+            507,
+            838,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "4The preliminary of TPR is presented in Appendix A.",
+        "bbox": [
+            134,
+            903,
+            463,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "6463",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "to a target node by unbinding the spatial filler from the source node memory (right part in Figure 3) using a self-determined key. The key can be the target node embedding itself if the target node can be easily recognized from the question, or some computationally extracted representation from the sequence of question token embeddings if the target node is hard to discern. We use the key to unbind the spatial relation from all nodes' memory and pass the concatenation of it with source and target node embeddings to a multilayer perceptron to get the updated node embeddings.",
+        "bbox": [
+            112,
+            84,
+            489,
+            277
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {f} _ {i [ \\text {k e y} ]} = M _ {i} V _ {[ \\text {k e y} ]} ^ {T} \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            228,
+            288,
+            487,
+            312
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {V} _ {i} = F F N ([ V _ {i}, \\mathbf {l a y e r n m} (f _ {i [ k e y ]}), V _ {[ k e y ]} ]) \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            124,
+            313,
+            487,
+            335
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The updated node embeddings are then passed to the prediction module to get the final result.",
+        "bbox": [
+            112,
+            347,
+            485,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            392,
+            260,
+            409
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Setups",
+        "text_level": 1,
+        "bbox": [
+            112,
+            419,
+            324,
+            435
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets & Evaluation Metrics We investigated our model on StepGame (Shi et al., 2022) and ReSQ (Mirzaee and Kordjamshidi, 2022) datasets, which were recently published for multi-hop spatial reasoning. StepGame is a synthetic textual QA dataset that has a number of relations  $(k)$  ranging from 1 to 10. In particular, we follow the experimental procedure in the original paper (Shi et al., 2022), which utilized the TrainVersion of the dataset containing 10,000/1,000 training/validation clean samples for each  $k$  value from 1 to 5 and 10,000 noisy test examples for  $k$  value from 1 to 10. The test set contains three kinds of distraction noise: disconnected, irrelevant, and supporting. ReSQ is a crowdsourcing benchmark that includes only Yes/No questions with 1008, 333, and 610 examples for training, validating, and testing respectively. Since it is human-generated, it can reflect the natural complexity of real-world spatial description. Following the setup in (Shi et al., 2022; Mirzaee and Kordjamshidi, 2022), we report the accuracy on corresponding test sets for all experiments.",
+        "bbox": [
+            112,
+            441,
+            489,
+            796
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines For StepGame, we select all traditional reasoning models used in the original paper (Shi et al., 2022) and three PLMs, i.e., BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and ALBERT (Lan et al., 2020) as our baselines. For ReSQ, we also follow the experiment setting described in (Mirzaee and Kordjamshidi, 2022),",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "which used BERT with or without further synthetic supervision as baselines.",
+        "bbox": [
+            507,
+            84,
+            880,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details For all experiments, we use the base version of corresponding PLMs, which has 768 embedding dimensions. The model was trained in an end-to-end manner using Adam optimizer (Kingma and Ba, 2015). The training was stopped if, up to 3 epochs, there is no improvement greater than 1e-3 on the cross-entropy loss for the validation set. We also applied a Pytorch training scheduler that reduces the learning rate with a factor of 0.1 if the improvement of cross-entropy loss on the validation set is lower than 1e-3 for 2 epochs. In terms of the determination of the key in the Spatial Relation Retrieval part, we used the target node embedding for StepGame since it can be easily recognized, and we employed a single linear layer to extract the key representation from the sum-aggregated question token embeddings for ReSQ. In the StepGame experiment, we fine-tune the model on the training set and test it on the test set. For ReSQ, we follow the procedure in (Mirzaee and Kordjamshidi, 2022) to test the model on ReSQ with or without further supervision from SPARTUN (Mirzaee and Kordjamshidi, 2022). Unless specified, all the experiments use LSTM (Hochreiter and Schmidhuber, 1997) as depth aggregator by default. The detailed hyperparameter settings are given in the Appendix B.",
+        "bbox": [
+            507,
+            128,
+            884,
+            562
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Overall Performance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            575,
+            721,
+            589
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1 and 2 report the experiment results on StepGame and ReSQ respectively. As shown in Table 1, PLMs overwhelmingly outperform all the traditional reasoning models and the proposed DeepWiNet overtakes the PLMs by a large margin, especially for the cases with greater  $k$  where the multi-hop reasoning capability plays a more important role. This aligns with the characteristics of our model architecture that the aggregation focuses on the depth dimension, which effectively avoids the problem of over-smoothing and the mixture of redundant information from the breadth aggregation. Despite the model being only trained on clean distraction-noise-absent samples with  $k$  from 1 to 5, it achieves impressive performance on the distraction-noise-present test data with  $k$  value from 6 to 10, demonstrating the more advanced generalization capability of our model. Moreover, our method also entails an improvement for the examples with lower  $k$  values for BERT and AL",
+        "bbox": [
+            507,
+            596,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6464",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/c4f4fba40f83db2041777c7d39bd9a2bf0e855ac4758d1d65b7d8be8fca2d93b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>k=1</td><td>k=2</td><td>k=3</td><td>k=4</td><td>k=5</td><td>k=6</td><td>k=7</td><td>k=8</td><td>k=9</td><td>k=10</td><td>Mean(k=1-5)</td><td>Mean(k=6-10)</td></tr><tr><td>RN (Santoro et al., 2017)</td><td>22.64</td><td>17.08</td><td>15.08</td><td>12.84</td><td>11.52</td><td>11.12</td><td>11.53</td><td>11.21</td><td>11.13</td><td>11.34</td><td>15.83</td><td>11.27</td></tr><tr><td>RRN (Palm et al., 2018)</td><td>24.05</td><td>19.98</td><td>16.03</td><td>13.22</td><td>12.31</td><td>11.62</td><td>11.40</td><td>11.83</td><td>11.22</td><td>11.69</td><td>17.12</td><td>11.56</td></tr><tr><td>UT (Dehghani et al., 2019)</td><td>45.11</td><td>28.36</td><td>17.41</td><td>14.07</td><td>13.45</td><td>12.73</td><td>12.11</td><td>11.40</td><td>11.41</td><td>11.74</td><td>23.68</td><td>11.88</td></tr><tr><td>STM (Le et al., 2020)</td><td>53.42</td><td>35.96</td><td>23.03</td><td>18.45</td><td>15.14</td><td>13.80</td><td>12.63</td><td>11.54</td><td>11.30</td><td>11.77</td><td>29.20</td><td>12.21</td></tr><tr><td>TPR-RNN (Schlag and Schmidhuber, 2018b)</td><td>70.29</td><td>46.03</td><td>36.14</td><td>26.82</td><td>24.77</td><td>22.25</td><td>19.88</td><td>15.45</td><td>13.01</td><td>12.65</td><td>40.81</td><td>16.65</td></tr><tr><td>TP-MANN (Shi et al., 2022)</td><td>85.77</td><td>60.31</td><td>50.18</td><td>37.45</td><td>31.25</td><td>28.53</td><td>26.45</td><td>23.67</td><td>22.52</td><td>21.46</td><td>52.99</td><td>24.53</td></tr><tr><td>BERT (Mirzaee and Kordjamshidi, 2022)</td><td>98.53</td><td>93.40</td><td>91.19</td><td>66.98</td><td>54.57</td><td>48.59</td><td>42.81</td><td>37.98</td><td>34.16</td><td>33.62</td><td>80.93</td><td>39.43</td></tr><tr><td>RoBERTa* (Mirzaee and Kordjamshidi, 2022)</td><td>98.68</td><td>97.05</td><td>95.66</td><td>79.59</td><td>74.89</td><td>70.67</td><td>66.01</td><td>61.42</td><td>57.20</td><td>54.53</td><td>89.17</td><td>61.93</td></tr><tr><td>ALBERT* (Mirzaee and Kordjamshidi, 2022)</td><td>98.56</td><td>97.56</td><td>96.08</td><td>90.01</td><td>83.33</td><td>77.24</td><td>71.57</td><td>64.47</td><td>60.02</td><td>56.18</td><td>93.11</td><td>65.90</td></tr><tr><td>DepWiNet (BERT)</td><td>98.44</td><td>96.57</td><td>94.75</td><td>81.41</td><td>70.68</td><td>62.80</td><td>57.46</td><td>49.14</td><td>45.56</td><td>44.01</td><td>88.37+9.2%</td><td>51.79+31.3%</td></tr><tr><td>DepWiNet (RoBERTa)</td><td>98.70</td><td>96.69</td><td>95.54</td><td>79.50</td><td>74.94</td><td>70.37</td><td>66.22</td><td>61.89</td><td>58.01</td><td>56.89</td><td>89.07-0.1%</td><td>62.68+1.2%</td></tr><tr><td>DepWiNet (ALBERT)</td><td>98.59</td><td>97.13</td><td>95.87</td><td>91.96</td><td>87.50</td><td>82.62</td><td>77.80</td><td>71.40</td><td>67.55</td><td>64.98</td><td>94.21+1.2%</td><td>72.87+10.6%</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            884,
+            288
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/7fb024c5e63d60a148c58f439467e8b2c1fd9d346904739a104768533635d234.jpg",
+        "table_caption": [
+            "Table 1: Experimental results on StepGame. * denotes the implementation of the same model as BERT (Mirzaee and Kordjamshidi, 2022) with RoBERTa and ALBERT. The other results are reported in (Shi et al., 2022)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>SynSup</td><td>ReSQ</td></tr><tr><td>Majority Baseline</td><td>-</td><td>50.21</td></tr><tr><td>BERT</td><td>-</td><td>57.37</td></tr><tr><td>BERT</td><td>SPARTQA-AUTO</td><td>55.08</td></tr><tr><td>BERT</td><td>StepGame</td><td>60.14</td></tr><tr><td>BERT</td><td>SPARTUN-S</td><td>58.03</td></tr><tr><td>BERT</td><td>SPARTUN</td><td>63.60</td></tr><tr><td>DepWiNet (BERT)</td><td>-</td><td>63.30</td></tr><tr><td>DepWiNet (BERT)</td><td>SPARTUN</td><td>65.54</td></tr></table>",
+        "bbox": [
+            144,
+            349,
+            458,
+            500
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Experimental results on ReSQ, under the transfer learning setting (Mirzaee and Kordjamshidi, 2022).",
+        "bbox": [
+            112,
+            511,
+            489,
+            541
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "BERT and an ignorable decrease for RoBERTa, which attests to the innocence of our model to the few-hop reasoning tasks.",
+        "bbox": [
+            112,
+            573,
+            487,
+            621
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Experimental results in Table 2 show that DeepWiNet reaches a new SOTA result on ReSQ for both cases where extra supervision from SPARTUN is present and absent. Excitingly, the performance of our model without extra supervision even overwhelms the performance of the BERT with extra supervision from SPARTQA-AUTO (Mirzaee et al., 2021), StepGame, and SPARTUN-S and approaches closely to the SPARTUN supervision case. All these phenomenons signify that our model has the potential to better tackle the natural intricacy of real-world spatial expressions.",
+        "bbox": [
+            112,
+            625,
+            489,
+            819
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Ablation study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            841,
+            278,
+            858
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct ablation studies on the impact of the three components of DepWiGNN and different depth aggregators, as presented in Table 3.",
+        "bbox": [
+            112,
+            871,
+            487,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of DepWiNet components The model performance experiences a drastic decrease, particularly for the mean score of  $k$  between 6 and 10, after the Long Dependency Collection (LDC) has been removed, verifying that this component serves a crucial role in the model. Note that the mean score for  $k(6-10)$  even drops below the ALBERT baseline (Table 1). This is reasonable as the LDC is directly responsible for collecting long dependencies. We further defunctionalized the Node Memory Initialization (NMI) and then Spatial Relation Retrieval (SRR) by setting the initial fillers (Eq.8) and key representation (Eq.14) to a random vector separately. Compared with the case where only LDC was removed, both of them lead to a further decrease in small and large  $k$  values.",
+        "bbox": [
+            507,
+            353,
+            884,
+            611
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impact of Different Aggregators The results show that the mean and max pooling depth aggregators fail to understand the spatial rule as achieved by LSTM. This may be caused by the relatively less expressiveness and generalization caliber of mean and max pooling operation.",
+        "bbox": [
+            507,
+            624,
+            882,
+            720
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Comparisons with Different GNNs",
+        "text_level": 1,
+        "bbox": [
+            507,
+            734,
+            828,
+            750
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To certify our model's capacity of collecting long dependencies as well as its immunity to the oversmoothing problem, we contrast it with four graph neural networks, namely, GCN (Kipf and Welling, 2017), GraphConv (Morris et al., 2019), GAT (Velickovic et al., 2017) and GCNII (Chen et al., 2020c). We consider the cases with the number of layers varying from 1 to 5 and select the best performance to compare. The plot of the accuracy metric is reported in Figure 4 and the corresponding best",
+        "bbox": [
+            507,
+            758,
+            884,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "6465",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/e46a55ae70fab0d85af269292d7d4c929fd1973f938e3a13b1a3678501045677.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>k=1</td><td>k=2</td><td>k=3</td><td>k=4</td><td>k=5</td><td>k=6</td><td>k=7</td><td>k=8</td><td>k=9</td><td>k=10</td><td>Mean(k=1-5)</td><td>Mean(k=6-10)</td></tr><tr><td>DepWiNet (ALBERT)</td><td>98,59</td><td>97.13</td><td>95.87</td><td>91.96</td><td>87.50</td><td>82.62</td><td>77.80</td><td>71.40</td><td>67.55</td><td>64.98</td><td>94.21</td><td>72.87</td></tr><tr><td>- w/o LDC</td><td>98.58</td><td>97.72</td><td>96.42</td><td>82.46</td><td>78.53</td><td>74.52</td><td>70.45</td><td>62.84</td><td>59.06</td><td>56.57</td><td>90.74-3.7%</td><td>64.69-11.2%</td></tr><tr><td>- w/o NMI &amp; LDC</td><td>98.48</td><td>96.88</td><td>94.83</td><td>79.14</td><td>75.46</td><td>70.51</td><td>67.19</td><td>61.88</td><td>57.65</td><td>55.02</td><td>88.96-5.6%</td><td>62.45-14.3%</td></tr><tr><td>- w/o SRR &amp; LDC</td><td>98.64</td><td>97.52</td><td>95.91</td><td>80.72</td><td>76.85</td><td>72.56</td><td>67.93</td><td>61.29</td><td>57.97</td><td>54.17</td><td>89.93-4.5%</td><td>62.78-13.8%</td></tr><tr><td>- w/ DepWiGNNmean</td><td>98.52</td><td>96.46</td><td>93.69</td><td>81.10</td><td>77.90</td><td>73.47</td><td>70.78</td><td>65.54</td><td>62.42</td><td>59.82</td><td>89.53-5.0%</td><td>66.41-8.9%</td></tr><tr><td>- w/ DepWiGNNmaxpooling</td><td>98.65</td><td>95.83</td><td>93.90</td><td>80.81</td><td>77.20</td><td>72.07</td><td>68.76</td><td>62.79</td><td>59.63</td><td>57.14</td><td>89.28-5.2%</td><td>64.08-12.1%</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            884,
+            211
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/06755425936a3d90c143fb9d926cd41dcf1ce993b38d0db71c6a40ee24e039ca.jpg",
+        "table_caption": [
+            "Table 3: Ablation study. NMI, LDC, and SRR denotes the three stages in DepWiGNN, i.e., Node Memory Initialization, Long Dependency Collection, and Spatial Relation Retrieval, respectively."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>k=1</td><td>k=2</td><td>k=3</td><td>k=4</td><td>k=5</td><td>k=6</td><td>k=7</td><td>k=8</td><td>k=9</td><td>k=10</td><td>Mean(k=1-5)</td><td>Mean(k=6-10)</td></tr><tr><td>ALBERT</td><td>98.56</td><td>97.56</td><td>96.08</td><td>90.01</td><td>83.33</td><td>77.24</td><td>71.57</td><td>64.47</td><td>60.02</td><td>56.18</td><td>93.11</td><td>65.90</td></tr><tr><td>w/ GCN5 (Kipf and Welling, 2017)</td><td>98.62</td><td>97.72</td><td>96.49</td><td>82.01</td><td>76.70</td><td>71.60</td><td>67.34</td><td>60.49</td><td>56.11</td><td>52.65</td><td>90.31-3.01%</td><td>61.64-6.46%</td></tr><tr><td>w/ GraphConv2 (Morris et al., 2019)</td><td>98.66</td><td>97.73</td><td>96.48</td><td>82.37</td><td>78.57</td><td>74.47</td><td>69.91</td><td>63.04</td><td>60.01</td><td>56.58</td><td>90.76-2.52%</td><td>64.80-1.67%</td></tr><tr><td>w/ GAT1 (Velickovic et al., 2017)</td><td>98.67</td><td>97.34</td><td>95.59</td><td>81.47</td><td>78.39</td><td>73.79</td><td>70.17</td><td>63.52</td><td>60.64</td><td>57.33</td><td>90.29-3.03%</td><td>65.09-1.23%</td></tr><tr><td>w/ GCNII4 (Chen et al., 2020c)</td><td>98.56</td><td>97.61</td><td>96.36</td><td>83.31</td><td>79.22</td><td>74.88</td><td>70.64</td><td>63.64</td><td>60.59</td><td>56.99</td><td>91.01-2.25%</td><td>65.35+0.01%</td></tr><tr><td>w/ DepWiGNN</td><td>98.59</td><td>97.13</td><td>95.87</td><td>91.96</td><td>87.50</td><td>82.62</td><td>77.80</td><td>71.40</td><td>67.55</td><td>64.98</td><td>94.21+1.18%</td><td>72.87+10.58%</td></tr></table>",
+        "bbox": [
+            115,
+            261,
+            882,
+            366
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Comparisons with different GNNs. The subscripts represent the number of GNN layers. We select the best mean performance among layer number 1 to 5.",
+        "bbox": [
+            112,
+            375,
+            882,
+            405
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d7e58ce4a4ad34b5cc46e1ade07941b6b47aab671665c5017a7fb0586e832ffc.jpg",
+        "image_caption": [
+            "Figure 4: Impact of the layer number in different GNNs on StepGame. The solid and dashed lines denote the mean score of  $(k = 1 - 5)$  and  $(k = 6 - 10)$  respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            430,
+            484,
+            605
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "cases are in Table 4. It is worth noting that almost all the baseline GNNs cause a reduction in the original PLM performance (Table 4). The reason for this may partially come from the breadth aggregation, which aggregates the neighbors round and round and leads to indistinguishability among entity embeddings such that the PLM reasoning process has been disturbed. The majority of baseline GNNs suffer from an apparent performance drop when increasing the number of layers (Figure 4), while our model consistently performs better and is not affected by the number of layers at all since it does not use breadth aggregation. Therefore, our model has immunity to over-smoothing problems.",
+        "bbox": [
+            112,
+            694,
+            490,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f56bfb1e39687845f8350abe77bd1573fa2c041f747dc0bdcfee147055f9092f.jpg",
+        "image_caption": [
+            "Figure 5: Cases with distracting noisy from StepGame."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            430,
+            880,
+            573
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In both small and large  $k$  cases, our model outperforms the best performance of all four GNNs (including GCNII, which is specifically designed for over-smoothing issues) by a large margin (Table 4), which serves as evidence of the superiority of our model in long dependency collection.",
+        "bbox": [
+            507,
+            623,
+            882,
+            720
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 Case study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            730,
+            640,
+            745
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present case studies to intuitively show how DepWiGNN mitigates the three kinds of distracting noise introduced in StepGame, namely, disconnected, irrelevant, and supporting.",
+        "bbox": [
+            507,
+            750,
+            884,
+            815
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- The disconnected noise is the set of entities and relations that forms a new independent chain in the graph (Figure 5(a)). The node memories constructed in DepWiGNN contain spatial information about the nodes if and only if that node stays in the same connected component; otherwise, it",
+        "bbox": [
+            507,
+            822,
+            885,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "6466",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "has no information about the node as there is no path between them. Hence, in this case, for the questioned source node  $\\mathbf{P}$ , its memory has no information for the disconnected noise  $\\mathbf{T}$  and  $\\mathbf{D}$ .",
+        "bbox": [
+            127,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The irrelevant noise branches the correct reasoning chain out with new entities and relations but results in no alternative reasoning path (Figure 5(b)). Hence the irrelevant noise entities will not be included in the reasoning path between the source and destination, which means that it will not affect the destination spatial filler stored in the source node memory. In this case, when the key representation (embedding of entity  $\\mathbf{E}$ ) is used to unbind the spatial filler from node memory of the source node  $\\mathbf{P}$ , it obtains the filler  $f_{PE} = FFN(Aggregator([f_{PX}; f_{XQ}; f_{QE}])$ , which is intact to the irrelevant entity  $\\mathbf{Y}$  and relation  $f_{xy}$  or  $f_{yx}$ .",
+            "- The supporting noise adds new entities and relations to the original reasoning chain that provides an alternative reasoning path (Figure 5(c)). DeepWiGNN is naturally exempted from this noise for two reasons: first, it finds the shortest path between two entities, therefore, will not include  $\\mathbf{I}$  and  $\\mathbf{M}$  in the path; Second, even if the longer path is considered, the depth aggregator should reach the same result as the shorted one since the source and destination are the same."
+        ],
+        "bbox": [
+            114,
+            162,
+            489,
+            561
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            576,
+            247,
+            590
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we introduce DepWiGNN, a novel graph-based method that facilitates depth-wise propagation in graphs, enabling effective capture of long dependencies while mitigating the challenges of over-smoothing and excessive layer stacking. Our approach incorporates a node memory scheme leveraging the TPR mechanism, enabling simple arithmetic operations for memory updating and retrieval, instead of relying on additional neural layers. Experiments on two recently released textual multi-hop spatial reasoning datasets demonstrate the superiority of DepWiGNN in collecting long dependencies over the other three typical GNNs and its immunity to the over-smoothing problem.",
+        "bbox": [
+            112,
+            604,
+            489,
+            829
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            844,
+            211,
+            858
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Unlike the classical GNNs, which use one-dimensional embedding as the node memory, our method applies a two-dimensional matrix-shaped",
+        "bbox": [
+            112,
+            871,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "node memory. This poses a direct increase in memory requirement. The system has to assign extra space to store a matrix with shape  $\\mathbb{R}^{d_h\\times d_h}$  for each node in the graph, which makes the method less scalable. However, it is a worthwhile trade-off because the two-dimensional node memory can potentially store  $d_{h} - 1$  number of spatial fillers bound with node embeddings while keeping the size fixed, and it does not suffer from information overloading as faced by the one-dimension memory since the spatial fillers can be straightforwardly retrieved from the memory.",
+        "bbox": [
+            507,
+            84,
+            884,
+            275
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Another issue is the time complexity of finding the shortest path between every pair of nodes. For all experiments in this paper, since the edges do not have weights, the complexity is  $\\mathrm{O}((\\mathrm{n} + \\mathrm{e})^{*}\\mathrm{n})$  where  $\\mathbf{n}$  and  $\\mathbf{e}$  are the numbers of nodes and edges, respectively. However, things will get worse if the edges in the graph are weighted. We believe this is a potential future research direction for the improvement of the algorithm.",
+        "bbox": [
+            507,
+            278,
+            884,
+            422
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            449,
+            608,
+            463
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. CoRR, abs/2302.04023.",
+            "Yu Cao, Meng Fang, and Dacheng Tao. 2019. BAG: bi-directional attention entity graph convolutional network for multi-hop reasoning question answering. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, pages 357-362.",
+            "Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. 2019. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 12538-12547. Computer Vision Foundation / IEEE.",
+            "Kezhen Chen, Qiuyuan Huang, Hamid Palangi, Paul Smolensky, Kenneth D. Forbus, and Jianfeng Gao. 2020a. Mapping natural-language problems to formal-language solutions using structured neural representations. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1566-1575. PMLR.",
+            "Kunlong Chen, Weidi Xu, Xingyi Cheng, Zou Xiaochuan, Yuyu Zhang, Le Song, Taifeng Wang, Yuan"
+        ],
+        "bbox": [
+            509,
+            470,
+            885,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "6467",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Qi, and Wei Chu. 2020b. Question directed graph attention network for numerical reasoning over text. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 6759-6768. Association for Computational Linguistics.",
+            "Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. 2020c. Simple and deep graph convolutional networks. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1725-1735. PMLR.",
+            "Surabhi Datta, Yuqi Si, Laritza Rodriguez, Sonya E. Shooshan, Dina Demner-Fushman, and Kirk Roberts. 2020. Understanding spatial language in radiology: Representation framework, annotation, and spatial relation extraction from chest x-ray reports using deep learning. J. Biomed. Informatics, 108:103473.",
+            "Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. 2019. Universal transformers. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.",
+            "Yang Deng, Shuaiyi Li, and Wai Lam. 2023a. Learning to ask clarification questions with spatial reasoning. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, pages 2113-2117.",
+            "Yang Deng, Yuexiang Xie, Yaliang Li, Min Yang, Wai Lam, and Ying Shen. 2022. Contextualized knowledge-aware attentive neural network: Enhancing answer selection with knowledge. ACM Trans. Inf. Syst., 40(1):2:1-2:33.",
+            "Yang Deng, Wenxuan Zhang, Weiwen Xu, Ying Shen, and Wai Lam. 2023b. Nonfactoid question answering as query-focused summarization with graph-enhanced multihop inference. IEEE Transactions on Neural Networks and Learning Systems.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.",
+            "Yuwei Fang, Siqi Sun, Zhe Gan, Rohit Pillai, Shuo-hang Wang, and Jingjing Liu. 2020. Hierarchical graph network for multi-hop question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, pages 8823-8838."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "William L. Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 1024-1034.",
+            "Yu-Jung Heo, Eun-Sol Kim, Woo Suk Choi, and Byoung-Tak Zhang. 2022. Hypergraph transformer: Weakly-supervised multi-hop reasoning for knowledge-based visual question answering. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, pages 373-390.",
+            "Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural Comput., 9(8):1735-1780.",
+            "Giwon Hong, Jeonghwan Kim, Junmo Kang, and Sung-Hyon Myaeng. 2022. Graph-induced transformers for efficient multi-hop question answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, pages 10288-10294.",
+            "Qiuyuan Huang, Paul Smolensky, Xiaodong He, Li Deng, and Dapeng Oliver Wu. 2018. Tensor product generation networks for deep NLP modeling. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 1263-1273. Association for Computational Linguistics.",
+            "Rui Huang and Ping Li. 2023. Hub-hub connections matter: Improving edge dropout to relieve oversmoothing in graph neural networks. Knowl. Based Syst., 270:110556.",
+            "Yongjie Huang and Meng Yang. 2021. Breadth first reasoning graph for multi-hop question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, pages 5810-5821.",
+            "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.",
+            "Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.",
+            "Yeskendir Koishekenov. 2023. Reducing oversmoothing in graph neural networks using relational embeddings. CoRR, abs/2301.02924."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "6468",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Parisa Kordjamshidi, Martijn van Otterlo, and Marie-Francine Moens. 2010. Spatial role labeling: Task definition and annotation scheme. In Proceedings of the International Conference on Language Resources and Evaluation, LREC 2010, 17-23 May 2010, Valletta, Malta. European Language Resources Association.",
+            "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. ALBERT: A lite BERT for self-supervised learning of language representations. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.",
+            "Hung Le, Truyen Tran, and Svetha Venkatesh. 2020. Self-attentive associative memory. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 5682-5691. PMLR.",
+            "Qimai Li, Zhichao Han, and Xiao-Ming Wu. 2018. Deeper insights into graph convolutional networks for semi-supervised learning. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 3538-3545. AAAI Press.",
+            "Yang Liu, Chuan Zhou, Shirui Pan, Jia Wu, Zhao Li, Hongyang Chen, and Peng Zhang. 2023. Curvdrop: A ricci curvature based approach to prevent graph neural networks from over-smoothing and oversquashing. In Proceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX, USA, 30 April 2023 - 4 May 2023, pages 221-230. ACM.",
+            "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692.",
+            "Qian Luo, Yunfei Li, and Yi Wu. 2023. Grounding object relations in language-conditioned robotic manipulation with semantic-spatial reasoning. CoRR, abs/2303.17919.",
+            "Wouter Massa, Parisa Kordjamshidi, Thomas Provoost, and Marie-Francine Moens. 2015. Machine reading of biological texts - bacteria-biotope extraction. In BIOINFORMATICS 2015 - Proceedings of the International Conference on Bioinformatics Models, Methods and Algorithms, Lisbon, Portugal, 12-15 January, 2015, pages 55-64. SciTePress.",
+            "Yimeng Min, Frederik Wenkel, and Guy Wolf. 2020. Scattering GCN: overcoming oversmoothness in graph convolutional networks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Roshanak Mirzaee, Hossein Rajaby Faghihi, Qiang Ning, and Parisa Kordjamshidi. 2021. SPARTQA: A textual question answering benchmark for spatial reasoning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021, pages 4582-4598. Association for Computational Linguistics.",
+            "Roshanak Mirzaee and Parisa Kordjamshidi. 2022. Transfer learning with synthetic corpora for spatial role labeling and reasoning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 6148-6165. Association for Computational Linguistics.",
+            "Christopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. 2019. Weisfeiler and leman go neural: Higher-order graph neural networks. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 4602-4609. AAAI Press.",
+            "Rasmus Berg Palm, Ulrich Paquet, and Ole Winther. 2018. Recurrent relational networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montreal, Canada, pages 3372-3382.",
+            "Lin Qiu, Yunxuan Xiao, Yanru Qu, Hao Zhou, Lei Li, Weinan Zhang, and Yong Yu. 2019. Dynamically fused graph network for multi-hop reasoning. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 6140-6150. Association for Computational Linguistics.",
+            "Adam Santoro, David Raposo, David G. T. Barrett, Mateusz Malinowski, Razvan Pascanu, Peter W. Battaglia, and Tim Lillicrap. 2017. A simple neural network module for relational reasoning. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 4967-4976.",
+            "Imanol Schlag, Tsendsuren Munkhdalai, and Jurgen Schmidhuber. 2021. Learning associative inference using fast weight memory. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.",
+            "Imanol Schlag and Jürgen Schmidhuber. 2018a. Learning to reason with third order tensor products. In Advances in Neural Information Processing Systems"
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "6469",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 10003-10014.",
+            "Imanol Schlag and Jürgen Schmidhuber. 2018b. Learning to reason with third order tensor products. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 10003-10014.",
+            "Zhengxiang Shi, Qiang Zhang, and Aldo Lipani. 2022. Stepgame: A new benchmark for robust multi-hop spatial reasoning in texts. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, pages 11321-11329.",
+            "Paul Smolensky. 1990. Tensor product variable binding and the representation of symbolic structures in connectionist systems. Artif. Intell., 46(1-2):159-216.",
+            "Yunchong Song, Chenghu Zhou, Xinbing Wang, and Zhouhan Lin. 2023. Ordered GNN: ordering message passing to deal with heterophily and oversmoothing. CoRR, abs/2302.01524.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.",
+            "Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. CoRR, abs/1710.10903.",
+            "Sagar Gubbi Venkatesh, Anirban Biswas, Raviteja Upadrashta, Vikram Srinivasan, Partha Talukdar, and Bharadwaj Amrutur. 2021. Spatial reasoning from natural language instructions for robot manipulation. In IEEE International Conference on Robotics and Automation, ICRA 2021, Xi'an, China, May 30 - June 5, 2021, pages 11196-11202. IEEE.",
+            "Guangtao Wang, Rex Ying, Jing Huang, and Jure Leskovec. 2021. Multi-hop attention graph neural networks. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJ-CAI 2021, pages 3089-3096. ijcai.org.",
+            "Jason Weston, Antoine Bordes, Sumit Chopra, and Tomás Mikolov. 2016. Towards ai-complete question answering: A set of prerequisite toy tasks. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings.",
+            "Gongce Wu, Shukuan Lin, Xiaoxue Shao, Peng Zhang, and Jianzhong Qiao. 2023. QPGCN: graph convolutional network with a quadratic polynomial filter for overcoming over-smoothing. Appl. Intell., 53(6):7216-7231."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Weiwen Xu, Yang Deng, Huihui Zhang, Deng Cai, and Wai Lam. 2021a. Exploiting reasoning chains for multi-hop science question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 1143-1156.",
+            "Weiwen Xu, Huihui Zhang, Deng Cai, and Wai Lam. 2021b. Dynamic semantic graph construction and reasoning for explainable multi-hop science question answering. In Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021, volume ACL/IJCNLP 2021 of Findings of ACL, pages 1044-1056. Association for Computational Linguistics.",
+            "Rui Yang, Wenrui Dai, Chenglin Li, Junni Zou, and Hongkai Xiong. 2023. Tackling over-smoothing in graph convolutional networks with em-based joint topology optimization and node classification. IEEE Trans. Signal Inf. Process. over Networks, 9:123-139.",
+            "Yue Zhang, Quan Guo, and Parisa Kordjamshidi. 2021. Towards navigation by reasoning over spatial configurations. CoRR, abs/2105.06839.",
+            "Yue Zhang and Parisa Kordjamshidi. 2022. Explicit object relation alignment for vision and language navigation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 322-331. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            493
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "text_level": 1,
+        "bbox": [
+            512,
+            517,
+            596,
+            533
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Preliminary of Tensor Product Representation",
+        "text_level": 1,
+        "bbox": [
+            512,
+            543,
+            811,
+            575
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Consider the implicit relation hidden in the sentence \"X is over K.\" Our method decomposes this sentence into two groups of role  $r \\in \\mathcal{R}$  and filler  $f \\in \\mathcal{F}$  symbol components, i.e.,  $\\mathcal{R} = \\{X, K\\}$  and  $\\mathcal{F} = \\{above, below\\}$ . The role and filler symbols are projected into role  $V_{\\mathcal{R}}$  and filler  $V_{\\mathcal{F}}$  vector space, respectively. The TPR of these symbol structures is defined as the tensor  $\\mathbf{T}$  within the vector space  $V_{\\mathcal{R}} \\otimes V_{\\mathcal{F}}$ , where  $\\otimes$  denotes the tensor product (equivalent to outer product when role and filler are one-dimension vectors). In this example, we can drive the equation of writing our example to  $\\mathbf{T}$ :",
+        "bbox": [
+            510,
+            583,
+            882,
+            776
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {T} = f _ {\\text {a b o v e}} \\otimes r _ {X} + f _ {\\text {b e l o w}} \\otimes r _ {K} \\tag {16} \\\\ = f _ {a b o v e} r _ {X} ^ {T} + f _ {b e l o w} r _ {K} ^ {T} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            791,
+            880,
+            827
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The tensor product  $\\otimes$  acts as a binding operator to bind the corresponding role and filler.",
+        "bbox": [
+            512,
+            839,
+            880,
+            870
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The tensor inner product serves as the unbinding operator in TPR, which is employed to reconstruct the previously reserved fillers from  $\\mathbf{T}$ . Given the",
+        "bbox": [
+            512,
+            871,
+            880,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "6470",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/af0ce1c260a15cc495fce2e55fa86bcc501806f469d3428f294746ecda8eb874.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Experiment</td><td>Dataset</td><td>epochs</td><td>batch size</td><td>PLM learning rate</td><td>non-PLM learning rate</td></tr><tr><td>BERT</td><td>StepGame</td><td>22</td><td>32</td><td>2e-05</td><td>-</td></tr><tr><td>ALBERT</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>-</td></tr><tr><td>RoBERTa</td><td>StepGame</td><td>16</td><td>32</td><td>6e-06</td><td>-</td></tr><tr><td>DepWiNet(BERT)</td><td>StepGame</td><td>17</td><td>32</td><td>2e-05</td><td>1e-04</td></tr><tr><td>DepWiNet(ALBERT)</td><td>StepGame</td><td>22</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(RoBERTa)</td><td>StepGame</td><td>19</td><td>32</td><td>6e-06</td><td>1e-04</td></tr><tr><td>DepWiNet(BERT)</td><td>SPARTUN</td><td>9</td><td>16</td><td>8e-06</td><td>1e-04</td></tr><tr><td>DepWiNet(BERT)</td><td>ReSQ(with SPARTUN supervision)</td><td>4</td><td>16</td><td>5e-05</td><td>4e-05</td></tr><tr><td>DepWiNet(BERT)</td><td>ReSQ(without SPARTUN supervision)</td><td>4</td><td>16</td><td>5e-05</td><td>4e-05</td></tr><tr><td>DepWiNet(w/o LDC)</td><td>StepGame</td><td>15</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/o NMI &amp; LDC)</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/o SRR &amp; LDC)</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ DepWiGNNmean)</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ DepWiGNNmaxpooling)</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ GCN1-5)</td><td>StepGame</td><td>(16,12,10,19,14)</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ GraphConv1-5)</td><td>StepGame</td><td>(14,14,16,20,15)</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ GAT1-5)</td><td>StepGame</td><td>(16,13,16,16,17)</td><td>32</td><td>6e-06</td><td>5e-05</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            884,
+            343
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Hyperparameters and setup information for each experiment.",
+        "bbox": [
+            260,
+            347,
+            734,
+            363
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "unbinding vector  $u_{K}$ , we can retrieve the stored filler  $f_{below}$  from the  $\\mathbf{T}$ . In the most ideal cases, if the role vectors are orthogonal to each other,  $u_{K}$  equals  $r_{K}$ . When  $\\mathbf{T} \\in \\mathbb{R}^2$ , it can be expressed as the matrix multiplication.:",
+        "bbox": [
+            112,
+            376,
+            489,
+            458
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {T} \\cdot u _ {K} = \\left(f _ {a b o v e} r _ {X} ^ {T} + f _ {b e l o w} r _ {K} ^ {T}\\right) \\cdot r _ {K} \\tag {17} \\\\ = \\alpha f _ {b e l o w} \\propto f _ {b e l o w} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            142,
+            467,
+            487,
+            506
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Hyperparameter Settings",
+        "text_level": 1,
+        "bbox": [
+            114,
+            518,
+            371,
+            535
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Detailed hyperparameter settings for each experiment are provided in Table 5. Epoch numbers for DepWiNet with the three classical GNNs are grouped (in order) for simplicity.",
+        "bbox": [
+            112,
+            543,
+            489,
+            609
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "6471",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/DepWiGNN_ A Depth-wise Graph Neural Network for Multi-hop Spatial Reasoning in Text/82978807-d92d-4dd8-83a8-263f59cfaa61_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55c8e623fa97d41083eb38a9c70506a7153ea560e540ac86ce2a3080643fa65d
+size 820483

2023/DepWiGNN_ A Depth-wise Graph Neural Network for Multi-hop Spatial Reasoning in Text/full.md

@@ -0,0 +1,343 @@
+# DepWiGNN: A Depth-wise Graph Neural Network for Multi-hop Spatial Reasoning in Text *
+
+Shuaiyi Li $^{1}$ , Yang Deng $^{2,\dagger}$ , Wai Lam $^{1}$
+
+<sup>1</sup> The Chinese University of Hong Kong, <sup>2</sup> National University of Singapore li.shuaiyi@link.cuhk.edu.hk, ydeng@nus.edu.sg, wlam@se.cuhk.edu.hk
+
+# Abstract
+
+Spatial reasoning in text plays a crucial role in various real-world applications. Existing approaches for spatial reasoning typically infer spatial relations from pure text, which overlook the gap between natural language and symbolic structures. Graph neural networks (GNNs) have showcased exceptional proficiency in inducing and aggregating symbolic structures. However, classical GNNs face challenges in handling multi-hop spatial reasoning due to the over-smoothing issue, i.e., the performance decreases substantially as the number of graph layers increases. To cope with these challenges, we propose a novel Depth-Wise Graph Neural Network (DepWiGNN). Specifically, we design a novel node memory scheme and aggregate the information over the depth dimension instead of the breadth dimension of the graph, which empowers the ability to collect long dependencies without stacking multiple layers. Experimental results on two challenging multi-hop spatial reasoning datasets show that DepWiGNN outperforms existing spatial reasoning methods. The comparisons with the other three GNNs further demonstrate its superiority in capturing long dependency in the graph.
+
+# 1 Introduction
+
+Spatial reasoning in text is crucial and indispensable in many areas, e.g., medical domain (Datta et al., 2020; Massa et al., 2015), navigations (Zhang et al., 2021; Zhang and Kordjamshidi, 2022; Chen et al., 2019) and robotics (Luo et al., 2023; Venkatesh et al., 2021). It has been demonstrated to be a challenging problem for both modern pretrained language models (PLMs) (Mirzaee et al., 2021; Deng et al., 2023a) and large language models (LLMs) like ChatGPT (Bang et al., 2023). However, early textual spatial reasoning datasets, e.g.,
+
+![](images/93f34138c5657851645495012225cbbb2be40164e2ca36b03e02f56c190b8c8c.jpg)  
+Figure 1: An example of multihop spatial reasoning in text from the StepGame dataset (Shi et al., 2022).
+
+bAbI (Weston et al., 2016), suffer from the issue of over-simplicity and, therefore, is not qualified for revealing real textual spatial reasoning scenario. Recently, researchers propose several new benchmarks (Shi et al., 2022; Mirzaee and Kordjamshidi, 2022; Mirzaee et al., 2021) with an increased level of complexity, which involve more required reasoning steps, enhanced variety of spatial relation expression, and more. As shown in Figure 1, 4 steps of reasoning are required to answer the question, and spatial relation descriptions and categories are diverse.
+
+To tackle the problem of multi-hop spatial reasoning, Shi et al. (2022) propose a recurrent memory network based on Tensor Product Representation (TPR) (Schlag and Schmidhuber, 2018a), which mimics the step-by-step reasoning by iteratively updating and removing episodes from memory. Specifically, TPR encodes symbolic knowledge hidden in natural language into distributed vector space to be used for deductive reasoning. Despite the effectiveness of applying TPR memory, the performance of this model is overwhelmed by modern PLMs (Mirzaee and Kordjamshidi, 2022). Moreover, these works typically overlook the gap between natural language and symbolic relations.
+
+Graph Neural Neural Networks (GNNs) have been considerably used in multi-hop reasoning (Xu et al., 2021b; Chen et al., 2020b; Qiu et al., 2019).
+
+These methods often treat a single graph convolutional layer of node information propagation (node to its immediate neighbors) in GNN as one step of reasoning and expand it to multi-hop reasoning by stacking multiple layers. However, increasing the number of graph convolutional layers in deep neural structures can have a detrimental effect on model performance (Li et al., 2018). This phenomenon, known as the over-smoothing problem, occurs because each layer of graph convolutions causes adjacent nodes to become more similar to each other. This paradox poses a challenge for multi-hop reasoning: although multiple layers are needed to capture multi-hop dependencies, implementing them can fail to capture these dependencies due to the over-smoothing problem. Furthermore, many chain-finding problems, e.g., multihop spatial reasoning, only require specific depth path information to solve a single question and do not demand full breadth aggregation for all neighbors (Figure 1). Nevertheless, existing methods (Palm et al., 2018; Xu et al., 2021b; Chen et al., 2020b; Deng et al., 2022) for solving this kind of problem usually lie in the propagation conducted by iterations of breadth aggregation, which brings superfluous and irrelevant information that may distract the model from the key information.
+
+In light of these challenges, we propose a novel graph-based method, named Depth-Wise Graph Neural Network (DepWiGNN), which operates over the depth instead of the breadth dimension of the graph to tackle the multi-hop spatial reasoning problem. It introduces a novel node memory implementation that only stores depth path information between nodes by applying the TPR technique. Specifically, it first initializes the node memory by filling the atomic information (spatial relation) between each pair of directly connected nodes, and then collects the relation between two indirectly connected nodes via depth-wisely retrieving and aggregating all atomic spatial relations reserved in the memory of each node in the path. The collected long-dependency information is further stored in the source node memory in the path and can be retrieved freely if the target node is given. Unlike typical existing GNNs (Morris et al., 2019; Velickovic et al., 2017; Hamilton et al., 2017; Kipf and Welling, 2017), DepWiGNN does not need to be stacked to gain long relationships between two distant nodes and, hence, is immune to the over-smoothing problem. Moreover, instead of
+
+aimlessly performing breadth aggregation on all immediate neighbors, it selectively prioritizes the key path information.
+
+Experiments on two challenging multi-hop spatial reasoning datasets show that DepWiGNN not only outperforms existing spatial reasoning methods, but also enhances the spatial reasoning capability of PLMs. The comparisons with three GNNs verify that DepWiGNN surpasses classical graph convolutional layers in capturing long dependencies by a noticeable margin without harming the performance of short dependency collection.
+
+Overall, our contributions are threefold:
+
+- We propose a novel graph-based method, DepWiGNN, to perform propagation over the depth dimension of a graph, which can capture long dependency without the need of stacking layers and avoid the issue of over-smoothing.  
+- We implement a novel node memory scheme, which takes advantage of TPR mechanism, enabling convenient memory updating and retrieval operations through simple arithmetic operations instead of neural layers.  
+- DepWiGNN excels in multi-hop spatial reasoning tasks, surpassing existing methods in experimental evaluations on two challenging datasets. Besides, comparisons with three other GNNs highlight its superior ability to capture long dependencies within the graph. Our code will be released via https://github.com/Syon-Li/DepWiGNN.
+
+# 2 Related works
+
+Spatial Reasoning In Text has experienced a thriving development in recent years, supported by several benchmark datasets. Weston et al. (2016) proposes the bAbI project, which contains 20 QA tasks, including one focusing on textual spatial reasoning. However, some issues exist in bAbI, such as data leakage, overly short reasoning steps, and monotony of spatial relation categories and descriptions, which makes it fails to reflect the intricacy of spatial reasoning in natural language (Shi et al., 2022). Targeting these shortages, StepGame (Shi et al., 2022) expands the spatial relation types, the diversity of relation descriptions, and the required reasoning steps. Besides, SPARTQA (Mirzaee et al., 2021) augments the number of question types from one to four: find relation (FR), find blocks (FB), choose objects (CO), and yes/no (YN), while
+
+SPARTUN (Mirzaee and Kordjamshidi, 2022) only has two question types (FR, YN) built with broader coverage of spatial relation types.
+
+Tensor Product Representation (TPR) (Schlag and Schmidhuber, 2018a) is a mechanism to encode symbolic knowledge into a vector space, which can be applied to various natural language reasoning tasks (Huang et al., 2018; Chen et al., 2020a). For example, Schlag and Schmidhuber (2018a) perform reasoning by deconstructing the language into combinatorial symbolic representations and binding them using Third-order TPR, which can be further combined with RNN to improve the model capability of making sequential inference (Schlag et al., 2021). Shi et al. (2022) used a paragraph-level, TPR memory-augmented way to implement complex multi-hop spatial reasoning. However, existing methods typically apply TPR to pure text, which neglects the gap between natural language and symbolic structures.
+
+Graph Neural Networks (GNNs) (Morris et al., 2019; Velickovic et al., 2017; Hamilton et al., 2017; Kipf and Welling, 2017) have been demonstrated to be effective in inducing and aggregating symbolic structures on other multi-hop question answering problems (Cao et al., 2019; Fang et al., 2020; Huang and Yang, 2021; Heo et al., 2022; Xu et al., 2021a; Deng et al., 2023b). In practice, the required number of graph layers grows with the multi-hop dependency between two distant nodes (Wang et al., 2021; Hong et al., 2022), which inevitably encounters the problem of oversmoothing (Li et al., 2018). Some researchers have conducted studies on relieving this problem (Wu et al., 2023; Min et al., 2020; Huang and Li, 2023; Yang et al., 2023; Liu et al., 2023; Koishekenov, 2023; Song et al., 2023). However, these methods are all breadth-aggregation-based, i.e., they only posed adjustments on breadth aggregation like improving the aggregation filters, scattering the aggregation target using the probabilistic tool, etc., but never jumping out it. In this work, we investigate a depth-wise aggregation approach that captures long-range dependencies across any distance without the need to increase the model depth.
+
+# 3 Method
+
+# 3.1 Problem Definition
+
+Following previous studies on spatial reasoning in text (Mirzaee et al., 2021; Shi et al., 2022; Mirzaee
+
+and Kordjamshidi, 2022), we define the problem as follows: Given a story description  $S$  consisting of multiple sentences, the system aims to answer a question  $Q$  based on the story  $S$  by selecting a correct answer from the fixed set of given candidate answers regarding the spatial relations.
+
+# 3.2 The DepWiNet
+
+As presented in Figure 2, the overall framework named DepWiNet consists of three modules: the entity representation extraction module, the DepWiGNN reasoning module, and the prediction module. The entity representation extraction module provides comprehensive entity embedding that is used in the reasoning module. A graph with recognized entities as nodes is constructed after obtaining entity representations, which later will be fed into DepWiGNN. The DepWiGNN reasons over the constructed graph and updates the node embedding correspondingly. The final prediction module adds the entity embedding from DepWiGNN to the embeddings from the first extraction module and applies a single step of attention (Vaswani et al., 2017) to generate the final result.
+
+Entity Representation Extraction Module We leverage PLMs to extract entity representations<sup>1</sup>. The model takes the concatenation of the story  $S$  and the question  $Q$  as the input and output embeddings of each token. The output embeddings are further projected using a single linear layer.
+
+$$
+\hat {\mathbf {S}}, \hat {\mathbf {Q}} = \mathbf {P L M} (S, Q) \tag {1}
+$$
+
+$$
+\hat {\mathbf {S}} _ {\alpha} = \hat {\mathbf {S}} \mathbf {W} _ {\alpha} ^ {T} \quad \hat {\mathbf {Q}} _ {\alpha} = \hat {\mathbf {Q}} \mathbf {W} _ {\alpha} ^ {T} \tag {2}
+$$
+
+where  $\mathbf{PLM}(\cdot)$  denotes the PLM-based encoder, e.g., BERT (Devlin et al., 2019).  $\hat{\mathbf{S}}_{\alpha}\in \mathbb{R}^{r_S\times d_h}$ ,  $\hat{\mathbf{Q}}_{\alpha}\in \mathbb{R}^{r_Q\times d_h}$  denotes the list of projected tokens of size  $d_h$  in the story and question, and  $\mathbf{W}_{\alpha}\in \mathbb{R}^{d_h\times d_h}$  is the shared projection matrix. The entity representation is just the mean pooling of all token embeddings belonging to that entity.
+
+Graph Construction The entities are first recognized from the input by employing rule-based entity recognition. Specifically, in StepGame (Shi et al., 2022), the entities are represented by a single capitalized letter, so we only need to locate all single capitalized letters. For SPARTUN
+
+![](images/b5117fa614d2e7653a82fb581fdd79f1966f76c00f16390fee945063b2b92bdf.jpg)  
+Figure 2: The DepWiNet framework. The entity representations are first extracted from the entity representation extraction module (left), and then a homogeneous graph is constructed based on the entity embeddings and fed into the DepWiNet reasoning module. The DepWiNet depth-wisely aggregates information for all indirectly connected node pairs, and stores it in node memories. The updated node embeddings are then passed to the prediction module.
+
+and ReSQ, we use nthk RegexpParser $^2$  with self-designed grammars $^3$  to recognize entities. Entities and their embeddings are treated as nodes of the graph, and an edge between two entities exists if and only if the two entities co-occur in the same sentence. We also add an edge feature for each edge.
+
+$$
+E _ {i j} = \left\{ \begin{array}{l l} 0 & \text {i f i = = j ,} \\ \mathbf {e} _ {[ C L S ]} & \text {o r t h e r w i s e .} \end{array} \right. \tag {3}
+$$
+
+If two entities are the same (self-loop), the edge feature is just a zero tensor with a size  $d_h$ ; otherwise, it is the sequence's last layer hidden state of the [CLS] token. The motivation behind treating [CLS] token as an edge feature is that it can help the model better understand the atomic relation  $(k = 1)$  between two nodes (entities) so as to facilitate later depth aggregation. A straightforward justification can be found in Table 1, where all three PLMs achieve very high accuracy at  $k = 1$  cases by using the [CLS] token, which demonstrates that the [CLS] token favors the understanding of the atomic relation.
+
+DepWiGNN Module The DepWiGNN module (middle part in Figure 2) is responsible for multi-hop reasoning and can collect arbitrarily distant dependency without layer stacking.
+
+$$
+\hat {V} = \operatorname {D e p} \operatorname {W i G N N} (\mathcal {G}; V, E) \tag {4}
+$$
+
+where  $V \in \mathbb{R}^{|V| \times d_h}$  and  $E \in \mathbb{R}^{|E| \times d_h}$  are the nodes and edges of the graph. It comprises three components: node memory initialization, long dependency collection, and spatial relation retrieval. Details of these three components will be discussed in Section 3.3.
+
+Prediction Module The node embedding updated in DepWiGNN is correspondingly added to entity embedding extracted from PLM.
+
+$$
+\hat {Z} _ {i} = \left\{ \begin{array}{l l} Z _ {i} + \hat {V} _ {i d x (i)} & \text {i f i - t h t o k e n} \in \hat {V} \\ Z _ {i} & \text {o r t h e r w i s e .} \end{array} \right. \tag {5}
+$$
+
+where  $Z_{i} = [\hat{S};\hat{Q}]$  is the ith token embedding from PLM and  $\hat{V}$  denotes the updated entity representation set.  $idx(i)$  is the index of  $i$ -th token in the graph nodes.
+
+Then, the sequence of token embeddings in the question and story are extracted separately to perform attention mechanism (Vaswani et al., 2017) with the query to be the sequence of question tokens embeddings  $\hat{Z}_{\hat{\mathbf{Q}}}$ , key and value to be the sequence of story token embeddings  $\hat{Z}_{\hat{\mathbf{S}}}$ .
+
+$$
+R = \sum_ {i = 0} ^ {r _ {Q}} \left(\operatorname {s o f t m a x} \left(\frac {\hat {Z} _ {\hat {\mathbf {Q}}} \hat {Z} _ {\hat {\mathbf {S}}} ^ {T}}{\sqrt {d _ {h}}}\right) \hat {Z} _ {\hat {\mathbf {S}}}\right) _ {i} \tag {6}
+$$
+
+$$
+C = F F N (\operatorname {l a y e r n m} (R)) \tag {7}
+$$
+
+where layernm means layer normalization and  $C$  is the final logits. The result embeddings are summed over the first dimension, layernormed, and
+
+![](images/9282c7cc833d54129685fc4ca0316422997a1261d856cb969918d36c9b338705.jpg)  
+Figure 3: The illustration of DepWiGNN.
+
+fed into a 3-layer feedforward neural network to acquire the final result. The overall framework is trained in an end-to-end manner to minimize the cross-entropy loss between the predicted candidate answer probabilities and the ground-truth answer.
+
+# 3.3 Depth-wise Graph Neural Network
+
+As illustrated in Figure 3, we introduce the proposed graph neural network, called DepWiGNN, i.e., the operation  $\hat{V} = \mathsf{DepWiGNN}(\mathcal{G};V,E)$ . Unlike existing GNNs (e.g. (Morris et al., 2019; Velickovic et al., 2017; Hamilton et al., 2017)), which counts the one-dimension node embedding itself as its memory to achieve the function of information reservation, updating, and retrieval, DepWiGNN employs a novel two-dimension node memory implementation approach that takes the advantage of TPR mechanism allowing the updating and retrieval operations of the memory to be conveniently realized by simple arithmetic operations like plus, minus and outer product. This essentially determines that the information propagation between any pair of nodes with any distance in the graph can be accomplished without the demand to iteratively apply neural layers.
+
+Node Memory Initialization At this stage, the node memories of all nodes are initialized with the relations to their immediate neighbors. In the multi-hop spatial reasoning cases, the relations will be the atomic spatial orientation relations (which only needs one hop) of the destination node relative to the source node. For example, "X is to the left of K and is on the same horizontal plane." We follow the TPR mechanism (Smolensky, 1990), which uses outer product operation to bind roles and fillers<sup>4</sup>. In this work, the calculated spatial vectors are considered to be the fillers. They will be bound with corresponding node embeddings and stored in the two-dimensional node memory. Explicitly, we first
+
+acquire spatial orientation filler  $f_{ij} \in \mathbb{R}^{d_h}$  by using a feedforward network, the input to FFN is concatenation in the form  $[V_i, E_{ij}, V_j]$ .  $V_i$  represents  $i$ -th node embedding.
+
+$$
+f _ {i j} = F F N \left(\left[ V _ {i}, E _ {i j}, V _ {j} \right]\right) \tag {8}
+$$
+
+$$
+M _ {i} = \sum_ {V _ {j} \in N (V _ {i})} f _ {i j} \otimes V _ {j} \tag {9}
+$$
+
+The filler is bound together with the corresponding destination node  $V_{j}$  using the outer product. The initial node memory  $M_{i}$  for node  $V_{i}$  is just the summation of all outer product results of the fillers and corresponding neighbors (left part in Figure 3).
+
+Long Dependency Collection We discuss how the model collects long dependencies in this section. Since the atomic relations are bound by corresponding destinations and have already been contained in the node memories, we can easily unbind all the atomic relation fillers in a path using the corresponding atomic destination node embedding (middle part in Figure 3). For each pair of indirectly connected nodes, we first find the shortest path between them using breadth-first search (BFS). Then, all the existing atomic relation fillers along the path are unbound using the embedding of each node in the path (Eq.10).
+
+$$
+\hat {f} _ {p _ {i} \left(p _ {i + 1}\right)} = M _ {p _ {i}} V _ {p _ {i + 1}} ^ {T} \tag {10}
+$$
+
+$$
+\hat {\mathbf {F}} = \operatorname {A g g r e g a t o r} (\mathbf {F}) \tag {11}
+$$
+
+$$
+f _ {p _ {0} p _ {n}} = \operatorname {l a y e r n m} (F F N (\hat {\mathbf {F}}) + \hat {\mathbf {F}}) \tag {12}
+$$
+
+$$
+M _ {p _ {0}} = M _ {p _ {0}} + f _ {p _ {0} p _ {n}} \otimes V _ {p _ {n}} ^ {T} \tag {13}
+$$
+
+where  $p_i$  denotes the  $i$ -th element in the path and  $\mathbf{F} = [\hat{f}_{p_0p_1};\dots;\hat{f}_{p_{n-1}p_n}]$  is the retrieved filler set along the path. The collected relation fillers are aggregated using a selected depth aggregator like LSTM (Hochreiter and Schmidhuber, 1997), etc, and passed to a feedforward neural network to reason the relation filler between the source and destination node in the path (Eq.12). The result spatial filler is then bound with the target node embedding and added to the source node memory (Eq.13). In this way, each node memory will finally contain the spatial orientation information to every other connected node in the graph.
+
+Spatial Relation Retrieval After the collection process is completed, every node memory contains spatial relation information to all other connected nodes in the graph. Therefore, we can conveniently retrieve the spatial information from a source node
+
+to a target node by unbinding the spatial filler from the source node memory (right part in Figure 3) using a self-determined key. The key can be the target node embedding itself if the target node can be easily recognized from the question, or some computationally extracted representation from the sequence of question token embeddings if the target node is hard to discern. We use the key to unbind the spatial relation from all nodes' memory and pass the concatenation of it with source and target node embeddings to a multilayer perceptron to get the updated node embeddings.
+
+$$
+\hat {f} _ {i [ \text {k e y} ]} = M _ {i} V _ {[ \text {k e y} ]} ^ {T} \tag {14}
+$$
+
+$$
+\hat {V} _ {i} = F F N ([ V _ {i}, \mathbf {l a y e r n m} (f _ {i [ k e y ]}), V _ {[ k e y ]} ]) \tag {15}
+$$
+
+The updated node embeddings are then passed to the prediction module to get the final result.
+
+# 4 Experiments
+
+# 4.1 Experimental Setups
+
+Datasets & Evaluation Metrics We investigated our model on StepGame (Shi et al., 2022) and ReSQ (Mirzaee and Kordjamshidi, 2022) datasets, which were recently published for multi-hop spatial reasoning. StepGame is a synthetic textual QA dataset that has a number of relations  $(k)$  ranging from 1 to 10. In particular, we follow the experimental procedure in the original paper (Shi et al., 2022), which utilized the TrainVersion of the dataset containing 10,000/1,000 training/validation clean samples for each  $k$  value from 1 to 5 and 10,000 noisy test examples for  $k$  value from 1 to 10. The test set contains three kinds of distraction noise: disconnected, irrelevant, and supporting. ReSQ is a crowdsourcing benchmark that includes only Yes/No questions with 1008, 333, and 610 examples for training, validating, and testing respectively. Since it is human-generated, it can reflect the natural complexity of real-world spatial description. Following the setup in (Shi et al., 2022; Mirzaee and Kordjamshidi, 2022), we report the accuracy on corresponding test sets for all experiments.
+
+Baselines For StepGame, we select all traditional reasoning models used in the original paper (Shi et al., 2022) and three PLMs, i.e., BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and ALBERT (Lan et al., 2020) as our baselines. For ReSQ, we also follow the experiment setting described in (Mirzaee and Kordjamshidi, 2022),
+
+which used BERT with or without further synthetic supervision as baselines.
+
+Implementation Details For all experiments, we use the base version of corresponding PLMs, which has 768 embedding dimensions. The model was trained in an end-to-end manner using Adam optimizer (Kingma and Ba, 2015). The training was stopped if, up to 3 epochs, there is no improvement greater than 1e-3 on the cross-entropy loss for the validation set. We also applied a Pytorch training scheduler that reduces the learning rate with a factor of 0.1 if the improvement of cross-entropy loss on the validation set is lower than 1e-3 for 2 epochs. In terms of the determination of the key in the Spatial Relation Retrieval part, we used the target node embedding for StepGame since it can be easily recognized, and we employed a single linear layer to extract the key representation from the sum-aggregated question token embeddings for ReSQ. In the StepGame experiment, we fine-tune the model on the training set and test it on the test set. For ReSQ, we follow the procedure in (Mirzaee and Kordjamshidi, 2022) to test the model on ReSQ with or without further supervision from SPARTUN (Mirzaee and Kordjamshidi, 2022). Unless specified, all the experiments use LSTM (Hochreiter and Schmidhuber, 1997) as depth aggregator by default. The detailed hyperparameter settings are given in the Appendix B.
+
+# 4.2 Overall Performance
+
+Table 1 and 2 report the experiment results on StepGame and ReSQ respectively. As shown in Table 1, PLMs overwhelmingly outperform all the traditional reasoning models and the proposed DeepWiNet overtakes the PLMs by a large margin, especially for the cases with greater  $k$  where the multi-hop reasoning capability plays a more important role. This aligns with the characteristics of our model architecture that the aggregation focuses on the depth dimension, which effectively avoids the problem of over-smoothing and the mixture of redundant information from the breadth aggregation. Despite the model being only trained on clean distraction-noise-absent samples with  $k$  from 1 to 5, it achieves impressive performance on the distraction-noise-present test data with  $k$  value from 6 to 10, demonstrating the more advanced generalization capability of our model. Moreover, our method also entails an improvement for the examples with lower  $k$  values for BERT and AL
+
+<table><tr><td>Method</td><td>k=1</td><td>k=2</td><td>k=3</td><td>k=4</td><td>k=5</td><td>k=6</td><td>k=7</td><td>k=8</td><td>k=9</td><td>k=10</td><td>Mean(k=1-5)</td><td>Mean(k=6-10)</td></tr><tr><td>RN (Santoro et al., 2017)</td><td>22.64</td><td>17.08</td><td>15.08</td><td>12.84</td><td>11.52</td><td>11.12</td><td>11.53</td><td>11.21</td><td>11.13</td><td>11.34</td><td>15.83</td><td>11.27</td></tr><tr><td>RRN (Palm et al., 2018)</td><td>24.05</td><td>19.98</td><td>16.03</td><td>13.22</td><td>12.31</td><td>11.62</td><td>11.40</td><td>11.83</td><td>11.22</td><td>11.69</td><td>17.12</td><td>11.56</td></tr><tr><td>UT (Dehghani et al., 2019)</td><td>45.11</td><td>28.36</td><td>17.41</td><td>14.07</td><td>13.45</td><td>12.73</td><td>12.11</td><td>11.40</td><td>11.41</td><td>11.74</td><td>23.68</td><td>11.88</td></tr><tr><td>STM (Le et al., 2020)</td><td>53.42</td><td>35.96</td><td>23.03</td><td>18.45</td><td>15.14</td><td>13.80</td><td>12.63</td><td>11.54</td><td>11.30</td><td>11.77</td><td>29.20</td><td>12.21</td></tr><tr><td>TPR-RNN (Schlag and Schmidhuber, 2018b)</td><td>70.29</td><td>46.03</td><td>36.14</td><td>26.82</td><td>24.77</td><td>22.25</td><td>19.88</td><td>15.45</td><td>13.01</td><td>12.65</td><td>40.81</td><td>16.65</td></tr><tr><td>TP-MANN (Shi et al., 2022)</td><td>85.77</td><td>60.31</td><td>50.18</td><td>37.45</td><td>31.25</td><td>28.53</td><td>26.45</td><td>23.67</td><td>22.52</td><td>21.46</td><td>52.99</td><td>24.53</td></tr><tr><td>BERT (Mirzaee and Kordjamshidi, 2022)</td><td>98.53</td><td>93.40</td><td>91.19</td><td>66.98</td><td>54.57</td><td>48.59</td><td>42.81</td><td>37.98</td><td>34.16</td><td>33.62</td><td>80.93</td><td>39.43</td></tr><tr><td>RoBERTa* (Mirzaee and Kordjamshidi, 2022)</td><td>98.68</td><td>97.05</td><td>95.66</td><td>79.59</td><td>74.89</td><td>70.67</td><td>66.01</td><td>61.42</td><td>57.20</td><td>54.53</td><td>89.17</td><td>61.93</td></tr><tr><td>ALBERT* (Mirzaee and Kordjamshidi, 2022)</td><td>98.56</td><td>97.56</td><td>96.08</td><td>90.01</td><td>83.33</td><td>77.24</td><td>71.57</td><td>64.47</td><td>60.02</td><td>56.18</td><td>93.11</td><td>65.90</td></tr><tr><td>DepWiNet (BERT)</td><td>98.44</td><td>96.57</td><td>94.75</td><td>81.41</td><td>70.68</td><td>62.80</td><td>57.46</td><td>49.14</td><td>45.56</td><td>44.01</td><td>88.37+9.2%</td><td>51.79+31.3%</td></tr><tr><td>DepWiNet (RoBERTa)</td><td>98.70</td><td>96.69</td><td>95.54</td><td>79.50</td><td>74.94</td><td>70.37</td><td>66.22</td><td>61.89</td><td>58.01</td><td>56.89</td><td>89.07-0.1%</td><td>62.68+1.2%</td></tr><tr><td>DepWiNet (ALBERT)</td><td>98.59</td><td>97.13</td><td>95.87</td><td>91.96</td><td>87.50</td><td>82.62</td><td>77.80</td><td>71.40</td><td>67.55</td><td>64.98</td><td>94.21+1.2%</td><td>72.87+10.6%</td></tr></table>
+
+Table 1: Experimental results on StepGame. * denotes the implementation of the same model as BERT (Mirzaee and Kordjamshidi, 2022) with RoBERTa and ALBERT. The other results are reported in (Shi et al., 2022).  
+
+<table><tr><td>Method</td><td>SynSup</td><td>ReSQ</td></tr><tr><td>Majority Baseline</td><td>-</td><td>50.21</td></tr><tr><td>BERT</td><td>-</td><td>57.37</td></tr><tr><td>BERT</td><td>SPARTQA-AUTO</td><td>55.08</td></tr><tr><td>BERT</td><td>StepGame</td><td>60.14</td></tr><tr><td>BERT</td><td>SPARTUN-S</td><td>58.03</td></tr><tr><td>BERT</td><td>SPARTUN</td><td>63.60</td></tr><tr><td>DepWiNet (BERT)</td><td>-</td><td>63.30</td></tr><tr><td>DepWiNet (BERT)</td><td>SPARTUN</td><td>65.54</td></tr></table>
+
+Table 2: Experimental results on ReSQ, under the transfer learning setting (Mirzaee and Kordjamshidi, 2022).
+
+BERT and an ignorable decrease for RoBERTa, which attests to the innocence of our model to the few-hop reasoning tasks.
+
+Experimental results in Table 2 show that DeepWiNet reaches a new SOTA result on ReSQ for both cases where extra supervision from SPARTUN is present and absent. Excitingly, the performance of our model without extra supervision even overwhelms the performance of the BERT with extra supervision from SPARTQA-AUTO (Mirzaee et al., 2021), StepGame, and SPARTUN-S and approaches closely to the SPARTUN supervision case. All these phenomenons signify that our model has the potential to better tackle the natural intricacy of real-world spatial expressions.
+
+# 4.3 Ablation study
+
+We conduct ablation studies on the impact of the three components of DepWiGNN and different depth aggregators, as presented in Table 3.
+
+Impact of DepWiNet components The model performance experiences a drastic decrease, particularly for the mean score of  $k$  between 6 and 10, after the Long Dependency Collection (LDC) has been removed, verifying that this component serves a crucial role in the model. Note that the mean score for  $k(6-10)$  even drops below the ALBERT baseline (Table 1). This is reasonable as the LDC is directly responsible for collecting long dependencies. We further defunctionalized the Node Memory Initialization (NMI) and then Spatial Relation Retrieval (SRR) by setting the initial fillers (Eq.8) and key representation (Eq.14) to a random vector separately. Compared with the case where only LDC was removed, both of them lead to a further decrease in small and large  $k$  values.
+
+Impact of Different Aggregators The results show that the mean and max pooling depth aggregators fail to understand the spatial rule as achieved by LSTM. This may be caused by the relatively less expressiveness and generalization caliber of mean and max pooling operation.
+
+# 4.4 Comparisons with Different GNNs
+
+To certify our model's capacity of collecting long dependencies as well as its immunity to the oversmoothing problem, we contrast it with four graph neural networks, namely, GCN (Kipf and Welling, 2017), GraphConv (Morris et al., 2019), GAT (Velickovic et al., 2017) and GCNII (Chen et al., 2020c). We consider the cases with the number of layers varying from 1 to 5 and select the best performance to compare. The plot of the accuracy metric is reported in Figure 4 and the corresponding best
+
+<table><tr><td>Method</td><td>k=1</td><td>k=2</td><td>k=3</td><td>k=4</td><td>k=5</td><td>k=6</td><td>k=7</td><td>k=8</td><td>k=9</td><td>k=10</td><td>Mean(k=1-5)</td><td>Mean(k=6-10)</td></tr><tr><td>DepWiNet (ALBERT)</td><td>98,59</td><td>97.13</td><td>95.87</td><td>91.96</td><td>87.50</td><td>82.62</td><td>77.80</td><td>71.40</td><td>67.55</td><td>64.98</td><td>94.21</td><td>72.87</td></tr><tr><td>- w/o LDC</td><td>98.58</td><td>97.72</td><td>96.42</td><td>82.46</td><td>78.53</td><td>74.52</td><td>70.45</td><td>62.84</td><td>59.06</td><td>56.57</td><td>90.74-3.7%</td><td>64.69-11.2%</td></tr><tr><td>- w/o NMI &amp; LDC</td><td>98.48</td><td>96.88</td><td>94.83</td><td>79.14</td><td>75.46</td><td>70.51</td><td>67.19</td><td>61.88</td><td>57.65</td><td>55.02</td><td>88.96-5.6%</td><td>62.45-14.3%</td></tr><tr><td>- w/o SRR &amp; LDC</td><td>98.64</td><td>97.52</td><td>95.91</td><td>80.72</td><td>76.85</td><td>72.56</td><td>67.93</td><td>61.29</td><td>57.97</td><td>54.17</td><td>89.93-4.5%</td><td>62.78-13.8%</td></tr><tr><td>- w/ DepWiGNNmean</td><td>98.52</td><td>96.46</td><td>93.69</td><td>81.10</td><td>77.90</td><td>73.47</td><td>70.78</td><td>65.54</td><td>62.42</td><td>59.82</td><td>89.53-5.0%</td><td>66.41-8.9%</td></tr><tr><td>- w/ DepWiGNNmaxpooling</td><td>98.65</td><td>95.83</td><td>93.90</td><td>80.81</td><td>77.20</td><td>72.07</td><td>68.76</td><td>62.79</td><td>59.63</td><td>57.14</td><td>89.28-5.2%</td><td>64.08-12.1%</td></tr></table>
+
+Table 3: Ablation study. NMI, LDC, and SRR denotes the three stages in DepWiGNN, i.e., Node Memory Initialization, Long Dependency Collection, and Spatial Relation Retrieval, respectively.  
+
+<table><tr><td>Method</td><td>k=1</td><td>k=2</td><td>k=3</td><td>k=4</td><td>k=5</td><td>k=6</td><td>k=7</td><td>k=8</td><td>k=9</td><td>k=10</td><td>Mean(k=1-5)</td><td>Mean(k=6-10)</td></tr><tr><td>ALBERT</td><td>98.56</td><td>97.56</td><td>96.08</td><td>90.01</td><td>83.33</td><td>77.24</td><td>71.57</td><td>64.47</td><td>60.02</td><td>56.18</td><td>93.11</td><td>65.90</td></tr><tr><td>w/ GCN5 (Kipf and Welling, 2017)</td><td>98.62</td><td>97.72</td><td>96.49</td><td>82.01</td><td>76.70</td><td>71.60</td><td>67.34</td><td>60.49</td><td>56.11</td><td>52.65</td><td>90.31-3.01%</td><td>61.64-6.46%</td></tr><tr><td>w/ GraphConv2 (Morris et al., 2019)</td><td>98.66</td><td>97.73</td><td>96.48</td><td>82.37</td><td>78.57</td><td>74.47</td><td>69.91</td><td>63.04</td><td>60.01</td><td>56.58</td><td>90.76-2.52%</td><td>64.80-1.67%</td></tr><tr><td>w/ GAT1 (Velickovic et al., 2017)</td><td>98.67</td><td>97.34</td><td>95.59</td><td>81.47</td><td>78.39</td><td>73.79</td><td>70.17</td><td>63.52</td><td>60.64</td><td>57.33</td><td>90.29-3.03%</td><td>65.09-1.23%</td></tr><tr><td>w/ GCNII4 (Chen et al., 2020c)</td><td>98.56</td><td>97.61</td><td>96.36</td><td>83.31</td><td>79.22</td><td>74.88</td><td>70.64</td><td>63.64</td><td>60.59</td><td>56.99</td><td>91.01-2.25%</td><td>65.35+0.01%</td></tr><tr><td>w/ DepWiGNN</td><td>98.59</td><td>97.13</td><td>95.87</td><td>91.96</td><td>87.50</td><td>82.62</td><td>77.80</td><td>71.40</td><td>67.55</td><td>64.98</td><td>94.21+1.18%</td><td>72.87+10.58%</td></tr></table>
+
+Table 4: Comparisons with different GNNs. The subscripts represent the number of GNN layers. We select the best mean performance among layer number 1 to 5.
+
+![](images/d7e58ce4a4ad34b5cc46e1ade07941b6b47aab671665c5017a7fb0586e832ffc.jpg)  
+Figure 4: Impact of the layer number in different GNNs on StepGame. The solid and dashed lines denote the mean score of  $(k = 1 - 5)$  and  $(k = 6 - 10)$  respectively.
+
+cases are in Table 4. It is worth noting that almost all the baseline GNNs cause a reduction in the original PLM performance (Table 4). The reason for this may partially come from the breadth aggregation, which aggregates the neighbors round and round and leads to indistinguishability among entity embeddings such that the PLM reasoning process has been disturbed. The majority of baseline GNNs suffer from an apparent performance drop when increasing the number of layers (Figure 4), while our model consistently performs better and is not affected by the number of layers at all since it does not use breadth aggregation. Therefore, our model has immunity to over-smoothing problems.
+
+![](images/f56bfb1e39687845f8350abe77bd1573fa2c041f747dc0bdcfee147055f9092f.jpg)  
+Figure 5: Cases with distracting noisy from StepGame.
+
+In both small and large  $k$  cases, our model outperforms the best performance of all four GNNs (including GCNII, which is specifically designed for over-smoothing issues) by a large margin (Table 4), which serves as evidence of the superiority of our model in long dependency collection.
+
+# 4.5 Case study
+
+In this section, we present case studies to intuitively show how DepWiGNN mitigates the three kinds of distracting noise introduced in StepGame, namely, disconnected, irrelevant, and supporting.
+
+- The disconnected noise is the set of entities and relations that forms a new independent chain in the graph (Figure 5(a)). The node memories constructed in DepWiGNN contain spatial information about the nodes if and only if that node stays in the same connected component; otherwise, it
+
+has no information about the node as there is no path between them. Hence, in this case, for the questioned source node  $\mathbf{P}$ , its memory has no information for the disconnected noise  $\mathbf{T}$  and  $\mathbf{D}$ .
+
+- The irrelevant noise branches the correct reasoning chain out with new entities and relations but results in no alternative reasoning path (Figure 5(b)). Hence the irrelevant noise entities will not be included in the reasoning path between the source and destination, which means that it will not affect the destination spatial filler stored in the source node memory. In this case, when the key representation (embedding of entity  $\mathbf{E}$ ) is used to unbind the spatial filler from node memory of the source node  $\mathbf{P}$ , it obtains the filler  $f_{PE} = FFN(Aggregator([f_{PX}; f_{XQ}; f_{QE}])$ , which is intact to the irrelevant entity  $\mathbf{Y}$  and relation  $f_{xy}$  or  $f_{yx}$ .  
+- The supporting noise adds new entities and relations to the original reasoning chain that provides an alternative reasoning path (Figure 5(c)). DeepWiGNN is naturally exempted from this noise for two reasons: first, it finds the shortest path between two entities, therefore, will not include  $\mathbf{I}$  and  $\mathbf{M}$  in the path; Second, even if the longer path is considered, the depth aggregator should reach the same result as the shorted one since the source and destination are the same.
+
+# 5 Conclusion
+
+In this work, we introduce DepWiGNN, a novel graph-based method that facilitates depth-wise propagation in graphs, enabling effective capture of long dependencies while mitigating the challenges of over-smoothing and excessive layer stacking. Our approach incorporates a node memory scheme leveraging the TPR mechanism, enabling simple arithmetic operations for memory updating and retrieval, instead of relying on additional neural layers. Experiments on two recently released textual multi-hop spatial reasoning datasets demonstrate the superiority of DepWiGNN in collecting long dependencies over the other three typical GNNs and its immunity to the over-smoothing problem.
+
+# Limitation
+
+Unlike the classical GNNs, which use one-dimensional embedding as the node memory, our method applies a two-dimensional matrix-shaped
+
+node memory. This poses a direct increase in memory requirement. The system has to assign extra space to store a matrix with shape  $\mathbb{R}^{d_h\times d_h}$  for each node in the graph, which makes the method less scalable. However, it is a worthwhile trade-off because the two-dimensional node memory can potentially store  $d_{h} - 1$  number of spatial fillers bound with node embeddings while keeping the size fixed, and it does not suffer from information overloading as faced by the one-dimension memory since the spatial fillers can be straightforwardly retrieved from the memory.
+
+Another issue is the time complexity of finding the shortest path between every pair of nodes. For all experiments in this paper, since the edges do not have weights, the complexity is  $\mathrm{O}((\mathrm{n} + \mathrm{e})^{*}\mathrm{n})$  where  $\mathbf{n}$  and  $\mathbf{e}$  are the numbers of nodes and edges, respectively. However, things will get worse if the edges in the graph are weighted. We believe this is a potential future research direction for the improvement of the algorithm.
+
+# References
+
+Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. CoRR, abs/2302.04023.  
+Yu Cao, Meng Fang, and Dacheng Tao. 2019. BAG: bi-directional attention entity graph convolutional network for multi-hop reasoning question answering. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, pages 357-362.  
+Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. 2019. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 12538-12547. Computer Vision Foundation / IEEE.  
+Kezhen Chen, Qiuyuan Huang, Hamid Palangi, Paul Smolensky, Kenneth D. Forbus, and Jianfeng Gao. 2020a. Mapping natural-language problems to formal-language solutions using structured neural representations. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1566-1575. PMLR.  
+Kunlong Chen, Weidi Xu, Xingyi Cheng, Zou Xiaochuan, Yuyu Zhang, Le Song, Taifeng Wang, Yuan
+
+Qi, and Wei Chu. 2020b. Question directed graph attention network for numerical reasoning over text. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 6759-6768. Association for Computational Linguistics.  
+Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. 2020c. Simple and deep graph convolutional networks. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1725-1735. PMLR.  
+Surabhi Datta, Yuqi Si, Laritza Rodriguez, Sonya E. Shooshan, Dina Demner-Fushman, and Kirk Roberts. 2020. Understanding spatial language in radiology: Representation framework, annotation, and spatial relation extraction from chest x-ray reports using deep learning. J. Biomed. Informatics, 108:103473.  
+Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. 2019. Universal transformers. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.  
+Yang Deng, Shuaiyi Li, and Wai Lam. 2023a. Learning to ask clarification questions with spatial reasoning. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, pages 2113-2117.  
+Yang Deng, Yuexiang Xie, Yaliang Li, Min Yang, Wai Lam, and Ying Shen. 2022. Contextualized knowledge-aware attentive neural network: Enhancing answer selection with knowledge. ACM Trans. Inf. Syst., 40(1):2:1-2:33.  
+Yang Deng, Wenxuan Zhang, Weiwen Xu, Ying Shen, and Wai Lam. 2023b. Nonfactoid question answering as query-focused summarization with graph-enhanced multihop inference. IEEE Transactions on Neural Networks and Learning Systems.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.  
+Yuwei Fang, Siqi Sun, Zhe Gan, Rohit Pillai, Shuo-hang Wang, and Jingjing Liu. 2020. Hierarchical graph network for multi-hop question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, pages 8823-8838.
+
+William L. Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 1024-1034.  
+Yu-Jung Heo, Eun-Sol Kim, Woo Suk Choi, and Byoung-Tak Zhang. 2022. Hypergraph transformer: Weakly-supervised multi-hop reasoning for knowledge-based visual question answering. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, pages 373-390.  
+Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural Comput., 9(8):1735-1780.  
+Giwon Hong, Jeonghwan Kim, Junmo Kang, and Sung-Hyon Myaeng. 2022. Graph-induced transformers for efficient multi-hop question answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, pages 10288-10294.  
+Qiuyuan Huang, Paul Smolensky, Xiaodong He, Li Deng, and Dapeng Oliver Wu. 2018. Tensor product generation networks for deep NLP modeling. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 1263-1273. Association for Computational Linguistics.  
+Rui Huang and Ping Li. 2023. Hub-hub connections matter: Improving edge dropout to relieve oversmoothing in graph neural networks. Knowl. Based Syst., 270:110556.  
+Yongjie Huang and Meng Yang. 2021. Breadth first reasoning graph for multi-hop question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, pages 5810-5821.  
+Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.  
+Yeskendir Koishekenov. 2023. Reducing oversmoothing in graph neural networks using relational embeddings. CoRR, abs/2301.02924.
+
+Parisa Kordjamshidi, Martijn van Otterlo, and Marie-Francine Moens. 2010. Spatial role labeling: Task definition and annotation scheme. In Proceedings of the International Conference on Language Resources and Evaluation, LREC 2010, 17-23 May 2010, Valletta, Malta. European Language Resources Association.  
+Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. ALBERT: A lite BERT for self-supervised learning of language representations. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.  
+Hung Le, Truyen Tran, and Svetha Venkatesh. 2020. Self-attentive associative memory. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 5682-5691. PMLR.  
+Qimai Li, Zhichao Han, and Xiao-Ming Wu. 2018. Deeper insights into graph convolutional networks for semi-supervised learning. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 3538-3545. AAAI Press.  
+Yang Liu, Chuan Zhou, Shirui Pan, Jia Wu, Zhao Li, Hongyang Chen, and Peng Zhang. 2023. Curvdrop: A ricci curvature based approach to prevent graph neural networks from over-smoothing and oversquashing. In Proceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX, USA, 30 April 2023 - 4 May 2023, pages 221-230. ACM.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692.  
+Qian Luo, Yunfei Li, and Yi Wu. 2023. Grounding object relations in language-conditioned robotic manipulation with semantic-spatial reasoning. CoRR, abs/2303.17919.  
+Wouter Massa, Parisa Kordjamshidi, Thomas Provoost, and Marie-Francine Moens. 2015. Machine reading of biological texts - bacteria-biotope extraction. In BIOINFORMATICS 2015 - Proceedings of the International Conference on Bioinformatics Models, Methods and Algorithms, Lisbon, Portugal, 12-15 January, 2015, pages 55-64. SciTePress.  
+Yimeng Min, Frederik Wenkel, and Guy Wolf. 2020. Scattering GCN: overcoming oversmoothness in graph convolutional networks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
+
+Roshanak Mirzaee, Hossein Rajaby Faghihi, Qiang Ning, and Parisa Kordjamshidi. 2021. SPARTQA: A textual question answering benchmark for spatial reasoning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021, pages 4582-4598. Association for Computational Linguistics.  
+Roshanak Mirzaee and Parisa Kordjamshidi. 2022. Transfer learning with synthetic corpora for spatial role labeling and reasoning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 6148-6165. Association for Computational Linguistics.  
+Christopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. 2019. Weisfeiler and leman go neural: Higher-order graph neural networks. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 4602-4609. AAAI Press.  
+Rasmus Berg Palm, Ulrich Paquet, and Ole Winther. 2018. Recurrent relational networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montreal, Canada, pages 3372-3382.  
+Lin Qiu, Yunxuan Xiao, Yanru Qu, Hao Zhou, Lei Li, Weinan Zhang, and Yong Yu. 2019. Dynamically fused graph network for multi-hop reasoning. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 6140-6150. Association for Computational Linguistics.  
+Adam Santoro, David Raposo, David G. T. Barrett, Mateusz Malinowski, Razvan Pascanu, Peter W. Battaglia, and Tim Lillicrap. 2017. A simple neural network module for relational reasoning. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 4967-4976.  
+Imanol Schlag, Tsendsuren Munkhdalai, and Jurgen Schmidhuber. 2021. Learning associative inference using fast weight memory. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.  
+Imanol Schlag and Jürgen Schmidhuber. 2018a. Learning to reason with third order tensor products. In Advances in Neural Information Processing Systems
+
+31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 10003-10014.  
+Imanol Schlag and Jürgen Schmidhuber. 2018b. Learning to reason with third order tensor products. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 10003-10014.  
+Zhengxiang Shi, Qiang Zhang, and Aldo Lipani. 2022. Stepgame: A new benchmark for robust multi-hop spatial reasoning in texts. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, pages 11321-11329.  
+Paul Smolensky. 1990. Tensor product variable binding and the representation of symbolic structures in connectionist systems. Artif. Intell., 46(1-2):159-216.  
+Yunchong Song, Chenghu Zhou, Xinbing Wang, and Zhouhan Lin. 2023. Ordered GNN: ordering message passing to deal with heterophily and oversmoothing. CoRR, abs/2302.01524.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.  
+Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. CoRR, abs/1710.10903.  
+Sagar Gubbi Venkatesh, Anirban Biswas, Raviteja Upadrashta, Vikram Srinivasan, Partha Talukdar, and Bharadwaj Amrutur. 2021. Spatial reasoning from natural language instructions for robot manipulation. In IEEE International Conference on Robotics and Automation, ICRA 2021, Xi'an, China, May 30 - June 5, 2021, pages 11196-11202. IEEE.  
+Guangtao Wang, Rex Ying, Jing Huang, and Jure Leskovec. 2021. Multi-hop attention graph neural networks. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJ-CAI 2021, pages 3089-3096. ijcai.org.  
+Jason Weston, Antoine Bordes, Sumit Chopra, and Tomás Mikolov. 2016. Towards ai-complete question answering: A set of prerequisite toy tasks. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings.  
+Gongce Wu, Shukuan Lin, Xiaoxue Shao, Peng Zhang, and Jianzhong Qiao. 2023. QPGCN: graph convolutional network with a quadratic polynomial filter for overcoming over-smoothing. Appl. Intell., 53(6):7216-7231.
+
+Weiwen Xu, Yang Deng, Huihui Zhang, Deng Cai, and Wai Lam. 2021a. Exploiting reasoning chains for multi-hop science question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 1143-1156.  
+Weiwen Xu, Huihui Zhang, Deng Cai, and Wai Lam. 2021b. Dynamic semantic graph construction and reasoning for explainable multi-hop science question answering. In Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021, volume ACL/IJCNLP 2021 of Findings of ACL, pages 1044-1056. Association for Computational Linguistics.  
+Rui Yang, Wenrui Dai, Chenglin Li, Junni Zou, and Hongkai Xiong. 2023. Tackling over-smoothing in graph convolutional networks with em-based joint topology optimization and node classification. IEEE Trans. Signal Inf. Process. over Networks, 9:123-139.  
+Yue Zhang, Quan Guo, and Parisa Kordjamshidi. 2021. Towards navigation by reasoning over spatial configurations. CoRR, abs/2105.06839.  
+Yue Zhang and Parisa Kordjamshidi. 2022. Explicit object relation alignment for vision and language navigation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 322-331. Association for Computational Linguistics.
+
+# Appendix
+
+# A Preliminary of Tensor Product Representation
+
+Consider the implicit relation hidden in the sentence "X is over K." Our method decomposes this sentence into two groups of role  $r \in \mathcal{R}$  and filler  $f \in \mathcal{F}$  symbol components, i.e.,  $\mathcal{R} = \{X, K\}$  and  $\mathcal{F} = \{above, below\}$ . The role and filler symbols are projected into role  $V_{\mathcal{R}}$  and filler  $V_{\mathcal{F}}$  vector space, respectively. The TPR of these symbol structures is defined as the tensor  $\mathbf{T}$  within the vector space  $V_{\mathcal{R}} \otimes V_{\mathcal{F}}$ , where  $\otimes$  denotes the tensor product (equivalent to outer product when role and filler are one-dimension vectors). In this example, we can drive the equation of writing our example to  $\mathbf{T}$ :
+
+$$
+\begin{array}{l} \mathbf {T} = f _ {\text {a b o v e}} \otimes r _ {X} + f _ {\text {b e l o w}} \otimes r _ {K} \tag {16} \\ = f _ {a b o v e} r _ {X} ^ {T} + f _ {b e l o w} r _ {K} ^ {T} \\ \end{array}
+$$
+
+The tensor product  $\otimes$  acts as a binding operator to bind the corresponding role and filler.
+
+The tensor inner product serves as the unbinding operator in TPR, which is employed to reconstruct the previously reserved fillers from  $\mathbf{T}$ . Given the
+
+<table><tr><td>Experiment</td><td>Dataset</td><td>epochs</td><td>batch size</td><td>PLM learning rate</td><td>non-PLM learning rate</td></tr><tr><td>BERT</td><td>StepGame</td><td>22</td><td>32</td><td>2e-05</td><td>-</td></tr><tr><td>ALBERT</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>-</td></tr><tr><td>RoBERTa</td><td>StepGame</td><td>16</td><td>32</td><td>6e-06</td><td>-</td></tr><tr><td>DepWiNet(BERT)</td><td>StepGame</td><td>17</td><td>32</td><td>2e-05</td><td>1e-04</td></tr><tr><td>DepWiNet(ALBERT)</td><td>StepGame</td><td>22</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(RoBERTa)</td><td>StepGame</td><td>19</td><td>32</td><td>6e-06</td><td>1e-04</td></tr><tr><td>DepWiNet(BERT)</td><td>SPARTUN</td><td>9</td><td>16</td><td>8e-06</td><td>1e-04</td></tr><tr><td>DepWiNet(BERT)</td><td>ReSQ(with SPARTUN supervision)</td><td>4</td><td>16</td><td>5e-05</td><td>4e-05</td></tr><tr><td>DepWiNet(BERT)</td><td>ReSQ(without SPARTUN supervision)</td><td>4</td><td>16</td><td>5e-05</td><td>4e-05</td></tr><tr><td>DepWiNet(w/o LDC)</td><td>StepGame</td><td>15</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/o NMI &amp; LDC)</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/o SRR &amp; LDC)</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ DepWiGNNmean)</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ DepWiGNNmaxpooling)</td><td>StepGame</td><td>14</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ GCN1-5)</td><td>StepGame</td><td>(16,12,10,19,14)</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ GraphConv1-5)</td><td>StepGame</td><td>(14,14,16,20,15)</td><td>32</td><td>6e-06</td><td>5e-05</td></tr><tr><td>DepWiNet(w/ GAT1-5)</td><td>StepGame</td><td>(16,13,16,16,17)</td><td>32</td><td>6e-06</td><td>5e-05</td></tr></table>
+
+Table 5: Hyperparameters and setup information for each experiment.
+
+unbinding vector  $u_{K}$ , we can retrieve the stored filler  $f_{below}$  from the  $\mathbf{T}$ . In the most ideal cases, if the role vectors are orthogonal to each other,  $u_{K}$  equals  $r_{K}$ . When  $\mathbf{T} \in \mathbb{R}^2$ , it can be expressed as the matrix multiplication.:
+
+$$
+\begin{array}{l} \mathbf {T} \cdot u _ {K} = \left(f _ {a b o v e} r _ {X} ^ {T} + f _ {b e l o w} r _ {K} ^ {T}\right) \cdot r _ {K} \tag {17} \\ = \alpha f _ {b e l o w} \propto f _ {b e l o w} \\ \end{array}
+$$
+
+# B Hyperparameter Settings
+
+Detailed hyperparameter settings for each experiment are provided in Table 5. Epoch numbers for DepWiNet with the three classical GNNs are grouped (in order) for simplicity.

2023/DepWiGNN_ A Depth-wise Graph Neural Network for Multi-hop Spatial Reasoning in Text/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdd6204e95fb659e7dc1366f86e2d34d6eece38c28a086fd16443f84c4ebfb5e
+size 661956

2023/Descriptive Prompt Paraphrasing for Target-Oriented Multimodal Sentiment Classification/9ad0a505-4aa3-4e82-936d-bde5e9f488e4_content_list.json

@@ -0,0 +1,1938 @@
+[
+    {
+        "type": "text",
+        "text": "Descriptive Prompt Paraphrasing for Target-Oriented Multimodal Sentiment Classification",
+        "text_level": 1,
+        "bbox": [
+            154,
+            87,
+            843,
+            127
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dan Liu  $^{1}$ , Lin Li  $^{1}$ , Xiaohui Tao  $^{2}$ , Jian Cui  $^{1}$ , Qing Xie",
+        "bbox": [
+            174,
+            140,
+            825,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup> Wuhan University of Technology, Wuhan, China",
+        "bbox": [
+            297,
+            159,
+            704,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>2</sup> The University of Southern Queensland, Toowoomba, Australia",
+        "bbox": [
+            236,
+            175,
+            764,
+            191
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$  {liunnew, cathylilin, 318971, felixxq} @whut.edu.cn",
+        "bbox": [
+            273,
+            192,
+            727,
+            209
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$  Xiaohui.Tao@unisq.edu.au",
+        "bbox": [
+            381,
+            209,
+            618,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Target-Oriented Multimodal Sentiment Classification (TMSC) aims to perform sentiment polarity on a target jointly considering its corresponding multiple modalities including text, image, and others. Current researches mainly work on either of two types of targets in a decentralized manner. One type is entity, such as a person name, a location name, etc. and the other is aspect, such as 'food', 'service', etc. We believe that this target type based division in task modelling is not necessary because the sentiment polarity of the specific target is not governed by its type but its context. For this reason, we propose a unified model for target-oriented multimodal sentiment classification, so called UnifiedTMSC. It is prompt-based language modelling and performs well on four datasets spanning the above two target types. Specifically, we design descriptive prompt paraphrasing to reformulate TMSC task via (1) task paraphrasing, which obtains paraphrased prompts based on the task description through a paraphrasing rule, and (2) image prefix tuning, which optimizes a small continuous image vector throughout the multimodal representation space of text and images. Conducted on two entity-level multimodal datasets: Twitter-2015 and Twitter-2017, and two aspect-level multimodal datasets: Multi-ZOL and MASAD, the experimental results show the effectiveness of our UnifiedTMSC.",
+        "bbox": [
+            141,
+            279,
+            460,
+            720
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            732,
+            258,
+            747
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "With the emergence of new media and advanced technology, the forms of information released by people are quietly changing, from mono-modality to multi-modality now, such as text, images, etc (Xu and Mao, 2017). This also pushes researchers to conduct multimodal learning (Yu et al., 2016; Xu et al., 2018; Long et al., 2022). For sentiment analysis, both text and image are highly correlated with sentiment polarity. Moreover, they can complement and reinforce each other (Xu et al., 2019).",
+        "bbox": [
+            112,
+            757,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e54dc4cfb903ad65d631dbed57c7543878be0a411450d87e3276d9be8c6b6450.jpg",
+        "image_caption": [
+            "(a) Entity-level MSC"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            249,
+            877,
+            382
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c2210b2c262129dd8d8214aadbfb6f10f8cd6bfd4ab9e3dad4adee45770a3648.jpg",
+        "image_caption": [
+            "(b) Aspect-level MSC",
+            "Figure 1: Two forms of target-oriented multimodal sentiment classification."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            406,
+            877,
+            539
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "At present, fine-grained Multimodal Sentiment Classification (MSC) includes two main tasks: entity-level MSC and aspect-level MSC, as shown in Figure 1. (1) For entity-level, the entity and its context are encoded as independent input text in some studies(Yu and Jiang, 2019; Yu et al., 2020; Wang et al., 2021; Zhao et al., 2022). Others jointly consider the encoding of an entity with its contextual learning, which is their main focus to achieve good MSC performance (Khan and Fu, 2021; Xiao et al., 2022; Yang et al., 2022). For example, Xiao et al. (2022) presented a dual stream adaptive multifeature extraction graph convolutional network to convert an image into its caption. (2) For aspect-level, the aspect and its context are being encoded individually due to the semantics involved in an aspect itself (Xu et al., 2019; Zhang et al., 2021; Zhou et al., 2021). For example, Xu et al. (2019)",
+        "bbox": [
+            507,
+            629,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "4174",
+        "bbox": [
+            478,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4174-4186",
+        "bbox": [
+            216,
+            945,
+            779,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            277,
+            958,
+            719,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "studied a multi-interactive memory network model. These TMSC works can promote human decisions by assisting users in knowing about certain targets.",
+        "bbox": [
+            112,
+            84,
+            490,
+            134
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As we know, an entity usually is a person name, or a location name, etc (Li et al., 2022b). Such entity target only has a specific meaning when connecting it with a specific modality content, which means it is difficult to accurately understand the entity without its context. On the contrary, an aspect itself in some degree can represent what it means even without its context. Because of this obvious difference, previous studies show their interest in differently modelling the two tasks to capture target related context (Xu et al., 2019; Yu et al., 2020; Zhang et al., 2021; Song et al., 2023). For example, in Figure 1(a), \"Chuck Bass\", \"MCM\", and \"Iran\" do not have any exact meaning when they are disconnected from the specific context. However, in Figure 1(b), both aspect term and aspect category have their own meanings with a hidden sentiment tendency. For the \"battery life\" of mobile phones, the first reaction is that the longer the battery life, the better. Despite the above differences among the TMSC tasks, from the perspective of sentiment classification, the goal of TMSC is to predict the sentiment polarity of a target no matter whether it is an entity or an aspect. Therefore, in our view, the boundary is unnecessary.",
+        "bbox": [
+            115,
+            137,
+            490,
+            539
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a unified TMSC model via prompt based language modelling, so called UnifiedTMSC, which is independent of the target type in TMSC. Our core is to reconstruct the two TMSC tasks through descriptive prompt paraphrasing. The prompts we design can place entity and aspect in their context, while also being close to the TMSC task description. To achieve this goal, we carry out our work from two aspects: (1) task paraphrasing. The task description is transformed into a seed prompt, and different paraphrased prompts are obtained by using our paraphrasing rule. They serve as discrete prompts for the text and fit into the Masked Language Modeling (MLM) format. And (2) image prefix tuning (Li and Liang, 2021). A segment vector is initialized for the image pretrained embedding as the prefix continuous prompt. In the subsequent multimodal continuous representation space, the image pre-trained embedding is fixed and only some segment of the initialized vector is optimized. In this way, sentiment labels are generated through the cloze-filling method.",
+        "bbox": [
+            115,
+            545,
+            489,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In our extensive experiments, our UnifiedTMSC",
+        "bbox": [
+            131,
+            903,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "model achieves state-of-the-art performance on two entity-level datasets: Twitter-2015 and Twitter-2017, and two aspect-level datasets: Multi-ZOL and MASAD. (1) The results of the task description based prompts are superior to those of arbitrary prompt templates. (2) On two entity-level datasets, our model improves Accuracy by  $1.0\\% - 2.8\\%$  and  $1.5\\% - 3.7\\%$ , macro-F1 by  $1.5\\% - 4.5\\%$  and  $1.7\\% - 7.5\\%$ , respectively. (3) On two aspect-level datasets, our model gains of  $6.46\\% - 8.87\\%$  and  $1.66\\% - 3.02\\%$  on Accuracy,  $3.77\\% - 5.01\\%$  and  $2.19\\% - 3.25\\%$  on macro-F1 are derived. The experimental results demonstrate the effectiveness of our UnifiedTMSC.",
+        "bbox": [
+            507,
+            84,
+            885,
+            309
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            322,
+            665,
+            338
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Entity-level MSC",
+        "text_level": 1,
+        "bbox": [
+            509,
+            349,
+            692,
+            363
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As a pioneer, Yu and Jiang (2019) proposed a BERT-based multimodal architecture to determine the sentiment polarity of an entity. Yu et al. (2020) introduced an entity-sensitive attention and fusion network. And Wang et al. (2021) put forward a recurrent attention network. Khan and Fu (2021) introduced an input space translation framework to construct image context from the image. Zhao et al. (2022) used the adjective-noun pairs extracted from images as the knowledge enhancement based on Yu and Jiang (2019) and Wang et al. (2021). Moreover, Yang et al. (2022) explored facial information in images to obtain visual and sentiment clues.",
+        "bbox": [
+            507,
+            370,
+            884,
+            579
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In the research stated above, the entity can be encoded as a distinct text input, or entity and context can be combined as the text input. Their goal is to more effectively learn the semantics related to entity sentiment.",
+        "bbox": [
+            507,
+            580,
+            884,
+            659
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Aspect-level MSC",
+        "text_level": 1,
+        "bbox": [
+            507,
+            671,
+            697,
+            687
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Aspect-level multimodal classification was first proposed by Xu et al. (2019), and they introduced a multi-interactive memory network to analyze multiple correlations in multimodal data. Zhang et al. (2021) presented a multimodal fusion discriminant attention network and designed a discriminant matrix to supervise the modality fusion. Zhou et al. (2021) conducted a multimodal interaction model that learns the relationships between text, image, and target aspect through interaction layers and adversarial training. One key difference between an aspect and an entity is that the aspect has its own semantics inferred from the aspect words. Therefore, existing research usually regards the aspect",
+        "bbox": [
+            507,
+            694,
+            885,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "4175",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/4b659a4531a1d0f63dac519f07ec64fb7a403f8763130ce00d627737033e3e45.jpg",
+        "image_caption": [
+            "Figure 2: The model of our proposed UnifiedTMSC. Seed prompt and keyword 'sentiment' are obtained based on the task description and are converted into diverse paraphrased prompts through the paraphrasing rule (Equation 1). The paraphrased prompt and the original text input are concatenated to form the new text input which is encoded by the encoder to gain text embedding. The image is divided into multiple regions and each of the regions is initialized as a vector that is used as the prefix prompt for image pre-trained embedding. The text embedding and image embedding are undergone cross attention to gain modality fusion embedding."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            85,
+            331,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/81a1fa7c04df0a379853669bcaa530a31807b2d6041a1dc1e2d6fd568e22104e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            332,
+            86,
+            586,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/01d7b8540d4cc6764f9510df13190097385a35470e0b4c9a597f90ff28553a16.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            589,
+            86,
+            880,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "itself as an additional input.",
+        "bbox": [
+            112,
+            478,
+            321,
+            493
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our focus: In contrast to previous studies, our model can run across TMSC tasks. It combines the target (entity or aspect) and its context as a text input using task description based paraphrased prompts. We can get sentiment-related semantics about the target by providing its context in task description based prompts.",
+        "bbox": [
+            112,
+            495,
+            485,
+            607
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Prompt paraphrasing",
+        "text_level": 1,
+        "bbox": [
+            112,
+            623,
+            331,
+            638
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Prompt tuning has received increasing attention recently (Radford et al., 2021; Yao et al., 2021) and has been successfully applied in many domains (Han et al., 2022; Liu et al., 2023b). For example, in the field of Question Answering (QA), Khashabi et al. (2020) reformulated many QA tasks as a text generation problem by fine-tuning seq2seq-based pre-trained models and appropriate prompts from the context and questions. For Information Extraction (IE), Chen et al. (2022) first explored the application of fixed-prompt LM Tuning in relation extraction and Lu et al. (2022) applied prompt to control the information to be extracted. In other research fields, Cui et al. (2023) used prompt learning in text input to conduct the meme mining task. Recently, the prompt has been used for the task involving fine-grained text sentiment analysis, and",
+        "bbox": [
+            112,
+            646,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the results are promising (Seoh et al., 2021; Li et al., 2021, 2022a; Gao et al., 2022; Liu et al., 2023a).",
+        "bbox": [
+            507,
+            478,
+            882,
+            508
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Inspired by the above studies, our unified model is through the task description based prompt paraphrasing with jointly soft and hard prompt tuning.",
+        "bbox": [
+            507,
+            510,
+            882,
+            558
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            509,
+            569,
+            611,
+            583
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Task Formulation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            594,
+            695,
+            608
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a multimodal samples  $D$ , for each sample  $d \\in D$ , it contains a sentence  $S$  with  $n$  words  $(w_{1}, w_{2}, \\ldots, w_{n})$  and one or more related images  $I$ , as well as a target  $T$  which contains  $m$  words  $(w_{1}, w_{2}, \\ldots, w_{m})$  and is a sub-sequence of  $S$  or predefined phrase. For the target  $T$ , it is associated with a sentiment label  $Y$ . In general,  $Y \\in \\{positive, neutral, negative\\}$ , and different tasks may have different sentiment labels. Our goal is to learn a target-oriented sentiment classifier that can correctly predict the sentiment label for each sample  $X = (S, I, T)$ .",
+        "bbox": [
+            507,
+            614,
+            882,
+            808
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Overview",
+        "text_level": 1,
+        "bbox": [
+            509,
+            818,
+            631,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 2, our model consists of two modules: task paraphrasing (hard prompt) and image prefix tuning (soft prompt). For the given multimodal data  $X = (S, I, T)$ , we obtain paraphrased prompts  $(P_{1}, P_{2} \\dots)$  through the task description of",
+        "bbox": [
+            507,
+            839,
+            882,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "4176",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/b5347ca45346e90a6665be6ce3b0b622319daaa771e8698c62c17aa1e2874631.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 1: Some examples of paraphrased prompts, as well as the relative positions of  $Y$  (sentiment labels, replaced by [MASK] token, marked with bold),  $T$  (the targets that need to recognize sentiment polarity, annotated with italics), and  $K$  (keywords extracted from seed prompt or replaced with synonyms, annotated with underline) in the prompts. 'Synonym of  $\\mathrm{K}$  ’ indicates whether to replace the keyword 'sentiment' extracted from the seed prompt with synonyms. 'B', 'M', and 'E' respectively represent the positions in the paraphrased prompt as: 'Beginning', 'Middle', and 'Ending'."
+        ],
+        "table_body": "<table><tr><td>Paraphrased Prompts</td><td>&lt; Y, Position &gt;</td><td>&lt; T,Position &gt;</td><td>&lt; K, Position &gt;</td><td>Synonym of K</td></tr><tr><td>{target} express a [MASK] sentiment.</td><td>&lt; [MASK],M &gt;</td><td>&lt; {target},B &gt;</td><td>&lt; sentiment,E &gt;</td><td>no</td></tr><tr><td>The emotion of {target} is [MASK].</td><td>&lt; [MASK],E &gt;</td><td>&lt; {target},M &gt;</td><td>&lt; emotion,B &gt;</td><td>yes</td></tr><tr><td>{target}&#x27;s emotion is [MASK].</td><td>&lt; [MASK],E &gt;</td><td>&lt; {target},B &gt;</td><td>&lt; emotion,M &gt;</td><td>yes</td></tr><tr><td>The [MASK] feeling is {target}&#x27;s sentiment.</td><td>&lt; [MASK],B &gt;</td><td>&lt; {target},M &gt;</td><td>&lt; sentiment,E &gt;</td><td>no</td></tr><tr><td>The sentiment is [MASK] and conveyed by {target}.</td><td>&lt; [MASK],M &gt;</td><td>&lt; {target},E &gt;</td><td>&lt; sentiment,B &gt;</td><td>no</td></tr><tr><td>The emotion expressed by {target} is [MASK].</td><td>&lt; [MASK],E &gt;</td><td>&lt; {target},M &gt;</td><td>&lt; emotion,B &gt;</td><td>yes</td></tr><tr><td>A [MASK] sentiment is expressed towards to {target}.</td><td>&lt; [MASK],B &gt;</td><td>&lt; {target},E &gt;</td><td>&lt; sentiment,M &gt;</td><td>no</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            885,
+            209
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "TMSC, which are used as prefixes for text input  $S$  in the task paraphrasing module (the left of Figure 2). Paraphrased prompt  $P_{i}$  contains a [MASK] token and is connected with text  $S$ . The text embedding  $R$  is obtained through an encoder in the bottom-right of Figure 2. In the image prefix tuning module, we apply an initialization vector  $V$  for pre-train image embedding  $E$  as the continuous prefix. After text embedding  $R$  and image embedding  $(V + E)$  are added to their respective position encoding, the fusion vector is obtained through the multimodal Transformer in the middle right of Figure 2. We take the hidden layer vector  $H$  and pass it through the MLM head to get the prediction score of [MASK] position for each word in the vocabulary  $C$ . Finally, the cross entropy loss  $L_{MLM}$  of the prediction result  $Output$  and the true sentiment label  $Y$  is calculated.",
+        "bbox": [
+            110,
+            329,
+            489,
+            618
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Task Paraphrasing",
+        "text_level": 1,
+        "bbox": [
+            112,
+            636,
+            310,
+            651
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to provide a specific context for a target, i.e., entity and aspect, we add the target-involved prompts to the text input. Since we change the original sentiment classification task in this way to the pre-trained MLM task, it is crucial to figure out how to develop prompts that are suitable for the original task, that is to say, the prompts should be consistent with the expression of the task description. Inspired by prompt tuning in various domains, such as the visual grounding problem (Yao et al., 2021) and visual question answering task (Liu et al., 2022), we propose the task paraphrasing module to draw paraphrased prompts as a solution to the above issue.",
+        "bbox": [
+            112,
+            659,
+            489,
+            883
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We get the seed prompt according to the task description composed of natural language, and take",
+        "bbox": [
+            112,
+            887,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "the task-related keyword  $K$  ('sentiment') from the task description. The seed prompt is transformed through the paraphrasing rule, and guides the generation of paraphrased prompts that are close to the original task description. Our paraphrasing rule can be formalized in the following form:",
+        "bbox": [
+            507,
+            329,
+            884,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (Y) \\wedge f (T) \\wedge f (K) \\wedge g (K) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            581,
+            435,
+            882,
+            453
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where the function  $f$  represents the relative position in the paraphrased prompts and the substitution of synonyms for a keyword ('sentiment') is illustrated by the function  $g$ .  $f(Y), f(T), f(K) \\in \\{B, M, E\\}$  (meaning 'Beginning', 'Middle', 'Ending'), which respectively stands for the relative position of the sentiment label, the target entity, and the keyword derived from the task description.  $g(K) \\in \\{yes, no\\}$  means whether to replace synonyms of keywords  $K$ . Moreover, synonyms are synonymous explanations for the keyword 'sentiment' in the dictionary. For example, in the Bing dictionary, the synonyms for 'sentiment' include 'emotion', 'feeling', 'opinion', etc.",
+        "bbox": [
+            507,
+            461,
+            884,
+            686
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The task paraphrasing module is shown in the left part of Figure 2. Generally speaking, based on the relative position combination of  $Y$ ,  $K$ , and  $T$ , one seed prompt can be paraphrased to gain multiple candidate prompts. If we replace the keyword with different synonyms, we will receive more paraphrased prompts. Some examples of paraphrased prompts  $P_{i}$  are listed in Table 1. When paraphrased prompt  $P_{i}$  is in the training phase, the position of  $Y$  is replaced by [MASK] token which is the prediction object. Finally, the  $P_{i}$  and text  $S$  are concatenated to obtain a new text input:",
+        "bbox": [
+            507,
+            687,
+            882,
+            879
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS ^ {\\text {n e w}} = [ C L S ] P _ {i} S [ S E P ] \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            588,
+            889,
+            882,
+            906
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "4177",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf2e1f42ed723ff41e703b4494bd19293b9129cdd41e4dbe23ab1ab6b77a403a.jpg",
+        "image_caption": [
+            "Figure 3: An example of image prefix tuning.  $v_{P_{idx}}$  is random initialization of the prefix sequence and  $e_{I_{idx}}$  is the image pre-trained embedding."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            84,
+            480,
+            195
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Image Prefix Tuning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            280,
+            319,
+            296
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Discrete prompts, i.e., hard prompts, in the text are natural and easy to understand. For images, directly applying discrete prompts cannot ensure alignment between text based prompts and images since there are modality gaps between different modalities. Therefore, finding suitable prompts for images is crucial. This module mainly focuses on how to add prompts to images to facilitate a better modality fusion of images and text.",
+        "bbox": [
+            112,
+            303,
+            487,
+            447
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inspired by (Tsimpoukelli et al., 2021) and (Li and Liang, 2021), we introduce a continuous vector as a prefix, i.e., soft prompt, to image pre-trained embedding, as shown in Figure 3. We first segment the image into  $r$  regions, and then initialize a vector  $v_{i}$ $(i\\in \\{1,2,\\dots r\\})$  for each region to form the prefix embedding  $V = \\{v_{1},v_{2},\\ldots v_{r}\\}$  and  $V\\in \\mathbb{R}^{r\\times 2048}$ . Here,  $P_{idx}$  denotes the indices of the prefix sequence.  $|P_{idx}|$  denotes the length of the prefix and  $|I_{idx}|$  indicates the length of the image pre-trained embedding. In  $E = \\{e_1,e_2,e_3,e_4,e_5\\}$ ,  $e_i\\in \\mathbb{R}^{2048}$  refers to the image embedding of the  $i$ -th region. The prefix embedding  $V = \\{v_{1},v_{2},v_{3}\\}$  and  $V\\in \\mathbb{R}^{3\\times 2048}$ . If there are multiple images corresponding to the text, we initialize a soft prompt for each image and their averaged vector is used as the final prefix prompt for the images.  $V$  and  $E$  are concatenated to obtain new image embedding  $E^{new}$ .",
+        "bbox": [
+            112,
+            449,
+            489,
+            753
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE ^ {\\text {n e w}} = V \\oplus E \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            238,
+            758,
+            487,
+            774
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {(i)} ^ {\\text {n e w}} = \\left\\{ \\begin{array}{c c} v _ {i} & i \\in P _ {i d x} \\\\ e _ {(i - | P _ {i d x} |)} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            164,
+            816,
+            485,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$E^{new}$  and  $S^{new}$  are image input and text input respectively, and the subsequent encoding process is carried out together.",
+        "bbox": [
+            112,
+            870,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/b647e129c1910d0b705b0ede820141cf937a719cf74910432da49383ad4966f9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Twitter-2015</td><td>Twitter-2017</td><td>Multi-ZOL</td><td>MASAD</td></tr><tr><td></td><td>Text-Image</td><td>Text-Image</td><td>Text-Image</td><td>Text-Image</td></tr><tr><td>Train</td><td>3179</td><td>3562</td><td>22775</td><td>26292</td></tr><tr><td>Dev</td><td>1122</td><td>1176</td><td>2847</td><td>2925</td></tr><tr><td>Test</td><td>1037</td><td>1234</td><td>2847</td><td>8850</td></tr><tr><td>Total</td><td>5338</td><td>5972</td><td>28469</td><td>38532</td></tr></table>",
+        "bbox": [
+            514,
+            80,
+            878,
+            154
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Statistics of four TMSC datasets.",
+        "bbox": [
+            551,
+            162,
+            836,
+            175
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.5 Multimodal Transformer With MLM",
+        "text_level": 1,
+        "bbox": [
+            507,
+            202,
+            848,
+            216
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training. Given a triplet  $(S, I, T, Y)$ , after the task paraphrasing module and image prefix tuning module, the updated text input  $S^{new}$  and image embeddings  $E^{new}$  are obtained. Text encoder encodes  $S^{new}$  to gain text embedding  $R$ .  $R$  and the image embedding  $E^{new}$  undergo cross-attention through a Transformer to obtain the fusion embedding  $H$ . In this multimodal Transformer, the image pre-trained embedding  $E$  is fixed and not updated, and only  $V$  is updated. Fusion embedding  $H$  passes through an MLP to obtain prediction scores Logit:",
+        "bbox": [
+            507,
+            222,
+            882,
+            399
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {L o g i t} = M L P _ {M L M} (H) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            600,
+            410,
+            882,
+            428
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The word with the highest prediction score is the prediction result:",
+        "bbox": [
+            507,
+            438,
+            880,
+            469
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {O u t p u t} = \\operatorname {a r g m a x} (\\operatorname {L o g i t}) \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            589,
+            481,
+            882,
+            498
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Finally, the predicted result and the true sentiment label  $Y$  are calculated in the cross-entropy loss to optimize our model.",
+        "bbox": [
+            507,
+            508,
+            882,
+            556
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {M L M} = - \\log \\left(\\frac {\\exp (L o g i t _ {i})}{\\sum_ {j = 0} ^ {| C |} \\exp (L o g i t _ {j})}\\right) \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            554,
+            565,
+            882,
+            606
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $i$  is the true sentiment label number and  $C$  is the size of the language model vocabulary.",
+        "bbox": [
+            507,
+            614,
+            880,
+            646
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inference. In the inference stage, given a triplet  $(S,I,T)$ , the sentiment polarity  $Y^{\\prime}$  of  $T$  is determined via the triplet.",
+        "bbox": [
+            507,
+            646,
+            882,
+            694
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nY ^ {\\prime} = \\operatorname {a r g m a x} \\left(\\operatorname {L o g i t} _ {[ M A S K ]}\\right) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            579,
+            705,
+            882,
+            724
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After obtaining the fusion embedding of text and images, we take the argmax of the logit of the [MASK] position to obtain the final prediction result. Finally, for non-label words generated by the model, the answer engineering is used to map them to sentiment labels.",
+        "bbox": [
+            507,
+            733,
+            880,
+            827
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            841,
+            648,
+            857
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            712,
+            882
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets. According to Zhou et al. (2021), there are four datasets for target-oriented multimodal",
+        "bbox": [
+            507,
+            887,
+            880,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "4178",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "sentiment classification, including two entity-level datasets: Twitter-2015 (Zhang et al., 2018; Yu and Jiang, 2019), Twitter-2017 (Lu et al., 2018; Yu and Jiang, 2019) and two aspect-level datasets: Multi-ZOL (Xu et al., 2019), MASDA (Zhou et al., 2021) The statistics information of these four datasets are shown in Table 2, and their details about the data partitioning and label types are presented in A.2.",
+        "bbox": [
+            112,
+            84,
+            487,
+            212
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics. To fairly compare with state-of-the-art approaches, our UnifiedTMSC is evaluated across two TMSC tasks and adopts the Accuracy (Acc) and Macro-F1 score (F1), following Yu and Jiang (2019) and Ling et al. (2022).",
+        "bbox": [
+            112,
+            215,
+            489,
+            294
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. For the Twitter-2015, Twitter-2017, and MASAD, the batch size is set to 16 and the epochs are set to 6. The text encoder used is BERT-base-uncased (Devlin et al., 2019). For the Multi-ZOL dataset, in order to ensure fairness in the experiment, we employ BERT-base-chinese (Devlin et al., 2019) as the text encoder, the batch size is 8, and the epochs as 6. Moreover, multilingual pre-trained models can also be used as our text encoders.",
+        "bbox": [
+            112,
+            297,
+            489,
+            456
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For all datasets, we apply Resnet-50 (He et al., 2016) as the image encoder to get the image prefix embedding, and the max length of the new text input  $S^{new}$  (in Eq. (2)) is 96. The model learning rate is set as 1e-5 and the dropout rate is 1e-2. Four layer Transformers (Vaswani et al., 2017) are aimed to perform cross attention between different modalities and the pre-training parameters are not loaded. Four NVIDIA Titan Xp GPUs, each with 12GB of memory, are employed in our experiments which are done on a CentOS computer. The deep learning framework is Pytorch, and AdamW is used as the optimizer.",
+        "bbox": [
+            112,
+            458,
+            489,
+            668
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Compared Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            686,
+            321,
+            701
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Because previous work has separated entity-level MSC and aspect-level MSC, the baseline models for each task are different. For the Twitter-2015 and Twitter-2017 datasets, we compare six baselines including TomBERT (IJCAI, Yu and Jiang (2019)), SaliencyBERT (PRCV, Wang et al. (2021)), CapTrBERT (ACM Multimedia, Khan and Fu (2021)), JML-MASC (EMNLP, Ju et al. (2021)), VLP-MABSA (ACL, Ling et al. (2022)), FITE-DE-Large (EMNLP, Yang et al. (2022)). For the Multi-ZOL and MASAD datasets, our model is compared with MIMN (AAAI, Xu et al. (2019)), ModalNet (WWW, Zhang et al. (2021)), MMAP",
+        "bbox": [
+            112,
+            709,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2b378f61ed1938c8a595675bec219d04da213360cc56ee742624cf1d5fcbdd9c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Twitter-2015</td><td colspan=\"3\">Twitter-2017</td></tr><tr><td>Acc</td><td>F1</td><td>P-Value</td><td>Acc</td><td>F1</td><td>P-Value</td></tr><tr><td>TomBERT</td><td>77.2</td><td>71.8</td><td>0.0194</td><td>70.5</td><td>68.0</td><td>0.0156</td></tr><tr><td>SaliencyBERT</td><td>77.0</td><td>72.4</td><td>0.0109</td><td>69.7</td><td>67.2</td><td>0.0156</td></tr><tr><td>CapTrBERT</td><td>78.0</td><td>73.3</td><td>0.0120</td><td>72.3</td><td>70.2</td><td>0.0134</td></tr><tr><td>JML-MASC</td><td>78.7</td><td>-</td><td>-</td><td>72.7</td><td>-</td><td>-</td></tr><tr><td>VLP-MABSA</td><td>78.6</td><td>73.8</td><td>0.0131</td><td>73.8</td><td>71.8</td><td>0.0111</td></tr><tr><td>FITDE-DE-Large</td><td>78.8</td><td>74.8</td><td>0.0048</td><td>73.9</td><td>73.0</td><td>0.0016</td></tr><tr><td>UnifiedTMSC</td><td>79.8</td><td>76.3</td><td>-</td><td>75.4</td><td>74.7</td><td>-</td></tr></table>",
+        "bbox": [
+            510,
+            80,
+            887,
+            170
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/a3cf520bdc7ca41f63e3de9243e2d474ff30f0079715b819b112f122c093c90a.jpg",
+        "table_caption": [
+            "Table 3: The experimental results on the multimodal entity-level datasets: Twitter-2015 and Twitter-2017. The results presented in the table are the average of different prompt results."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Multi-ZOL</td><td colspan=\"3\">MASAD</td></tr><tr><td>Acc</td><td>F1</td><td>P-Value</td><td>Acc</td><td>F1</td><td>P-Value</td></tr><tr><td>MIMN</td><td>61.59</td><td>60.51</td><td>0.0316</td><td>94.52</td><td>93.55</td><td>0.0019</td></tr><tr><td>ModalNet</td><td>62.71</td><td>60.94</td><td>0.0266</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MMAP</td><td>-</td><td>-</td><td>-</td><td>95.29</td><td>94.61</td><td>0.0049</td></tr><tr><td>CLIP</td><td>64.00</td><td>59.70</td><td>0.0092</td><td>95.88</td><td>94.21</td><td>0.0079</td></tr><tr><td>UnifiedTMSC</td><td>70.46</td><td>64.71</td><td>-</td><td>97.54</td><td>96.80</td><td>-</td></tr></table>",
+        "bbox": [
+            510,
+            250,
+            882,
+            321
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 4: The experimental results on the multimodal aspect-level datasets: Multi-ZOL and MASAD.",
+        "bbox": [
+            507,
+            331,
+            880,
+            359
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "(Neurocomputing, Zhou et al. (2021)) and CLIP (INT J COMPUT VISION, Zhou et al. (2022)). The detailed introduction of all the baseline models mentioned above is in Section A.1.",
+        "bbox": [
+            507,
+            385,
+            882,
+            449
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Experimental Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            460,
+            828,
+            475
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3.1 Overall Performance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            481,
+            734,
+            495
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The experimental results on multimodal entity-level and aspect-level datasets are presented in Table 3 and Table 4 respectively. The best results on each metric are marked in bold and the second best results are highlighted with an underline.",
+        "bbox": [
+            505,
+            500,
+            882,
+            579
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Multimodal Entity-level Datasets Results. As reported in Table 3, compared with the baselines, our UnifiedTMSC makes significant improvements in entity-level MSC. On the Twitter-2015 dataset, our improvement is approximately  $1.0\\%$  on Accuracy and  $1.5\\%$  on the macro-F1 compared to the FITE-DE-Large. On the Twitter-2017 dataset, we achieve improvements over multimodal baseline VLP-MABSA on the Accuracy by  $1.5\\%$  and  $1.7\\%$  on the macro-F1, which indicates that using prompt tuning to fuse entity and context can achieve a better sentiment-related semantic understanding of an entity, resulting in better classification results.",
+        "bbox": [
+            505,
+            581,
+            882,
+            788
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Multimodal Aspect-level Datasets Results. Our UnifiedTMSC performs best in two datasets among all multimodal baselines, as listed in Table 4. This demonstrates the effectiveness of our proposed unified model based on prompt paraphrasing. Especially on the Multi-ZOL dataset, our UnifiedTMSC outperforms the ModalNet on the Accuracy by  $7.75\\%$  and  $3.77\\%$  on the macro-F1. On",
+        "bbox": [
+            507,
+            790,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "4179",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/fbda33860a9a1e793b02d606fe121567f39ffd7a31c0200e4ab56a7bc807190f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td></td><td colspan=\"2\">Food</td><td colspan=\"2\">Buildings</td><td colspan=\"2\">Goods</td><td colspan=\"2\">Human</td><td colspan=\"2\">Scenery</td></tr><tr><td></td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>MIMN</td><td>AAAI 2019</td><td>94.72</td><td>91.39</td><td>96.26</td><td>95.80</td><td>95.93</td><td>95.87</td><td>92.54</td><td>92.31</td><td>93.17</td><td>92.38</td></tr><tr><td>MMAP</td><td>Neucomp 2021</td><td>95.75</td><td>92.89</td><td>96.86</td><td>96.85</td><td>96.55</td><td>96.44</td><td>92.74</td><td>92.74</td><td>94.57</td><td>94.15</td></tr><tr><td>CLIP</td><td>IJCV 2022</td><td>96.15</td><td>93.05</td><td>95.84</td><td>94.27</td><td>97.36</td><td>96.88</td><td>94.20</td><td>93.24</td><td>95.84</td><td>93.59</td></tr><tr><td>UnifiedTMSC</td><td></td><td>98.00</td><td>96.02</td><td>98.26</td><td>98.25</td><td>98.08</td><td>98.01</td><td>97.82</td><td>96.49</td><td>95.56</td><td>95.24</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            885,
+            158
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the MASAD dataset, compared to the MMAP, ours improve performance by about  $2.25\\%$  on Accuracy and  $2.19\\%$  on the macro-F1.",
+        "bbox": [
+            112,
+            234,
+            485,
+            281
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Specifically, there are multiple domains in the MASAD dataset, we conduct experiments on each of them, and the experimental results are shown in Table 5. It is clear from the results that our model has achieved large improvements in each domain.",
+        "bbox": [
+            112,
+            282,
+            485,
+            362
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Finally, t-tests are conducted to demonstrate the effectiveness of UnifiedTMSC. From the P-value of other models in Table 3 and Table 4, it can be found that all P-values are less than 0.05. This shows a significant difference in statistics between UnifiedTMSC and other models.",
+        "bbox": [
+            112,
+            363,
+            485,
+            458
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "After comparing and analyzing the experimental results, we can summarize the following two points from our prompt tuning:",
+        "bbox": [
+            112,
+            460,
+            485,
+            508
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "I. Our prompt paraphrasing method delivers the target's context and fits the TMSC job effectively, and it produces good results, demonstrating the efficacy of our unified model.",
+            "II. Utilizing the target as a separate input has worse results than taking the target and context together as text input. This shows that contextual information affects the target's semantics, and a contextual content that is appropriate for the task will result in a well understanding of a target with semantics."
+        ],
+        "bbox": [
+            127,
+            517,
+            489,
+            705
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3.2 The Effect of Prompt Designs",
+        "text_level": 1,
+        "bbox": [
+            112,
+            722,
+            405,
+            739
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For the Twitter-2015 and Twitter-2017 datasets, we select three paraphrased prompts from Table 1 to conduct the experiments. The three selected paraphrased prompts are as follows:",
+        "bbox": [
+            112,
+            741,
+            487,
+            806
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "-  $P_{1}$ : {target} express a [MASK] sentiment.",
+            "-  $P_{2}$ : The emotion of {target} is [MASK].",
+            "-  $P_{3}$ : A [MASK] sentiment is expressed towards to {target}."
+        ],
+        "bbox": [
+            112,
+            807,
+            487,
+            870
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In addition, to verify the performance of the task paraphrasing module. We design three arbitrary prompt templates and compare them with the above",
+        "bbox": [
+            112,
+            871,
+            487,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/56ffffbbe3965b0c8e8f167075befa055c04461dba33623e5c334162bab5a8da.jpg",
+        "table_caption": [
+            "Table 5: The experimental results of each domain in the MASAD dataset. Because the accuracy and Macro F1 score in the domain of animal and plant reach  $99.07\\% - 99.22\\%$  due to the duplicated samples, they will not be displayed here."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\"></td><td colspan=\"2\">Twitter-2015</td><td colspan=\"2\">Twitter-2017</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td rowspan=\"3\">paraphrased prompts</td><td>P1</td><td>80.2</td><td>77.6</td><td>76.9</td><td>76.7</td></tr><tr><td>P2</td><td>81.1</td><td>77.1</td><td>76.6</td><td>75.9</td></tr><tr><td>P3</td><td>79.8</td><td>77.5</td><td>76.2</td><td>75.9</td></tr><tr><td rowspan=\"3\">arbitrary prompts</td><td>P1&#x27;</td><td>79.1</td><td>75.5</td><td>73.8</td><td>73.7</td></tr><tr><td>P2&#x27;</td><td>79.0</td><td>75.1</td><td>74.3</td><td>73.6</td></tr><tr><td>P3&#x27;</td><td>79.4</td><td>75.1</td><td>74.5</td><td>72.9</td></tr></table>",
+        "bbox": [
+            510,
+            231,
+            880,
+            338
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 6: The experimental results of different prompts. The results of paraphrased prompts are superior to the arbitrary prompts, and  $P_1'$ ,  $P_2'$  and  $P_3'$  are comparable. In the experiments, one of the paraphrased prompts can be selected for experimentation.",
+        "bbox": [
+            507,
+            348,
+            882,
+            419
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "three paraphrased prompts. The three arbitrary prompts are as follows:",
+        "bbox": [
+            507,
+            443,
+            880,
+            474
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "-  $P_1'$ : I feel the {target} is [MASK].",
+            "-  $P_2'$ : The (target) made me feel [MASK].",
+            "-  $P_3^7$ : I [MASK] the {target}."
+        ],
+        "bbox": [
+            527,
+            475,
+            845,
+            524
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "where  $\\{target\\}$  is the entity that needs to determine sentiment polarity, and [MASK] represents the masked word, i.e. sentiment label. The masked word in the  $P_1', P_2', P_3'$  is {good, ok, bad}, {good, indifferent, bad} and {love, dislike, hate} respectively. After the masked words are generated, we perform answer engineering to map the predicted results to the sentiment polarity set, that is, the probabilities of these predicted words are made to be equal to the probabilities of being Positive, Neutral, and Negative.",
+        "bbox": [
+            507,
+            525,
+            882,
+            700
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The results of several different prompts are shown in Table 6. Through analysis and comparison, we can obtain the following summaries:",
+        "bbox": [
+            507,
+            701,
+            882,
+            749
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "I. Our paraphrased prompts created by using the task description are much superior to arbitrary prompt templates, demonstrating the value of our task paraphrasing module in producing paraphrased prompts that are appropriate for the original sentiment classification task. In addition, the performance of different paraphrased prompts is comparable, and in subsequent experiments, anyone can be selected for training and inference.",
+        "bbox": [
+            527,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "4180",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/764ff1b9df2cec2552b3f40892e2f89d9737331c214a6c0c7d22ed2b84b70efc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Image</td><td>(a) Fan Throws Water Bottle at Justin Bieber After He Says He Doesn&#x27;t Know the ...</td><td>(b) Kim Kardashian goes all out for Kanye West&#x27;s 40 th birthday in the Bahamas</td><td>(c) RT @ TrumpDoral : Congratulations to the new # Miss Universe , Miss Colombia , Paulina Vega !</td><td>(d) Nba - The Cavs are Shocked Draymond Green Keeps Getting Away With Kic -</td></tr><tr><td>(Text,Label)</td><td>(Justin Bieber, Negative)</td><td>(Kim Kardashian, Positive) (Kanye West, Positive) (Bahamas, Neutral)</td><td>(Miss Colombia, Positive) (Paulina Vega, Positive)</td><td>(Nba, Neutral) (Cavs, Neutral) (Draymond Green, Negative)</td></tr><tr><td>w/o \nParaphrased \nPrompt</td><td>(Justin Bieber, Neutral) ×</td><td>(Kim Kardashian, Neutral) × (Kanye West, Neutral) × (Bahamas, Neutral) √</td><td>(Miss Colombia, Neutral) × (Paulina Vega, Neutral) ×</td><td>(Nba, Neutral) √ (Cavs, Negative) × (Draymond Green, Positive) ×</td></tr><tr><td>w/o Image \nPrefix</td><td>(Justin Bieber, Positive) ×</td><td>(Kim Kardashian, Positive) √ (Kanye West, Neutral) × (Bahamas, Neutral) √</td><td>(Miss Colombia, Neutral) × (Paulina Vega, Positive) √</td><td>(Nba, Neutral) √ (Cavs, Negative) × (Draymond Green, Negative) √</td></tr><tr><td>UnifiedTMSC</td><td>(Justin Bieber, Negative) √</td><td>(Kim Kardashian, Positive) √ (Kanye West, Positive) √ (Bahamas, Neutral) √</td><td>(Miss Colombia, Positive) √ (Paulina Vega, Positive) √</td><td>(Nba, Neutral) √ (Cavs, Neutral) √ (Draymond Green, Negative) √</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            880,
+            360
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/bc3aa62f7e701e7f164daca42e3de4bedfccc8abb9f8d540b36b1228d59038df.jpg",
+        "table_caption": [
+            "Figure 4: Case study on four test samples. Red font indicates correctly predicted labels."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"2\">Twitter-2015</td><td colspan=\"2\">Twitter-2017</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>UnifiedTMSC</td><td>79.8</td><td>76.3</td><td>75.4</td><td>74.7</td></tr><tr><td>w/o Paraphrased Prompts</td><td>75.7</td><td>72.0</td><td>68</td><td>65.9</td></tr><tr><td>P1 + w/o Image Prefix</td><td>78.9</td><td>75.2</td><td>74.1</td><td>73.7</td></tr><tr><td>P2 + w/o Image Prefix</td><td>78.6</td><td>75.7</td><td>74.4</td><td>73.8</td></tr><tr><td>P3 + w/o Image Prefix</td><td>78.5</td><td>75.7</td><td>74.3</td><td>73.9</td></tr><tr><td>AVG w/o Image Prefix</td><td>78.7</td><td>75.5</td><td>74.3</td><td>73.8</td></tr></table>",
+        "bbox": [
+            119,
+            404,
+            482,
+            502
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Ablation study of our UnifiedTMSC model.",
+        "bbox": [
+            119,
+            510,
+            478,
+            525
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "II. The position of [MASK] in the paraphrased prompts can also have an impact on the experimental results. In our case, the effect is best when the relative position of [MASK] is \"Middle\" rather than \"Beginning\" or \"Ending\". Therefore, when meeting a new task, the position of [MASK] may be a factor to affect task performance.",
+        "bbox": [
+            127,
+            550,
+            489,
+            678
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            689,
+            280,
+            703
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To further investigate the effects of paraphrased prompts and image prefix, because entity-level MSC is more challenging than aspect-level MSC, we conduct ablation analysis on the multimodal entity-level datasets: Twitter-2015 and Twitter-2017. The results of the ablation experiments are shown in Table 7.",
+        "bbox": [
+            112,
+            709,
+            487,
+            820
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Paraphrased Prompts.  $P_{i}$  is omitted from Eq. (2) and we just add the image prefix  $V$  in Eq. (3) for the experiment in order to examine the effects of the paraphrased prompts. The linear classification layer uses the fusion vector derived by multimodal Transformers as input to estimate the sentiment la",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "bel of the target. The results are shown in Table 7. The absence of paraphrased prompts has resulted in a considerable performance decrease. On the Twitter-2015 dataset, the Accuracy and macro-F1 are dropped by approximately  $4.1\\%$  and  $4.3\\%$ , while on the Twitter-2017 dataset, the Accuracy declines by about  $7.4\\%$  and the macro-F1 drops by about  $8.8\\%$ . This demonstrates how using text paraphrased prompts can provide an entity with its task-related semantics.",
+        "bbox": [
+            507,
+            407,
+            884,
+            567
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Image Prefix. We only use the paraphrased prompt  $P_{i}$  in Eq. (2) without applying the image prefix prompt  $V$  in Eq. (3) to study the importance of the image prefix. For text input,  $P_{1}, P_{2}$ , and  $P_{3}$  are proceeded for the ablation study. From Table 7, it can be seen that the performance has dropped after taking out the image prefix  $V$  (in Eq. (3)). The Accuracy decreases by  $1.1\\%$ , and the macro-F1 drops by  $0.8\\%$ , according to the average results for the Twitter-2015 dataset. On the Twitter-2017 dataset, the Accuracy and macro-F1 decline by roughly  $1.0\\%$ . This illustrates the effectiveness of the image prefix. Moreover, the ablation study also shows that different paraphrased prompts have varied outcomes, demonstrating the language models' sensitivity to prompts.",
+        "bbox": [
+            507,
+            571,
+            884,
+            828
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 Case Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            846,
+            643,
+            860
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In our case study, the compared methods are only image prefix (denoted by w/o Paraphrased Prompt), only text prompt (denoted by w/o Image Prefix),",
+        "bbox": [
+            507,
+            871,
+            882,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "4181",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and our UnifiedTMSC model with soft and hard prompts. We apply  $P_{1}$  for both w/o Image Prefix and UnifiedTMSC.",
+        "bbox": [
+            112,
+            84,
+            487,
+            131
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 4, for example (a), when there is no paraphrased prompt, the result obtained from text and image information is Neutral. When there is no image prefix, the image pre-trained embedding dominates the prediction results as Positive. Both of these two prediction results do not match the correct sentiment label Negative. For example (b), there are multiple targets that require sentiment classification. The sentiment labels predicted for place name Bahamas are Neutral, and adding appropriate prompts to both image and text can be predicted correctly. For examples (c) and (d), they are similar to example (b).",
+        "bbox": [
+            115,
+            133,
+            489,
+            341
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "These four samples further confirm the usefulness of our unified model. It can assign specific sentiment-related semantics to an entity via applying a paraphrased prompt. And prefix tuning of images can obtain better task-specific image embedding than image pre-trained embedding.",
+        "bbox": [
+            112,
+            342,
+            489,
+            439
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            451,
+            401,
+            467
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "There are currently two formats for target-oriented multimodal sentiment classification: entity-level and aspect-level. Our analysis shows that this barrier is superfluous. By incorporating paraphrased prompt and prefix vector into the multimodal input, the proposed model, i.e., UnifiedTMSC, unifies the two types of TMSC tasks. We conduct experiments on four datasets, and the results demonstrate the superiority and efficacy of our UnifiedTMSC.",
+        "bbox": [
+            112,
+            478,
+            487,
+            621
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our ongoing effort will primarily concentrate on two issues. One is to investigate how to design a paraphrasing rule to automatically generate paraphrased prompts without depending on human labor. The other is to investigate the generalizability of our model to see whether it can be used in other multimodal studies. In addition, we notice that the auto-regressive model XLNet can alleviate the problem of generating non-label words, and our future work will consider this.",
+        "bbox": [
+            112,
+            623,
+            489,
+            783
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            796,
+            220,
+            810
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our model has three limitations. The first one is that the design of paraphrased prompts relies on human experience. Although our paraphrasing rule is designed based on the relative position and synonym substitution, manual experience is still required to obtain paraphrased prompts that comply",
+        "bbox": [
+            112,
+            822,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "with grammar rules, and paraphrased prompts that comply with grammar rules may not necessarily be the best. The second limitation is that more attempts are needed to conduct experiments on multilingual pre-trained models. Furthermore, the last is that we have not explored whether our model can be extended to other multimodal research fields, which will be our future research direction.",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            509,
+            225,
+            672,
+            240
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is partially supported by NSFC, China (No.62276196).",
+        "bbox": [
+            507,
+            250,
+            880,
+            281
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            309,
+            608,
+            323
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Damian Borth, Rongrong Ji, Tao Chen, Thomas M. Breuel, and Shih-Fu Chang. 2013. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In ACM Multimedia Conference, MM '13, Barcelona, Spain, October 21-25, 2013, pages 223-232. ACM.",
+            "Xiang Chen, Ningyu Zhang, Xin Xie, Shumin Deng, Yunzhi Yao, Chuanqi Tan, Fei Huang, Luo Si, and Huajun Chen. 2022. Knowprompt: Knowledge-aware prompt-tuning with synergistic optimization for relation extraction. In WWW '22: The ACM Web Conference 2022, Virtual Event, Lyon, France, April 25 - 29, 2022, pages 2778-2788. ACM.",
+            "Jian Cui, Lin Li, and Xiaohui Tao. 2023. Be-ornot prompt enhanced hard negatives generating for memes category detection. In IEEE International Conference on Multimedia and Expo, ICME 2023, Brisbane, Australia, July 10-14, 2023, pages 174-179. IEEE.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. pages 4171-4186. Association for Computational Linguistics.",
+            "Tianhao Gao, Jun Fang, Hanyu Liu, Zhiyuan Liu, Chao Liu, Pengzhang Liu, Yongjun Bao, and Weipeng Yan. 2022. LEGO-ABSA: A prompt-based task assembly unified generative framework for multi-task aspect-based sentiment analysis. In Proceedings of the 29th International Conference on Computational Linguistics, COLING 2022, Gyeongju, Republic of Korea, October 12-17, 2022, pages 7002-7012. International Committee on Computational Linguistics.",
+            "Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. 2022. PTR: prompt tuning with rules for text classification. AI Open, 3:182-192.",
+            "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas,"
+        ],
+        "bbox": [
+            510,
+            332,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "4182",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society.",
+            "Brendan Jou, Tao Chen, Nikolaos Pappas, Miriam Redi, Mercan Topkara, and Shih-Fu Chang. 2015. Visual affect around the world: A large-scale multilingual visual sentiment ontology. In Proceedings of the 23rd Annual ACM Conference on Multimedia Conference, MM '15, Brisbane, Australia, October 26 - 30, 2015, pages 159-168. ACM.",
+            "Xincheng Ju, Dong Zhang, Rong Xiao, Junhui Li, Shoushan Li, Min Zhang, and Guodong Zhou. 2021. Joint multi-modal aspect-sentiment analysis with auxiliary cross-modal relation detection. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 4395-4405. Association for Computational Linguistics.",
+            "Zaid Khan and Yun Fu. 2021. Exploiting BERT for multimodal target sentiment classification through input space translation. In MM '21: ACM Multimedia Conference, Virtual Event, China, October 20 - 24, 2021, pages 3034-3042. ACM.",
+            "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, Online Event, 16-20 November 2020, volume EMNLP 2020 of Findings of ACL, pages 1896-1907. Association for Computational Linguistics.",
+            "Chengxi Li, Feiyu Gao, Jiajun Bu, Lu Xu, Xiang Chen, Yu Gu, Zirui Shao, Qi Zheng, Ningyu Zhang, Yongpan Wang, and Zhi Yu. 2021. Sentiprompt: Sentiment knowledge enhanced prompt-tuning for aspect-based sentiment analysis. CoRR, abs/2109.08306.",
+            "Guowei Li, Fuqiang Lin, WangQun Chen, DiWen Dong, and Bo Liu. 2022a. Prompt-based learning for aspect-level sentiment classification. In Neural Information Processing - 29th International Conference, ICONIP 2022, Virtual Event, November 22-26, 2022, Proceedings, Part III, volume 13625 of Lecture Notes in Computer Science, pages 509-520. Springer.",
+            "Jing Li, Aixin Sun, Jianglei Han, and Chenliang Li. 2022b. A survey on deep learning for named entity recognition. IEEE Trans. Knowl. Data Eng., 34(1):50-70.",
+            "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, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 4582-4597. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yan Ling, Jianfei Yu, and Rui Xia. 2022. Vision-language pre-training for multimodal aspect-based sentiment analysis. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 2149-2159. Association for Computational Linguistics.",
+            "Bin Liu, Tao Lin, and Ming Li. 2023a. Enhancing aspect-category sentiment analysis via syntactic data augmentation and knowledge enhancement. Knowl. Based Syst., 264:110339.",
+            "Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023b. Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 55(9):195:1-195:35.",
+            "Yuhang Liu, Wei Wei, Daowan Peng, and Feida Zhu. 2022. Declaration-based prompt tuning for visual question answering. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, pages 3264-3270. ijcai.org.",
+            "Sijie Long, Lin Li, Jingling Yuan, and Jianquan Liu. 2022. Momnet: Gender prediction using mechanism of working memory. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022, pages 4259-4263. ACM.",
+            "Di Lu, Leonardo Neves, Vitor Carvalho, Ning Zhang, and Heng Ji. 2018. Visual attention model for name tagging in multimodal social media. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 1990-1999. Association for Computational Linguistics.",
+            "Yaojie Lu, Qing Liu, Dai Dai, Xinyan Xiao, Hongyu Lin, Xianpei Han, Le Sun, and Hua Wu. 2022. Unified structure generation for universal information extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 5755-5772. Association for Computational Linguistics.",
+            "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR.",
+            "Ronald Seoh, Ian Birle, Mrinal Tak, Haw-Shiuan Chang, Brian Pinette, and Alfred Hough. 2021. Open aspect target sentiment classification with natural language"
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "4183",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "prompts. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 6311-6322. Association for Computational Linguistics.",
+            "Zhengxin Song, Yun Xue, Donghong Gu, Haolan Zhang, and Weiping Ding. 2023. Target-oriented multimodal sentiment classification by using topic model and gating mechanism. Int. J. Mach. Learn. Cybern., 14(7):2289-2299.",
+            "Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. 2021. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 200-212.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.",
+            "Jiawei Wang, Zhe Liu, Victor S. Sheng, Yuqing Song, and Chenjian Qiu. 2021. Saliencybert: Recurrent attention network for target-oriented multimodal sentiment classification. In Pattern Recognition and Computer Vision - 4th Chinese Conference, PRCV 2021, Beijing, China, October 29 - November 1, 2021, Proceedings, Part III, volume 13021 of Lecture Notes in Computer Science, pages 3-15. Springer.",
+            "Luwei Xiao, Ejian Zhou, Xingjiao Wu, Shuwen Yang, Tianlong Ma, and Liang He. 2022. Adaptive multi-feature extraction graph convolutional networks for multimodal target sentiment analysis. In IEEE International Conference on Multimedia and Expo, ICME 2022, Taipei, Taiwan, July 18-22, 2022, pages 1-6. IEEE.",
+            "Nan Xu and Wenji Mao. 2017. Multisentinent: A deep semantic network for multimodal sentiment analysis. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, CIKM 2017, Singapore, November 06 - 10, 2017. ACM.",
+            "Nan Xu, Wenji Mao, and Guandan Chen. 2018. A comemory network for multimodal sentiment analysis. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR 2018, Ann Arbor, MI, USA, July 08-12, 2018, pages 929-932. ACM.",
+            "Nan Xu, Wenji Mao, and Guandan Chen. 2019. Multi-interactive memory network for aspect based multimodal sentiment analysis. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 371-378. AAAI Press.",
+            "Hao Yang, Yanyan Zhao, and Bing Qin. 2022. Face-sensitive image-to-emotional-text cross-modal translation for multimodal aspect-based sentiment analysis. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 3324-3335. Association for Computational Linguistics.",
+            "Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. 2021. CPT: colorful prompt tuning for pre-trained vision-language models. CoRR, abs/2109.11797.",
+            "Jianfei Yu and Jing Jiang. 2019. Adapting BERT for target-oriented multimodal sentiment classification. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 5408-5414. ijcai.org.",
+            "Jianfei Yu, Jing Jiang, and Rui Xia. 2020. Entity-sensitive attention and fusion network for entity-level multimodal sentiment classification. IEEE ACM Trans. Audio Speech Lang. Process., 28:429-439.",
+            "Yuhai Yu, Hongfei Lin, Jiana Meng, and Zhehuan Zhao. 2016. Visual and textual sentiment analysis of a microblog using deep convolutional neural networks. Algorithms, 9(2):41.",
+            "Qi Zhang, Jinlan Fu, Xiaoyu Liu, and Xuanjing Huang. 2018. Adaptive co-attention network for named entity recognition in tweets. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 5674-5681. AAAI Press.",
+            "Zhe Zhang, Zhu Wang, Xiaona Li, Nannan Liu, Bin Guo, and Zhiwen Yu. 2021. Modalnet: an aspect-level sentiment classification model by exploring multimodal data with fusion discriminant attentional network. World Wide Web, 24(6):1957-1974.",
+            "Fei Zhao, Zhen Wu, Siyu Long, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2022. Learning from adjective-noun pairs: A knowledge-enhanced framework for target-oriented multimodal sentiment classification. In Proceedings of the 29th International Conference on Computational Linguistics, COLING 2022, Gyeongju, Republic of Korea, October 12-17, 2022, pages 6784-6794. International Committee on Computational Linguistics.",
+            "Jie Zhou, Jiabao Zhao, Jimmy Xiangji Huang, Qin-min Vivian Hu, and Liang He. 2021. MASAD: A large-scale dataset for multimodal aspect-based sentiment analysis. Neurocomputing, 455:47-58."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "4184",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130(9):2337-2348.",
+        "bbox": [
+            115,
+            85,
+            489,
+            137
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            172,
+            236,
+            187
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.1 Baselines",
+        "text_level": 1,
+        "bbox": [
+            114,
+            200,
+            236,
+            215
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) TomBERT (IJCAI, Yu and Jiang (2019)), which employ BERT for the inter-modal interactions and Target-Image matching layer to obtain a target-sensitive visual.",
+            "(2) SaliencyBERT (PRCV, Wang et al. (2021)), which design a recurrent attention mechanism to capture the inter-modality dynamics.",
+            "(3) CapTrBERT (ACM Multimedia, Khan and Fu (2021)), a BERT-based model translating the image into caption and fusing the caption and input text-entity pair.",
+            "(4) JML-MASC (EMNLP, Ju et al. (2021)), which is a multi-task learning method with the cross-modal relation detection.",
+            "(5) VLP-MABSA (ACL, Ling et al. (2022)), a task-specific pre-training vision-language model.",
+            "(6) FITE-DE-Large (EMNLP, Yang et al. (2022)), introducing a FITE method to focus on capturing emotional cues through facial expressions.",
+            "(7) MIMN (AAAI, Xu et al. (2019)), using memory network to model multimodal data and learn the interactive influences in cross-modality and self-modality.",
+            "(8) MMAP (Neurocomputing, Zhou et al. (2021)), learning the interaction between text and image, text and aspect, and image and aspect through three interactive mechanisms.",
+            "(9) ModalNet (WWW, Zhang et al. (2021)), designing a discriminant matrix to supervise the fusion of inter-modal information.",
+            "(10) CLIP (INT J COMPUT VISION, Zhou et al. (2022)), a multimodal pre-training model, converts image classification tasks into image-text matching tasks using comparative learning."
+        ],
+        "bbox": [
+            105,
+            225,
+            487,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.2 Dataset Details",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            677,
+            98
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) Twitter-2015 and Twitter-2017. The two Twitter datasets include user tweets released during 2014-2015 and 2016-2017, collected by Zhang et al. (2018) and Lu et al. (2018). Since the two publicly available multimodal datasets Twitter-2015 and Twitter-2017 only provide annotated targets in each tweet, Yu and Jiang (2019) ask three domain experts to annotate the sentiment towards each target, and take the majority label among the three annotators as the gold label. These two datasets contain multimodal tweets and annotated mentioned entities in the text, as well as the sentiment polarity of each entity, including positive, neutral, and negative. Each multimodal tweet consists of a text and a corresponding image. We follow their official splitting for training, validation and testing. The statistics of Twitter-2015 and Twitter-2017 are in Table 8.",
+            "(2) Multi-ZOL. This dataset is collected by Xu et al. (2019) from the Chinese website ZOL.com. The website consists of 40 large channels, including news, shopping malls, hardware, mobile phones, and more. They searched from pages 1 to 50 in the mobile phones channel. For each phone, only the reviews from the first 20 pages were crawled. The Multi-ZOL dataset contains 5288 multimodal reviews and each multimodal review contains a textual content, an image set, and at least one but no more than six aspects. The six aspects are cost performance, performance configuration, battery life, appearance, photography effect, and screen. For each aspect, the review has an integer sentiment score from 1 to 10, which is regarded as the sentiment label in our experiment. Actually, we convert digital sentiment labels into Chinese characters, such as '10' to 'ten'. Combining each aspect with the multimodal review, 28469 pairs of aspect review samples can be obtained. We randomly partition the dataset into training  $(80\\%)$ , development  $(10\\%)$ , and testing  $(10\\%)$  sets. The statistical information after dividing the dataset is shown in Table 9.",
+            "(3) MASAD. This dataset was collected and published by Zhou et al. (2021) based on the publicly available Visual Sentiment Ontology (VSO) dataset (Borth et al., 2013) and Multi-"
+        ],
+        "bbox": [
+            509,
+            105,
+            882,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "4185",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/cdfe4596d7e7322abcf1cb417892359505e25858fe4900a7b5630d85223d4f02.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"4\">Twitter-2015</td><td colspan=\"4\">Twitter-2017</td></tr><tr><td>#Positive</td><td>#Neutral</td><td>#Negative</td><td>Total</td><td>#Positive</td><td>#Neutral</td><td>#Negative</td><td>Total</td></tr><tr><td>Train</td><td>928</td><td>1883</td><td>368</td><td>3179</td><td>1508</td><td>416</td><td>1638</td><td>3562</td></tr><tr><td>Dev</td><td>303</td><td>679</td><td>149</td><td>1122</td><td>515</td><td>144</td><td>517</td><td>1176</td></tr><tr><td>Test</td><td>317</td><td>607</td><td>113</td><td>1037</td><td>493</td><td>168</td><td>573</td><td>1234</td></tr></table>",
+        "bbox": [
+            189,
+            80,
+            803,
+            143
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/307a928357e89963df42cc38a5ec0cb8f683c372d76cbcd6359a0460902e43eb.jpg",
+        "table_caption": [
+            "Table 8: The statistics of Twitter-2015 and Twitter-2017 datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>#1</td><td>#2</td><td>#3</td><td>#4</td><td>#5</td><td>#6</td><td>#7</td><td>#8</td><td>#9</td><td>#10</td><td>Total</td><td>Max #Image</td><td>Min #Image</td></tr><tr><td>Train</td><td>6</td><td>1163</td><td>6</td><td>1190</td><td>3</td><td>3509</td><td>0</td><td>6952</td><td>0</td><td>9948</td><td>22775</td><td>111</td><td>1</td></tr><tr><td>Dev</td><td>0</td><td>139</td><td>0</td><td>150</td><td>0</td><td>433</td><td>0</td><td>872</td><td>0</td><td>1253</td><td>2847</td><td>111</td><td>1</td></tr><tr><td>Test</td><td>0</td><td>132</td><td>0</td><td>151</td><td>0</td><td>413</td><td>0</td><td>871</td><td>0</td><td>1280</td><td>2847</td><td>72</td><td>1</td></tr></table>",
+        "bbox": [
+            147,
+            187,
+            853,
+            239
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/e886fd5f91d426d768d9d70cd493c5ab1f89b71cd7169028c2b39e8ac6c25aa5.jpg",
+        "table_caption": [
+            "Table 9: The statistics of Multi-ZOL dataset after partitioning. 'Max #Image' refers to the maximum number of image corresponding to the review, while 'Min #Image' is the minimum quantity."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">Train</td><td colspan=\"3\">Dev</td><td colspan=\"3\">Test</td></tr><tr><td>#Positive</td><td>#Negative</td><td>Total</td><td>#Positive</td><td>#Negative</td><td>Total</td><td>#Positive</td><td>#Negative</td><td>Total</td></tr><tr><td>Food</td><td>2122</td><td>392</td><td>2514</td><td>239</td><td>41</td><td>280</td><td>592</td><td>109</td><td>701</td></tr><tr><td>Buildings</td><td>969</td><td>874</td><td>1843</td><td>109</td><td>96</td><td>205</td><td>273</td><td>245</td><td>518</td></tr><tr><td>Goods</td><td>2398</td><td>1512</td><td>3910</td><td>273</td><td>162</td><td>435</td><td>743</td><td>512</td><td>1255</td></tr><tr><td>Animal</td><td>2728</td><td>1979</td><td>4707</td><td>295</td><td>229</td><td>524</td><td>1126</td><td>670</td><td>1796</td></tr><tr><td>Human</td><td>1805</td><td>1648</td><td>3453</td><td>194</td><td>190</td><td>384</td><td>503</td><td>464</td><td>967</td></tr><tr><td>Plant</td><td>2542</td><td>2341</td><td>4883</td><td>277</td><td>266</td><td>543</td><td>1269</td><td>947</td><td>2216</td></tr><tr><td>Scenery</td><td>3222</td><td>1760</td><td>4982</td><td>378</td><td>176</td><td>554</td><td>907</td><td>490</td><td>1397</td></tr><tr><td>Total</td><td>15786</td><td>10506</td><td>26292</td><td>1765</td><td>1160</td><td>2925</td><td>5413</td><td>3437</td><td>8850</td></tr></table>",
+        "bbox": [
+            157,
+            287,
+            845,
+            407
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 10: The statistical information of MASAD datasets. The seven domains are Food, Buildings, Goods, Animal, Human, Plant, Scenery.",
+        "bbox": [
+            129,
+            416,
+            873,
+            444
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "lingual Visual Sentiment Ontology (MVSO) dataset (Jou et al., 2015). Zhou et al. (2021) selected the samples from a partial VSO dataset (approximately 120k samples) that can express significant sentiments (about 38k samples) and categorized them into 7 domains, resulting in the MASAD dataset. The seven domains of the MASAD are food, buildings, goods, animal, human, plant, and scenery. Each domain encompasses multiple aspects, such as the animal domain, including cat, dog, horse, and so on. According to our statistics, there are a total of 57 predefined aspects. This dataset only includes training and testing sets, both containing positive and negative samples. We partition each domain in the training set into a new training set and a validation set in a 9:1 ratio, keeping the original testing set unchanged. The statistics of MASAD are in Table 10.",
+        "bbox": [
+            142,
+            470,
+            489,
+            777
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "In addition, there are instances of duplicate data in the testing and training sets in this dataset. In the domain of animal and plant, according to our statistics, about  $81.6\\%$  and  $62.2\\%$  of the data in the testing set has appeared in the training set. In order to compare more fairly with other baseline models, we conduct experiments in other domains besides ani",
+        "bbox": [
+            142,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "mal and plant.",
+        "bbox": [
+            539,
+            472,
+            650,
+            487
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "4186",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Descriptive Prompt Paraphrasing for Target-Oriented Multimodal Sentiment Classification/9ad0a505-4aa3-4e82-936d-bde5e9f488e4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b03e26b800a92c879d3e3fff3f54d45dc228a49c5c0b7a4d27638553392ead32
+size 790474

2023/Descriptive Prompt Paraphrasing for Target-Oriented Multimodal Sentiment Classification/full.md

@@ -0,0 +1,378 @@
+# Descriptive Prompt Paraphrasing for Target-Oriented Multimodal Sentiment Classification
+
+Dan Liu  $^{1}$ , Lin Li  $^{1}$ , Xiaohui Tao  $^{2}$ , Jian Cui  $^{1}$ , Qing Xie
+
+<sup>1</sup> Wuhan University of Technology, Wuhan, China
+
+<sup>2</sup> The University of Southern Queensland, Toowoomba, Australia
+
+$^{1}$  {liunnew, cathylilin, 318971, felixxq} @whut.edu.cn
+
+$^{2}$  Xiaohui.Tao@unisq.edu.au
+
+# Abstract
+
+Target-Oriented Multimodal Sentiment Classification (TMSC) aims to perform sentiment polarity on a target jointly considering its corresponding multiple modalities including text, image, and others. Current researches mainly work on either of two types of targets in a decentralized manner. One type is entity, such as a person name, a location name, etc. and the other is aspect, such as 'food', 'service', etc. We believe that this target type based division in task modelling is not necessary because the sentiment polarity of the specific target is not governed by its type but its context. For this reason, we propose a unified model for target-oriented multimodal sentiment classification, so called UnifiedTMSC. It is prompt-based language modelling and performs well on four datasets spanning the above two target types. Specifically, we design descriptive prompt paraphrasing to reformulate TMSC task via (1) task paraphrasing, which obtains paraphrased prompts based on the task description through a paraphrasing rule, and (2) image prefix tuning, which optimizes a small continuous image vector throughout the multimodal representation space of text and images. Conducted on two entity-level multimodal datasets: Twitter-2015 and Twitter-2017, and two aspect-level multimodal datasets: Multi-ZOL and MASAD, the experimental results show the effectiveness of our UnifiedTMSC.
+
+# 1 Introduction
+
+With the emergence of new media and advanced technology, the forms of information released by people are quietly changing, from mono-modality to multi-modality now, such as text, images, etc (Xu and Mao, 2017). This also pushes researchers to conduct multimodal learning (Yu et al., 2016; Xu et al., 2018; Long et al., 2022). For sentiment analysis, both text and image are highly correlated with sentiment polarity. Moreover, they can complement and reinforce each other (Xu et al., 2019).
+
+![](images/e54dc4cfb903ad65d631dbed57c7543878be0a411450d87e3276d9be8c6b6450.jpg)  
+(a) Entity-level MSC
+
+![](images/c2210b2c262129dd8d8214aadbfb6f10f8cd6bfd4ab9e3dad4adee45770a3648.jpg)  
+(b) Aspect-level MSC  
+Figure 1: Two forms of target-oriented multimodal sentiment classification.
+
+At present, fine-grained Multimodal Sentiment Classification (MSC) includes two main tasks: entity-level MSC and aspect-level MSC, as shown in Figure 1. (1) For entity-level, the entity and its context are encoded as independent input text in some studies(Yu and Jiang, 2019; Yu et al., 2020; Wang et al., 2021; Zhao et al., 2022). Others jointly consider the encoding of an entity with its contextual learning, which is their main focus to achieve good MSC performance (Khan and Fu, 2021; Xiao et al., 2022; Yang et al., 2022). For example, Xiao et al. (2022) presented a dual stream adaptive multifeature extraction graph convolutional network to convert an image into its caption. (2) For aspect-level, the aspect and its context are being encoded individually due to the semantics involved in an aspect itself (Xu et al., 2019; Zhang et al., 2021; Zhou et al., 2021). For example, Xu et al. (2019)
+
+studied a multi-interactive memory network model. These TMSC works can promote human decisions by assisting users in knowing about certain targets.
+
+As we know, an entity usually is a person name, or a location name, etc (Li et al., 2022b). Such entity target only has a specific meaning when connecting it with a specific modality content, which means it is difficult to accurately understand the entity without its context. On the contrary, an aspect itself in some degree can represent what it means even without its context. Because of this obvious difference, previous studies show their interest in differently modelling the two tasks to capture target related context (Xu et al., 2019; Yu et al., 2020; Zhang et al., 2021; Song et al., 2023). For example, in Figure 1(a), "Chuck Bass", "MCM", and "Iran" do not have any exact meaning when they are disconnected from the specific context. However, in Figure 1(b), both aspect term and aspect category have their own meanings with a hidden sentiment tendency. For the "battery life" of mobile phones, the first reaction is that the longer the battery life, the better. Despite the above differences among the TMSC tasks, from the perspective of sentiment classification, the goal of TMSC is to predict the sentiment polarity of a target no matter whether it is an entity or an aspect. Therefore, in our view, the boundary is unnecessary.
+
+In this paper, we propose a unified TMSC model via prompt based language modelling, so called UnifiedTMSC, which is independent of the target type in TMSC. Our core is to reconstruct the two TMSC tasks through descriptive prompt paraphrasing. The prompts we design can place entity and aspect in their context, while also being close to the TMSC task description. To achieve this goal, we carry out our work from two aspects: (1) task paraphrasing. The task description is transformed into a seed prompt, and different paraphrased prompts are obtained by using our paraphrasing rule. They serve as discrete prompts for the text and fit into the Masked Language Modeling (MLM) format. And (2) image prefix tuning (Li and Liang, 2021). A segment vector is initialized for the image pretrained embedding as the prefix continuous prompt. In the subsequent multimodal continuous representation space, the image pre-trained embedding is fixed and only some segment of the initialized vector is optimized. In this way, sentiment labels are generated through the cloze-filling method.
+
+In our extensive experiments, our UnifiedTMSC
+
+model achieves state-of-the-art performance on two entity-level datasets: Twitter-2015 and Twitter-2017, and two aspect-level datasets: Multi-ZOL and MASAD. (1) The results of the task description based prompts are superior to those of arbitrary prompt templates. (2) On two entity-level datasets, our model improves Accuracy by  $1.0\% - 2.8\%$  and  $1.5\% - 3.7\%$ , macro-F1 by  $1.5\% - 4.5\%$  and  $1.7\% - 7.5\%$ , respectively. (3) On two aspect-level datasets, our model gains of  $6.46\% - 8.87\%$  and  $1.66\% - 3.02\%$  on Accuracy,  $3.77\% - 5.01\%$  and  $2.19\% - 3.25\%$  on macro-F1 are derived. The experimental results demonstrate the effectiveness of our UnifiedTMSC.
+
+# 2 Related Work
+
+# 2.1 Entity-level MSC
+
+As a pioneer, Yu and Jiang (2019) proposed a BERT-based multimodal architecture to determine the sentiment polarity of an entity. Yu et al. (2020) introduced an entity-sensitive attention and fusion network. And Wang et al. (2021) put forward a recurrent attention network. Khan and Fu (2021) introduced an input space translation framework to construct image context from the image. Zhao et al. (2022) used the adjective-noun pairs extracted from images as the knowledge enhancement based on Yu and Jiang (2019) and Wang et al. (2021). Moreover, Yang et al. (2022) explored facial information in images to obtain visual and sentiment clues.
+
+In the research stated above, the entity can be encoded as a distinct text input, or entity and context can be combined as the text input. Their goal is to more effectively learn the semantics related to entity sentiment.
+
+# 2.2 Aspect-level MSC
+
+Aspect-level multimodal classification was first proposed by Xu et al. (2019), and they introduced a multi-interactive memory network to analyze multiple correlations in multimodal data. Zhang et al. (2021) presented a multimodal fusion discriminant attention network and designed a discriminant matrix to supervise the modality fusion. Zhou et al. (2021) conducted a multimodal interaction model that learns the relationships between text, image, and target aspect through interaction layers and adversarial training. One key difference between an aspect and an entity is that the aspect has its own semantics inferred from the aspect words. Therefore, existing research usually regards the aspect
+
+![](images/4b659a4531a1d0f63dac519f07ec64fb7a403f8763130ce00d627737033e3e45.jpg)  
+Figure 2: The model of our proposed UnifiedTMSC. Seed prompt and keyword 'sentiment' are obtained based on the task description and are converted into diverse paraphrased prompts through the paraphrasing rule (Equation 1). The paraphrased prompt and the original text input are concatenated to form the new text input which is encoded by the encoder to gain text embedding. The image is divided into multiple regions and each of the regions is initialized as a vector that is used as the prefix prompt for image pre-trained embedding. The text embedding and image embedding are undergone cross attention to gain modality fusion embedding.
+
+![](images/81a1fa7c04df0a379853669bcaa530a31807b2d6041a1dc1e2d6fd568e22104e.jpg)
+
+![](images/01d7b8540d4cc6764f9510df13190097385a35470e0b4c9a597f90ff28553a16.jpg)
+
+itself as an additional input.
+
+Our focus: In contrast to previous studies, our model can run across TMSC tasks. It combines the target (entity or aspect) and its context as a text input using task description based paraphrased prompts. We can get sentiment-related semantics about the target by providing its context in task description based prompts.
+
+# 2.3 Prompt paraphrasing
+
+Prompt tuning has received increasing attention recently (Radford et al., 2021; Yao et al., 2021) and has been successfully applied in many domains (Han et al., 2022; Liu et al., 2023b). For example, in the field of Question Answering (QA), Khashabi et al. (2020) reformulated many QA tasks as a text generation problem by fine-tuning seq2seq-based pre-trained models and appropriate prompts from the context and questions. For Information Extraction (IE), Chen et al. (2022) first explored the application of fixed-prompt LM Tuning in relation extraction and Lu et al. (2022) applied prompt to control the information to be extracted. In other research fields, Cui et al. (2023) used prompt learning in text input to conduct the meme mining task. Recently, the prompt has been used for the task involving fine-grained text sentiment analysis, and
+
+the results are promising (Seoh et al., 2021; Li et al., 2021, 2022a; Gao et al., 2022; Liu et al., 2023a).
+
+Inspired by the above studies, our unified model is through the task description based prompt paraphrasing with jointly soft and hard prompt tuning.
+
+# 3 Method
+
+# 3.1 Task Formulation
+
+Given a multimodal samples  $D$ , for each sample  $d \in D$ , it contains a sentence  $S$  with  $n$  words  $(w_{1}, w_{2}, \ldots, w_{n})$  and one or more related images  $I$ , as well as a target  $T$  which contains  $m$  words  $(w_{1}, w_{2}, \ldots, w_{m})$  and is a sub-sequence of  $S$  or predefined phrase. For the target  $T$ , it is associated with a sentiment label  $Y$ . In general,  $Y \in \{positive, neutral, negative\}$ , and different tasks may have different sentiment labels. Our goal is to learn a target-oriented sentiment classifier that can correctly predict the sentiment label for each sample  $X = (S, I, T)$ .
+
+# 3.2 Overview
+
+As shown in Figure 2, our model consists of two modules: task paraphrasing (hard prompt) and image prefix tuning (soft prompt). For the given multimodal data  $X = (S, I, T)$ , we obtain paraphrased prompts  $(P_{1}, P_{2} \dots)$  through the task description of
+
+<table><tr><td>Paraphrased Prompts</td><td>&lt; Y, Position &gt;</td><td>&lt; T,Position &gt;</td><td>&lt; K, Position &gt;</td><td>Synonym of K</td></tr><tr><td>{target} express a [MASK] sentiment.</td><td>&lt; [MASK],M &gt;</td><td>&lt; {target},B &gt;</td><td>&lt; sentiment,E &gt;</td><td>no</td></tr><tr><td>The emotion of {target} is [MASK].</td><td>&lt; [MASK],E &gt;</td><td>&lt; {target},M &gt;</td><td>&lt; emotion,B &gt;</td><td>yes</td></tr><tr><td>{target}&#x27;s emotion is [MASK].</td><td>&lt; [MASK],E &gt;</td><td>&lt; {target},B &gt;</td><td>&lt; emotion,M &gt;</td><td>yes</td></tr><tr><td>The [MASK] feeling is {target}&#x27;s sentiment.</td><td>&lt; [MASK],B &gt;</td><td>&lt; {target},M &gt;</td><td>&lt; sentiment,E &gt;</td><td>no</td></tr><tr><td>The sentiment is [MASK] and conveyed by {target}.</td><td>&lt; [MASK],M &gt;</td><td>&lt; {target},E &gt;</td><td>&lt; sentiment,B &gt;</td><td>no</td></tr><tr><td>The emotion expressed by {target} is [MASK].</td><td>&lt; [MASK],E &gt;</td><td>&lt; {target},M &gt;</td><td>&lt; emotion,B &gt;</td><td>yes</td></tr><tr><td>A [MASK] sentiment is expressed towards to {target}.</td><td>&lt; [MASK],B &gt;</td><td>&lt; {target},E &gt;</td><td>&lt; sentiment,M &gt;</td><td>no</td></tr></table>
+
+Table 1: Some examples of paraphrased prompts, as well as the relative positions of  $Y$  (sentiment labels, replaced by [MASK] token, marked with bold),  $T$  (the targets that need to recognize sentiment polarity, annotated with italics), and  $K$  (keywords extracted from seed prompt or replaced with synonyms, annotated with underline) in the prompts. 'Synonym of  $\mathrm{K}$  ’ indicates whether to replace the keyword 'sentiment' extracted from the seed prompt with synonyms. 'B', 'M', and 'E' respectively represent the positions in the paraphrased prompt as: 'Beginning', 'Middle', and 'Ending'.
+
+TMSC, which are used as prefixes for text input  $S$  in the task paraphrasing module (the left of Figure 2). Paraphrased prompt  $P_{i}$  contains a [MASK] token and is connected with text  $S$ . The text embedding  $R$  is obtained through an encoder in the bottom-right of Figure 2. In the image prefix tuning module, we apply an initialization vector  $V$  for pre-train image embedding  $E$  as the continuous prefix. After text embedding  $R$  and image embedding  $(V + E)$  are added to their respective position encoding, the fusion vector is obtained through the multimodal Transformer in the middle right of Figure 2. We take the hidden layer vector  $H$  and pass it through the MLM head to get the prediction score of [MASK] position for each word in the vocabulary  $C$ . Finally, the cross entropy loss  $L_{MLM}$  of the prediction result  $Output$  and the true sentiment label  $Y$  is calculated.
+
+# 3.3 Task Paraphrasing
+
+In order to provide a specific context for a target, i.e., entity and aspect, we add the target-involved prompts to the text input. Since we change the original sentiment classification task in this way to the pre-trained MLM task, it is crucial to figure out how to develop prompts that are suitable for the original task, that is to say, the prompts should be consistent with the expression of the task description. Inspired by prompt tuning in various domains, such as the visual grounding problem (Yao et al., 2021) and visual question answering task (Liu et al., 2022), we propose the task paraphrasing module to draw paraphrased prompts as a solution to the above issue.
+
+We get the seed prompt according to the task description composed of natural language, and take
+
+the task-related keyword  $K$  ('sentiment') from the task description. The seed prompt is transformed through the paraphrasing rule, and guides the generation of paraphrased prompts that are close to the original task description. Our paraphrasing rule can be formalized in the following form:
+
+$$
+f (Y) \wedge f (T) \wedge f (K) \wedge g (K) \tag {1}
+$$
+
+where the function  $f$  represents the relative position in the paraphrased prompts and the substitution of synonyms for a keyword ('sentiment') is illustrated by the function  $g$ .  $f(Y), f(T), f(K) \in \{B, M, E\}$  (meaning 'Beginning', 'Middle', 'Ending'), which respectively stands for the relative position of the sentiment label, the target entity, and the keyword derived from the task description.  $g(K) \in \{yes, no\}$  means whether to replace synonyms of keywords  $K$ . Moreover, synonyms are synonymous explanations for the keyword 'sentiment' in the dictionary. For example, in the Bing dictionary, the synonyms for 'sentiment' include 'emotion', 'feeling', 'opinion', etc.
+
+The task paraphrasing module is shown in the left part of Figure 2. Generally speaking, based on the relative position combination of  $Y$ ,  $K$ , and  $T$ , one seed prompt can be paraphrased to gain multiple candidate prompts. If we replace the keyword with different synonyms, we will receive more paraphrased prompts. Some examples of paraphrased prompts  $P_{i}$  are listed in Table 1. When paraphrased prompt  $P_{i}$  is in the training phase, the position of  $Y$  is replaced by [MASK] token which is the prediction object. Finally, the  $P_{i}$  and text  $S$  are concatenated to obtain a new text input:
+
+$$
+S ^ {\text {n e w}} = [ C L S ] P _ {i} S [ S E P ] \tag {2}
+$$
+
+![](images/bf2e1f42ed723ff41e703b4494bd19293b9129cdd41e4dbe23ab1ab6b77a403a.jpg)  
+Figure 3: An example of image prefix tuning.  $v_{P_{idx}}$  is random initialization of the prefix sequence and  $e_{I_{idx}}$  is the image pre-trained embedding.
+
+# 3.4 Image Prefix Tuning
+
+Discrete prompts, i.e., hard prompts, in the text are natural and easy to understand. For images, directly applying discrete prompts cannot ensure alignment between text based prompts and images since there are modality gaps between different modalities. Therefore, finding suitable prompts for images is crucial. This module mainly focuses on how to add prompts to images to facilitate a better modality fusion of images and text.
+
+Inspired by (Tsimpoukelli et al., 2021) and (Li and Liang, 2021), we introduce a continuous vector as a prefix, i.e., soft prompt, to image pre-trained embedding, as shown in Figure 3. We first segment the image into  $r$  regions, and then initialize a vector  $v_{i}$ $(i\in \{1,2,\dots r\})$  for each region to form the prefix embedding  $V = \{v_{1},v_{2},\ldots v_{r}\}$  and  $V\in \mathbb{R}^{r\times 2048}$ . Here,  $P_{idx}$  denotes the indices of the prefix sequence.  $|P_{idx}|$  denotes the length of the prefix and  $|I_{idx}|$  indicates the length of the image pre-trained embedding. In  $E = \{e_1,e_2,e_3,e_4,e_5\}$ ,  $e_i\in \mathbb{R}^{2048}$  refers to the image embedding of the  $i$ -th region. The prefix embedding  $V = \{v_{1},v_{2},v_{3}\}$  and  $V\in \mathbb{R}^{3\times 2048}$ . If there are multiple images corresponding to the text, we initialize a soft prompt for each image and their averaged vector is used as the final prefix prompt for the images.  $V$  and  $E$  are concatenated to obtain new image embedding  $E^{new}$ .
+
+$$
+E ^ {\text {n e w}} = V \oplus E \tag {3}
+$$
+
+$$
+E _ {(i)} ^ {\text {n e w}} = \left\{ \begin{array}{c c} v _ {i} & i \in P _ {i d x} \\ e _ {(i - | P _ {i d x} |)} & \text {o t h e r w i s e} \end{array} \right. \tag {4}
+$$
+
+$E^{new}$  and  $S^{new}$  are image input and text input respectively, and the subsequent encoding process is carried out together.
+
+<table><tr><td></td><td>Twitter-2015</td><td>Twitter-2017</td><td>Multi-ZOL</td><td>MASAD</td></tr><tr><td></td><td>Text-Image</td><td>Text-Image</td><td>Text-Image</td><td>Text-Image</td></tr><tr><td>Train</td><td>3179</td><td>3562</td><td>22775</td><td>26292</td></tr><tr><td>Dev</td><td>1122</td><td>1176</td><td>2847</td><td>2925</td></tr><tr><td>Test</td><td>1037</td><td>1234</td><td>2847</td><td>8850</td></tr><tr><td>Total</td><td>5338</td><td>5972</td><td>28469</td><td>38532</td></tr></table>
+
+Table 2: Statistics of four TMSC datasets.
+
+# 3.5 Multimodal Transformer With MLM
+
+Training. Given a triplet  $(S, I, T, Y)$ , after the task paraphrasing module and image prefix tuning module, the updated text input  $S^{new}$  and image embeddings  $E^{new}$  are obtained. Text encoder encodes  $S^{new}$  to gain text embedding  $R$ .  $R$  and the image embedding  $E^{new}$  undergo cross-attention through a Transformer to obtain the fusion embedding  $H$ . In this multimodal Transformer, the image pre-trained embedding  $E$  is fixed and not updated, and only  $V$  is updated. Fusion embedding  $H$  passes through an MLP to obtain prediction scores Logit:
+
+$$
+\operatorname {L o g i t} = M L P _ {M L M} (H) \tag {5}
+$$
+
+The word with the highest prediction score is the prediction result:
+
+$$
+\text {O u t p u t} = \operatorname {a r g m a x} (\operatorname {L o g i t}) \tag {6}
+$$
+
+Finally, the predicted result and the true sentiment label  $Y$  are calculated in the cross-entropy loss to optimize our model.
+
+$$
+L _ {M L M} = - \log \left(\frac {\exp (L o g i t _ {i})}{\sum_ {j = 0} ^ {| C |} \exp (L o g i t _ {j})}\right) \tag {7}
+$$
+
+where  $i$  is the true sentiment label number and  $C$  is the size of the language model vocabulary.
+
+Inference. In the inference stage, given a triplet  $(S,I,T)$ , the sentiment polarity  $Y^{\prime}$  of  $T$  is determined via the triplet.
+
+$$
+Y ^ {\prime} = \operatorname {a r g m a x} \left(\operatorname {L o g i t} _ {[ M A S K ]}\right) \tag {8}
+$$
+
+After obtaining the fusion embedding of text and images, we take the argmax of the logit of the [MASK] position to obtain the final prediction result. Finally, for non-label words generated by the model, the answer engineering is used to map them to sentiment labels.
+
+# 4 Experiment
+
+# 4.1 Experimental Setup
+
+Datasets. According to Zhou et al. (2021), there are four datasets for target-oriented multimodal
+
+sentiment classification, including two entity-level datasets: Twitter-2015 (Zhang et al., 2018; Yu and Jiang, 2019), Twitter-2017 (Lu et al., 2018; Yu and Jiang, 2019) and two aspect-level datasets: Multi-ZOL (Xu et al., 2019), MASDA (Zhou et al., 2021) The statistics information of these four datasets are shown in Table 2, and their details about the data partitioning and label types are presented in A.2.
+
+Evaluation Metrics. To fairly compare with state-of-the-art approaches, our UnifiedTMSC is evaluated across two TMSC tasks and adopts the Accuracy (Acc) and Macro-F1 score (F1), following Yu and Jiang (2019) and Ling et al. (2022).
+
+Implementation Details. For the Twitter-2015, Twitter-2017, and MASAD, the batch size is set to 16 and the epochs are set to 6. The text encoder used is BERT-base-uncased (Devlin et al., 2019). For the Multi-ZOL dataset, in order to ensure fairness in the experiment, we employ BERT-base-chinese (Devlin et al., 2019) as the text encoder, the batch size is 8, and the epochs as 6. Moreover, multilingual pre-trained models can also be used as our text encoders.
+
+For all datasets, we apply Resnet-50 (He et al., 2016) as the image encoder to get the image prefix embedding, and the max length of the new text input  $S^{new}$  (in Eq. (2)) is 96. The model learning rate is set as 1e-5 and the dropout rate is 1e-2. Four layer Transformers (Vaswani et al., 2017) are aimed to perform cross attention between different modalities and the pre-training parameters are not loaded. Four NVIDIA Titan Xp GPUs, each with 12GB of memory, are employed in our experiments which are done on a CentOS computer. The deep learning framework is Pytorch, and AdamW is used as the optimizer.
+
+# 4.2 Compared Baselines
+
+Because previous work has separated entity-level MSC and aspect-level MSC, the baseline models for each task are different. For the Twitter-2015 and Twitter-2017 datasets, we compare six baselines including TomBERT (IJCAI, Yu and Jiang (2019)), SaliencyBERT (PRCV, Wang et al. (2021)), CapTrBERT (ACM Multimedia, Khan and Fu (2021)), JML-MASC (EMNLP, Ju et al. (2021)), VLP-MABSA (ACL, Ling et al. (2022)), FITE-DE-Large (EMNLP, Yang et al. (2022)). For the Multi-ZOL and MASAD datasets, our model is compared with MIMN (AAAI, Xu et al. (2019)), ModalNet (WWW, Zhang et al. (2021)), MMAP
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Twitter-2015</td><td colspan="3">Twitter-2017</td></tr><tr><td>Acc</td><td>F1</td><td>P-Value</td><td>Acc</td><td>F1</td><td>P-Value</td></tr><tr><td>TomBERT</td><td>77.2</td><td>71.8</td><td>0.0194</td><td>70.5</td><td>68.0</td><td>0.0156</td></tr><tr><td>SaliencyBERT</td><td>77.0</td><td>72.4</td><td>0.0109</td><td>69.7</td><td>67.2</td><td>0.0156</td></tr><tr><td>CapTrBERT</td><td>78.0</td><td>73.3</td><td>0.0120</td><td>72.3</td><td>70.2</td><td>0.0134</td></tr><tr><td>JML-MASC</td><td>78.7</td><td>-</td><td>-</td><td>72.7</td><td>-</td><td>-</td></tr><tr><td>VLP-MABSA</td><td>78.6</td><td>73.8</td><td>0.0131</td><td>73.8</td><td>71.8</td><td>0.0111</td></tr><tr><td>FITDE-DE-Large</td><td>78.8</td><td>74.8</td><td>0.0048</td><td>73.9</td><td>73.0</td><td>0.0016</td></tr><tr><td>UnifiedTMSC</td><td>79.8</td><td>76.3</td><td>-</td><td>75.4</td><td>74.7</td><td>-</td></tr></table>
+
+Table 3: The experimental results on the multimodal entity-level datasets: Twitter-2015 and Twitter-2017. The results presented in the table are the average of different prompt results.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Multi-ZOL</td><td colspan="3">MASAD</td></tr><tr><td>Acc</td><td>F1</td><td>P-Value</td><td>Acc</td><td>F1</td><td>P-Value</td></tr><tr><td>MIMN</td><td>61.59</td><td>60.51</td><td>0.0316</td><td>94.52</td><td>93.55</td><td>0.0019</td></tr><tr><td>ModalNet</td><td>62.71</td><td>60.94</td><td>0.0266</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MMAP</td><td>-</td><td>-</td><td>-</td><td>95.29</td><td>94.61</td><td>0.0049</td></tr><tr><td>CLIP</td><td>64.00</td><td>59.70</td><td>0.0092</td><td>95.88</td><td>94.21</td><td>0.0079</td></tr><tr><td>UnifiedTMSC</td><td>70.46</td><td>64.71</td><td>-</td><td>97.54</td><td>96.80</td><td>-</td></tr></table>
+
+Table 4: The experimental results on the multimodal aspect-level datasets: Multi-ZOL and MASAD.
+
+(Neurocomputing, Zhou et al. (2021)) and CLIP (INT J COMPUT VISION, Zhou et al. (2022)). The detailed introduction of all the baseline models mentioned above is in Section A.1.
+
+# 4.3 Experimental Results and Analysis
+
+# 4.3.1 Overall Performance
+
+The experimental results on multimodal entity-level and aspect-level datasets are presented in Table 3 and Table 4 respectively. The best results on each metric are marked in bold and the second best results are highlighted with an underline.
+
+Multimodal Entity-level Datasets Results. As reported in Table 3, compared with the baselines, our UnifiedTMSC makes significant improvements in entity-level MSC. On the Twitter-2015 dataset, our improvement is approximately  $1.0\%$  on Accuracy and  $1.5\%$  on the macro-F1 compared to the FITE-DE-Large. On the Twitter-2017 dataset, we achieve improvements over multimodal baseline VLP-MABSA on the Accuracy by  $1.5\%$  and  $1.7\%$  on the macro-F1, which indicates that using prompt tuning to fuse entity and context can achieve a better sentiment-related semantic understanding of an entity, resulting in better classification results.
+
+Multimodal Aspect-level Datasets Results. Our UnifiedTMSC performs best in two datasets among all multimodal baselines, as listed in Table 4. This demonstrates the effectiveness of our proposed unified model based on prompt paraphrasing. Especially on the Multi-ZOL dataset, our UnifiedTMSC outperforms the ModalNet on the Accuracy by  $7.75\%$  and  $3.77\%$  on the macro-F1. On
+
+<table><tr><td rowspan="2">Method</td><td></td><td colspan="2">Food</td><td colspan="2">Buildings</td><td colspan="2">Goods</td><td colspan="2">Human</td><td colspan="2">Scenery</td></tr><tr><td></td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>MIMN</td><td>AAAI 2019</td><td>94.72</td><td>91.39</td><td>96.26</td><td>95.80</td><td>95.93</td><td>95.87</td><td>92.54</td><td>92.31</td><td>93.17</td><td>92.38</td></tr><tr><td>MMAP</td><td>Neucomp 2021</td><td>95.75</td><td>92.89</td><td>96.86</td><td>96.85</td><td>96.55</td><td>96.44</td><td>92.74</td><td>92.74</td><td>94.57</td><td>94.15</td></tr><tr><td>CLIP</td><td>IJCV 2022</td><td>96.15</td><td>93.05</td><td>95.84</td><td>94.27</td><td>97.36</td><td>96.88</td><td>94.20</td><td>93.24</td><td>95.84</td><td>93.59</td></tr><tr><td>UnifiedTMSC</td><td></td><td>98.00</td><td>96.02</td><td>98.26</td><td>98.25</td><td>98.08</td><td>98.01</td><td>97.82</td><td>96.49</td><td>95.56</td><td>95.24</td></tr></table>
+
+the MASAD dataset, compared to the MMAP, ours improve performance by about  $2.25\%$  on Accuracy and  $2.19\%$  on the macro-F1.
+
+Specifically, there are multiple domains in the MASAD dataset, we conduct experiments on each of them, and the experimental results are shown in Table 5. It is clear from the results that our model has achieved large improvements in each domain.
+
+Finally, t-tests are conducted to demonstrate the effectiveness of UnifiedTMSC. From the P-value of other models in Table 3 and Table 4, it can be found that all P-values are less than 0.05. This shows a significant difference in statistics between UnifiedTMSC and other models.
+
+After comparing and analyzing the experimental results, we can summarize the following two points from our prompt tuning:
+
+I. Our prompt paraphrasing method delivers the target's context and fits the TMSC job effectively, and it produces good results, demonstrating the efficacy of our unified model.  
+II. Utilizing the target as a separate input has worse results than taking the target and context together as text input. This shows that contextual information affects the target's semantics, and a contextual content that is appropriate for the task will result in a well understanding of a target with semantics.
+
+# 4.3.2 The Effect of Prompt Designs
+
+For the Twitter-2015 and Twitter-2017 datasets, we select three paraphrased prompts from Table 1 to conduct the experiments. The three selected paraphrased prompts are as follows:
+
+-  $P_{1}$ : {target} express a [MASK] sentiment.  
+-  $P_{2}$ : The emotion of {target} is [MASK].  
+-  $P_{3}$ : A [MASK] sentiment is expressed towards to {target}.
+
+In addition, to verify the performance of the task paraphrasing module. We design three arbitrary prompt templates and compare them with the above
+
+Table 5: The experimental results of each domain in the MASAD dataset. Because the accuracy and Macro F1 score in the domain of animal and plant reach  $99.07\% - 99.22\%$  due to the duplicated samples, they will not be displayed here.  
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="2">Twitter-2015</td><td colspan="2">Twitter-2017</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td rowspan="3">paraphrased prompts</td><td>P1</td><td>80.2</td><td>77.6</td><td>76.9</td><td>76.7</td></tr><tr><td>P2</td><td>81.1</td><td>77.1</td><td>76.6</td><td>75.9</td></tr><tr><td>P3</td><td>79.8</td><td>77.5</td><td>76.2</td><td>75.9</td></tr><tr><td rowspan="3">arbitrary prompts</td><td>P1&#x27;</td><td>79.1</td><td>75.5</td><td>73.8</td><td>73.7</td></tr><tr><td>P2&#x27;</td><td>79.0</td><td>75.1</td><td>74.3</td><td>73.6</td></tr><tr><td>P3&#x27;</td><td>79.4</td><td>75.1</td><td>74.5</td><td>72.9</td></tr></table>
+
+Table 6: The experimental results of different prompts. The results of paraphrased prompts are superior to the arbitrary prompts, and  $P_1'$ ,  $P_2'$  and  $P_3'$  are comparable. In the experiments, one of the paraphrased prompts can be selected for experimentation.
+
+three paraphrased prompts. The three arbitrary prompts are as follows:
+
+-  $P_1'$ : I feel the {target} is [MASK].  
+-  $P_2'$ : The (target) made me feel [MASK].  
+-  $P_3^7$ : I [MASK] the {target}.
+
+where  $\{target\}$  is the entity that needs to determine sentiment polarity, and [MASK] represents the masked word, i.e. sentiment label. The masked word in the  $P_1', P_2', P_3'$  is {good, ok, bad}, {good, indifferent, bad} and {love, dislike, hate} respectively. After the masked words are generated, we perform answer engineering to map the predicted results to the sentiment polarity set, that is, the probabilities of these predicted words are made to be equal to the probabilities of being Positive, Neutral, and Negative.
+
+The results of several different prompts are shown in Table 6. Through analysis and comparison, we can obtain the following summaries:
+
+I. Our paraphrased prompts created by using the task description are much superior to arbitrary prompt templates, demonstrating the value of our task paraphrasing module in producing paraphrased prompts that are appropriate for the original sentiment classification task. In addition, the performance of different paraphrased prompts is comparable, and in subsequent experiments, anyone can be selected for training and inference.
+
+<table><tr><td>Image</td><td>(a) Fan Throws Water Bottle at Justin Bieber After He Says He Doesn&#x27;t Know the ...</td><td>(b) Kim Kardashian goes all out for Kanye West&#x27;s 40 th birthday in the Bahamas</td><td>(c) RT @ TrumpDoral : Congratulations to the new # Miss Universe , Miss Colombia , Paulina Vega !</td><td>(d) Nba - The Cavs are Shocked Draymond Green Keeps Getting Away With Kic -</td></tr><tr><td>(Text,Label)</td><td>(Justin Bieber, Negative)</td><td>(Kim Kardashian, Positive) (Kanye West, Positive) (Bahamas, Neutral)</td><td>(Miss Colombia, Positive) (Paulina Vega, Positive)</td><td>(Nba, Neutral) (Cavs, Neutral) (Draymond Green, Negative)</td></tr><tr><td>w/o 
+Paraphrased 
+Prompt</td><td>(Justin Bieber, Neutral) ×</td><td>(Kim Kardashian, Neutral) × (Kanye West, Neutral) × (Bahamas, Neutral) √</td><td>(Miss Colombia, Neutral) × (Paulina Vega, Neutral) ×</td><td>(Nba, Neutral) √ (Cavs, Negative) × (Draymond Green, Positive) ×</td></tr><tr><td>w/o Image 
+Prefix</td><td>(Justin Bieber, Positive) ×</td><td>(Kim Kardashian, Positive) √ (Kanye West, Neutral) × (Bahamas, Neutral) √</td><td>(Miss Colombia, Neutral) × (Paulina Vega, Positive) √</td><td>(Nba, Neutral) √ (Cavs, Negative) × (Draymond Green, Negative) √</td></tr><tr><td>UnifiedTMSC</td><td>(Justin Bieber, Negative) √</td><td>(Kim Kardashian, Positive) √ (Kanye West, Positive) √ (Bahamas, Neutral) √</td><td>(Miss Colombia, Positive) √ (Paulina Vega, Positive) √</td><td>(Nba, Neutral) √ (Cavs, Neutral) √ (Draymond Green, Negative) √</td></tr></table>
+
+Figure 4: Case study on four test samples. Red font indicates correctly predicted labels.  
+
+<table><tr><td rowspan="2"></td><td colspan="2">Twitter-2015</td><td colspan="2">Twitter-2017</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>UnifiedTMSC</td><td>79.8</td><td>76.3</td><td>75.4</td><td>74.7</td></tr><tr><td>w/o Paraphrased Prompts</td><td>75.7</td><td>72.0</td><td>68</td><td>65.9</td></tr><tr><td>P1 + w/o Image Prefix</td><td>78.9</td><td>75.2</td><td>74.1</td><td>73.7</td></tr><tr><td>P2 + w/o Image Prefix</td><td>78.6</td><td>75.7</td><td>74.4</td><td>73.8</td></tr><tr><td>P3 + w/o Image Prefix</td><td>78.5</td><td>75.7</td><td>74.3</td><td>73.9</td></tr><tr><td>AVG w/o Image Prefix</td><td>78.7</td><td>75.5</td><td>74.3</td><td>73.8</td></tr></table>
+
+Table 7: Ablation study of our UnifiedTMSC model.
+
+II. The position of [MASK] in the paraphrased prompts can also have an impact on the experimental results. In our case, the effect is best when the relative position of [MASK] is "Middle" rather than "Beginning" or "Ending". Therefore, when meeting a new task, the position of [MASK] may be a factor to affect task performance.
+
+# 4.4 Ablation Study
+
+To further investigate the effects of paraphrased prompts and image prefix, because entity-level MSC is more challenging than aspect-level MSC, we conduct ablation analysis on the multimodal entity-level datasets: Twitter-2015 and Twitter-2017. The results of the ablation experiments are shown in Table 7.
+
+Paraphrased Prompts.  $P_{i}$  is omitted from Eq. (2) and we just add the image prefix  $V$  in Eq. (3) for the experiment in order to examine the effects of the paraphrased prompts. The linear classification layer uses the fusion vector derived by multimodal Transformers as input to estimate the sentiment la
+
+bel of the target. The results are shown in Table 7. The absence of paraphrased prompts has resulted in a considerable performance decrease. On the Twitter-2015 dataset, the Accuracy and macro-F1 are dropped by approximately  $4.1\%$  and  $4.3\%$ , while on the Twitter-2017 dataset, the Accuracy declines by about  $7.4\%$  and the macro-F1 drops by about  $8.8\%$ . This demonstrates how using text paraphrased prompts can provide an entity with its task-related semantics.
+
+Image Prefix. We only use the paraphrased prompt  $P_{i}$  in Eq. (2) without applying the image prefix prompt  $V$  in Eq. (3) to study the importance of the image prefix. For text input,  $P_{1}, P_{2}$ , and  $P_{3}$  are proceeded for the ablation study. From Table 7, it can be seen that the performance has dropped after taking out the image prefix  $V$  (in Eq. (3)). The Accuracy decreases by  $1.1\%$ , and the macro-F1 drops by  $0.8\%$ , according to the average results for the Twitter-2015 dataset. On the Twitter-2017 dataset, the Accuracy and macro-F1 decline by roughly  $1.0\%$ . This illustrates the effectiveness of the image prefix. Moreover, the ablation study also shows that different paraphrased prompts have varied outcomes, demonstrating the language models' sensitivity to prompts.
+
+# 4.5 Case Study
+
+In our case study, the compared methods are only image prefix (denoted by w/o Paraphrased Prompt), only text prompt (denoted by w/o Image Prefix),
+
+and our UnifiedTMSC model with soft and hard prompts. We apply  $P_{1}$  for both w/o Image Prefix and UnifiedTMSC.
+
+As shown in Figure 4, for example (a), when there is no paraphrased prompt, the result obtained from text and image information is Neutral. When there is no image prefix, the image pre-trained embedding dominates the prediction results as Positive. Both of these two prediction results do not match the correct sentiment label Negative. For example (b), there are multiple targets that require sentiment classification. The sentiment labels predicted for place name Bahamas are Neutral, and adding appropriate prompts to both image and text can be predicted correctly. For examples (c) and (d), they are similar to example (b).
+
+These four samples further confirm the usefulness of our unified model. It can assign specific sentiment-related semantics to an entity via applying a paraphrased prompt. And prefix tuning of images can obtain better task-specific image embedding than image pre-trained embedding.
+
+# 5 Conclusion and Future Work
+
+There are currently two formats for target-oriented multimodal sentiment classification: entity-level and aspect-level. Our analysis shows that this barrier is superfluous. By incorporating paraphrased prompt and prefix vector into the multimodal input, the proposed model, i.e., UnifiedTMSC, unifies the two types of TMSC tasks. We conduct experiments on four datasets, and the results demonstrate the superiority and efficacy of our UnifiedTMSC.
+
+Our ongoing effort will primarily concentrate on two issues. One is to investigate how to design a paraphrasing rule to automatically generate paraphrased prompts without depending on human labor. The other is to investigate the generalizability of our model to see whether it can be used in other multimodal studies. In addition, we notice that the auto-regressive model XLNet can alleviate the problem of generating non-label words, and our future work will consider this.
+
+# Limitations
+
+Our model has three limitations. The first one is that the design of paraphrased prompts relies on human experience. Although our paraphrasing rule is designed based on the relative position and synonym substitution, manual experience is still required to obtain paraphrased prompts that comply
+
+with grammar rules, and paraphrased prompts that comply with grammar rules may not necessarily be the best. The second limitation is that more attempts are needed to conduct experiments on multilingual pre-trained models. Furthermore, the last is that we have not explored whether our model can be extended to other multimodal research fields, which will be our future research direction.
+
+# Acknowledgments
+
+This work is partially supported by NSFC, China (No.62276196).
+
+# References
+
+Damian Borth, Rongrong Ji, Tao Chen, Thomas M. Breuel, and Shih-Fu Chang. 2013. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In ACM Multimedia Conference, MM '13, Barcelona, Spain, October 21-25, 2013, pages 223-232. ACM.  
+Xiang Chen, Ningyu Zhang, Xin Xie, Shumin Deng, Yunzhi Yao, Chuanqi Tan, Fei Huang, Luo Si, and Huajun Chen. 2022. Knowprompt: Knowledge-aware prompt-tuning with synergistic optimization for relation extraction. In WWW '22: The ACM Web Conference 2022, Virtual Event, Lyon, France, April 25 - 29, 2022, pages 2778-2788. ACM.  
+Jian Cui, Lin Li, and Xiaohui Tao. 2023. Be-ornot prompt enhanced hard negatives generating for memes category detection. In IEEE International Conference on Multimedia and Expo, ICME 2023, Brisbane, Australia, July 10-14, 2023, pages 174-179. IEEE.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. pages 4171-4186. Association for Computational Linguistics.  
+Tianhao Gao, Jun Fang, Hanyu Liu, Zhiyuan Liu, Chao Liu, Pengzhang Liu, Yongjun Bao, and Weipeng Yan. 2022. LEGO-ABSA: A prompt-based task assembly unified generative framework for multi-task aspect-based sentiment analysis. In Proceedings of the 29th International Conference on Computational Linguistics, COLING 2022, Gyeongju, Republic of Korea, October 12-17, 2022, pages 7002-7012. International Committee on Computational Linguistics.  
+Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. 2022. PTR: prompt tuning with rules for text classification. AI Open, 3:182-192.  
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas,
+
+NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society.  
+Brendan Jou, Tao Chen, Nikolaos Pappas, Miriam Redi, Mercan Topkara, and Shih-Fu Chang. 2015. Visual affect around the world: A large-scale multilingual visual sentiment ontology. In Proceedings of the 23rd Annual ACM Conference on Multimedia Conference, MM '15, Brisbane, Australia, October 26 - 30, 2015, pages 159-168. ACM.  
+Xincheng Ju, Dong Zhang, Rong Xiao, Junhui Li, Shoushan Li, Min Zhang, and Guodong Zhou. 2021. Joint multi-modal aspect-sentiment analysis with auxiliary cross-modal relation detection. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 4395-4405. Association for Computational Linguistics.  
+Zaid Khan and Yun Fu. 2021. Exploiting BERT for multimodal target sentiment classification through input space translation. In MM '21: ACM Multimedia Conference, Virtual Event, China, October 20 - 24, 2021, pages 3034-3042. ACM.  
+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, Online Event, 16-20 November 2020, volume EMNLP 2020 of Findings of ACL, pages 1896-1907. Association for Computational Linguistics.  
+Chengxi Li, Feiyu Gao, Jiajun Bu, Lu Xu, Xiang Chen, Yu Gu, Zirui Shao, Qi Zheng, Ningyu Zhang, Yongpan Wang, and Zhi Yu. 2021. Sentiprompt: Sentiment knowledge enhanced prompt-tuning for aspect-based sentiment analysis. CoRR, abs/2109.08306.  
+Guowei Li, Fuqiang Lin, WangQun Chen, DiWen Dong, and Bo Liu. 2022a. Prompt-based learning for aspect-level sentiment classification. In Neural Information Processing - 29th International Conference, ICONIP 2022, Virtual Event, November 22-26, 2022, Proceedings, Part III, volume 13625 of Lecture Notes in Computer Science, pages 509-520. Springer.  
+Jing Li, Aixin Sun, Jianglei Han, and Chenliang Li. 2022b. A survey on deep learning for named entity recognition. IEEE Trans. Knowl. Data Eng., 34(1):50-70.  
+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, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 4582-4597. Association for Computational Linguistics.
+
+Yan Ling, Jianfei Yu, and Rui Xia. 2022. Vision-language pre-training for multimodal aspect-based sentiment analysis. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 2149-2159. Association for Computational Linguistics.  
+Bin Liu, Tao Lin, and Ming Li. 2023a. Enhancing aspect-category sentiment analysis via syntactic data augmentation and knowledge enhancement. Knowl. Based Syst., 264:110339.  
+Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023b. Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 55(9):195:1-195:35.  
+Yuhang Liu, Wei Wei, Daowan Peng, and Feida Zhu. 2022. Declaration-based prompt tuning for visual question answering. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, pages 3264-3270. ijcai.org.  
+Sijie Long, Lin Li, Jingling Yuan, and Jianquan Liu. 2022. Momnet: Gender prediction using mechanism of working memory. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022, pages 4259-4263. ACM.  
+Di Lu, Leonardo Neves, Vitor Carvalho, Ning Zhang, and Heng Ji. 2018. Visual attention model for name tagging in multimodal social media. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 1990-1999. Association for Computational Linguistics.  
+Yaojie Lu, Qing Liu, Dai Dai, Xinyan Xiao, Hongyu Lin, Xianpei Han, Le Sun, and Hua Wu. 2022. Unified structure generation for universal information extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 5755-5772. Association for Computational Linguistics.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR.  
+Ronald Seoh, Ian Birle, Mrinal Tak, Haw-Shiuan Chang, Brian Pinette, and Alfred Hough. 2021. Open aspect target sentiment classification with natural language
+
+prompts. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 6311-6322. Association for Computational Linguistics.  
+Zhengxin Song, Yun Xue, Donghong Gu, Haolan Zhang, and Weiping Ding. 2023. Target-oriented multimodal sentiment classification by using topic model and gating mechanism. Int. J. Mach. Learn. Cybern., 14(7):2289-2299.  
+Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. 2021. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 200-212.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.  
+Jiawei Wang, Zhe Liu, Victor S. Sheng, Yuqing Song, and Chenjian Qiu. 2021. Saliencybert: Recurrent attention network for target-oriented multimodal sentiment classification. In Pattern Recognition and Computer Vision - 4th Chinese Conference, PRCV 2021, Beijing, China, October 29 - November 1, 2021, Proceedings, Part III, volume 13021 of Lecture Notes in Computer Science, pages 3-15. Springer.  
+Luwei Xiao, Ejian Zhou, Xingjiao Wu, Shuwen Yang, Tianlong Ma, and Liang He. 2022. Adaptive multi-feature extraction graph convolutional networks for multimodal target sentiment analysis. In IEEE International Conference on Multimedia and Expo, ICME 2022, Taipei, Taiwan, July 18-22, 2022, pages 1-6. IEEE.  
+Nan Xu and Wenji Mao. 2017. Multisentinent: A deep semantic network for multimodal sentiment analysis. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, CIKM 2017, Singapore, November 06 - 10, 2017. ACM.  
+Nan Xu, Wenji Mao, and Guandan Chen. 2018. A comemory network for multimodal sentiment analysis. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR 2018, Ann Arbor, MI, USA, July 08-12, 2018, pages 929-932. ACM.  
+Nan Xu, Wenji Mao, and Guandan Chen. 2019. Multi-interactive memory network for aspect based multimodal sentiment analysis. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI
+
+Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 371-378. AAAI Press.  
+Hao Yang, Yanyan Zhao, and Bing Qin. 2022. Face-sensitive image-to-emotional-text cross-modal translation for multimodal aspect-based sentiment analysis. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 3324-3335. Association for Computational Linguistics.  
+Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. 2021. CPT: colorful prompt tuning for pre-trained vision-language models. CoRR, abs/2109.11797.  
+Jianfei Yu and Jing Jiang. 2019. Adapting BERT for target-oriented multimodal sentiment classification. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 5408-5414. ijcai.org.  
+Jianfei Yu, Jing Jiang, and Rui Xia. 2020. Entity-sensitive attention and fusion network for entity-level multimodal sentiment classification. IEEE ACM Trans. Audio Speech Lang. Process., 28:429-439.  
+Yuhai Yu, Hongfei Lin, Jiana Meng, and Zhehuan Zhao. 2016. Visual and textual sentiment analysis of a microblog using deep convolutional neural networks. Algorithms, 9(2):41.  
+Qi Zhang, Jinlan Fu, Xiaoyu Liu, and Xuanjing Huang. 2018. Adaptive co-attention network for named entity recognition in tweets. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 5674-5681. AAAI Press.  
+Zhe Zhang, Zhu Wang, Xiaona Li, Nannan Liu, Bin Guo, and Zhiwen Yu. 2021. Modalnet: an aspect-level sentiment classification model by exploring multimodal data with fusion discriminant attentional network. World Wide Web, 24(6):1957-1974.  
+Fei Zhao, Zhen Wu, Siyu Long, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2022. Learning from adjective-noun pairs: A knowledge-enhanced framework for target-oriented multimodal sentiment classification. In Proceedings of the 29th International Conference on Computational Linguistics, COLING 2022, Gyeongju, Republic of Korea, October 12-17, 2022, pages 6784-6794. International Committee on Computational Linguistics.  
+Jie Zhou, Jiabao Zhao, Jimmy Xiangji Huang, Qin-min Vivian Hu, and Liang He. 2021. MASAD: A large-scale dataset for multimodal aspect-based sentiment analysis. Neurocomputing, 455:47-58.
+
+Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130(9):2337-2348.
+
+# A Appendix
+
+# A.1 Baselines
+
+(1) TomBERT (IJCAI, Yu and Jiang (2019)), which employ BERT for the inter-modal interactions and Target-Image matching layer to obtain a target-sensitive visual.  
+(2) SaliencyBERT (PRCV, Wang et al. (2021)), which design a recurrent attention mechanism to capture the inter-modality dynamics.  
+(3) CapTrBERT (ACM Multimedia, Khan and Fu (2021)), a BERT-based model translating the image into caption and fusing the caption and input text-entity pair.  
+(4) JML-MASC (EMNLP, Ju et al. (2021)), which is a multi-task learning method with the cross-modal relation detection.  
+(5) VLP-MABSA (ACL, Ling et al. (2022)), a task-specific pre-training vision-language model.  
+(6) FITE-DE-Large (EMNLP, Yang et al. (2022)), introducing a FITE method to focus on capturing emotional cues through facial expressions.  
+(7) MIMN (AAAI, Xu et al. (2019)), using memory network to model multimodal data and learn the interactive influences in cross-modality and self-modality.  
+(8) MMAP (Neurocomputing, Zhou et al. (2021)), learning the interaction between text and image, text and aspect, and image and aspect through three interactive mechanisms.  
+(9) ModalNet (WWW, Zhang et al. (2021)), designing a discriminant matrix to supervise the fusion of inter-modal information.  
+(10) CLIP (INT J COMPUT VISION, Zhou et al. (2022)), a multimodal pre-training model, converts image classification tasks into image-text matching tasks using comparative learning.
+
+# A.2 Dataset Details
+
+(1) Twitter-2015 and Twitter-2017. The two Twitter datasets include user tweets released during 2014-2015 and 2016-2017, collected by Zhang et al. (2018) and Lu et al. (2018). Since the two publicly available multimodal datasets Twitter-2015 and Twitter-2017 only provide annotated targets in each tweet, Yu and Jiang (2019) ask three domain experts to annotate the sentiment towards each target, and take the majority label among the three annotators as the gold label. These two datasets contain multimodal tweets and annotated mentioned entities in the text, as well as the sentiment polarity of each entity, including positive, neutral, and negative. Each multimodal tweet consists of a text and a corresponding image. We follow their official splitting for training, validation and testing. The statistics of Twitter-2015 and Twitter-2017 are in Table 8.  
+(2) Multi-ZOL. This dataset is collected by Xu et al. (2019) from the Chinese website ZOL.com. The website consists of 40 large channels, including news, shopping malls, hardware, mobile phones, and more. They searched from pages 1 to 50 in the mobile phones channel. For each phone, only the reviews from the first 20 pages were crawled. The Multi-ZOL dataset contains 5288 multimodal reviews and each multimodal review contains a textual content, an image set, and at least one but no more than six aspects. The six aspects are cost performance, performance configuration, battery life, appearance, photography effect, and screen. For each aspect, the review has an integer sentiment score from 1 to 10, which is regarded as the sentiment label in our experiment. Actually, we convert digital sentiment labels into Chinese characters, such as '10' to 'ten'. Combining each aspect with the multimodal review, 28469 pairs of aspect review samples can be obtained. We randomly partition the dataset into training  $(80\%)$ , development  $(10\%)$ , and testing  $(10\%)$  sets. The statistical information after dividing the dataset is shown in Table 9.  
+(3) MASAD. This dataset was collected and published by Zhou et al. (2021) based on the publicly available Visual Sentiment Ontology (VSO) dataset (Borth et al., 2013) and Multi-
+
+<table><tr><td rowspan="2"></td><td colspan="4">Twitter-2015</td><td colspan="4">Twitter-2017</td></tr><tr><td>#Positive</td><td>#Neutral</td><td>#Negative</td><td>Total</td><td>#Positive</td><td>#Neutral</td><td>#Negative</td><td>Total</td></tr><tr><td>Train</td><td>928</td><td>1883</td><td>368</td><td>3179</td><td>1508</td><td>416</td><td>1638</td><td>3562</td></tr><tr><td>Dev</td><td>303</td><td>679</td><td>149</td><td>1122</td><td>515</td><td>144</td><td>517</td><td>1176</td></tr><tr><td>Test</td><td>317</td><td>607</td><td>113</td><td>1037</td><td>493</td><td>168</td><td>573</td><td>1234</td></tr></table>
+
+Table 8: The statistics of Twitter-2015 and Twitter-2017 datasets.  
+
+<table><tr><td></td><td>#1</td><td>#2</td><td>#3</td><td>#4</td><td>#5</td><td>#6</td><td>#7</td><td>#8</td><td>#9</td><td>#10</td><td>Total</td><td>Max #Image</td><td>Min #Image</td></tr><tr><td>Train</td><td>6</td><td>1163</td><td>6</td><td>1190</td><td>3</td><td>3509</td><td>0</td><td>6952</td><td>0</td><td>9948</td><td>22775</td><td>111</td><td>1</td></tr><tr><td>Dev</td><td>0</td><td>139</td><td>0</td><td>150</td><td>0</td><td>433</td><td>0</td><td>872</td><td>0</td><td>1253</td><td>2847</td><td>111</td><td>1</td></tr><tr><td>Test</td><td>0</td><td>132</td><td>0</td><td>151</td><td>0</td><td>413</td><td>0</td><td>871</td><td>0</td><td>1280</td><td>2847</td><td>72</td><td>1</td></tr></table>
+
+Table 9: The statistics of Multi-ZOL dataset after partitioning. 'Max #Image' refers to the maximum number of image corresponding to the review, while 'Min #Image' is the minimum quantity.  
+
+<table><tr><td rowspan="2"></td><td colspan="3">Train</td><td colspan="3">Dev</td><td colspan="3">Test</td></tr><tr><td>#Positive</td><td>#Negative</td><td>Total</td><td>#Positive</td><td>#Negative</td><td>Total</td><td>#Positive</td><td>#Negative</td><td>Total</td></tr><tr><td>Food</td><td>2122</td><td>392</td><td>2514</td><td>239</td><td>41</td><td>280</td><td>592</td><td>109</td><td>701</td></tr><tr><td>Buildings</td><td>969</td><td>874</td><td>1843</td><td>109</td><td>96</td><td>205</td><td>273</td><td>245</td><td>518</td></tr><tr><td>Goods</td><td>2398</td><td>1512</td><td>3910</td><td>273</td><td>162</td><td>435</td><td>743</td><td>512</td><td>1255</td></tr><tr><td>Animal</td><td>2728</td><td>1979</td><td>4707</td><td>295</td><td>229</td><td>524</td><td>1126</td><td>670</td><td>1796</td></tr><tr><td>Human</td><td>1805</td><td>1648</td><td>3453</td><td>194</td><td>190</td><td>384</td><td>503</td><td>464</td><td>967</td></tr><tr><td>Plant</td><td>2542</td><td>2341</td><td>4883</td><td>277</td><td>266</td><td>543</td><td>1269</td><td>947</td><td>2216</td></tr><tr><td>Scenery</td><td>3222</td><td>1760</td><td>4982</td><td>378</td><td>176</td><td>554</td><td>907</td><td>490</td><td>1397</td></tr><tr><td>Total</td><td>15786</td><td>10506</td><td>26292</td><td>1765</td><td>1160</td><td>2925</td><td>5413</td><td>3437</td><td>8850</td></tr></table>
+
+Table 10: The statistical information of MASAD datasets. The seven domains are Food, Buildings, Goods, Animal, Human, Plant, Scenery.
+
+lingual Visual Sentiment Ontology (MVSO) dataset (Jou et al., 2015). Zhou et al. (2021) selected the samples from a partial VSO dataset (approximately 120k samples) that can express significant sentiments (about 38k samples) and categorized them into 7 domains, resulting in the MASAD dataset. The seven domains of the MASAD are food, buildings, goods, animal, human, plant, and scenery. Each domain encompasses multiple aspects, such as the animal domain, including cat, dog, horse, and so on. According to our statistics, there are a total of 57 predefined aspects. This dataset only includes training and testing sets, both containing positive and negative samples. We partition each domain in the training set into a new training set and a validation set in a 9:1 ratio, keeping the original testing set unchanged. The statistics of MASAD are in Table 10.
+
+In addition, there are instances of duplicate data in the testing and training sets in this dataset. In the domain of animal and plant, according to our statistics, about  $81.6\%$  and  $62.2\%$  of the data in the testing set has appeared in the training set. In order to compare more fairly with other baseline models, we conduct experiments in other domains besides ani
+
+mal and plant.

2023/Descriptive Prompt Paraphrasing for Target-Oriented Multimodal Sentiment Classification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9301dfa9b411ef635f27cfe1a08502457d9ad30d3870d82b84b5a139310f474f
+size 683394

2023/DetectLLM_ Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text/b6e08278-e610-46f1-92c0-6af873c348c2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb56611d42af4bcdadc5b51fcad9b69a75cc0626b80ccede56c370a0e9d3d1d8
+size 587869

2023/DetectLLM_ Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text/full.md

@@ -0,0 +1,381 @@
+# DetectLLM: Leveraging Log-Rank Information for Zero-Shot Detection of Machine-Generated Text
+
+Jinyan Su $^{1,2}$ , Terry Yue Zhuo $^{3}$ , Di Wang $^{4}$ , Preslav Nakov $^{1}$
+
+<sup>1</sup>Mohamed bin Zayed University of Artificial Intelligence, <sup>2</sup>Cornell University
+
+<sup>3</sup>Monash University and CSIRO's Data61
+
+$^{4}$ King Abdullah University of Science and Technology
+
+{Jinyan.Su, preslav.nakov}@mbzuai.ac.ae
+
+terry.zhuo@monash.edu,di.wang@kaust.edu.sa
+
+# Abstract
+
+With the rapid progress of Large language models (LLMs) and the huge amount of text they generate, it becomes impractical to manually distinguish whether a text is machine-generated. The growing use of LLMs in social media and education, prompts us to develop methods to detect machine-generated text, preventing malicious use such as plagiarism, misinformation, and propaganda. In this paper, we introduce two novel zero-shot methods for detecting machine-generated text by leveraging the Log-Rank information. One is called DetectLLM-LRR, which is fast and efficient, and the other is called DetectLLM-NPR, which is more accurate, but slower due to the need for perturbations. Our experiments on three datasets and seven language models show that our proposed methods improve over the state of the art by 3.9 and 1.75 AUROC points absolute. Moreover, DetectLLM-NPR needs fewer perturbations than previous work to achieve the same level of performance, which makes it more practical for real-world use. We also investigate the efficiency-performance trade-off based on users' preference for these two measures and provide intuition for using them in practice effectively. We release the data and the code of both methods in https://github.com/mbzuai-nlp/DetectLLM.
+
+# 1 Introduction
+
+Large language models (LLMs) have made rapid advancements in recent years, and are now able to generate text with significantly improved diversity, fluency, and quality. Models such as ChatGPT (OpenAI, 2022), GPT-3 (Brown et al., 2020), LLaMa (Touvron et al., 2023) and BLOOM (Scao et al., 2022) demonstrate exceptional performance in answering questions (Robinson et al., 2022), writing stories (Fan et al., 2018; Yuan et al., 2022) and thus facilitating daily life and improving work efficiency. However, LLMs can also be misused for generating plagiarized text, misinformation,
+
+and propaganda, which can lead to negative consequences (Zhuo et al., 2023). For instance, students might use LLMs to write assignments (Rosenblatt, 2023), making fair evaluation difficult for teachers, and in the long run, undermining the integrity of the entire education system. Malicious actors might generate fake news articles to spread misinformation and propaganda or to manipulate public opinion, which is dangerous, especially when it comes to politics (Floridi and Chiriatti, 2020; Stokel-Walker, 2022).
+
+With the proliferation of LLMs and the increasing amount of texts they produce, it is challenging for humans to accurately identify machine-generated texts (Gehrmann et al., 2019). Moreover, it is unrealistic to hire humans to manually identify machine-generated text at scale due to the prohibitively high costs and the efficiency requirements in real-time applications, e.g., in social media. Thus, it is essential to develop tools and strategies to automatically identify machine-generated text and to mitigate the potential negative impact of LLMs.
+
+The problem of distinguishing machine-generated from human-written text is commonly formulated as a binary task (Jawahar et al., 2020). Most previous work has focused on the black-box scenario, where the detector has access to the output of the LLMs only and cannot make use of its internals. Such methods lack flexibility since they need to be retrained from scratch to be able to recognize the output of a new LLM (Mitchell et al., 2023). Given the speed at which new LLMs are developed, black-box methods are becoming more and more expensive and impractical. In cases when the access to the LLM is via an API only, one possibility is for the LLM owner to record all content it has generated, or to watermark all texts it has generated (Kirchenbauer et al., 2023; Zhao et al., 2023). However, such solutions are not feasible for third parties.
+
+We therefore consider a white-box setting, where the detector has full access to the LLMs. We focus on zero-shot methods, where we use the LLM without additional training. Generally speaking, zero-shot methods use the source LLM to extract statistics such as the average per-token log probability or the average rank of each token in the ranked list of possible choices to make a prediction by comparing it to a threshold (Solaiman et al., 2019; Ippolito et al., 2019; Gehrmann et al., 2019; Mitchell et al., 2023). Based on whether the queried statistics are only about the target texts, we can roughly categorize them as perturbation-free and perturbation-based. Perturbation-free methods only query LLMs about the statistics on the target text  $x$ , while perturbation-based methods such as Mitchell et al. (2023) queries also the statistics of additional perturbed texts, which achieves better performance but is 50-100 times more costly than perturbation-free methods. Thus, there exists a trade-off between performance and efficiency among zero-shot methods.
+
+To mitigate the gap of these two categories and design zero-shot methods with better performance-efficiency balance, we should either improve the accuracy of perturbation-free methods or reduce their cost. Thus, we propose two novel zero-shot methods, one perturbation-free, but more accurate than previous methods, and one perturbation-based method, but with better efficiency.
+
+In the perturbation-free method we proposed, we apply Log-Likelihood Log-Rank ratio (LRR), which enhances Log-Likelihood information with Log-Rank information as the discerning feature that achieves better performance than existing perturbation-free methods and can even surpass perturbation-based methods on some LLMs. For the perturbation-based method, we use Normalized perturbed log rank (NPR), which is based on the intuition that machine-generated texts are more sensitive to minor rewrites (or small perturbations). Compared to existing perturbation-based methods such as DetectGPT (Mitchell et al., 2023), this approach takes less time and computational resources and is more efficient. We denote these two methods DetectLLM-LRR and DetectLLM-NPR respectively. Our contributions are as follows:
+
+- We propose two novel zero-shot approaches based on Log-Rank statistics, which improve over the state of the art. On average, these methods improved upon the previous best
+
+![](images/1749326a3275bb686abf4bb824ab6f9c771df4a9f30750d7ba4a3cbb95136d25.jpg)  
+Figure 1: Distribution of LRR and NPR visualized on 300 human-written texts (HG) from the WritingPrompts dataset (Fan et al., 2018) as well as 300 texts generated with GPT-2-x1 (MG) by prompting it with the first 30 tokens from human-written texts.
+
+zero-shot methods by 3.9 and 1.75 AUROC points absolute.
+
+- We investigate the efficacy of existing zero-shot methods and explore their limits as the size of the LLMs increases from 1.5 to 20 billion.  
+- We conduct comprehensive experiments to better understand the efficiency-performance trade-offs in zero-shot methods, thereby providing interesting insights on how to choose among different categories of zero-shot methods based on users' preference for performance or efficiency.
+
+# 2 Related Work
+
+The detection of machine-generated text is commonly formulated as a classification task (Jawahar et al., 2020; Fagni et al., 2021; Bakhtin et al., 2019; Sadasivan et al., 2023; Wang et al., 2023). One way of solving it is to use supervised learning, where a classification model is trained on a dataset containing both machine-generated and human-written texts. For example, GPT-2 Detector (Solaiman et al., 2019) fine-tunes RoBERTa (Liu et al., 2019) on the output of GPT-2, while the ChatGPT Detector (Guo et al., 2023) fine-tunes RoBERTa on the HC3 (Guo et al., 2023) dataset. However, models trained explicitly to detect machine-generated texts may overfit their training distribution of the domains (Bakhtin et al., 2019; Uchendu et al., 2020).
+
+Another stream of work attempts to distinguish machine-generated from human-written texts based on statistical irregularities in the entropy (Lavergne et al., 2008), perplexity (Beresneva, 2016) or in the  $n$ -gram frequencies (Badaskar et al., 2008). Gehrmann et al. (2019) introduced hand-crafted statistical features to assist humans in detecting-machine generated texts. Moreover, (Solaiman et al., 2019) noted the efficacy of simple zero-shot
+
+![](images/6a551d32cbb8ddf2f934a9b8dbbd13ae3943b846c784b1283bb4c061bdf5352c.jpg)  
+Figure 2: Comparison of DetectGPT to NPR averaged across six models (in terms of AUROC). (The full results are given in Figure 6 in the Appendix).
+
+methods for detecting machine-generated text by evaluating the per-token log probability of texts and using thresholding. Mitchell et al. (2023) observed that machine-generated texts tend to lie in the local curvature of the log probability and proposed DetectGPT, whose prominent performance can only be guaranteed by the large size of the perturbation function and by a large number of perturbations, and thus costs more computational resources.
+
+Other work explored watermarking, which imprints specific patterns of the LLM-output text that can be detected by an algorithm while being imperceptible to humans. Grinbaum and Adomaitis (2022) and Abdelnabi and Fritz (2021) watermarked machine-generated text using syntax tree manipulation, while Kirchenbauer et al. (2023) required access to the LLM's logits at each time step.
+
+# 3 Improved Zero-Shot Approaches by Leveraging Log-Rank Information
+
+In this section, we introduce the Log-Likelihood Log-Rank Ratio (LRR) and the Normalized Perturbed log-Rank (NPR). LRR combines Log-Rank and Log-Likelihood as they provide complementary information about the text. NPR uses the idea that the Log-Rank of machine-generated texts should be more sensitive to smaller perturbations.
+
+# 3.1 Log-Likelihood Log-Rank Ratio (LRR)
+
+We define the Log-Likelihood Log-Rank Ratio as
+
+$$
+\begin{array}{l} \mathrm {L R R} = \left| \begin{array}{c} \frac {1}{t} \sum_ {i = 1} ^ {t} \log p _ {\theta} (x _ {i} | x _ {<   i}) \\ \hline \frac {1}{t} \sum_ {i = 1} ^ {t} \log r _ {\theta} (x _ {i} | x _ {<   i}) \end{array} \right| \\ = - \frac {\sum_ {i = 1} ^ {t} \log p _ {\theta} (x _ {i} | x _ {<   i})}{\sum_ {i = 1} ^ {t} \log r _ {\theta} (x _ {i} | x _ {<   i})}, \\ \end{array}
+$$
+
+where  $r_{\theta}(x_i|x_{< i}) \geq 1$  is the rank of token  $x_{i}$  conditioned on the previous tokens.
+
+The Log-Likelihood in the numerator represents the absolute confidence for the correct token, while the Log-Rank in the denominator accounts for
+
+the relative confidence, which reveals complementary information about the texts. As illustrated in Figure 1, LRR is generally larger for machine-generated text, which can be used for distinguishing machine-generated from human-written text. One plausible reason might be that for machine-generated text, the Log-Rank is more discernible than the Log-Likelihood, so LRR illustrates this pattern for machine-generated text. In Sections 4 and 6, we experimentally show that LRR is a better discriminator than either the Log-Likelihood or the Log-Rank. We call the zero-shot method using LRR as a detection feature as DetectLLM-LRR, and use the abbreviation LRR in the rest of the paper.
+
+# 3.2 Normalized Log-Rank Perturbation (NPR)
+
+We define the normalized perturbed Log-Rank as
+
+$$
+\mathrm {N P R} = \frac {\frac {1}{n} \sum_ {p = 1} ^ {n} \log r _ {\theta} (\tilde {x} _ {p})}{\log r _ {\theta} (x)},
+$$
+
+where small perturbations are applied on the target text  $x$  to produce the perturbed text  $\tilde{x}_p$ . Here, a perturbation means minor rewrites, such as replacing some of the words. We call the zero-shot method using NPR as a detection feature DetectLLM-NPR, and use the abbreviation NPR in the rest of the paper.
+
+The motivation for NPR is that machine-generated and human-written texts are both negatively affected by small perturbations, i.e., the Log-Rank score will increase after perturbations, but the machine-generated text is more susceptible to perturbations and thus increase more on Log-Rank score after perturbation, which suggests higher NPR score for machine-generated texts. As shown in Figure 1, NPR can be a discernible signal for distinguishing machine-generated from human-written text. DetectGPT (Mitchell et al., 2023) uses a similar idea, but experimentally, we find NPR to be more efficient and to perform better. Details and comparisons are given in Section 4.
+
+# 4 Experimental Setup
+
+In this section, we conduct comprehensive experiments to evaluate the performance of LRR and NPR in comparison to several methods previously proposed in the literature. We experiment with LLM sizes varying from 1.5B to 20B parameters, probing the boundary of zero-shot methods when
+
+LLMs continue to grow in size. We further study the impact of the perturbation function, the number of perturbations (especially for NPR and Detect-GPT), the decoding strategy, and the temperature.
+
+Data Following (Mitchell et al., 2023), we use three datasets: XSum (Narayan et al., 2018), SQuAD (Rajpurkar et al., 2016), WritingPrompts (Fan et al., 2018), containing news articles, Wikipedia paragraphs and prompted stories, respectively, as human-written texts and we produce machine-generated texts using LLMs. These datasets are chosen to represent the areas where LLMs could have a negative impact. For each experiment, we evaluate 300 machine-generated and human-written texts pairs by prompting the LLMs with the first 30 tokens of the human-written text. We release the code for this.
+
+Evaluation Measure Following previous work (Mitchell et al., 2023; He et al., 2023; Krishna et al., 2023), we use the area under the receiver operating characteristic curve (AUROC), which is the probability that a classifier correctly ranks the machine-generated example higher than human-written example. Since for zero-shot methods, detection rates are heavily dependent on the threshold when using discriminative statistics, AUROC is commonly used to measure zero-shot detector performance, which considers the range of all possible thresholds (Krishna et al., 2023).
+
+# 4.1 Methods
+
+Zero-Shot Methods We compare the following:
+
+-  $\log p(x)$ : the idea is that a passage with a high average log probability is more likely to have been generated by the target LLM;  
+- Rank: the idea is that a passage with a higher average rank is more likely to have been generated by the target LLM;  
+- Log-Rank: passage with a higher average observed Log-Rank is more likely to have been generated by the target LLM;  
+- Entropy: machine-generated text has higher entropy;  
+- DetectGPT: machine-generated text has more negative log probability curvature.
+
+More detail and exact definitions of these methods can be found in Appendix A.
+
+These zero-shot baselines, along with our newly proposed LRR and NPR, can be categorized as
+
+- Perturbation-free:  $\log p(x)$ , Rank, Log-Rank, Entropy, LRR. They only query the
+
+LLM for statistics about the target text  $x$ .
+
+- Perturbation Based: DetectGPT and NPR. These methods query the LLM not only for the target text  $x$ , but also for perturbed versions thereof  $\tilde{x}_1, \dots, \tilde{x}_p$ .
+
+As perturbation-based methods perform better (but are also more time-consuming), for fair comparison, we compare them within their own group.
+
+Supervised Methods We also experiment with two supervised detectors: RoBERTa-base and RoBERTa-large. As these are not central to our narrative, we put the results in Appendix B.
+
+Experimental Details For the perturbation-based methods (DetectGPT and NPR), we use T5-3B for perturbation and we perturb the input text 50 times for all the experiments, unless specified otherwise. For all zero-shot methods, we use sampling with a temperature of 1, unless specified otherwise. More detail are given in Appendix A.
+
+# 5 Evaluation Results
+
+# 5.1 Zero-Shot Results
+
+Table 1 shows a comparison of the five baseline zero-shot approaches to our proposed LRR and NPR, grouped as perturbation-based and perturbation-free. We can see that for the perturbation-based methods, NPR consistently outperforms DetectGPT on all datasets and LLMs, except for one case, with an average improvement of 0.90, 2.03, 2.32 AUROC points absolute on XSum, SQuAD, and WritingPrompts, respectively, (using the same perturbation function and the same number of perturbations). For the experiments among perturbation-free methods, on average, our method achieves the best performance and improves by 2.15, 8.27, 1.28 AUROC points absolute over the second-best perturbation-free method (i.e., Log-Rank) on XSum, SQuAD, and WritingPrompts, respectively. Moreover, we find that in some cases, LRR can even perform better than perturbation-based methods, e.g., on SQuAD, LRR outperforms DetectGPT by 4.23 AUROC point absolute and outperforms NPR by 2.20 AUROC points.
+
+# 5.2 Comparing DetectGPT to NPR
+
+Equipped with large perturbation functions and an adequate amount of perturbations, perturbation-based methods generally outperform perturbation-free ones, e.g., using T5-3b as the perturbation function and perturb 50 times as in Table 1. However, in
+
+<table><tr><td>Dataset</td><td>Perturbation</td><td>Method</td><td>GPT-2-x1</td><td>Neo-2.7</td><td>OPT-2.7</td><td>GPT-j</td><td>OPT-13</td><td>Llama-13</td><td>NeoX</td><td>Avg.</td></tr><tr><td rowspan="7">XSum</td><td rowspan="5">w/o</td><td>log p</td><td>89.16</td><td>87.69</td><td>86.98</td><td>83.10</td><td>83.90</td><td>56.89</td><td>78.16</td><td>80.84</td></tr><tr><td>Rank</td><td>79.79</td><td>77.87</td><td>76.07</td><td>76.28</td><td>74.10</td><td>48.81</td><td>72.44</td><td>72.19</td></tr><tr><td>Log-Rank</td><td>91.75</td><td>90.79</td><td>89.18</td><td>86.42</td><td>85.88</td><td>61.33</td><td>81.44</td><td>83.83</td></tr><tr><td>Entropy</td><td>56.78</td><td>55.14</td><td>50.34</td><td>55.51</td><td>50.98</td><td>69.43</td><td>60.84</td><td>57.00</td></tr><tr><td>LRR (ours)</td><td>93.47</td><td>92.24</td><td>88.70</td><td>88.68</td><td>83.79</td><td>71.07</td><td>83.89</td><td>85.98</td></tr><tr><td rowspan="2">w/</td><td>DetectGPT</td><td>98.80</td><td>99.11</td><td>96.02</td><td>95.88</td><td>92.65</td><td>73.55</td><td>93.58</td><td>92.80</td></tr><tr><td>NPR (ours)</td><td>99.40</td><td>99.46</td><td>97.09</td><td>95.76</td><td>94.63</td><td>75.51</td><td>94.08</td><td>93.70</td></tr><tr><td rowspan="7">SQuAD</td><td rowspan="5">w/o</td><td>log p</td><td>90.72</td><td>84.18</td><td>87.84</td><td>78.20</td><td>80.65</td><td>42.91</td><td>68.78</td><td>76.18</td></tr><tr><td>Rank</td><td>83.46</td><td>79.77</td><td>81.85</td><td>79.46</td><td>77.47</td><td>54.44</td><td>73.10</td><td>75.65</td></tr><tr><td>Log-Rank</td><td>94.33</td><td>89.52</td><td>91.76</td><td>83.37</td><td>85.05</td><td>48.28</td><td>73.88</td><td>80.88</td></tr><tr><td>Entropy</td><td>57.97</td><td>58.48</td><td>53.29</td><td>58.26</td><td>57.14</td><td>69.71</td><td>59.97</td><td>59.26</td></tr><tr><td>LRR (ours)</td><td>97.42</td><td>95.74</td><td>95.89</td><td>91.59</td><td>91.36</td><td>68.78</td><td>83.31</td><td>89.15</td></tr><tr><td rowspan="2">w/</td><td>DetectGPT</td><td>98.52</td><td>95.86</td><td>96.91</td><td>88.66</td><td>90.60</td><td>47.03</td><td>76.84</td><td>84.92</td></tr><tr><td>NPR (ours)</td><td>99.40</td><td>97.56</td><td>98.39</td><td>91.88</td><td>93.04</td><td>48.67</td><td>79.73</td><td>86.95</td></tr><tr><td rowspan="7">WritingP</td><td rowspan="5">w/o</td><td>log p</td><td>96.71</td><td>95.63</td><td>95.05</td><td>94.43</td><td>92.53</td><td>83.54</td><td>93.27</td><td>93.02</td></tr><tr><td>Rank</td><td>87.62</td><td>82.79</td><td>83.89</td><td>83.21</td><td>83.52</td><td>77.64</td><td>81.64</td><td>82.90</td></tr><tr><td>Log-Rank</td><td>98.02</td><td>97.15</td><td>96.32</td><td>96.06</td><td>94.34</td><td>88.11</td><td>95.14</td><td>95.02</td></tr><tr><td>Entropy</td><td>36.45</td><td>34.07</td><td>39.75</td><td>36.93</td><td>42.49</td><td>47.64</td><td>37.89</td><td>39.32</td></tr><tr><td>LRR (ours)</td><td>98.34</td><td>98.02</td><td>96.45</td><td>96.97</td><td>95.09</td><td>92.66</td><td>96.56</td><td>96.30</td></tr><tr><td rowspan="2">w/</td><td>DetectGPT</td><td>99.30</td><td>98.71</td><td>98.33</td><td>95.52</td><td>96.46</td><td>83.01</td><td>92.94</td><td>94.90</td></tr><tr><td>NPR (ours)</td><td>99.78</td><td>99.59</td><td>98.87</td><td>98.07</td><td>98.14</td><td>89.39</td><td>96.72</td><td>97.22</td></tr></table>
+
+Table 1: Zero-shot experiments. Comparison of the proposed LRR and NPR to other zero-shot methods in terms of AUROC. For fair comparison, we show in bold the best results, both with and without perturbations.  
+
+<table><tr><td colspan="2"></td><td colspan="5">w/o Perturbation</td><td colspan="2">w/ Perturbation</td></tr><tr><td>Decoding</td><td>Dataset</td><td>log p</td><td>Rank</td><td>Log-Rank</td><td>Entropy</td><td>LRR (ours)</td><td>DetectGPT</td><td>NPR (ours)</td></tr><tr><td rowspan="3">top-k</td><td>XSum</td><td>81.64</td><td>70.68</td><td>85.19</td><td>55.47</td><td>89.25</td><td>91.34</td><td>92.93</td></tr><tr><td>SQuAD</td><td>76.31</td><td>74.31</td><td>81.28</td><td>57.96</td><td>90.61</td><td>82.42</td><td>84.99</td></tr><tr><td>WritingP</td><td>93.80</td><td>82.15</td><td>95.72</td><td>37.26</td><td>97.10</td><td>93.89</td><td>96.33</td></tr><tr><td rowspan="3">top-p</td><td>XSum</td><td>86.94</td><td>70.86</td><td>88.65</td><td>53.89</td><td>88.29</td><td>92.74</td><td>93.42</td></tr><tr><td>SQuAD</td><td>82.07</td><td>75.03</td><td>85.49</td><td>55.86</td><td>91.09</td><td>83.98</td><td>86.19</td></tr><tr><td>WritingP</td><td>96.51</td><td>82.48</td><td>97.44</td><td>33.92</td><td>97.25</td><td>94.20</td><td>96.55</td></tr></table>
+
+Table 2: Decoding strategy analysis. Shown are the AUROC scores for methods with top-  $k$  (  $k = {40}$  ) and top-  $p$ $\left( {p = {0.96}}\right)$  sampling averaged across four LLMs: Neo-2.7, OPT-2.7, GPT-j, Llama-13.
+
+practice, due to time and resource constraints, not all users can afford these models and large amounts of perturbations. Thus, it is important to investigate how NPR and DetectGPT behave with smaller perturbation function size and fewer perturbations.
+
+Different Number of Perturbations. Figure 2 shows the averaged performance of DetectGPT and NPR with varying number of perturbations. We can see that NPR consistently performs better than DetectGPT when using the same number of perturbations. In other words, NPR can achieve a comparable or better performance but with significantly fewer perturbations. For example, in SQuAD and WritingPrompts dataset, NPR achieves 85 points and 95 points using approximately 10 perturbations while DetectGPT requires around 100 perturbations, which highlights the effectiveness and efficiency of NPR. More complete results for each dataset and model can be found in Figure 6 and Figure 7 of Appendix C.
+
+Different Perturbation Functions. In Table 3, we compare NPR to DetectGPT using a smaller perturbation model T5-large, and the result is averaged over 6 LLMs and 3 datasets. We found
+
+that replacing T5-3b with smaller models harms the performance of both NPR and DetectGPT, and the performance degradation can't be mitigated by increasing the number of perturbations. For both NPR and DetectGPT, the average performance of 100 perturbations with T5-large is still worse than 10 perturbations with T5-3b (emphasized with the grey box in Table 3). Moreover, one can observe that, NPR is less affected by the reduced perturbation function size: when replacing T5-3b to T5-large, the performance degradation averaged over 10, 20, 50, 100 perturbations for NPR is 4.40 points, much smaller compared to that of 8.06 points for DetectGPT. The complete results on 6 LLMs and 3 datasets can be found in Figure 8 of Appendix C.
+
+# 5.3 Different Decoding Strategy and Temperature
+
+Alternative Decoding Strategies. In line with prior work (Pagnoni et al., 2022), we experimented with top- $k$  sampling (Fan et al., 2018) and top- $p$  sampling (Holtzman et al., 2019). Top- $k$  sampling generates from top- $k$  most likely words according to the LLM. Top- $p$  sampling (nucleus sampling) samples from the set of words that collectively ac
+
+<table><tr><td rowspan="2">Perturbation Function</td><td rowspan="2">Dataset</td><td colspan="4"># (Perturbations)</td></tr><tr><td>10</td><td>20</td><td>50</td><td>100</td></tr><tr><td rowspan="3">NPR (ours)</td><td>T5-large</td><td>86.69</td><td>88.00</td><td>88.74</td><td>88.94</td></tr><tr><td>T5-3b</td><td>91.39</td><td>92.35</td><td>93.04</td><td>93.20</td></tr><tr><td>Diff</td><td>4.70</td><td>4.35</td><td>4.30</td><td>4.26</td></tr><tr><td rowspan="3">DetectGPT</td><td>T5-large</td><td>77.94</td><td>81.12</td><td>83.90</td><td>84.54</td></tr><tr><td>T5-3b</td><td>86.70</td><td>89.57</td><td>91.38</td><td>92.10</td></tr><tr><td>Diff</td><td>8.76</td><td>8.45</td><td>7.48</td><td>7.56</td></tr></table>
+
+Table 3: Perturbation analysis. Comparing DetectGPT to NPR using different perturbations (AUROC scores).
+
+count for a total mass probability  $p$ . The results (averaged across 4 LLMs) are shown in Table 2, and complete results can be found in Table 9 of Appendix D. We find that, although almost all the zero-shot methods perform better when using top- $k$  and top- $p$  sampling than temperature sampling, Log-Rank and Log-Likelihood methods are more in favour of top- $p$  sampling, while LRR is stable in both top- $p$  and top- $k$  sampling. For top- $k$  decoding, LRR improves 4.06, 9.33, and 1.38 points over the second-best zero-shot method baseline on three datasets, respectively. LRR performance also improves when using top- $p$  decoding strategy, but due to the unstable performance surge of the Log-Rank method, LRR become slightly behind the Log-Rank method, with a minor difference of 0.36 and 0.19 points on the XSum and WritingPrompts datasets, respectively. For perturbation-based methods, their behaviour is consistent with previous results, where NPR outperforms DetectGPT for both top- $p$  and top- $k$  sampling strategies.
+
+Different Temperature. Temperature controls the degree of randomness of the generation process. Increasing the temperature leads to more randomness and creativity while reducing it leads to more conservation and less novelty. In practice, people adjust the temperature for their specific purposes. For example, students might set a high temperature to encourage more original and diverse output when writing a creative essay, whereas fake news producers might set lower temperatures to generate seemingly convincing news articles for their deceptive purposes. Based on our experiments in Table 4, we found that Log-Likelihood  $(\log p)$ , Log-Rank and LRR is highly sensitive to the temperature and can get even better results than perturbation-based methods when the temperature is relatively low. In addition, the performance improvement of the Rank method with the increased temperature is negligible compared to Log-Likelihood, Log-Rank and LRR, while the performance of the entropy method seems to be positively correlated to the temperature. We conjure that the abnormal behaviour of
+
+the Entropy method might be because of the assumption that "machine-generated text has higher entropy" (Mitchell et al., 2023), which, from our experiments, doesn't stand for high temperature. As for the perturbation-based method, the impact of temperature is not as clear as a perturbation-free method. But in general, the results suggest the temperature has only minor effects on DetectGPT while it improves the performance of NPR. Another observation is that the perturbation-free method performs better than the perturbation-based method in low temperatures, for example, if the temperature is smaller than 0.95, perturbation-based methods get better detection accuracy while being efficient.
+
+# 6 Analysis of the Efficiency
+
+Though in Table 1, perturbation-based methods appear to be significantly better than perturbation-free methods, it is important to note that their superior performances can only be achieved with large perturbation functions and multiple number of perturbations, which leads to intensive demand for computational resources and longer computational time. Thus, while performance is an important factor, it is crucial to consider the efficiency of these zero-shot methods as well.
+
+# 6.1 Computational Cost Analysis
+
+To get an idea of how costly different zero-shot methods are to achieve their performance in Table 1, we estimated the computational time (per sample) for each zero-shot method in Table 5. The time is estimated over the average of 10 samples. For perturbation-based methods, since the time depends on the perturbation function and the number of perturbations, we used T5-3b as the perturbation function and use 50 perturbations since this is the setting used for the main results in Table 1, we want to provide an idea of how much more it costs for perturbation based method to achieve exceptional performance in Table 1. We observed that the computational time of Log-Likelihood, Rank, Log-Rank and Entropy are almost the same, while LRR runs approximately 2 times longer than these methods since it requests both the Log-Rank and Log-Likelihood statistics. For perturbation-based methods, the running time is at least 50 times longer compared to Log-Likelihood, Rank, Log-Rank, and Entropy method, since they calculate the Log-Likelihood or Log-Rank for not only the target text but also perturbed samples.
+
+<table><tr><td rowspan="2">Temperature</td><td colspan="5">w/o Perturbation</td><td colspan="2">w/ Perturbation</td></tr><tr><td>log p</td><td>Rank</td><td>Log-Rank</td><td>Entropy</td><td>LRR (ours)</td><td>DetectGPT</td><td>NPR (ours)</td></tr><tr><td>0.5</td><td>98.72</td><td>77.87</td><td>99.29</td><td>25.90</td><td>99.23</td><td>86.14</td><td>95.76</td></tr><tr><td>0.7</td><td>97.01</td><td>76.98</td><td>98.05</td><td>38.28</td><td>98.84</td><td>90.28</td><td>95.61</td></tr><tr><td>0.9</td><td>90.04</td><td>75.82</td><td>92.28</td><td>47.14</td><td>94.50</td><td>90.33</td><td>92.89</td></tr><tr><td>0.95</td><td>86.15</td><td>75.43</td><td>88.88</td><td>50.42</td><td>92.04</td><td>89.97</td><td>91.98</td></tr><tr><td>1</td><td>81.48</td><td>74.85</td><td>84.81</td><td>52.37</td><td>89.15</td><td>89.02</td><td>90.86</td></tr></table>
+
+Table 4: Temperature experiments. Results of using different temperatures (AUROC scores).  
+
+<table><tr><td>Perturbation</td><td>Method</td><td>GPT-2-x1</td><td>Neo-2.7</td><td>OPT-2.7</td><td>GPT-j</td><td>OPT-13</td><td>Llama-13</td><td>NeoX</td></tr><tr><td rowspan="5">w/o</td><td>log p</td><td>0.06</td><td>0.09</td><td>0.10</td><td>0.04</td><td>0.07</td><td>0.07</td><td>0.60</td></tr><tr><td>Rank</td><td>0.07</td><td>0.10</td><td>0.09</td><td>0.04</td><td>0.05</td><td>0.07</td><td>0.60</td></tr><tr><td>Log-Rank</td><td>0.06</td><td>0.09</td><td>0.10</td><td>0.04</td><td>0.05</td><td>0.06</td><td>0.60</td></tr><tr><td>Entropy</td><td>0.06</td><td>0.09</td><td>0.09</td><td>0.04</td><td>0.05</td><td>0.06</td><td>0.60</td></tr><tr><td>LRR (ours)</td><td>0.12</td><td>0.19</td><td>0.18</td><td>0.08</td><td>0.10</td><td>0.14</td><td>1.20</td></tr><tr><td rowspan="2">w/</td><td>DetectGPT</td><td>8.07</td><td>9.60</td><td>9.80</td><td>7.03</td><td>7.98</td><td>8.14</td><td>35.56</td></tr><tr><td>NPR (ours)</td><td>8.15</td><td>9.69</td><td>9.90</td><td>7.12</td><td>7.83</td><td>7.98</td><td>35.67</td></tr></table>
+
+Table 5: Computational time (seconds) for different zero-shot methods on different LLMs (averaged over 10 reruns).  
+
+<table><tr><td colspan="5">tp(s)</td><td colspan="6">tm(s)</td></tr><tr><td>T5-3b</td><td>T5-large</td><td>T5-base</td><td>T5-small</td><td>GPT-2-x1</td><td>Neo-2.7</td><td>OPT-2.7</td><td>GPT-j</td><td>OPT-13</td><td>Llama-13</td><td>NeoX</td></tr><tr><td>0.10</td><td>0.08</td><td>0.04</td><td>0.03</td><td>0.06</td><td>0.09</td><td>0.10</td><td>0.04</td><td>0.07</td><td>0.07</td><td>0.60</td></tr></table>
+
+Table 6: Computation time. Estimated computation time for one perturbation  $(t_p)$  and for calculating the target statistics on the text  $(t_m)$ : shown in seconds.
+
+Composition of the Computational Time. In general, for perturbation-free zero-shot methods, the computational time only depends on the size of LLM and the complexity of statistics. LRR is twice as complex as simple statistics such as Log-Rank and Log-Likelihood, so it takes approximately twice as long to compute. As for LLM size, intuitively, larger models usually take more time to compute, which can also be observed in Table 5. The additional computational time of perturbation-based methods comes from two folds: (1) The total time for perturbation, which depends on the perturbation function we use and the number of perturbations. (2) The total time for calculating statistics of the perturbed texts, which depends on the number of perturbations, the size of LLM and the complexity of statistics. To reduce the computational time of the perturbation-based method, we could either choose a smaller size of the perturbation function or reduce the number of perturbations.
+
+# Formula for Estimating the Computational
+
+Time. Let  $t_p$  be the time of perturbing 1 sample,  $t_m$  be the time of calculating simple statistics (such as Log-Likelihood) of one sample for a particular LLM and  $n$  be the number of perturbations. The computational time for Log-Likelihood, rank, Log-Rank, and entropy is approximately  $t_m$ , the estimated time for LRR is  $2 \cdot t_m$ , while the estimated computational time for the perturbation-based method is  $n \cdot t_p + (n + 1) \cdot t_m$ . The estimated
+
+values of  $t_p$  and  $t_m$  are illustrated in Table 6, which can help us estimate the total running time (in seconds) of different zero-shot methods.
+
+# 6.2 Balancing Efficiency and Performance
+
+In this subsection, we provide additional experiments on LRR (the best perturbation-free method) and NPR (the best perturbation-based method, more time-consuming than LRR but also rather satisfactory performance) to provide users with some intuition on setting parameters of NPR and choosing among between these two methods according to user's preference of efficiency and performance.
+
+First, we study the perturbation function used for NPR. Different from Section 5.2, where the focus is to illustrate the advanced performance of NPR compared with DetectGPT, here, we mainly focus on the efficiency performance trade-off perspective and provide some intuition on choosing perturbation functions.
+
+T5-small and T5-base are not good candidates for perturbation functions. T5-small and T5-base are 2 or 3 times faster than larger models such as T5-large (as shown in Table 6), one might wonder if it is possible to trade the saved time with more perturbations for a better performance? We give a negative answer to this. We observe in Figure 3 that using T5-base and T5-small performs worse than LRR even with 50 to 100 perturbations, which suggests that LRR can be at least 50 to 100 times faster
+
+![](images/a6e18abb2c34c73693514aaa18530dabac1368d487497fc0c24bc645689d5b69.jpg)  
+Figure 3: Comparing LRR and NPR when T5-small and T5-base are used for perturbation in NPR (AUROC scores).
+
+![](images/56f3745589aaecf5deeae93b94360e335d5e4decb81f7fd58f48e3d8a591aaa7.jpg)  
+Figure 4: Comparing LRR and NPR using T5-3b and T5-large for perturbation in NPR (AUROC scores).
+
+while outperforming perturbation based methods. So, if the user can only afford T5-small or T5-base as a perturbation function, they should choose LRR with no hesitation since it achieves both better efficiency and better performance.
+
+Cost-Effectiveness on More Perturbations and Larger Perturbation Function. In Figure 4, we illustrate the effectiveness of LRR compared to NPR with T5-large and T5-3b as perturbation function respectively, from which, we find that (1) T5-3b has a higher performance upper limits compared with T5-large. So, if resources are allowed (enough memory and adequate perturbation time), t5-3b would be a better choice, especially for users who prioritize performance. (2) To achieve the same performance as LRR, generally, we only need less than 10 perturbations using T5-3b as the perturbation function. This estimate could help us choose whether to use NPR or LRR on the validation set: setting the number of perturbations to be 10, if LRR outperforms NPR, we would suggest using LRR, otherwise, NPR would be a better option. (3) To achieve the same performance, using T5-large takes more than 2 times perturbations than using
+
+T5-3b, while the perturbation time using T5-3b is less than twice the time using T5-large, so using large perturbation functions such as T5-3b is much more efficient than using smaller ones such as T5-large. The only concern is the memory.
+
+In summary, we suggest using the larger perturbation functions if memory permits, which is more cost-effective: and less time-consuming for the same performance and has a high-performance upper limit. Moreover, setting the number of perturbations to 10 would be a good threshold on the validation set to decide whether to use NPR or LRR.
+
+# 7 Conclusion
+
+In this paper, we proposed two simple but effective zero-shot machine-generated text detection methods by leveraging the Log-Rank information. The methods we proposed —LRR and NPR—, achieve state-of-the-art performance within their respective category. In addition, we explored different settings such as decoding strategy and temperatures, as well as different perturbation functions and number of perturbations to better understand the advantages
+
+and the disadvantages of different zero-shot methods. Then, we analyzed the computational costs of these methods, and we provided guidance on balancing efficiency and performance.
+
+# Limitations
+
+One of the limitations of zero-shot methods is the white box assumption that we can have some statistics about the source model. This induces two problems: for closed-source models (such as GPT-3), these statistics might not have been provided. Moreover, in practice, the detector might have to run the model locally to get the statistics for the purpose of detection, which requires that the detector have enough resources to use the LLM for inference. Based on the limitations of zero-shot methods, we consider weakly supervised learning (Ratner et al., 2017) as an important direction for future work. Though many papers in detecting machine-generated text assume knowing the source LLM where the text is generated from, in realistic, the source LLM might be unknown, so it is worth combining weak supervised learning as well as weak supervision sources (other LLMs at hand that might not be the target LLM) to weakly train a classifier. With the flexibility of the weak supervision sources, the limitations of our work could possibly be addressed: (1) Since the weak supervision sources do not have to be from the same target model, there is no need to assume that the target LLM is known. (2) Since the weak supervision sources are classifiers, we could only use statistics that are within reach, or even statistics from other open-source LLMs. (3) The weak supervision sources can be from smaller LLMs, rather than the target LLM, this relaxes the requirement for running an extremely large LLM locally.
+
+In addition, our conclusions hugely rely on our English-centric experiments. It is worth noting that the detection of machine-generation text in other languages is also important, especially for low-source languages. We encourage future studies on the investigation of zero-shot detection in multilingual settings.
+
+# Ethics and Broader Impact
+
+Although our paper focuses on the malicious use of LLMs such as spreading misinformation and propaganda, or dishonesty in the education system, it's necessary to recognize that LLMs also have a wide range of potential benefits, and we would like
+
+to point out that most of the people apply LLMs for good conducts such as improving their work efficiency.
+
+Moreover, even though our detectors achieve high AUC scores, it should be recognized that every machine-generated text detector, including ours, has its limitations and can make mistakes, we can't guarantee  $100\%$  accuracy for every sample. As such, when deciding whether a text, such as a student's essay, is written by a human or machine, our results are for reference only and should not be used as concrete evidence for punishment. For ethical concerns, our detector should only assist humans to make decisions, rather than directly make decisions for users, thus, we recommend users take these results as one of many pieces in a holistic assessment of texts.
+
+# References
+
+Sahar Abdelnabi and Mario Fritz. 2021. Adversarial watermarking transformer: Towards tracing text provenance with data hiding. In 2021 IEEE Symposium on Security and Privacy (SP), pages 121-140. IEEE.  
+Sameer Badaskar, Sachin Agarwal, and Shilpa Arora. 2008. Identifying real or fake articles: Towards better language modeling. In Proceedings of the Third International Joint Conference on Natural Language Processing: Volume-II.  
+Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc'Aurelio Ranzato, and Arthur Szlam. 2019. Real or fake? learning to discriminate machine from human generated text. arXiv preprint arXiv:1906.03351.  
+Daria Beresneva. 2016. Computer-generated text detection using machine learning: A systematic review. In *Natural Language Processing and Information Systems: 21st International Conference on Applications of Natural Language to Information Systems*, NLDB 2016, Salford, UK, June 22-24, 2016, Proceedings 21, pages 421-426. Springer.  
+Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, and Samuel Weinbach. 2022. Gpt-neox-20b: An opensource autoregressive language model.  
+Sid Black, Gao Leo, Phil Wang, Connor Leahy, and Stella Biderman. 2021. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow. If you use this software, please cite it using these metadata.
+
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Tiziano Fagni, Fabrizio Falchi, Margherita Gambini, Antonio Martella, and Maurizio Tesconi. 2021. Tweepfake: About detecting deepfake tweets. Plos one, 16(5):e0251415.  
+Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833.  
+Luciano Floridi and Massimo Chiriatti. 2020. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30:681–694.  
+Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027.  
+Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. 2019. Gltr: Statistical detection and visualization of generated text. arXiv preprint arXiv:1906.04043.  
+Alexei Grinbaum and Laurynas Adomaitis. 2022. The ethical need for watermarks in machine-generated language. arXiv preprint arXiv:2209.03118.  
+Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. 2023. How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597.  
+Xinlei He, Xinyue Shen, Zeyuan Chen, Michael Backes, and Yang Zhang. 2023. Mgtbench: Benchmarking machine-generated text detection. arXiv preprint arXiv:2303.14822.  
+Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751.  
+Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck. 2019. Automatic detection of generated text is easiest when humans are fooled. arXiv preprint arXiv:1911.00650.  
+Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks VS Lakshmanan. 2020. Automatic detection of machine generated text: A critical survey. arXiv preprint arXiv:2011.01314.  
+John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. 2023. A watermark for large language models. arXiv preprint arXiv:2301.10226.
+
+Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. 2023. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. arXiv preprint arXiv:2303.13408.  
+Thomas Lavergne, Tanguy Urvoy, and François Yvon. 2008. Detecting fake content with relative entropy scoring. PAN, 8:27-31.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.  
+Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. 2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. arXiv preprint arXiv:2301.11305.  
+Shashi Narayan, Shay B Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. arXiv preprint arXiv:1808.08745.  
+OpenAI. 2022. gpt: Optimizing language models for dialogue. http://web.archive.org/web/20230109000707/https://openai.com/blog/chatgpt/.  
+Artidoro Pagnoni, Martin Graciarena, and Yulia Tsvetkov. 2022. Threat scenarios and best practices to detect neural fake news. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1233-1249.  
+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.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad:  $100,000+$  questions for machine comprehension of text. arXiv preprint arXiv:1606.05250.  
+Alexander Ratner, Stephen H Bach, Henry Ehrenberg, Jason Fries, Sen Wu, and Christopher Ré. 2017. Snorkel: Rapid training data creation with weak supervision. In Proceedings of the VLDB Endowment. International Conference on Very Large Data Bases, volume 11, page 269. NIH Public Access.  
+Joshua Robinson, Christopher Michael Ryting, and David Wingate. 2022. Leveraging large language models for multiple choice question answering. arXiv preprint arXiv:2210.12353.  
+Kalhan Rosenblatt. 2023. Chatgpt passes mba exam given by a wharton professor. Retrieved Jan, 25:2023.
+
+Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. 2023. Can ai-generated text be reliably detected? arXiv preprint arXiv:2303.11156.  
+Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. 2022. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100.  
+Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. 2019. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203.  
+Chris Stokel-Walker. 2022. Ai bot chatgpt writes smart essays-should academics worry? Nature.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. 2020. Authorship attribution for neural text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8384-8395.  
+Ben Wang and Aran Komatsuzaki. 2021. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model. https://github.com/kingoflolz/mesh-transformer-jax.  
+Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Chenxi Whitehouse, Osama Mohammed Afzal, Tarek Mahmoud, Alham Fikri Aji, et al. 2023. M4: Multi-generator, multi-domain, and multi-lingual black-box machine-generated text detection. arXiv preprint arXiv:2305.14902.  
+Ann Yuan, Andy Coenen, Emily Reif, and Daphne Ippolito. 2022. Wordcraft: story writing with large language models. In 27th International Conference on Intelligent User Interfaces, pages 841-852.  
+Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. 2022. Opt: Open pretrained transformer language models.  
+Xuandong Zhao, Yu-Xiang Wang, and Lei Li. 2023. Protecting language generation models via invisible watermarking. arXiv preprint arXiv:2302.03162.
+
+Terry Yue Zhuo, Yujin Huang, Chunyang Chen, and Zhenchang Xing. 2023. Red teaming chatgpt via jailbreaking: Bias, robustness, reliability and toxicity. arXiv preprint arXiv:2301.12867, pages 12-2.
+
+# A Experimental Details and Baselines
+
+Details on Baselines. We mainly compare the proposed methods with zero-shot methods, which utilize the source model itself to extract distinguishable statistic features, including:
+
+- Log-Likelihood  $(\log p)$  (Solaiman et al., 2019): This approach evaluates the average token-wise log probability of the text and classifies text with higher Log-Likelihood to be machine-generated.  
+- Rank (Gehrmann et al., 2019): This approach evaluates the average rank of each token of the text and classifies text with a smaller average rank to be machine-generated.  
+- Log-Rank (Mitchell et al., 2023): Instead of using the Rank score directly, this approach evaluates the average Log-Rank of each token of the text and classifies text with a smaller average Log-Rank to be machine-generated.  
+- Entropy (Gehrmann et al., 2019): This approach is inspired by the hypothesis that machine-generated texts are more likely to have over-confident (thus low entropy) predictive distributions. In practice, (Mitchell et al., 2023) discovered that entropy is positively correlated with passage fakeness, therefore, following their convention, we use high average entropy as a signal of machine-generated text.  
+- DetectGPT (Mitchell et al., 2023): DetectGPT is based on the hypothesis that when applying small perturbations to a passage  $x$  and produce the perturbed text  $\tilde{x}$ , the quantity  $\log p_{\theta}(x) - \log p_{\theta}(\tilde{x})$  is relatively larger for machine-generated samples than human written one. In practice, the performance of this approach depends heavily on the external perturbation function and the number of perturbations.
+
+Details on LLMs used. We used 7 LLMs ranging from 1.5B parameters to 20B parameters in our main experiments.
+
+- GPT-2-x1 (Radford et al., 2019) is the 1.5B parameter version of GPT-2 trained on a dataset of 8 million web pages called WebText (Radford et al., 2019), whose objective is to predict the next word given previous words within the text. GPT-2-x1 surpasses many other language models trained on specific domains (such as books, news, Wikipedia) without using domain-specific training datasets.  
+- GPT-Neo-2.7B (Black et al., 2021) was trained as an autoregressive language model on Pile (Gao et al., 2020) dataset with EleutherAI's replication of the GPT-3 architecture.  
+- OPT-2.7B and OPT-13B are two models among a collection of decoder-only pre-trained transformers introduced in (Zhang et al., 2022), with the performance roughly matching GPT-3 of the same size.  
+- GPT-j-6B (Wang and Komatsuzaki, 2021), which was also trained on Pile (Gao et al., 2020), exhibits zero-shot performance roughly comparable to GPT-3 of comparable size. In addition, the performance gap from GPT-3 of similar size is closer than the GPT-Neo models.  
+- Llama-13b is the 13B parameter model from Llama models (Touvron et al., 2023): a collection of models ranging from 7B to 65B parameters trained with publicly available datasets. Llama-13B outperforms GPT-3 (175B) on most benchmarks, and all the models are released to the research community.  
+- NeoX-20B (Black et al., 2022) is a 20B autoregressive model trained on Pile, whose weights have been released openly to the public.
+
+Experimental Details. For small models such as GPT-2-xl, Neo-2.7, OPT-2.7, GPT-j, we use 1 NVIDIA A100 GPU (with total memory 40G) in our experiments; for larger models such as OPT-13b and Llama-13, we use 3 A100 GPUs (total memory 120 G) while using 4 A100 GPUs (total memory 160 G) for the largest model NeoX-20.
+
+<table><tr><td colspan="2"></td><td>GPT-2-xl</td><td>Neo-2.7</td><td>OPT-2.7</td><td>GPT-j</td><td>OPT-13</td><td>Llama-13</td><td>NeoX</td><td>Avg.</td></tr><tr><td rowspan="2">XSum</td><td>RoBERTa-base</td><td>97.57</td><td>96.82</td><td>94.86</td><td>90.37</td><td>88.62</td><td>79.18</td><td>88.96</td><td>90.91</td></tr><tr><td>RoBERTa-large</td><td>99.74</td><td>99.73</td><td>98.37</td><td>97.58</td><td>93.85</td><td>85.93</td><td>95.13</td><td>95.76</td></tr><tr><td rowspan="2">SQuAD</td><td>RoBERTa-base</td><td>97.65</td><td>94.42</td><td>92.56</td><td>87.57</td><td>88.96</td><td>76.98</td><td>84.37</td><td>88.93</td></tr><tr><td>RoBERTa-large</td><td>99.01</td><td>98.30</td><td>96.53</td><td>93.31</td><td>91.62</td><td>82.59</td><td>88.37</td><td>92.82</td></tr><tr><td rowspan="2">WritingP</td><td>RoBERTa-base</td><td>96.88</td><td>95.23</td><td>89.57</td><td>93.26</td><td>86.18</td><td>83.49</td><td>88.92</td><td>90.50</td></tr><tr><td>RoBERTa-large</td><td>98.75</td><td>98.80</td><td>94.98</td><td>97.11</td><td>88.75</td><td>88.32</td><td>93.72</td><td>94.35</td></tr></table>
+
+Table 7: Complete results for the supervised methods (AUROC score).  
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="4">top-k</td><td colspan="4">top-p</td></tr><tr><td>Neo-2.7</td><td>OPT-2.7</td><td>GPT-j</td><td>Llama-13</td><td>Neo-2.7</td><td>OPT-2.7</td><td>GPT-j</td><td>Llama-13</td></tr><tr><td rowspan="2">XSum</td><td>RoBERTa-base</td><td>96.48</td><td>94.15</td><td>92.72</td><td>82.47</td><td>98.30</td><td>97.30</td><td>96.71</td><td>85.84</td></tr><tr><td>RoBERTa-large</td><td>99.74</td><td>98.06</td><td>98.29</td><td>87.33</td><td>99.84</td><td>98.97</td><td>98.82</td><td>89.35</td></tr><tr><td rowspan="2">SQuAD</td><td>RoBERTa-base</td><td>93.55</td><td>93.27</td><td>87.60</td><td>76.79</td><td>96.34</td><td>97.65</td><td>92.26</td><td>84.05</td></tr><tr><td>RoBERTa-large</td><td>98.35</td><td>96.88</td><td>93.97</td><td>82.42</td><td>98.21</td><td>98.39</td><td>95.09</td><td>86.46</td></tr><tr><td rowspan="2">WritingP</td><td>RoBERTa-base</td><td>97.27</td><td>90.14</td><td>93.86</td><td>83.24</td><td>98.33</td><td>94.09</td><td>96.55</td><td>88.78</td></tr><tr><td>RoBERTa-large</td><td>99.34</td><td>96.34</td><td>97.12</td><td>87.05</td><td>99.68</td><td>96.06</td><td>97.94</td><td>89.59</td></tr></table>
+
+Table 8: Complete results for the supervised methods using top-  $k$  (  $k = {40}$  ) and top-  $p\left( {p = {0.96}}\right)$  sampling across four models (AUROC scores).
+
+# B Supervised Methods
+
+Main results for supervised methods. Comparing Table 1 with Table 7, we found that, on average, our best zero shot method (either LRR on SQuAD dataset or NPR on XSum and WritingPrompts dataset) can exceed supervised model fine-tuned on RoBERTa-base. For the larger model RoBERTa-large, only on writing dataset, perturbation-based method DetectGPT and NPR outperform RoBERTa-large model, by a margin of  $0.55\%$  and  $2.87\%$  respectively.
+
+Supervised Method with Different Decoding Strategy. We experimented the 4 models used in zero-shot methods with top- $p$  and top- $k$  decoding strategy for the supervised method and found that using top- $p$  decoding strategy performs better than using top- $k$ . (See Table 8). Compared to zero-shot methods, the best zero-shot method NPR can outperform the RoBERTa-base model while being comparable to the RoBERTa-large model.
+
+Supervised Method with Different Temperature. Supervised methods also perform better with lower temperature, but zero-shot methods such as Log-Rank and Log-Likelihood methods might exceed supervised methods in low temperature. Moreover, we found that the performance gap of RoBERTa-base and RoBERTa-large would be narrowed with lower temperature. The results are illustrated in Figure 5.
+
+# C Comparing NPR and DetectGPT
+
+Different Number of Perturbations. The results for models smaller than or equal to 13B parameters are shown in Figure 6. For the NeoX-20b model, we don't have enough computation resources to perform 100 perturbations, so we show it separately in Figure 7 with 1, 10, 20, and 50 perturbations. For XSum dataset, NPR and DetectGPT almost coverages with 100 perturbations, but for the SQuAD and WritingPrompts dataset, NPR still outperforms DetectGPT even with 100 perturbations. For the SQuAD dataset with the Llama-13b model, DetectGPT exhibits abnormality while NPR maintains stably improved performance as the number of perturbations increases. In addition, in nearly all the datasets and models, NPR outperforms DetectGPT except GPT-j on the XSum dataset, demonstrating the effectiveness of NPR compared to DetectGPT.
+
+Using T5-large as Perturbation Function. We illustrate the performance of NPR and DetectGPT in Figure 8 with different combinations of dataset and LLMs using T5-large as a perturbation function. Compared to T5-3b illustrated in Figure 6, the superiority of NPR over DetectGPT becomes more distinct with T5-large being the perturbation function, where in almost all the LLMs, datasets and different numbers of perturbations (except with Llama-13b on SQuAD), NPR outperforms DetectGPT by a large margin. In addition, we could also observe that NPR achieves comparable or even better results with only
+
+![](images/7cf54f1013660f1036598b686aa4d22057fd7430231111ed574ced0c18bbdcd7.jpg)  
+Figure 5: Comparing supervised methods with different temperature (AUROC score).
+
+10 perturbations to that of DetectGPT with 100 perturbations, which indicates that NPR is more efficient and can achieve a similar level of performance with significantly fewer perturbations.
+
+# D Alternative Sampling Strategies and Temperature
+
+Different Sampling Strategy. In Table 9, we illustrate the complete results with different zero-shot methods with four LLMs using top- $p$  and top- $k$  sampling. For perturbation-based methods, even with different sampling strategies, NPR provides a clearer signal for machine-generated text detection than DetectGPT. Moreover, we find that although LRR is more stable than Log-Rank and Log-Likelihood methods: when replacing temperature sampling to top- $p$  and top- $k$  sampling, all the above-mentioned three zero-shot methods' performance improves, however, LRR improves approximately the same for both top- $k$  and top- $p$  sampling while the other two is more in favour of top- $p$  sampling.
+
+Different Temperature. Here, we investigate how the temperature used for machine-generated texts affects the detection accuracy of different zero-shot methods. From Figure 9, we find that all the perturbation-free zero-shot methods improved their performance with the decreasing temperature. In particular, for the Log-Rank and Log-Likelihood method, the performance can become extremely high
+
+![](images/2c4623bea0a7db4825fb0d935fedbd89343f26ff676447bd96bb5f46d3c45c99.jpg)  
+Figure 6: Comparing DetectGPT and NPR (AUROC score).
+
+![](images/1cc0a51456d733dedb4adeaa93a62d28f473740a9a85810521a5e1ab208d9280.jpg)  
+Figure 7: Comparing DetectGPT and NPR on NeoX-20b (AUROC score).
+
+when the temperature drops, even exceeding NPR and achieving approximately 100 points detection accuracy. For example, in Neo-2.7 and OPT-13 with temperature 0.5,  $\log p$  method and Log-Rank method achieve an accuracy of 100 points on WritingPrompts dataset, this prevalent performance can be observed notably in smaller models with relatively high temperature (such as GPT-2-xl and Neo-2.7 with high temperature such as 0.7) or in large models with relatively lower temperature such as OPT-13 with temperature 0.5 as we demonstrated in Figure 9. Though we omit the entropy method because it gets an accuracy worse than random guessing, one of the observations from our experiments is that using the assumption "machine-generated text has higher entropy" suggested in (Mitchell et al., 2023), the performance of the entropy method improves with the increasing temperature with absolute accuracy smaller than 50 points, which suggests that for low temperature, we should use the assumption "machine-generated text has lower entropy" for detection machine-generated text. In general, the Entropy method performs worse than random and is not an implementable detection method.
+
+For perturbation-based methods (Figure 10), while DetectGPT does not exhibit a clear trend with respect to temperature, the performance of NPR improves with the decreasing temperature most of the time. However, this trend is not as clear as the Log-Rank and Log-Likelihood methods, especially when the temperature becomes too low. This behaviour suggests that the perturbation-based method is more suitable for high temperatures, while the perturbation-free method is more suitable for low temperature.
+
+![](images/fd4f5bb7e8384b8bba3031489ca8e312eb76aef27fbe2e448e716f8ebf5c52b5.jpg)  
+Figure 8: Comparing DetectGPT and NPR using t5-large (AUROC score).
+
+<table><tr><td colspan="3"></td><td colspan="4">top-k</td><td colspan="4">top-p</td></tr><tr><td>Dataset</td><td>Perturbation</td><td>Method</td><td>Neo-2.7</td><td>OPT-2.7</td><td>GPT-j</td><td>Llama-13</td><td>Neo-2.7</td><td>OPT-2.7</td><td>GPT-j</td><td>Llama-13</td></tr><tr><td rowspan="7">XSum</td><td rowspan="5">w/o</td><td>log p</td><td>91.27</td><td>90.19</td><td>85.95</td><td>59.14</td><td>95.52</td><td>93.27</td><td>91.13</td><td>67.86</td></tr><tr><td>Rank</td><td>78.79</td><td>76.75</td><td>77.25</td><td>49.94</td><td>78.58</td><td>76.89</td><td>77.18</td><td>50.77</td></tr><tr><td>Log-Rank</td><td>94.20</td><td>92.30</td><td>89.18</td><td>65.09</td><td>96.71</td><td>93.93</td><td>92.53</td><td>71.44</td></tr><tr><td>Entropy</td><td>53.07</td><td>47.80</td><td>53.23</td><td>67.76</td><td>49.05</td><td>46.41</td><td>52.16</td><td>67.94</td></tr><tr><td>LRR (ours)</td><td>95.50</td><td>92.35</td><td>91.14</td><td>77.99</td><td>95.64</td><td>90.68</td><td>91.14</td><td>75.72</td></tr><tr><td rowspan="2">w/</td><td>DetectGPT</td><td>98.94</td><td>96.63</td><td>96.56</td><td>73.22</td><td>98.82</td><td>97.72</td><td>96.58</td><td>77.82</td></tr><tr><td>NPR (ours)</td><td>99.61</td><td>98.23</td><td>96.41</td><td>77.48</td><td>99.27</td><td>98.40</td><td>97.35</td><td>78.67</td></tr><tr><td rowspan="7">SQuAD</td><td rowspan="5">w/o</td><td>log p</td><td>87.85</td><td>91.00</td><td>81.32</td><td>45.06</td><td>91.20</td><td>94.24</td><td>86.69</td><td>56.16</td></tr><tr><td>Rank</td><td>80.10</td><td>82.14</td><td>79.81</td><td>55.21</td><td>80.56</td><td>82.40</td><td>80.28</td><td>56.89</td></tr><tr><td>Log-Rank</td><td>92.58</td><td>94.40</td><td>86.94</td><td>51.21</td><td>94.48</td><td>96.37</td><td>90.44</td><td>60.66</td></tr><tr><td>Entropy</td><td>54.62</td><td>50.83</td><td>56.89</td><td>69.52</td><td>54.51</td><td>50.01</td><td>55.67</td><td>63.26</td></tr><tr><td>LRR (ours)</td><td>97.79</td><td>97.58</td><td>94.55</td><td>72.52</td><td>97.48</td><td>98.11</td><td>94.38</td><td>74.38</td></tr><tr><td rowspan="2">w/</td><td>DetectGPT</td><td>97.04</td><td>97.53</td><td>87.59</td><td>47.52</td><td>97.50</td><td>97.48</td><td>88.90</td><td>52.06</td></tr><tr><td>NPR (ours)</td><td>98.56</td><td>99.35</td><td>91.21</td><td>50.83</td><td>98.32</td><td>99.18</td><td>92.99</td><td>54.28</td></tr><tr><td rowspan="7">WritingP</td><td rowspan="5">w/o</td><td>log p</td><td>96.62</td><td>95.99</td><td>95.67</td><td>86.93</td><td>98.16</td><td>98.10</td><td>97.11</td><td>92.68</td></tr><tr><td>Rank</td><td>82.67</td><td>83.96</td><td>83.49</td><td>78.49</td><td>82.89</td><td>84.45</td><td>83.55</td><td>79.01</td></tr><tr><td>Log-Rank</td><td>97.90</td><td>97.23</td><td>97.20</td><td>90.57</td><td>98.73</td><td>98.60</td><td>97.89</td><td>94.56</td></tr><tr><td>Entropy</td><td>32.37</td><td>38.22</td><td>34.37</td><td>44.09</td><td>27.08</td><td>36.77</td><td>32.82</td><td>39.03</td></tr><tr><td>LRR (ours)</td><td>98.58</td><td>97.97</td><td>98.06</td><td>93.80</td><td>98.46</td><td>97.97</td><td>97.76</td><td>94.79</td></tr><tr><td rowspan="2">w/</td><td>DetectGPT</td><td>99.05</td><td>98.65</td><td>96.05</td><td>81.83</td><td>98.80</td><td>98.62</td><td>96.67</td><td>82.70</td></tr><tr><td>NPR (ours)</td><td>99.58</td><td>99.46</td><td>98.27</td><td>87.99</td><td>99.36</td><td>99.04</td><td>97.85</td><td>89.96</td></tr></table>
+
+Table 9: Complete result for the zero-shot methods using top-  $k$  and top-  $p$  sampling across four models (AUROC score).
+
+![](images/dd9319c58baae417d970e9336efffd19c2ae24079ede5b5c18738b813eb3dc88.jpg)
+
+![](images/91733fbbc7732635854ce27e460821f10fee3054cdd8df41e3f7867355dc1d70.jpg)
+
+![](images/1e32709dfaaab0d85ca0d8b495583bc6175b4a6c7e8783d7bd7f0ace4265444f.jpg)
+
+![](images/4739a7d52fea07254a7492ff151ca0cb4d4c5ea7f287ba3657097c680e552ac4.jpg)
+
+![](images/5706955fb570a7c2e7e8d3633ae9256ef38645b303f290ce8cd2ed12278f53d8.jpg)
+
+![](images/cf58a433eafb1f29c35b9da5b97a39817021ee633e5df65698d05ef94ec809af.jpg)
+
+![](images/1bc9bc1a47e10227ca263ea8d9e2314542ff171cfb24fffaf8a59b3b26e82d44.jpg)
+
+![](images/ffbf232f965cacdb5f2656e88d677cb938e3a4d2abec17d8042a267ca1eedb62.jpg)
+
+![](images/6fa70f4172873ddf452ee576074f291a9b70647d13541360c1ccc039632a58b5.jpg)
+
+![](images/c1cb7d76c5a2850b5c802630ecef4a0cb9374453f9b3774bf603e0188a09aca8.jpg)  
+Figure 9: Comparison of perturbation-free methods using different temperatures (AUROC score).
+
+![](images/ed46c63e27aae9339386962a602123e410a037fdafd61ce5a141189365ac6446.jpg)
+
+![](images/36bb38488ea9e8b225b9797a5301702e5acf1f7845f545c26dc5960a4e68b37a.jpg)
+
+![](images/1545adfdbadc669bc5fbd7b651071f03036af1f4e23996e7124ba6e8f5da2656.jpg)  
+Figure 10: Comparison of perturbation methods using different temperature (AUROC score).

2023/DetectLLM_ Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb805fc1fedad66dbc5fe9d3538c9b0f5fa9c4c87ed6f0d9441dab47b40b3a92
+size 1374587

2023/Detecting Erroneously Recognized Handwritten Byzantine Text/e23fb75f-6ceb-4c26-9972-d5c23aa51bcb_content_list.json

@@ -0,0 +1,1752 @@
+[
+    {
+        "type": "text",
+        "text": "Detecting Erroneous Handwritten Byzantine Text Recognition",
+        "text_level": 1,
+        "bbox": [
+            176,
+            87,
+            818,
+            109
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "John Pavlopoulos",
+        "bbox": [
+            255,
+            123,
+            410,
+            139
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Informatics",
+        "bbox": [
+            223,
+            141,
+            442,
+            156
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Athens University of Economics and Business",
+        "bbox": [
+            146,
+            156,
+            521,
+            172
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Greece",
+        "bbox": [
+            302,
+            174,
+            363,
+            187
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "annis@aueb.gr",
+        "bbox": [
+            267,
+            191,
+            400,
+            205
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Vasiliki Kougia",
+        "bbox": [
+            598,
+            123,
+            732,
+            139
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Faculty of Computer Science",
+        "bbox": [
+            546,
+            140,
+            784,
+            156
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Vienna, Austria",
+        "bbox": [
+            546,
+            156,
+            786,
+            172
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "UniVie Doctoral School Computer Science",
+        "bbox": [
+            490,
+            174,
+            840,
+            189
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "vasiliki.kougia@univie.ac.at",
+        "bbox": [
+            524,
+            190,
+            806,
+            205
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Paraskevi Platanou",
+        "bbox": [
+            245,
+            227,
+            416,
+            241
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Philology, School of Philosophy",
+        "bbox": [
+            137,
+            243,
+            524,
+            261
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Athens, Greece",
+        "bbox": [
+            213,
+            262,
+            448,
+            277
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "pplatanou@phil.uoa.gr",
+        "bbox": [
+            223,
+            278,
+            436,
+            293
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Holger Essler",
+        "bbox": [
+            605,
+            227,
+            727,
+            243
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Humanities",
+        "bbox": [
+            557,
+            244,
+            774,
+            260
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ca'Foscari University of Venice, Italy",
+        "bbox": [
+            512,
+            261,
+            820,
+            277
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "holger.essler@unive.it",
+        "bbox": [
+            554,
+            278,
+            776,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            319,
+            339,
+            335
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Handwritten text recognition (HTR) yields textual output that comprises errors, which are considerably more compared to that of recognised printed (OCRed) text. Post-correcting methods can eliminate such errors but may also introduce errors. In this study, we investigate the issues arising from this reality in Byzantine Greek. We investigate the properties of the texts that lead post-correction systems to this adversarial behaviour and we experiment with text classification systems that learn to detect incorrect recognition output. A large masked language model, pre-trained in modern and fine-tuned in Byzantine Greek, achieves an Average Precision score of  $95\\%$ . The score improves to  $97\\%$  when using a model that is pretrained in modern and then in ancient Greek, the two language forms Byzantine Greek combines elements from. A century-based analysis shows that the advantage of the classifier that is further-pre-trained in ancient Greek concerns texts of older centuries. The application of this classifier before a neural post-corrector on HTRed text reduced significantly the post-correction mistakes.",
+        "bbox": [
+            141,
+            347,
+            460,
+            703
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            715,
+            258,
+            731
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Handwritten Text Recognition (HTR) concerns the automated transcription of a handwritten text, extracting its computerised form from an image. Applying HTR on an old manuscript can lead to a high character error rate (CER), often due to the lack of training data (Pavlopoulos et al., 2023). In the real world, this erroneous HTR output is post-corrected by experts, who get involved, typically, in a tedious and time-consuming task. Delaying the delivery of manually curated data, hinders both: the preservation of digitised manuscripts that have not yet been",
+        "bbox": [
+            112,
+            741,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1a6900ee3055eb9c40c6505bf00db057e7a0f2d9f507a3f97878c5d891264007.jpg",
+        "image_caption": [
+            "Figure 1: Number of HTRed-lines according to their CER as presented by Pavlopoulos et al. (2023)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            319,
+            867,
+            513
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "transcribed; and the progress of HTR for historical manuscripts, which struggles with low resources. This is why researchers from Digital Humanities have started encouraging the development of Natural Language Processing (NLP) solutions, in an attempt to assist the experts toward this goal, since improving the accuracy of recognition will be exciting for the disciplines struggling with vast corpora and little man power (McGillivray et al., 2020).",
+        "bbox": [
+            507,
+            580,
+            884,
+            726
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1.1 Background",
+        "text_level": 1,
+        "bbox": [
+            509,
+            737,
+            653,
+            753
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Although detecting erroneous recognised output is not a novel task (Chiron et al., 2017; Amrhein and Clematide, 2018; Rigaud et al., 2019; Jatowt et al., 2019; Nguyen et al., 2020; Schaefer and Neudecker, 2020), existing published studies focus on printed material (i.e., OCRed). Handwritten text, on the other hand, increases the challenge in recognition compared to printed material and hence, leads to more errors. The different hands, the greater time period, morphology and spelling",
+        "bbox": [
+            507,
+            758,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "7818",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7818-7828",
+        "bbox": [
+            216,
+            945,
+            779,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            277,
+            958,
+            719,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/63a6b54475e5dd4f3f50b4af50255ffa12a8de313a3385506f4015b035ca1c27.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Transcription</td><td>Recognition</td></tr><tr><td>a.</td><td>tov ποδων</td><td>τωη ποδων</td></tr><tr><td>b.</td><td>tov ποδων</td><td>αχοασδδεδ</td></tr><tr><td rowspan=\"2\">a.</td><td>έγγινουμένα πάθη</td><td>έγγινουμένα πάθη</td></tr><tr><td>μή σβενύντε</td><td>μή σβενύντε</td></tr><tr><td rowspan=\"2\">b.</td><td>έγγινουμένα πάθη</td><td>έγγινουμένδπάθη</td></tr><tr><td>μή σβενύντε</td><td>μαε σβανησοτε</td></tr><tr><td rowspan=\"3\">a.</td><td>τες εμπυρίζουσι</td><td>τες εμπυρίζουσι</td></tr><tr><td>tόν ἀμπελώνα ἀλλὰ</td><td>tόν ἀμπελΟνα ἀλλὰ</td></tr><tr><td>xαλό διὰ τες</td><td>xαλό διὰ τες</td></tr><tr><td rowspan=\"3\">b.</td><td>τες εμπυρίζουσι</td><td>τεσέμπυρίσουβοι</td></tr><tr><td>tόν ἀμπελώνα ἀλλὰ</td><td>xδα ἀμπελοαα xδλὰ</td></tr><tr><td>xαλό διὰ τες</td><td>xxxxαλ αδιοτες</td></tr></table>",
+        "bbox": [
+            115,
+            82,
+            487,
+            294
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Synthetic transcription (human-generated) and recognition (system-generated) output with few (a) and many (b) errors. Finding the differences between the transcription and the recognition is easier for the three examples shown in red (b).",
+        "bbox": [
+            112,
+            305,
+            489,
+            376
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/a4a3b7d2d163bed7f10abba2e1ed69e1a9eed264f270f5bc58f9693cb4b0948f.jpg",
+        "image_caption": [
+            "Figure 2: Image of the handwritten text 'των ποδων' - in the recognised text, the third character was misspelled, written as eta ('η') instead of 'v' (Table 1, 1st row)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            392,
+            475,
+            435
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "(see Section 3), can explain the greater diversity of error rates (Fig. 1). The recognition quality, however, is related to the detection quality of erroneous output. To give an example, the detection of an erroneous recognition is easier when the recognition output is scrambled compared with a single incorrect character. This is shown in Table 1, where classifying recognised text (as erroneous/not) is easier for the human eye in (b) compared to that of (a). We hypothesise that this applies also to models.",
+        "bbox": [
+            112,
+            514,
+            489,
+            677
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We argue that detecting erroneously recognised text can improve automatic error correction. Current approaches to post-correcting the often flawed recognised text, disregard the likelihood of a flawless recognition, assuming that the correct text can be simply copied (e.g., with an encoder-decoder-based correction). Consequently, they are faced with the possibility of introducing, instead of mitigating, errors that would be absent if post-correction hadn't been applied. Taking the case of a misspelled article (e.g., see Fig. 2), for instance, there is a variety of letters to choose from, for replacement, always accordingly to the declension and the case of the noun. Even if grammatical rules are set (e.g., in a rule-based system), an",
+        "bbox": [
+            112,
+            678,
+            490,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "HTR system working at the character level is likely examining morphological aspects and often disregards semantic and pragmatic levels. A vowel replacement function addressing the morphological level, however, may omit important semantic aspects and relationships between sentence elements at the syntactic level and thus, misinterpret the case. For example, the line “λεντιον διεζωσεν εαυτόν” was HTRed as “λεντιον διεζωσενεαυτόν”. A neural corrector, then, which received as input the HTRed output, yielded “λεντιον διεζωσεν εαυτόν” (Pavlopoulos et al., 2023), correctly separating the two last words but incorrectly removing the accent. In this way and in this instance, post-correcting involves the risk of introducing errors by transforming probably correct output cases.",
+        "bbox": [
+            507,
+            84,
+            884,
+            342
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1.2 The focus of the study",
+        "text_level": 1,
+        "bbox": [
+            509,
+            353,
+            726,
+            368
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This study quantifies the hazards of the above-mentioned adversarial post-correction, by focusing on the Byzantine Greek language. The development of this language entails a comprehensive examination of written language throughout the centuries under study. In essence, Byzantine Greek emerged as an academic style known as Atticised Greek, incorporating religious topics. As time progressed, elements of contemporary spoken Greek gradually became integrated into the language (Horrocks, 2014). As the results of our study also show, the combination of pre-trained data on Modern Greek, resembling what was considered as spoken language back then, and Ancient Greek, including the above-mentioned Atticised Greek, is beneficial for text classifiers detecting incorrect recognition output. Furthermore, by covering a period of more than seven centuries, the already challenging nature of recognition (Platanou et al., 2022) is increased, hence allowing a more diverse material for our investigation on post-correction. Errors in handwritten text recognition are often due to different hands (scribes) and scripts that evolved for many centuries. Printed text, on the other hand, exists for a much narrower period, there are no hands to cause similar errors, $^{1}$  and no clitics. As we show in this work, lower recognition error rate means more difficult classification of erroneous lines. By contrast, recognised handwritten text is often hard to parse even for humans, which makes an easier classification challenge and a much more difficult",
+        "bbox": [
+            507,
+            373,
+            884,
+            872
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>Fonts pose a similar problem, yet less challenging when compared to hands, due to their non-standardised form and high variation.",
+        "bbox": [
+            507,
+            881,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "7819",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "error-correction task, as shown in Table 1.",
+        "bbox": [
+            114,
+            85,
+            426,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "1.3 The contributions",
+        "text_level": 1,
+        "bbox": [
+            114,
+            114,
+            302,
+            128
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We explore the extent to which classification to flawless/flawed texts can improve the post-correction of recognised output, by avoiding adversarial errors introduced when applied to flawless texts. We benchmark machine- and deep-learning-based text classifiers, showing that a Transformer that is pre-trained in modern Greek and then finetuned to detect erroneously recognised Byzantine Greek (10th-16th CE) text achieves an Average Precision of  $95\\%$ . When the same Transformer is further pre-trained in ancient Greek prior to fine-tuning, the score increases to  $97\\%$ . We quantify the benefits of employing our text classifier in the real world, by training an MT5 (Xue et al., 2021) encoder-decoder error-corrector, which altered  $80\\%$  of the flawless texts. If our best-performing classifier had been applied prior to correction, the percentage would have been decreased to  $2\\%$  with a trade-off of  $2\\%$  texts being incorrectly classified as flawless.",
+        "bbox": [
+            115,
+            136,
+            489,
+            456
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The better performance of a Transformer that is not pre-trained only in modern Greek can be explained by the fact that Byzantine Greek combines language elements from both, ancient and modern Greek. Since our study covers a broad historical period, we separated the evaluation dataset based on the century in order to investigate this hypothesis. Our results show that the detection of erroneous recognition depends on the chronology of the manuscript and that further pre-training in ancient Greek benefits the classification of lines from manuscripts in the following three centuries: 10thCE, 11thCE, and 13thCE. By following a similar segmentation-based evaluation method for the error rate level, instead of the century, we show that higher error rates make an easier classification goal. By contrast, the problem is harder for low error rates, when flawed and flawless texts look alike.",
+        "bbox": [
+            115,
+            458,
+            489,
+            747
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2 Related work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            763,
+            265,
+            778
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our work focuses on the text classification of recognised handwritten text. In the absence of published studies for handwritten text, however, we summarise work focused on the recognition of printed text. Printed and handwritten materials differ, because the latter often yields scrambled recognition output (Table 1). This is probably what motivated Ströbel et al. (2022) in using the perplexity of lan",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "guage models to detect the erroneous output in an unsupervised manner. On the other hand, although we consider that recognition errors are overall more for handwritten text compared to printed material, the quality of recognition can vary significantly for the former (Hodel et al., 2021), as is also shown in Fig. 1, and does not always come with a high error rate. Therefore, the detection of erroneously recognised handwritten text is also related to that for printed text, regarding their low error output.",
+        "bbox": [
+            507,
+            84,
+            884,
+            244
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Erroneous recognised printed text detection",
+        "text_level": 1,
+        "bbox": [
+            509,
+            256,
+            858,
+            271
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Initial work using rule-based approaches could not address all errors, and was followed by machine and deep learning approaches (Lyu et al., 2021). Standard approaches include support vector machines on top of n-gram-based features (Dannélls and Virk, 2021; Virk et al., 2021). Jatowt et al. (2019) used a gradient tree boosting binary classifier on top of various features, including character and word n-grams, part-of-speech, token frequency based on automatically created resources.",
+        "bbox": [
+            507,
+            275,
+            884,
+            436
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Amrhein and Clematide (2018) performed detailed analyses and experiments of error detection and post-correction with various statistical and neural machine-translation methods. In general, they found that the former perform better in error correction, while the latter models are better in error detection. However, they concluded that there is not a single method that works best on all datasets while the results are highly affected by the data a model is trained on. Furthermore, they suggested that post-correction of recognised printed text should focus more on the improvement of error detection.",
+        "bbox": [
+            507,
+            437,
+            885,
+            629
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Nguyen et al. (2020) used BERT pre-trained on named-entity recognition for token classification in order to perform error detection. After substring tokenization they obtain Glove or Fasttext word embeddings, which are combined with segment and positional embeddings and given as input to BERT (Devlin et al., 2018). The hidden states are fed to a dense layer on top that classifies each token as erroneous or not. Although interesting, token-level detection is not suited for high-error settings.",
+        "bbox": [
+            507,
+            631,
+            884,
+            790
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The ICDAR challenge",
+        "text_level": 1,
+        "bbox": [
+            509,
+            802,
+            687,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ICDAR organised a competition focused on post-correcting text recognised from newspapers, shopping receipts, and other printed sources, including two subtasks: error detection, i.e., to detect the position and the length of the errors, and error correction (Chiron et al., 2017; Rigaud et al., 2019).",
+        "bbox": [
+            507,
+            822,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "7820",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In 2017, the best-performing method for the error detection subtask employed probabilistic character error models and weighted finite-state transducers while in 2019, the best was a pre-trained multilingual BERT with convolutional and fully-connected layers on top that classified each sub-token as erroneous or not. Schaefer and Neudecker (2020) followed a similar classification scheme, but focused on characters and not sub-tokens. Then, by excluding lines with at least one classified character from post-correction, they showed that false alteration (i.e., of correctly OCR red lines) by an encoder-decoder model could be avoided. Our work shows that casting this problem as a text classification task can lead to high accuracy for handwritten text, although similar benefits may apply to printed text.",
+        "bbox": [
+            112,
+            84,
+            492,
+            344
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 Data",
+        "text_level": 1,
+        "bbox": [
+            112,
+            354,
+            191,
+            368
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The written language of the manuscripts and papyri used in this study is Byzantine Greek.3 Within these texts, morphological categories such as the optative, the pluperfect, and the perfect have disappeared, while others such as the dative case have gradually decreased. Infinitives and participles are still there in the texts, as reminiscents of the classical tradition, encouraging one to treat the language as a unique variant, different from modern Greek. There are several spelling conventions that deviate from the older orthographic rules while the ancient punctuation signs are still in use, albeit not always with the same function. Therefore, following Classical and preceding Modern Greek, this language can be considered a combination of the two.",
+        "bbox": [
+            112,
+            379,
+            489,
+            620
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The transcriptions and the HTR output were provided by the organisers of the recent HTREC challenge, which regarded the automated correction of HTR errors in Byzantine Greek (Pavlopoulos et al., 2023). Selected images of handwritten text in respective manuscripts (10th to 16th CE) were transcribed by a human expert. The training dataset of the challenge consists of 1,875 lines of transcribed text by a human expert and by an HTR model (examples shown in Table 2). Transkribus (Kahle et al., 2017) was used by the organisers to produce the system transcriptions, trained on seven images one per century. Our investigation of the CER of the opted HTR model reveals",
+        "bbox": [
+            112,
+            621,
+            489,
+            845
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/bdab47f54557d25fe59dce788ce6afe429b7b6ed1b8acf3a66b466b791bdf41b.jpg",
+        "image_caption": [
+            "Figure 3: Violinplot of CER per century"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            531,
+            93,
+            865,
+            241
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "that a diversity of errors is produced, from very few to many per line (Fig. 1). This fact makes this dataset optimal for our training purposes, because we need a diversity of errors to train our classifiers. Furthermore, the errors are not uniformly distributed across centuries (Fig. 3), motivating a time/CER-based segmentation for evaluation purposes (discussed in Section 4.2).",
+        "bbox": [
+            507,
+            297,
+            884,
+            425
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our classification dataset consists of the 1,875 human-transcribed texts (lines), treated as the flawless negative (0) class, and the respective system-transcribed texts, treated as the flawed, positive (1) class. This resulted in total to an almost perfectly balanced dataset of 3,744 text lines, which we split to train, development and test subsets by using a 80/10/10 ratio. The average length of the transcription is 40.10 characters (6.83 words), and that of the HTR output is 37.09 characters (5.89 words).",
+        "bbox": [
+            507,
+            426,
+            884,
+            588
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            602,
+            737,
+            619
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We opted for machine and deep learning benchmarks, with different transfer learning settings.",
+        "bbox": [
+            507,
+            630,
+            882,
+            663
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Machine learning benchmarks",
+        "text_level": 1,
+        "bbox": [
+            507,
+            676,
+            756,
+            690
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Machine learning methods include a variant of Naive Bayes and Support Vector Machines (NB-SVM), which is often used as a baseline for text classification (Wang and Manning, 2012). Also, we experimented with a Random Forest and a multi-layer perceptron (MLP) using character n-grams",
+        "bbox": [
+            507,
+            697,
+            884,
+            796
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2The difference lies in using a sigmoid function (or so) on top of the text representation, instead of one per character.",
+        "bbox": [
+            112,
+            854,
+            487,
+            879
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "3Also known as Medieval Greek or simply Byzantine.",
+        "bbox": [
+            137,
+            879,
+            463,
+            892
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "4https://www.aicrowd.com/challenges/",
+        "bbox": [
+            137,
+            892,
+            410,
+            904
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "htrec-2022",
+        "bbox": [
+            115,
+            904,
+            196,
+            916
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "Fewer errors will harm class balance while similar errors will harm generalisation of the classifiers.",
+        "bbox": [
+            507,
+            806,
+            882,
+            832
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "Six lines were excluded from the positive class, because a perfect transcription was achieved.",
+        "bbox": [
+            507,
+            832,
+            882,
+            856
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "7All experiments were performed using Google's Colaboratory, using a Tesla T4 GPU card. Our large language models comprise 113 million parameters.",
+        "bbox": [
+            507,
+            856,
+            882,
+            892
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "We used the implementation of KTRAIN with default parameters (Maiya, 2020).",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "7821",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/9644d1193de75c46d8ffcc0e791bcce60d45e65162cb254442c746aefeeac595.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Transcription by expert</td><td>HTR output</td></tr><tr><td>èγγινόμενα πάθη μή σβενύντες ἀλλὰ τή εχλύσει (the born-in passions not extinguishing but the release)</td><td>èγγενομεναπαδημησεννωτες ἀλλατήε κλησει</td></tr><tr><td>του βίου του καῦ ́ εαυτός πολλα γένεσθουσυ υμχωρουν (of the life of themselves many happening forgive)</td><td>του β ou του καλεαυτός πολλαγινεσθουσι συγχωρόν</td></tr><tr><td>τες ἄμυριζουσι του ὄμελώνα ἀλλα καὶ ὅ διὰ τες (- set on fire the vineyard but and the due to the)</td><td>εμμυριζου σιμαμπελώνα ἀλλακαι ὅδξα</td></tr></table>",
+        "bbox": [
+            142,
+            80,
+            857,
+            205
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$(n \\in 1,5)$ , term-frequency inverse document frequency, and default parameters otherwise.",
+        "bbox": [
+            112,
+            268,
+            489,
+            300
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Deep learning benchmarks",
+        "text_level": 1,
+        "bbox": [
+            112,
+            312,
+            331,
+            326
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Deep learning methods comprise a recurrent neural network using Gated Recurrent Units (GRU) and trainable word embeddings of 300 dimensions, originally pre-trained on (modern) Greek (Joulin et al., 2017). Also, we experimented with the state-of-the-art in text classification, which is BERT (Devlin et al., 2018) and its variants (Li et al., 2022; Minaee et al., 2021). In specific, we also fine-tuned two BERT models pre-trained in Greek. The first (GreekBERT) is pre-trained in modern Greek (Koutsikakis et al., 2020) while the second extends the first by being further pre-trained in ancient Greek (Singh et al., 2021).",
+        "bbox": [
+            112,
+            332,
+            489,
+            541
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation measures",
+        "text_level": 1,
+        "bbox": [
+            112,
+            552,
+            284,
+            565
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We opted for the  $F1$  score per class, the Area under the Receiver Operating Characteristic Curve (AUC), and Average Precision (AP) that is preferred in class-imbalanced settings. All evaluation measures were implemented in SCIKIT-LEARN. $^{12}$",
+        "bbox": [
+            112,
+            573,
+            489,
+            653
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Text classification benchmark",
+        "text_level": 1,
+        "bbox": [
+            112,
+            664,
+            394,
+            678
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 3 shows the results of all our text classification benchmarks, including a random baseline (responding uniformly) that draws the lower performance limit. NBSVM is the worst machine learning model, followed by Forest and MLP. The latter achieves promising results, close to GRU (better in AP and  $F1^{-}$ ). GRU, on the other hand, is clearly",
+        "bbox": [
+            112,
+            684,
+            489,
+            797
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/446f8158475dba57b0c7bb1bc0fde2bd6081f4dc087fc8006210b30ac75a61bb.jpg",
+        "table_caption": [
+            "Table 2: Human (on the left) and HTR-based (on the right) transcription (lines) from the 'Commentary on Isaiah' by Basil the Great (fol. 75r, Oxford, Bodleian Library MS. Barocci 102)"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>AP</td><td>AUC</td><td>F1+</td><td>F1-</td></tr><tr><td>Random</td><td>0.52</td><td>0.50</td><td>0.49</td><td>0.47</td></tr><tr><td>NBSVM</td><td>0.66</td><td>0.65</td><td>0.60</td><td>0.51</td></tr><tr><td>Forest</td><td>0.64</td><td>0.65</td><td>0.64</td><td>0.50</td></tr><tr><td>MLP</td><td>0.79</td><td>0.79</td><td>0.73</td><td>0.69</td></tr><tr><td>GRU</td><td>0.79</td><td>0.79</td><td>0.68</td><td>0.71</td></tr><tr><td>GreekBERT:M</td><td>0.95</td><td>0.94</td><td>0.88</td><td>0.88</td></tr><tr><td>GreekBERT:M+A</td><td>0.97</td><td>0.97</td><td>0.90</td><td>0.91</td></tr></table>",
+        "bbox": [
+            510,
+            265,
+            890,
+            399
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Evaluation with Average Precision (AP), AUC,  $F1$  for the classification of HTRed texts to flawed (+) or flawless (-) transcription. Random (uniform) classification was used as a baseline. GreekBERT was pre-trained in modern (:M) or in modern and ancient Greek (:A).",
+        "bbox": [
+            507,
+            407,
+            884,
+            481
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "outperformed in all metrics by GreekBERT. Two GreekBERT models were used, one pre-trained in modern Greek (:M) and one in modern and then in ancient Greek (:M+A). Byzantine Greek, as already discussed (Section 3), combines language elements found in both ancient and modern Greek. This is probably why GreekBERT:M+A, which was pretrained in modern and then in ancient Greek, was better compared to GreekBERT:M, which was pretrained only in modern Greek.[13]",
+        "bbox": [
+            507,
+            504,
+            882,
+            664
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Evaluation per century",
+        "text_level": 1,
+        "bbox": [
+            507,
+            675,
+            739,
+            690
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We separated the lines of the test data based on the century of the manuscript that they came from. For each century, then, we evaluated the two Transformers, GreekBERT:M and GreekBERT:M+A. As can be seen in Fig. 4, $^{14}$  GreekBERT:M+A is better on the 10th (five percent units) and 11th century CE, with models performing similarly on the 12th and after the 13th century, when both models achieve a",
+        "bbox": [
+            507,
+            696,
+            882,
+            825
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>13</sup>We verified the consistency of this finding by repeating the experiments three times, reporting an average AUC of 0.97 over 0.94 (st.d. lower than 0.05).",
+        "bbox": [
+            507,
+            832,
+            882,
+            868
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "For statistical significance, we sampled twenty lines per century, repeating the sampling process ten times. We focus on Average Precision, but similar findings were observed with the other metrics.",
+        "bbox": [
+            507,
+            869,
+            880,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{9}$ https://scikit-learn.org/stable/modules/ generated/sklearn.neural_network  ${ }^{10}$ https://fasttext.cc/docs/en/crawl-vectors. htm1",
+        "bbox": [
+            112,
+            804,
+            460,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "11 Ancient Greek texts, which can be used for pre-training, are much fewer in number compared to publicly available modern Greek data. Hence, the option of pre-training directly in ancient Greek was not considered viable.",
+        "bbox": [
+            112,
+            856,
+            487,
+            903
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "12https://scikit-learn.org/stable/",
+        "bbox": [
+            132,
+            904,
+            386,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "7822",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7f33b797aff5e59ccfcfed607a70a076263648627bbfa4090df6949c069b0456.jpg",
+        "image_caption": [
+            "Figure 4: Average Precision per c. CE of GreekBERT:M, pretrained on modern Greek (solid), and GreekBERT:M+A, further pre-trained on ancient Greek (dashed) - error bands show the  $95\\%$  bootstrap confidence intervals"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            99,
+            811,
+            249
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "higher score. On the 13th century, the performance of both models drops, but considerably more for GreekBERT:M. The 14th, 15th and 16th century were more challenging in HTR (Platanou et al., 2022), which could explain the better performance of the classifiers. That is, more errors in the recognised output could potentially help the classifier distinguish between the two classes (correct/not).<sup>15</sup> Based on the previous observations, we report three findings:",
+        "bbox": [
+            112,
+            313,
+            489,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "F1. The classification performance depends on the chronology of the manuscript.",
+            "F2. Older manuscripts are better handled by further pre-training on ancient Greek.",
+            "F3. Lines from recent manuscripts pose an easier classification challenge to both GreekBERT models."
+        ],
+        "bbox": [
+            117,
+            486,
+            489,
+            620
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Evaluation per error zone",
+        "text_level": 1,
+        "bbox": [
+            112,
+            633,
+            366,
+            648
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our dataset is balanced regarding whether the recognition output is erroneous (positive class) or not, yet it remains imbalanced regarding the CER level (see Fig. 1). In order to assess the two Transformers across different input types, we rolled a window of ten units, from low to high CER values, and we assessed the models per window (Fig. 5).<sup>16</sup>",
+        "bbox": [
+            112,
+            653,
+            489,
+            765
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "High-CER lines (i.e., lines with CER of thirty or more) appear to set an easy classification task to both models. Performance increases as the CER increases. This is reasonable considering that lines with a high CER comprise many errors and are",
+        "bbox": [
+            112,
+            766,
+            489,
+            847
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "hard to parse even by humans. This characteristic, however, makes them easier to distinguish from flawless lines, which comprise no errors from the recognition.",
+        "bbox": [
+            507,
+            313,
+            884,
+            376
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Low-CER lines (i.e., lines with CER of zero to twenty) lead to a different picture, since Greek-BERT:M+A is consistently better, overall, than GreekBERT:M. Low CER means small differences between the two texts (flawless-flawed), which makes the classification task more difficult. In this zone, further pre-training on ancient Greek had a bigger impact, compared to zones with a higher error rate.",
+        "bbox": [
+            507,
+            379,
+            885,
+            524
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Discussion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            539,
+            636,
+            556
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, first, we assess the possibly adversarial nature of post-correction. Then, we describe a synthetic error detection application, followed by a century-based error analysis we performed.",
+        "bbox": [
+            507,
+            569,
+            885,
+            633
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Substantiality assessment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            646,
+            756,
+            663
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Error-correction systems may introduce errors, even when applied to flawless HTRed texts. To assess the substantiality of this claim, we experimented with a neural encoder-decoder postcorrector. In specific, we employed a multilingual variant of the T5 Text-to-Text Transfer Transformer (Raffel et al., 2019), called MT5 (Xue et al., 2020), which was pre-trained on a dataset covering 101 languages.[17] MT5 has been used for Grammatical Error Correction in modern Greek and sets a strong baseline for HTR Error Correction (Korre and Pavlopoulos, 2022). We fine-tuned this model on the 1,875 training instances of the HTREC",
+        "bbox": [
+            505,
+            671,
+            884,
+            879
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "17The Greek language was on the 20th position with 43 billion tokens extracted from 42 million pages",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>15</sup>It remains unclear to the authors whether more hands were present during these centuries, explaining the performance of the HTR and the classification models, or if this was due to another reason; e.g., writing became more cursive over time. <sup>16</sup>We sampled thirty lines, repeating sampling ten times.",
+        "bbox": [
+            112,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "7823",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1b1fc96a65a97018d7f67907f01cd0a6068fe460d5c6b6f4d4c969915a3566d6.jpg",
+        "image_caption": [
+            "Figure 5: Average Precision of GreekBERT:M, pretrained on modern Greek (blue, solid), and GreekBERT:M+A, further pre-trained on ancient Greek (orange, dashed), using a rolling window over CER (left edge shown horizontally) - error bands depict  $95\\%$  bootstrap confidence intervals"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            99,
+            811,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "challenge, by encoding the system- and decoding the human-transcriptions. $^{18}$  This model is a neural error correction baseline that reduced the CER in 19 out of the 180 HTRed texts in the official HTREC2022 evaluation set. We applied this corrector on the respective 180 transcriptions of these data (the ground truth evaluation instances in HTREC2022) and found that 144  $(80\\%)$  were mistakenly altered by our MT5 corrector. For example, the line 'αν τους μηδους οινες έτε αργυρτω παρα' was altered to 'αν τους μηδους οινες έτε αργυρτω παρα' (word division, 4th word). Similarly, 'χευσης μη φονευσης μη κλε' was altered to 'χλευσεως μη φονευσεως μη κλεπτ'. Most of these adversarial mistakes (all but the top three of Table 4) would have been avoided if our GreekBERT:M+A classifier had been applied prior to the application of this neural corrector. Conversely, only four (i.e., the rest in Table 4) would have been mistakenly found as correct from the system-transcribed instances.",
+        "bbox": [
+            115,
+            328,
+            490,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 A synthetic application of error detection",
+        "text_level": 1,
+        "bbox": [
+            112,
+            678,
+            482,
+            694
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We experimented with the HTREC2022 evaluation dataset, which was used to assess the participating systems. This dataset comprises 180 lines that were transcribed from nine images. Out of these lines, one had zero CER achieved by the HTR model (perfect transcription), which was correctly classified by our GreekBERT:M+A model. We then performed two analyses, first by applying our model on the flawless human transcriptions (as if a transcription was the HTR output) and then by applying",
+        "bbox": [
+            112,
+            699,
+            489,
+            859
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "it on the flawed system recognitions (the same lines were used). The human transcriptions were correctly classified as flawless, all except three  $(1.7\\%)$  error rate), which are shown in the top three rows of Table 4. The system recognitions were also correctly classified as flawed, all except four  $(2.3\\%)$  error rate), shown in the last four rows of Table 4.",
+        "bbox": [
+            507,
+            328,
+            884,
+            441
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Error analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            464,
+            670,
+            479
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "An error detection by experts in Table 4, $^{19}$  showed that no clue (indicating an error) was found for the first line (except from the rare ‘φιλαχόλουνός’ with double accent). As is shown in Figure 6 (a), which depicts the attention on the subwords of this line, this was indeed the network’s focus.",
+        "bbox": [
+            505,
+            491,
+            882,
+            588
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "On the second line, word 'iσμαηλήτι' was considered as the outcome of misspelling of the word 'iσμαηλιτη', matching the algorithm's decision. In this case, it was the ending of the word that attracted attention (Fig. 6), but the network focused more on the article 'των' preceding the word rather than on the word itself. The article is in plural form, not matching a word in singular. Apparently, 'iσμαηλήτι' is not a misspelling, but part of the word 'iσμαηλιτιχων' that was corrupted.",
+        "bbox": [
+            505,
+            592,
+            882,
+            752
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In the last four lines, that were classified as correct, several flaws were raised, indicating a failure of the algorithm. For example, word divisions, such as in the case of  $\\omega \\sigma x\\alpha$  instead of  $\\omega \\varsigma \\chi \\alpha \\iota$ ; added characters, such as in the case of  $\\varepsilon \\dot{\\upsilon}\\upsilon \\chi \\eta \\varsigma$  instead of  $\\varepsilon \\dot{\\upsilon}\\chi \\eta \\varsigma$ ; character replacements, such as in the case of  $\\zeta \\kappa \\tau \\varepsilon \\iota \\alpha \\nu$  instead of  $\\iota \\kappa \\varepsilon \\tau \\varepsilon \\iota \\alpha \\nu$ .",
+        "bbox": [
+            507,
+            757,
+            884,
+            869
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>18</sup>We used the SimpleT5 implementation (https://github.com/Shivanandroy/simpleT5) with 100 epochs, early stopping (patience of 2 epochs), 15 tokens max. length, batch size of 4, and best loss found at the 12th epoch (2.64).",
+        "bbox": [
+            112,
+            868,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{19}$ No access was given to the manuscript or the human transcription, for fair comparison with the setting of the algorithm.",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "7824",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/3d80b8cc0fcc53cfd920a7b99a4eddab0d31a0f381045f95879284e656341f9a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>TRANSCRIPTIONS AND HTR OUTPUT</td><td>GREEKBERT:M+A</td><td>GT</td></tr><tr><td>èγώ δ ἀεί πως φιλαχόλουθός εἰμι(I am always readily following)</td><td>F</td><td>N</td></tr><tr><td>λα και γέδεων εκ τών σχύλων τών ομαηλητι(- and Gedeon from the dogs the - )</td><td>F</td><td>N</td></tr><tr><td>δραμων εἰς και γόνυπετη(having run to and falling on the knee)</td><td>F</td><td>N</td></tr><tr><td>τιθεις τας χειρας επαυτα(having put the hands on them)</td><td>F</td><td>N</td></tr><tr><td>ζετείαν χέδεσποτα, και ευμενώς Ἁκουσου τών τής προς εύυχης(- - - and favourably listen to the of the prayer)</td><td>N</td><td>F</td></tr><tr><td>èγενόμην icos. αραή τής μεό τιεγεζήθης ἐλειεμοήτιριος(became - - - - - )</td><td>N</td><td>F</td></tr><tr><td>ωσααι τής συνεχους διδασκαλιας τους βα(- of the continuous instruction the - )</td><td>N</td><td>F</td></tr></table>",
+        "bbox": [
+            126,
+            80,
+            870,
+            329
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: HTRed lines and transcriptions that GreekBERT:M+A mistakenly predicted as flawed (F) or not (N) - the ground truth (GT) on the right shows whether there was an error or not",
+        "bbox": [
+            112,
+            338,
+            882,
+            368
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Error patterns",
+        "text_level": 1,
+        "bbox": [
+            112,
+            394,
+            235,
+            407
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Next, we proceed with an analysis of the recognition errors, using data from three centuries. We opt for 16th CE, the century on which both our models perform considerably well (Fig. 4). Next, we focus on the 14th CE, going back two centuries, but still observing good performance by both models. Last is shown the 13th CE, where GreekBERT:M performs much worse than GreekBERT:M+A.",
+        "bbox": [
+            112,
+            413,
+            489,
+            541
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "16th CE This is the data group with the highest average CER (see Figure 3). What seems to be challenging in this data group is the fact that there is a variety of punctuation marks which are sometimes read by the model as letters because of their shape. An illustrative example of this case is the word 'πλοῦς' which is read by the model as 'πλούου'. This is due to a comma following the word 'πλοῦς', which is very similar to the last letter of the word (ς). There are also many cases of ligatures in which the model fails to recognise both letters, especially the one written on the top of the ligature, such as the letter 'τ' in the combination 'τρ' as one can see in the following instance; 'ω πατρις' read as 'ως ωπαρις'.",
+        "bbox": [
+            112,
+            552,
+            489,
+            796
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "14th CE The data from the 14th century show the second highest CER on average (see Figure 3). This is partly because of a great number of ligatures appearing in this data group. A common ligature among this group is the one which represents the group of the letters  $\\varepsilon$  and  $\\rho$ . In this ligature it happens that the letter  $\\varepsilon$  is written above the letter",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "$\\varphi^{\\prime}$ . The way the model interprets this structure can be seen in the following case;  $\\varphi \\acute{\\varepsilon} \\rho^{\\prime} \\dot{\\varepsilon} \\pi^{\\prime} \\omega \\tau \\eta \\nu$ $\\iota \\omega$ , where  $\\varphi \\acute{\\varepsilon} \\rho^{\\prime}$  consists of one letter  $(\\varphi)$  and one ligature  $(\\varepsilon \\rho)$  is read by the model as  $\\varphi \\varepsilon \\pi \\alpha \\omega \\tau \\eta \\nu$ $\\iota \\varsigma \\omega$ . The model fails reading correctly the letter combination and returns only the letter  $\\varphi^{\\prime}$  and one of the two letters involved in the  $\\varepsilon \\rho^{\\prime}$  combination  $(\\rho)$ , which is actually the one written in the position next to the letter  $\\varphi^{\\prime}$ . The same instance shows clearly that there is a tendency towards adding characters because of different drawing representations of characters found in this data group. The letter  $\\omega$  can be considered such a case. In the previous example, we can see that the word  $\\iota \\omega$  is read by the model as  $\\iota \\varsigma \\omega$ . This is because, in this instance, the letter  $\\omega$  has the shape of a laid eight. The same letter has the usual shape in other cases in the text of this century, which was quite confusing for the model and led the model to decide that it should be a combination of both  $\\sigma^{\\prime}$  and  $\\omega^{\\prime}$ , since the former is quite close to the half laid-eight shape.",
+        "bbox": [
+            505,
+            393,
+            884,
+            732
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "13th CE The data group that is easiest to read for the model seems to be the 13th century one. In this group, errors appear mainly because of special letter shapes and abbreviations. Such a letter shape is that of the letter  $\\varepsilon$  which is written as the letter  $\\delta$  with an added line pointing upwards. A characteristic example of the failure of the recognition task is this one;  $\\tilde{\\varepsilon}\\omega\\varsigma$  is read as  $\\delta\\omega\\varsigma$ . Here the model has clearly classified the  $\\tilde{\\varepsilon}$  as a  $\\delta$ . Regarding the abbreviations representing character groups, one can find among others the following",
+        "bbox": [
+            507,
+            741,
+            885,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "7825",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b2b7a82b084099725c5d3a11f87f09630850cb3ae1b16ac4f0a3d88595c1dd97.jpg",
+        "image_caption": [
+            "(c)  $\\delta \\rho \\alpha \\mu \\omega \\nu \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\vareeta$",
+            "Figure 6: The attention of GreekBERT:M+A (darker is more) for the three top (missclassified) lines of Table 4 - the symbol ## is used to concatenate a subword with the preceding one during tokenisation ( $\\delta \\rho \\alpha \\mu \\omega \\nu$  is  $\\delta \\rho \\alpha \\# \\mu \\omega \\nu$ )"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            83,
+            878,
+            204
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "one;  $\\alpha v$  . In the case of ouxioov, where the last two letters are represented by the abbreviation, the model reads ouxoX, where  $\\omega$  has replaced the abbreviation because of similarity in shape.",
+        "bbox": [
+            112,
+            294,
+            487,
+            357
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.4 Language and script generalisation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            372,
+            436,
+            387
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our work can potentially inspire analogous pretraining in other languages with a long documented history. To provide an example: Old, Middle, and Modern English and Japanese (Lyovin, 1997, pp. 2) can qualify for pre-training on both Old and Modern data, as in the Greek language in our case. The scripts our results are based on, however, are very common both in Greek and Latin manuscripts while common handwritten character shapes, e.g., epsilon (E) (Chambers and Chambers, 1891, pp. 703), may illuminate error patterns in Latin scripts.",
+        "bbox": [
+            112,
+            394,
+            489,
+            571
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            586,
+            245,
+            600
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work showed that detecting erroneous recognised lines improves the accuracy of handwritten Byzantine text recognition, aiming to assist the disciplines struggling with vast corpora and little man power. By contrast to printed text, errors in handwritten text vary greatly, some of which are easier and others difficult to catch. Our experiments showed that two BERT models, one pre-trained in modern and the other pre-trained in modern then in ancient Greek perform well for the task of erroneous handwritten recognised text classification. The latter performs better, especially for manuscripts from older centuries and for lines with higher recognition error rate. By using a neural text-to-text Transformer for error correction along with our best performing classifier, we found that adversarial behaviour is not uncommon in the neural corrector's output which could have been significantly limited if a BERT classifier, pre-trained and",
+        "bbox": [
+            112,
+            613,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "fine-tuned, had classified the input first.",
+        "bbox": [
+            507,
+            294,
+            803,
+            309
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Directions of future work comprise the development of a larger dataset, which, although not easily developed in our domain of focus, it could potentially allow us to draw more conclusions and lead to further findings. Furthermore, we plan to extend our dataset with calibrated evaluation subsets, in order to gain more perspective on the performance of systems for specific error types, including ones that occur naturally and are not meant to be corrected.",
+        "bbox": [
+            507,
+            310,
+            882,
+            455
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            468,
+            613,
+            482
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Our study is focused on Byzantine Greek, from the 10th to the 16th century. Our hypothesis, that text classification could assist the post correction of HTRd output, should be studied for more languages and covering a longer period to draw general conclusions.",
+            "- By contrast to OCR, where low CER values are reported, HTRd output can significantly differ from the human transcription. For example, correct text may be misplaced (e.g., due to mistaken word division) or by the introduction of an extra character (e.g., the last word of the fifth line in Table 4). Hence, the creation of ground truth at the character level is very challenging in this context, hindering experimentation at the token or the character level, as in OCR (Schulz and Kuhn, 2017; Schaefer and Neudecker, 2020).",
+            "- Sensitivity analysis has not been performed. Although we repeated our experiments to verify the validity of our findings regarding GreekBERT, we have not experimented with cross validation over the centuries, as in (Platanou et al., 2022). Training, however, an algorithm on texts from specific centuries, then as"
+        ],
+        "bbox": [
+            531,
+            493,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "7826",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "sessing on ones from previous or future ones, could potentially reveal interesting patterns.",
+        "bbox": [
+            149,
+            85,
+            487,
+            116
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "- Diplomatic transcriptions may lead to naturally-occurring errors (i.e., ones the editors would want to keep, to stay as close to the original as possible), which our proposed classifiers could and probably would classify as incorrect recognition output. This limitation could be bypassed if the proposed systems are used complementary with an expert, not trying to substitute one. For example, the expert could view only lines the classifier labelled as incorrect, bypassing or acknowledging errors which were meant to be there.",
+        "bbox": [
+            136,
+            127,
+            489,
+            335
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            375,
+            213,
+            390
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Chantal Amrhein and Simon Clematide. 2018. SupervisedOCR error detection and correction using statistical and neural machine translation methods. Journal for Language Technology and Computational Linguistics (JLCL), 33(1):49-76.",
+            "William Chambers and Robert Chambers. 1891. Chambers' Encyclopaedia: A Dictionary of Universal Knowledge, volume 8. W. & R. Chambers.",
+            "Guillaume Chiron, Antoine Doucet, Mickaël Coustaty, and Jean-Philippe Moreux. 2017. Icdar2017 competition on post-ocr text correction. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 1, pages 1423-1428, Kyoto, Japan.",
+            "Dana Dannélls and Shafqat Virk. 2021. A supervised machine learning approach for post-ocr error detection for historical text. In Swedish Language Technology Conference and NLP4CALL, pages 13-20.",
+            "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.",
+            "Tobias Mathias Hodel, David Selim Schoch, Christa Schneider, and Jake Purcell. 2021. General models for handwritten text recognition: feasibility and state-of-the-art. german current as an example. Journal of open humanities data, 7(13):1-10.",
+            "Geoffrey Horrocks. 2014. Greek: A History of the Language and its Speakers. John Wiley & Sons.",
+            "Adam Jatowt, Mickaël Coustaty, Nhu-Van Nguyen, Antoine Doucet, et al. 2019. Post-ocr error detection by generating plausible candidates. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 876-881, Sydney, Australia."
+        ],
+        "bbox": [
+            115,
+            398,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431. Association for Computational Linguistics.",
+            "Philip Kahle, Sebastian Colutto, Günter Hackl, and Günter Mühlberger. 2017. Transkribus - a service platform for transcription, recognition and retrieval of historical documents. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 04, pages 19-24.",
+            "Katerina Korre and John Pavlopoulos. 2022. Enriching grammatical error correction resources for Modern Greek. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4984-4991, Marseille, France. European Language Resources Association.",
+            "John Koutsikakis, Ilias Chalkidis, Prodromos Malakasiotis, and Ion Androutsopoulos. 2020. Greek-bert: The greeks visiting sesame street. In 11th Hellenic Conference on Artificial Intelligence, pages 110-117.",
+            "Qian Li, Hao Peng, Jianxin Li, Congying Xia, Renyu Yang, Lichao Sun, Philip S Yu, and Lifang He. 2022. A survey on text classification: From traditional to deep learning. ACM Transactions on Intelligent Systems and Technology (TIST), 13(2):1-41.",
+            "Anatole Lyovin. 1997. An Introduction to the Languages of the World. Oxford University Press, USA.",
+            "Lijun Lyu, Maria Koutraki, Martin Krickl, and Besnik Fetahu. 2021. NeuralOCR post-hoc correction of historical corpora. Transactions of the Association for Computational Linguistics, pages 479-493.",
+            "Arun S. Maiya. 2020. ktrain: A low-code library for augmented machine learning. arXiv preprint arXiv:2004.10703.",
+            "Barbara McGillivray, Thierry Poibeaun, and Pablo Ruiz Fabo. 2020. Digital humanities and natural language processing: \"je t'aime... moi non plus\".",
+            "Shervin Minaee, Nal Kalchbrenner, Erik Cambria, Narjes Nikzad, Meysam Chenaghlu, and Jianfeng Gao. 2021. Deep learning-based text classification: a comprehensive review. ACM Computing Surveys (CSUR), 54(3):1-40.",
+            "Thi Tuyet Hai Nguyen, Adam Jatowt, Nhu-Van Nguyen, Mickael Coustaty, and Antoine Doucet. 2020. Neural machine translation with bert for post-ocr error detection and correction. In Proceedings of the ACM/IEEE joint conference on digital libraries in 2020, pages 333-336.",
+            "John Pavlopoulos, Vasiliki Kougia, Paraskevi Platanou, Stepan Shabalin, Konstantina Liagkou, Emmanouil Papadatos, Holger Essler, Jean-Baptiste Camps, and Franz Fischer. 2023. Error correcting htr'ed byzantine text."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "7827",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Paraskevi Platanou, John Pavlopoulos, and Georgios Papaioannou. 2022. Handwritten paleographic greek text recognition: A century-based approach. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6585-6589.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683.",
+            "Christophe Rigaud, Antoine Doucet, Mickaël Coustaty, and Jean-Philippe Moreux. 2019. Icdar 2019 competition on post-ocr text correction. In 2019 international conference on document analysis and recognition (ICDAR), pages 1588-1593, Sydney, Australia.",
+            "Robin Schaefer and Clemens Neudecker. 2020. A two-step approach for automaticOCR post-correction. In Proceedings of the The 4th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 52-57, held online.",
+            "Sarah Schulz and Jonas Kuhn. 2017. Multi-modular domain-tailoredOCR post-correction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2716-2726, Copenhagen, Denmark.",
+            "Pranaydeep Singh, Gorik Rutten, and Els Lefever. 2021. A pilot study for bert language modelling and morphological analysis for ancient and medieval greek. In The 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2021).",
+            "Phillip Benjamin Ströbel, Simon Clematide, Martin Volk, Raphael Schwitter, Tobias Hodel, and David Schoch. 2022. Evaluation of htr models without ground truth material. arXiv preprint arXiv:2201.06170.",
+            "Shafqat Mumtaz Virk, Dana Dannélls, and Azam Sheikh Muhammad. 2021. A novel machine learning based approach for post-ocr error detection. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1463–1470, Varna, Bulgaria.",
+            "Sida Wang and Christopher Manning. 2012. Baselines and bigrams: Simple, good sentiment and topic classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 90-94, Jeju Island, Korea. Association for Computational Linguistics.",
+            "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2020. mt5: A massively multilingual pre-trained text-to-text transformer. arXiv preprint arXiv:2010.11934."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics.",
+        "bbox": [
+            510,
+            85,
+            882,
+            190
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "7828",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2023/Detecting Erroneously Recognized Handwritten Byzantine Text/e23fb75f-6ceb-4c26-9972-d5c23aa51bcb_model.json

@@ -0,0 +1,2191 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.177,
+                0.089,
+                0.82,
+                0.11
+            ],
+            "angle": 0,
+            "content": "Detecting Erroneous Handwritten Byzantine Text Recognition"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.257,
+                0.124,
+                0.411,
+                0.14
+            ],
+            "angle": 0,
+            "content": "John Pavlopoulos"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.225,
+                0.142,
+                0.443,
+                0.157
+            ],
+            "angle": 0,
+            "content": "Department of Informatics"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.147,
+                0.158,
+                0.522,
+                0.173
+            ],
+            "angle": 0,
+            "content": "Athens University of Economics and Business"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.304,
+                0.175,
+                0.364,
+                0.188
+            ],
+            "angle": 0,
+            "content": "Greece"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.268,
+                0.192,
+                0.401,
+                0.206
+            ],
+            "angle": 0,
+            "content": "annis@aueb.gr"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.599,
+                0.124,
+                0.734,
+                0.14
+            ],
+            "angle": 0,
+            "content": "Vasiliki Kougia"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.547,
+                0.141,
+                0.786,
+                0.157
+            ],
+            "angle": 0,
+            "content": "Faculty of Computer Science"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.547,
+                0.158,
+                0.787,
+                0.173
+            ],
+            "angle": 0,
+            "content": "University of Vienna, Austria"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.492,
+                0.175,
+                0.841,
+                0.19
+            ],
+            "angle": 0,
+            "content": "UniVie Doctoral School Computer Science"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.525,
+                0.191,
+                0.808,
+                0.206
+            ],
+            "angle": 0,
+            "content": "vasiliki.kougia@univie.ac.at"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.246,
+                0.228,
+                0.417,
+                0.242
+            ],
+            "angle": 0,
+            "content": "Paraskevi Platanou"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.139,
+                0.244,
+                0.525,
+                0.262
+            ],
+            "angle": 0,
+            "content": "Department of Philology, School of Philosophy"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.214,
+                0.263,
+                0.45,
+                0.278
+            ],
+            "angle": 0,
+            "content": "University of Athens, Greece"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.224,
+                0.279,
+                0.438,
+                0.294
+            ],
+            "angle": 0,
+            "content": "pplatanou@phil.uoa.gr"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.606,
+                0.228,
+                0.728,
+                0.244
+            ],
+            "angle": 0,
+            "content": "Holger Essler"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.558,
+                0.245,
+                0.776,
+                0.261
+            ],
+            "angle": 0,
+            "content": "Department of Humanities"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.513,
+                0.262,
+                0.821,
+                0.278
+            ],
+            "angle": 0,
+            "content": "Ca'Foscari University of Venice, Italy"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.555,
+                0.279,
+                0.778,
+                0.293
+            ],
+            "angle": 0,
+            "content": "holger.essler@unive.it"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.32,
+                0.341,
+                0.336
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.348,
+                0.461,
+                0.704
+            ],
+            "angle": 0,
+            "content": "Handwritten text recognition (HTR) yields textual output that comprises errors, which are considerably more compared to that of recognised printed (OCRed) text. Post-correcting methods can eliminate such errors but may also introduce errors. In this study, we investigate the issues arising from this reality in Byzantine Greek. We investigate the properties of the texts that lead post-correction systems to this adversarial behaviour and we experiment with text classification systems that learn to detect incorrect recognition output. A large masked language model, pre-trained in modern and fine-tuned in Byzantine Greek, achieves an Average Precision score of \\(95\\%\\). The score improves to \\(97\\%\\) when using a model that is pretrained in modern and then in ancient Greek, the two language forms Byzantine Greek combines elements from. A century-based analysis shows that the advantage of the classifier that is further-pre-trained in ancient Greek concerns texts of older centuries. The application of this classifier before a neural post-corrector on HTRed text reduced significantly the post-correction mistakes."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.717,
+                0.26,
+                0.732
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.743,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Handwritten Text Recognition (HTR) concerns the automated transcription of a handwritten text, extracting its computerised form from an image. Applying HTR on an old manuscript can lead to a high character error rate (CER), often due to the lack of training data (Pavlopoulos et al., 2023). In the real world, this erroneous HTR output is post-corrected by experts, who get involved, typically, in a tedious and time-consuming task. Delaying the delivery of manually curated data, hinders both: the preservation of digitised manuscripts that have not yet been"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.526,
+                0.32,
+                0.868,
+                0.514
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.527,
+                0.883,
+                0.556
+            ],
+            "angle": 0,
+            "content": "Figure 1: Number of HTRed-lines according to their CER as presented by Pavlopoulos et al. (2023)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.581,
+                0.885,
+                0.727
+            ],
+            "angle": 0,
+            "content": "transcribed; and the progress of HTR for historical manuscripts, which struggles with low resources. This is why researchers from Digital Humanities have started encouraging the development of Natural Language Processing (NLP) solutions, in an attempt to assist the experts toward this goal, since improving the accuracy of recognition will be exciting for the disciplines struggling with vast corpora and little man power (McGillivray et al., 2020)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.738,
+                0.655,
+                0.754
+            ],
+            "angle": 0,
+            "content": "1.1 Background"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.759,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Although detecting erroneous recognised output is not a novel task (Chiron et al., 2017; Amrhein and Clematide, 2018; Rigaud et al., 2019; Jatowt et al., 2019; Nguyen et al., 2020; Schaefer and Neudecker, 2020), existing published studies focus on printed material (i.e., OCRed). Handwritten text, on the other hand, increases the challenge in recognition compared to printed material and hence, leads to more errors. The different hands, the greater time period, morphology and spelling"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7818"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.78,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7818-7828"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.278,
+                0.959,
+                0.72,
+                0.972
+            ],
+            "angle": 0,
+            "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.083,
+                0.488,
+                0.295
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Transcription</td><td>Recognition</td></tr><tr><td>a.</td><td>tov ποδων</td><td>τωη ποδων</td></tr><tr><td>b.</td><td>tov ποδων</td><td>αχοασδδεδ</td></tr><tr><td rowspan=\"2\">a.</td><td>έγγινουμένα πάθη</td><td>έγγινουμένα πάθη</td></tr><tr><td>μή σβενύντε</td><td>μή σβενύντε</td></tr><tr><td rowspan=\"2\">b.</td><td>έγγινουμένα πάθη</td><td>έγγινουμένδπάθη</td></tr><tr><td>μή σβενύντε</td><td>μαε σβανησοτε</td></tr><tr><td rowspan=\"3\">a.</td><td>τες εμπυρίζουσι</td><td>τες εμπυρίζουσι</td></tr><tr><td>tόν ἀμπελώνα ἀλλὰ</td><td>tόν ἀμπελΟνα ἀλλὰ</td></tr><tr><td>xαλό διὰ τες</td><td>xαλό διὰ τες</td></tr><tr><td rowspan=\"3\">b.</td><td>τες εμπυρίζουσι</td><td>τεσέμπυρίσουβοι</td></tr><tr><td>tόν ἀμπελώνα ἀλλὰ</td><td>xδα ἀμπελοαα xδλὰ</td></tr><tr><td>xαλό διὰ τες</td><td>xxxxαλ αδιοτες</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.306,
+                0.49,
+                0.378
+            ],
+            "angle": 0,
+            "content": "Table 1: Synthetic transcription (human-generated) and recognition (system-generated) output with few (a) and many (b) errors. Finding the differences between the transcription and the recognition is easier for the three examples shown in red (b)."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.128,
+                0.393,
+                0.476,
+                0.436
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.445,
+                0.49,
+                0.489
+            ],
+            "angle": 0,
+            "content": "Figure 2: Image of the handwritten text 'των ποδων' - in the recognised text, the third character was misspelled, written as eta ('η') instead of 'v' (Table 1, 1st row)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.516,
+                0.49,
+                0.678
+            ],
+            "angle": 0,
+            "content": "(see Section 3), can explain the greater diversity of error rates (Fig. 1). The recognition quality, however, is related to the detection quality of erroneous output. To give an example, the detection of an erroneous recognition is easier when the recognition output is scrambled compared with a single incorrect character. This is shown in Table 1, where classifying recognised text (as erroneous/not) is easier for the human eye in (b) compared to that of (a). We hypothesise that this applies also to models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.679,
+                0.491,
+                0.92
+            ],
+            "angle": 0,
+            "content": "We argue that detecting erroneously recognised text can improve automatic error correction. Current approaches to post-correcting the often flawed recognised text, disregard the likelihood of a flawless recognition, assuming that the correct text can be simply copied (e.g., with an encoder-decoder-based correction). Consequently, they are faced with the possibility of introducing, instead of mitigating, errors that would be absent if post-correction hadn't been applied. Taking the case of a misspelled article (e.g., see Fig. 2), for instance, there is a variety of letters to choose from, for replacement, always accordingly to the declension and the case of the noun. Even if grammatical rules are set (e.g., in a rule-based system), an"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.343
+            ],
+            "angle": 0,
+            "content": "HTR system working at the character level is likely examining morphological aspects and often disregards semantic and pragmatic levels. A vowel replacement function addressing the morphological level, however, may omit important semantic aspects and relationships between sentence elements at the syntactic level and thus, misinterpret the case. For example, the line “λεντιον διεζωσεν εαυτόν” was HTRed as “λεντιον διεζωσενεαυτόν”. A neural corrector, then, which received as input the HTRed output, yielded “λεντιον διεζωσεν εαυτόν” (Pavlopoulos et al., 2023), correctly separating the two last words but incorrectly removing the accent. In this way and in this instance, post-correcting involves the risk of introducing errors by transforming probably correct output cases."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.354,
+                0.727,
+                0.369
+            ],
+            "angle": 0,
+            "content": "1.2 The focus of the study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.374,
+                0.885,
+                0.873
+            ],
+            "angle": 0,
+            "content": "This study quantifies the hazards of the above-mentioned adversarial post-correction, by focusing on the Byzantine Greek language. The development of this language entails a comprehensive examination of written language throughout the centuries under study. In essence, Byzantine Greek emerged as an academic style known as Atticised Greek, incorporating religious topics. As time progressed, elements of contemporary spoken Greek gradually became integrated into the language (Horrocks, 2014). As the results of our study also show, the combination of pre-trained data on Modern Greek, resembling what was considered as spoken language back then, and Ancient Greek, including the above-mentioned Atticised Greek, is beneficial for text classifiers detecting incorrect recognition output. Furthermore, by covering a period of more than seven centuries, the already challenging nature of recognition (Platanou et al., 2022) is increased, hence allowing a more diverse material for our investigation on post-correction. Errors in handwritten text recognition are often due to different hands (scribes) and scripts that evolved for many centuries. Printed text, on the other hand, exists for a much narrower period, there are no hands to cause similar errors,\\(^{1}\\) and no clitics. As we show in this work, lower recognition error rate means more difficult classification of erroneous lines. By contrast, recognised handwritten text is often hard to parse even for humans, which makes an easier classification challenge and a much more difficult"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.882,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>Fonts pose a similar problem, yet less challenging when compared to hands, due to their non-standardised form and high variation."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7819"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.086,
+                0.427,
+                0.099
+            ],
+            "angle": 0,
+            "content": "error-correction task, as shown in Table 1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.115,
+                0.303,
+                0.129
+            ],
+            "angle": 0,
+            "content": "1.3 The contributions"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.137,
+                0.49,
+                0.457
+            ],
+            "angle": 0,
+            "content": "We explore the extent to which classification to flawless/flawed texts can improve the post-correction of recognised output, by avoiding adversarial errors introduced when applied to flawless texts. We benchmark machine- and deep-learning-based text classifiers, showing that a Transformer that is pre-trained in modern Greek and then finetuned to detect erroneously recognised Byzantine Greek (10th-16th CE) text achieves an Average Precision of \\(95\\%\\). When the same Transformer is further pre-trained in ancient Greek prior to fine-tuning, the score increases to \\(97\\%\\). We quantify the benefits of employing our text classifier in the real world, by training an MT5 (Xue et al., 2021) encoder-decoder error-corrector, which altered \\(80\\%\\) of the flawless texts. If our best-performing classifier had been applied prior to correction, the percentage would have been decreased to \\(2\\%\\) with a trade-off of \\(2\\%\\) texts being incorrectly classified as flawless."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.46,
+                0.49,
+                0.749
+            ],
+            "angle": 0,
+            "content": "The better performance of a Transformer that is not pre-trained only in modern Greek can be explained by the fact that Byzantine Greek combines language elements from both, ancient and modern Greek. Since our study covers a broad historical period, we separated the evaluation dataset based on the century in order to investigate this hypothesis. Our results show that the detection of erroneous recognition depends on the chronology of the manuscript and that further pre-training in ancient Greek benefits the classification of lines from manuscripts in the following three centuries: 10thCE, 11thCE, and 13thCE. By following a similar segmentation-based evaluation method for the error rate level, instead of the century, we show that higher error rates make an easier classification goal. By contrast, the problem is harder for low error rates, when flawed and flawless texts look alike."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.764,
+                0.267,
+                0.779
+            ],
+            "angle": 0,
+            "content": "2 Related work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.791,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Our work focuses on the text classification of recognised handwritten text. In the absence of published studies for handwritten text, however, we summarise work focused on the recognition of printed text. Printed and handwritten materials differ, because the latter often yields scrambled recognition output (Table 1). This is probably what motivated Ströbel et al. (2022) in using the perplexity of lan"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.246
+            ],
+            "angle": 0,
+            "content": "guage models to detect the erroneous output in an unsupervised manner. On the other hand, although we consider that recognition errors are overall more for handwritten text compared to printed material, the quality of recognition can vary significantly for the former (Hodel et al., 2021), as is also shown in Fig. 1, and does not always come with a high error rate. Therefore, the detection of erroneously recognised handwritten text is also related to that for printed text, regarding their low error output."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.257,
+                0.859,
+                0.272
+            ],
+            "angle": 0,
+            "content": "Erroneous recognised printed text detection"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.277,
+                0.885,
+                0.437
+            ],
+            "angle": 0,
+            "content": "Initial work using rule-based approaches could not address all errors, and was followed by machine and deep learning approaches (Lyu et al., 2021). Standard approaches include support vector machines on top of n-gram-based features (Dannélls and Virk, 2021; Virk et al., 2021). Jatowt et al. (2019) used a gradient tree boosting binary classifier on top of various features, including character and word n-grams, part-of-speech, token frequency based on automatically created resources."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.438,
+                0.886,
+                0.63
+            ],
+            "angle": 0,
+            "content": "Amrhein and Clematide (2018) performed detailed analyses and experiments of error detection and post-correction with various statistical and neural machine-translation methods. In general, they found that the former perform better in error correction, while the latter models are better in error detection. However, they concluded that there is not a single method that works best on all datasets while the results are highly affected by the data a model is trained on. Furthermore, they suggested that post-correction of recognised printed text should focus more on the improvement of error detection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.632,
+                0.885,
+                0.791
+            ],
+            "angle": 0,
+            "content": "Nguyen et al. (2020) used BERT pre-trained on named-entity recognition for token classification in order to perform error detection. After substring tokenization they obtain Glove or Fasttext word embeddings, which are combined with segment and positional embeddings and given as input to BERT (Devlin et al., 2018). The hidden states are fed to a dense layer on top that classifies each token as erroneous or not. Although interesting, token-level detection is not suited for high-error settings."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.803,
+                0.689,
+                0.818
+            ],
+            "angle": 0,
+            "content": "The ICDAR challenge"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "ICDAR organised a competition focused on post-correcting text recognised from newspapers, shopping receipts, and other printed sources, including two subtasks: error detection, i.e., to detect the position and the length of the errors, and error correction (Chiron et al., 2017; Rigaud et al., 2019)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7820"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.345
+            ],
+            "angle": 0,
+            "content": "In 2017, the best-performing method for the error detection subtask employed probabilistic character error models and weighted finite-state transducers while in 2019, the best was a pre-trained multilingual BERT with convolutional and fully-connected layers on top that classified each sub-token as erroneous or not. Schaefer and Neudecker (2020) followed a similar classification scheme, but focused on characters and not sub-tokens. Then, by excluding lines with at least one classified character from post-correction, they showed that false alteration (i.e., of correctly OCR red lines) by an encoder-decoder model could be avoided. Our work shows that casting this problem as a text classification task can lead to high accuracy for handwritten text, although similar benefits may apply to printed text."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.355,
+                0.192,
+                0.369
+            ],
+            "angle": 0,
+            "content": "3 Data"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.38,
+                0.49,
+                0.621
+            ],
+            "angle": 0,
+            "content": "The written language of the manuscripts and papyri used in this study is Byzantine Greek.3 Within these texts, morphological categories such as the optative, the pluperfect, and the perfect have disappeared, while others such as the dative case have gradually decreased. Infinitives and participles are still there in the texts, as reminiscents of the classical tradition, encouraging one to treat the language as a unique variant, different from modern Greek. There are several spelling conventions that deviate from the older orthographic rules while the ancient punctuation signs are still in use, albeit not always with the same function. Therefore, following Classical and preceding Modern Greek, this language can be considered a combination of the two."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.622,
+                0.49,
+                0.846
+            ],
+            "angle": 0,
+            "content": "The transcriptions and the HTR output were provided by the organisers of the recent HTREC challenge, which regarded the automated correction of HTR errors in Byzantine Greek (Pavlopoulos et al., 2023). Selected images of handwritten text in respective manuscripts (10th to 16th CE) were transcribed by a human expert. The training dataset of the challenge consists of 1,875 lines of transcribed text by a human expert and by an HTR model (examples shown in Table 2). Transkribus (Kahle et al., 2017) was used by the organisers to produce the system transcriptions, trained on seven images one per century. Our investigation of the CER of the opted HTR model reveals"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.532,
+                0.094,
+                0.867,
+                0.242
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.558,
+                0.256,
+                0.835,
+                0.272
+            ],
+            "angle": 0,
+            "content": "Figure 3: Violinplot of CER per century"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.298,
+                0.885,
+                0.426
+            ],
+            "angle": 0,
+            "content": "that a diversity of errors is produced, from very few to many per line (Fig. 1). This fact makes this dataset optimal for our training purposes, because we need a diversity of errors to train our classifiers. Furthermore, the errors are not uniformly distributed across centuries (Fig. 3), motivating a time/CER-based segmentation for evaluation purposes (discussed in Section 4.2)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.428,
+                0.885,
+                0.589
+            ],
+            "angle": 0,
+            "content": "Our classification dataset consists of the 1,875 human-transcribed texts (lines), treated as the flawless negative (0) class, and the respective system-transcribed texts, treated as the flawed, positive (1) class. This resulted in total to an almost perfectly balanced dataset of 3,744 text lines, which we split to train, development and test subsets by using a 80/10/10 ratio. The average length of the transcription is 40.10 characters (6.83 words), and that of the HTR output is 37.09 characters (5.89 words)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.603,
+                0.738,
+                0.62
+            ],
+            "angle": 0,
+            "content": "4 Experimental analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.631,
+                0.884,
+                0.664
+            ],
+            "angle": 0,
+            "content": "We opted for machine and deep learning benchmarks, with different transfer learning settings."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.677,
+                0.757,
+                0.692
+            ],
+            "angle": 0,
+            "content": "Machine learning benchmarks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.699,
+                0.885,
+                0.797
+            ],
+            "angle": 0,
+            "content": "Machine learning methods include a variant of Naive Bayes and Support Vector Machines (NB-SVM), which is often used as a baseline for text classification (Wang and Manning, 2012). Also, we experimented with a Random Forest and a multi-layer perceptron (MLP) using character n-grams"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.855,
+                0.488,
+                0.88
+            ],
+            "angle": 0,
+            "content": "2The difference lies in using a sigmoid function (or so) on top of the text representation, instead of one per character."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.138,
+                0.881,
+                0.465,
+                0.893
+            ],
+            "angle": 0,
+            "content": "3Also known as Medieval Greek or simply Byzantine."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.138,
+                0.894,
+                0.411,
+                0.906
+            ],
+            "angle": 0,
+            "content": "4https://www.aicrowd.com/challenges/"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.114,
+                0.855,
+                0.488,
+                0.906
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.116,
+                0.906,
+                0.197,
+                0.917
+            ],
+            "angle": 0,
+            "content": "htrec-2022"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.807,
+                0.883,
+                0.833
+            ],
+            "angle": 0,
+            "content": "Fewer errors will harm class balance while similar errors will harm generalisation of the classifiers."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.833,
+                0.883,
+                0.857
+            ],
+            "angle": 0,
+            "content": "Six lines were excluded from the positive class, because a perfect transcription was achieved."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.857,
+                0.883,
+                0.893
+            ],
+            "angle": 0,
+            "content": "7All experiments were performed using Google's Colaboratory, using a Tesla T4 GPU card. Our large language models comprise 113 million parameters."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.894,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We used the implementation of KTRAIN with default parameters (Maiya, 2020)."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.509,
+                0.807,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7821"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.144,
+                0.081,
+                0.858,
+                0.206
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Transcription by expert</td><td>HTR output</td></tr><tr><td>èγγινόμενα πάθη μή σβενύντες ἀλλὰ τή εχλύσει (the born-in passions not extinguishing but the release)</td><td>èγγενομεναπαδημησεννωτες ἀλλατήε κλησει</td></tr><tr><td>του βίου του καῦ ́ εαυτός πολλα γένεσθουσυ υμχωρουν (of the life of themselves many happening forgive)</td><td>του β ou του καλεαυτός πολλαγινεσθουσι συγχωρόν</td></tr><tr><td>τες ἄμυριζουσι του ὄμελώνα ἀλλα καὶ ὅ διὰ τες (- set on fire the vineyard but and the due to the)</td><td>εμμυριζου σιμαμπελώνα ἀλλακαι ὅδξα</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.215,
+                0.884,
+                0.245
+            ],
+            "angle": 0,
+            "content": "Table 2: Human (on the left) and HTR-based (on the right) transcription (lines) from the 'Commentary on Isaiah' by Basil the Great (fol. 75r, Oxford, Bodleian Library MS. Barocci 102)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.269,
+                0.49,
+                0.302
+            ],
+            "angle": 0,
+            "content": "\\((n \\in 1,5)\\), term-frequency inverse document frequency, and default parameters otherwise."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.313,
+                0.332,
+                0.327
+            ],
+            "angle": 0,
+            "content": "Deep learning benchmarks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.333,
+                0.49,
+                0.542
+            ],
+            "angle": 0,
+            "content": "Deep learning methods comprise a recurrent neural network using Gated Recurrent Units (GRU) and trainable word embeddings of 300 dimensions, originally pre-trained on (modern) Greek (Joulin et al., 2017). Also, we experimented with the state-of-the-art in text classification, which is BERT (Devlin et al., 2018) and its variants (Li et al., 2022; Minaee et al., 2021). In specific, we also fine-tuned two BERT models pre-trained in Greek. The first (GreekBERT) is pre-trained in modern Greek (Koutsikakis et al., 2020) while the second extends the first by being further pre-trained in ancient Greek (Singh et al., 2021)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.553,
+                0.285,
+                0.567
+            ],
+            "angle": 0,
+            "content": "Evaluation measures"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.574,
+                0.49,
+                0.655
+            ],
+            "angle": 0,
+            "content": "We opted for the \\( F1 \\) score per class, the Area under the Receiver Operating Characteristic Curve (AUC), and Average Precision (AP) that is preferred in class-imbalanced settings. All evaluation measures were implemented in SCIKIT-LEARN.\\(^{12}\\)"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.665,
+                0.395,
+                0.68
+            ],
+            "angle": 0,
+            "content": "4.1 Text classification benchmark"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.686,
+                0.49,
+                0.798
+            ],
+            "angle": 0,
+            "content": "Table 3 shows the results of all our text classification benchmarks, including a random baseline (responding uniformly) that draws the lower performance limit. NBSVM is the worst machine learning model, followed by Forest and MLP. The latter achieves promising results, close to GRU (better in AP and \\( F1^{-} \\)). GRU, on the other hand, is clearly"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.511,
+                0.266,
+                0.891,
+                0.4
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>AP</td><td>AUC</td><td>F1+</td><td>F1-</td></tr><tr><td>Random</td><td>0.52</td><td>0.50</td><td>0.49</td><td>0.47</td></tr><tr><td>NBSVM</td><td>0.66</td><td>0.65</td><td>0.60</td><td>0.51</td></tr><tr><td>Forest</td><td>0.64</td><td>0.65</td><td>0.64</td><td>0.50</td></tr><tr><td>MLP</td><td>0.79</td><td>0.79</td><td>0.73</td><td>0.69</td></tr><tr><td>GRU</td><td>0.79</td><td>0.79</td><td>0.68</td><td>0.71</td></tr><tr><td>GreekBERT:M</td><td>0.95</td><td>0.94</td><td>0.88</td><td>0.88</td></tr><tr><td>GreekBERT:M+A</td><td>0.97</td><td>0.97</td><td>0.90</td><td>0.91</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.409,
+                0.885,
+                0.482
+            ],
+            "angle": 0,
+            "content": "Table 3: Evaluation with Average Precision (AP), AUC, \\( F1 \\) for the classification of HTRed texts to flawed (+) or flawless (-) transcription. Random (uniform) classification was used as a baseline. GreekBERT was pre-trained in modern (:M) or in modern and ancient Greek (:A)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.505,
+                0.884,
+                0.665
+            ],
+            "angle": 0,
+            "content": "outperformed in all metrics by GreekBERT. Two GreekBERT models were used, one pre-trained in modern Greek (:M) and one in modern and then in ancient Greek (:M+A). Byzantine Greek, as already discussed (Section 3), combines language elements found in both ancient and modern Greek. This is probably why GreekBERT:M+A, which was pretrained in modern and then in ancient Greek, was better compared to GreekBERT:M, which was pretrained only in modern Greek.[13]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.676,
+                0.74,
+                0.692
+            ],
+            "angle": 0,
+            "content": "4.2 Evaluation per century"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.697,
+                0.884,
+                0.826
+            ],
+            "angle": 0,
+            "content": "We separated the lines of the test data based on the century of the manuscript that they came from. For each century, then, we evaluated the two Transformers, GreekBERT:M and GreekBERT:M+A. As can be seen in Fig. 4,\\(^{14}\\) GreekBERT:M+A is better on the 10th (five percent units) and 11th century CE, with models performing similarly on the 12th and after the 13th century, when both models achieve a"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.833,
+                0.883,
+                0.869
+            ],
+            "angle": 0,
+            "content": "<sup>13</sup>We verified the consistency of this finding by repeating the experiments three times, reporting an average AUC of 0.97 over 0.94 (st.d. lower than 0.05)."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.87,
+                0.882,
+                0.918
+            ],
+            "angle": 0,
+            "content": "For statistical significance, we sampled twenty lines per century, repeating the sampling process ten times. We focus on Average Precision, but similar findings were observed with the other metrics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.509,
+                0.833,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.806,
+                0.461,
+                0.856
+            ],
+            "angle": 0,
+            "content": "\\(^{9}\\)https://scikit-learn.org/stable/modules/ generated/sklearn.neural_network \\({ }^{10}\\)https://fasttext.cc/docs/en/crawl-vectors. htm1"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.857,
+                0.488,
+                0.904
+            ],
+            "angle": 0,
+            "content": "11 Ancient Greek texts, which can be used for pre-training, are much fewer in number compared to publicly available modern Greek data. Hence, the option of pre-training directly in ancient Greek was not considered viable."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.133,
+                0.905,
+                0.388,
+                0.918
+            ],
+            "angle": 0,
+            "content": "12https://scikit-learn.org/stable/"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.114,
+                0.806,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7822"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.175,
+                0.101,
+                0.812,
+                0.25
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.26,
+                0.885,
+                0.292
+            ],
+            "angle": 0,
+            "content": "Figure 4: Average Precision per c. CE of GreekBERT:M, pretrained on modern Greek (solid), and GreekBERT:M+A, further pre-trained on ancient Greek (dashed) - error bands show the \\(95\\%\\) bootstrap confidence intervals"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.314,
+                0.49,
+                0.475
+            ],
+            "angle": 0,
+            "content": "higher score. On the 13th century, the performance of both models drops, but considerably more for GreekBERT:M. The 14th, 15th and 16th century were more challenging in HTR (Platanou et al., 2022), which could explain the better performance of the classifiers. That is, more errors in the recognised output could potentially help the classifier distinguish between the two classes (correct/not).<sup>15</sup> Based on the previous observations, we report three findings:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.118,
+                0.487,
+                0.489,
+                0.52
+            ],
+            "angle": 0,
+            "content": "F1. The classification performance depends on the chronology of the manuscript."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.118,
+                0.531,
+                0.49,
+                0.563
+            ],
+            "angle": 0,
+            "content": "F2. Older manuscripts are better handled by further pre-training on ancient Greek."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.118,
+                0.574,
+                0.49,
+                0.621
+            ],
+            "angle": 0,
+            "content": "F3. Lines from recent manuscripts pose an easier classification challenge to both GreekBERT models."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.118,
+                0.487,
+                0.49,
+                0.621
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.634,
+                0.367,
+                0.649
+            ],
+            "angle": 0,
+            "content": "4.3 Evaluation per error zone"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.655,
+                0.49,
+                0.766
+            ],
+            "angle": 0,
+            "content": "Our dataset is balanced regarding whether the recognition output is erroneous (positive class) or not, yet it remains imbalanced regarding the CER level (see Fig. 1). In order to assess the two Transformers across different input types, we rolled a window of ten units, from low to high CER values, and we assessed the models per window (Fig. 5).<sup>16</sup>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.768,
+                0.49,
+                0.848
+            ],
+            "angle": 0,
+            "content": "High-CER lines (i.e., lines with CER of thirty or more) appear to set an easy classification task to both models. Performance increases as the CER increases. This is reasonable considering that lines with a high CER comprise many errors and are"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.314,
+                0.885,
+                0.378
+            ],
+            "angle": 0,
+            "content": "hard to parse even by humans. This characteristic, however, makes them easier to distinguish from flawless lines, which comprise no errors from the recognition."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.38,
+                0.887,
+                0.525
+            ],
+            "angle": 0,
+            "content": "Low-CER lines (i.e., lines with CER of zero to twenty) lead to a different picture, since Greek-BERT:M+A is consistently better, overall, than GreekBERT:M. Low CER means small differences between the two texts (flawless-flawed), which makes the classification task more difficult. In this zone, further pre-training on ancient Greek had a bigger impact, compared to zones with a higher error rate."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.541,
+                0.638,
+                0.557
+            ],
+            "angle": 0,
+            "content": "5 Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.57,
+                0.886,
+                0.634
+            ],
+            "angle": 0,
+            "content": "In this section, first, we assess the possibly adversarial nature of post-correction. Then, we describe a synthetic error detection application, followed by a century-based error analysis we performed."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.648,
+                0.757,
+                0.664
+            ],
+            "angle": 0,
+            "content": "5.1 Substantiality assessment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.672,
+                0.885,
+                0.88
+            ],
+            "angle": 0,
+            "content": "Error-correction systems may introduce errors, even when applied to flawless HTRed texts. To assess the substantiality of this claim, we experimented with a neural encoder-decoder postcorrector. In specific, we employed a multilingual variant of the T5 Text-to-Text Transfer Transformer (Raffel et al., 2019), called MT5 (Xue et al., 2020), which was pre-trained on a dataset covering 101 languages.[17] MT5 has been used for Grammatical Error Correction in modern Greek and sets a strong baseline for HTR Error Correction (Korre and Pavlopoulos, 2022). We fine-tuned this model on the 1,875 training instances of the HTREC"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.893,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "17The Greek language was on the 20th position with 43 billion tokens extracted from 42 million pages"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.856,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>15</sup>It remains unclear to the authors whether more hands were present during these centuries, explaining the performance of the HTR and the classification models, or if this was due to another reason; e.g., writing became more cursive over time. <sup>16</sup>We sampled thirty lines, repeating sampling ten times."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7823"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.175,
+                0.101,
+                0.812,
+                0.252
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.26,
+                0.885,
+                0.304
+            ],
+            "angle": 0,
+            "content": "Figure 5: Average Precision of GreekBERT:M, pretrained on modern Greek (blue, solid), and GreekBERT:M+A, further pre-trained on ancient Greek (orange, dashed), using a rolling window over CER (left edge shown horizontally) - error bands depict \\(95\\%\\) bootstrap confidence intervals"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.329,
+                0.491,
+                0.668
+            ],
+            "angle": 0,
+            "content": "challenge, by encoding the system- and decoding the human-transcriptions.\\(^{18}\\) This model is a neural error correction baseline that reduced the CER in 19 out of the 180 HTRed texts in the official HTREC2022 evaluation set. We applied this corrector on the respective 180 transcriptions of these data (the ground truth evaluation instances in HTREC2022) and found that 144 \\((80\\%)\\) were mistakenly altered by our MT5 corrector. For example, the line 'αν τους μηδους οινες έτε αργυρτω παρα' was altered to 'αν τους μηδους οινες έτε αργυρτω παρα' (word division, 4th word). Similarly, 'χευσης μη φονευσης μη κλε' was altered to 'χλευσεως μη φονευσεως μη κλεπτ'. Most of these adversarial mistakes (all but the top three of Table 4) would have been avoided if our GreekBERT:M+A classifier had been applied prior to the application of this neural corrector. Conversely, only four (i.e., the rest in Table 4) would have been mistakenly found as correct from the system-transcribed instances."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.679,
+                0.484,
+                0.695
+            ],
+            "angle": 0,
+            "content": "5.2 A synthetic application of error detection"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.7,
+                0.49,
+                0.86
+            ],
+            "angle": 0,
+            "content": "We experimented with the HTREC2022 evaluation dataset, which was used to assess the participating systems. This dataset comprises 180 lines that were transcribed from nine images. Out of these lines, one had zero CER achieved by the HTR model (perfect transcription), which was correctly classified by our GreekBERT:M+A model. We then performed two analyses, first by applying our model on the flawless human transcriptions (as if a transcription was the HTR output) and then by applying"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.329,
+                0.885,
+                0.442
+            ],
+            "angle": 0,
+            "content": "it on the flawed system recognitions (the same lines were used). The human transcriptions were correctly classified as flawless, all except three \\((1.7\\%)\\) error rate), which are shown in the top three rows of Table 4. The system recognitions were also correctly classified as flawed, all except four \\((2.3\\%)\\) error rate), shown in the last four rows of Table 4."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.465,
+                0.671,
+                0.48
+            ],
+            "angle": 0,
+            "content": "5.3 Error analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.492,
+                0.884,
+                0.589
+            ],
+            "angle": 0,
+            "content": "An error detection by experts in Table 4,\\(^{19}\\) showed that no clue (indicating an error) was found for the first line (except from the rare ‘φιλαχόλουνός’ with double accent). As is shown in Figure 6 (a), which depicts the attention on the subwords of this line, this was indeed the network’s focus."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.593,
+                0.884,
+                0.753
+            ],
+            "angle": 0,
+            "content": "On the second line, word 'iσμαηλήτι' was considered as the outcome of misspelling of the word 'iσμαηλιτη', matching the algorithm's decision. In this case, it was the ending of the word that attracted attention (Fig. 6), but the network focused more on the article 'των' preceding the word rather than on the word itself. The article is in plural form, not matching a word in singular. Apparently, 'iσμαηλήτι' is not a misspelling, but part of the word 'iσμαηλιτιχων' that was corrupted."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.758,
+                0.885,
+                0.87
+            ],
+            "angle": 0,
+            "content": "In the last four lines, that were classified as correct, several flaws were raised, indicating a failure of the algorithm. For example, word divisions, such as in the case of \\(\\omega \\sigma x\\alpha\\) instead of \\(\\omega \\varsigma \\chi \\alpha \\iota\\); added characters, such as in the case of \\(\\varepsilon \\dot{\\upsilon}\\upsilon \\chi \\eta \\varsigma\\) instead of \\(\\varepsilon \\dot{\\upsilon}\\chi \\eta \\varsigma\\); character replacements, such as in the case of \\(\\zeta \\kappa \\tau \\varepsilon \\iota \\alpha \\nu\\) instead of \\(\\iota \\kappa \\varepsilon \\tau \\varepsilon \\iota \\alpha \\nu\\)."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.869,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>18</sup>We used the SimpleT5 implementation (https://github.com/Shivanandroy/simpleT5) with 100 epochs, early stopping (patience of 2 epochs), 15 tokens max. length, batch size of 4, and best loss found at the 12th epoch (2.64)."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.893,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(^{19}\\)No access was given to the manuscript or the human transcription, for fair comparison with the setting of the algorithm."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7824"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.127,
+                0.082,
+                0.871,
+                0.33
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>TRANSCRIPTIONS AND HTR OUTPUT</td><td>GREEKBERT:M+A</td><td>GT</td></tr><tr><td>èγώ δ ἀεί πως φιλαχόλουθός εἰμι(I am always readily following)</td><td>F</td><td>N</td></tr><tr><td>λα και γέδεων εκ τών σχύλων τών ομαηλητι(- and Gedeon from the dogs the - )</td><td>F</td><td>N</td></tr><tr><td>δραμων εἰς και γόνυπετη(having run to and falling on the knee)</td><td>F</td><td>N</td></tr><tr><td>τιθεις τας χειρας επαυτα(having put the hands on them)</td><td>F</td><td>N</td></tr><tr><td>ζετείαν χέδεσποτα, και ευμενώς Ἁκουσου τών τής προς εύυχης(- - - and favourably listen to the of the prayer)</td><td>N</td><td>F</td></tr><tr><td>èγενόμην icos. αραή τής μεό τιεγεζήθης ἐλειεμοήτιριος(became - - - - - )</td><td>N</td><td>F</td></tr><tr><td>ωσααι τής συνεχους διδασκαλιας τους βα(- of the continuous instruction the - )</td><td>N</td><td>F</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.339,
+                0.883,
+                0.369
+            ],
+            "angle": 0,
+            "content": "Table 4: HTRed lines and transcriptions that GreekBERT:M+A mistakenly predicted as flawed (F) or not (N) - the ground truth (GT) on the right shows whether there was an error or not"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.395,
+                0.236,
+                0.409
+            ],
+            "angle": 0,
+            "content": "Error patterns"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.414,
+                0.49,
+                0.542
+            ],
+            "angle": 0,
+            "content": "Next, we proceed with an analysis of the recognition errors, using data from three centuries. We opt for 16th CE, the century on which both our models perform considerably well (Fig. 4). Next, we focus on the 14th CE, going back two centuries, but still observing good performance by both models. Last is shown the 13th CE, where GreekBERT:M performs much worse than GreekBERT:M+A."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.554,
+                0.49,
+                0.797
+            ],
+            "angle": 0,
+            "content": "16th CE This is the data group with the highest average CER (see Figure 3). What seems to be challenging in this data group is the fact that there is a variety of punctuation marks which are sometimes read by the model as letters because of their shape. An illustrative example of this case is the word 'πλοῦς' which is read by the model as 'πλούου'. This is due to a comma following the word 'πλοῦς', which is very similar to the last letter of the word (ς). There are also many cases of ligatures in which the model fails to recognise both letters, especially the one written on the top of the ligature, such as the letter 'τ' in the combination 'τρ' as one can see in the following instance; 'ω πατρις' read as 'ως ωπαρις'."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.807,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "14th CE The data from the 14th century show the second highest CER on average (see Figure 3). This is partly because of a great number of ligatures appearing in this data group. A common ligature among this group is the one which represents the group of the letters \\(\\varepsilon\\) and \\(\\rho\\). In this ligature it happens that the letter \\(\\varepsilon\\) is written above the letter"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.394,
+                0.885,
+                0.733
+            ],
+            "angle": 0,
+            "content": "\\(\\varphi^{\\prime}\\). The way the model interprets this structure can be seen in the following case; \\(\\varphi \\acute{\\varepsilon} \\rho^{\\prime} \\dot{\\varepsilon} \\pi^{\\prime} \\omega \\tau \\eta \\nu\\) \\(\\iota \\omega\\), where \\(\\varphi \\acute{\\varepsilon} \\rho^{\\prime}\\) consists of one letter \\((\\varphi)\\) and one ligature \\((\\varepsilon \\rho)\\) is read by the model as \\(\\varphi \\varepsilon \\pi \\alpha \\omega \\tau \\eta \\nu\\) \\(\\iota \\varsigma \\omega\\). The model fails reading correctly the letter combination and returns only the letter \\(\\varphi^{\\prime}\\) and one of the two letters involved in the \\(\\varepsilon \\rho^{\\prime}\\) combination \\((\\rho)\\), which is actually the one written in the position next to the letter \\(\\varphi^{\\prime}\\). The same instance shows clearly that there is a tendency towards adding characters because of different drawing representations of characters found in this data group. The letter \\(\\omega\\) can be considered such a case. In the previous example, we can see that the word \\(\\iota \\omega\\) is read by the model as \\(\\iota \\varsigma \\omega\\). This is because, in this instance, the letter \\(\\omega\\) has the shape of a laid eight. The same letter has the usual shape in other cases in the text of this century, which was quite confusing for the model and led the model to decide that it should be a combination of both \\(\\sigma^{\\prime}\\) and \\(\\omega^{\\prime}\\), since the former is quite close to the half laid-eight shape."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.743,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "13th CE The data group that is easiest to read for the model seems to be the 13th century one. In this group, errors appear mainly because of special letter shapes and abbreviations. Such a letter shape is that of the letter \\(\\varepsilon\\) which is written as the letter \\(\\delta\\) with an added line pointing upwards. A characteristic example of the failure of the recognition task is this one; \\(\\tilde{\\varepsilon}\\omega\\varsigma\\) is read as \\(\\delta\\omega\\varsigma\\). Here the model has clearly classified the \\(\\tilde{\\varepsilon}\\) as a \\(\\delta\\). Regarding the abbreviations representing character groups, one can find among others the following"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7825"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.117,
+                0.084,
+                0.88,
+                0.205
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.407,
+                0.212,
+                0.593,
+                0.225
+            ],
+            "angle": 0,
+            "content": "(c) \\(\\delta \\rho \\alpha \\mu \\omega \\nu \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\varepsilon \\vareeta\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.24,
+                0.882,
+                0.27
+            ],
+            "angle": 0,
+            "content": "Figure 6: The attention of GreekBERT:M+A (darker is more) for the three top (missclassified) lines of Table 4 - the symbol ## is used to concatenate a subword with the preceding one during tokenisation (\\(\\delta \\rho \\alpha \\mu \\omega \\nu\\) is \\(\\delta \\rho \\alpha \\# \\mu \\omega \\nu\\))"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.295,
+                0.489,
+                0.359
+            ],
+            "angle": 0,
+            "content": "one; \\(\\alpha v\\) . In the case of ouxioov, where the last two letters are represented by the abbreviation, the model reads ouxoX, where \\(\\omega\\) has replaced the abbreviation because of similarity in shape."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.373,
+                0.437,
+                0.388
+            ],
+            "angle": 0,
+            "content": "5.4 Language and script generalisation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.395,
+                0.49,
+                0.572
+            ],
+            "angle": 0,
+            "content": "Our work can potentially inspire analogous pretraining in other languages with a long documented history. To provide an example: Old, Middle, and Modern English and Japanese (Lyovin, 1997, pp. 2) can qualify for pre-training on both Old and Modern data, as in the Greek language in our case. The scripts our results are based on, however, are very common both in Greek and Latin manuscripts while common handwritten character shapes, e.g., epsilon (E) (Chambers and Chambers, 1891, pp. 703), may illuminate error patterns in Latin scripts."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.587,
+                0.247,
+                0.601
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.614,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "This work showed that detecting erroneous recognised lines improves the accuracy of handwritten Byzantine text recognition, aiming to assist the disciplines struggling with vast corpora and little man power. By contrast to printed text, errors in handwritten text vary greatly, some of which are easier and others difficult to catch. Our experiments showed that two BERT models, one pre-trained in modern and the other pre-trained in modern then in ancient Greek perform well for the task of erroneous handwritten recognised text classification. The latter performs better, especially for manuscripts from older centuries and for lines with higher recognition error rate. By using a neural text-to-text Transformer for error correction along with our best performing classifier, we found that adversarial behaviour is not uncommon in the neural corrector's output which could have been significantly limited if a BERT classifier, pre-trained and"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.295,
+                0.805,
+                0.31
+            ],
+            "angle": 0,
+            "content": "fine-tuned, had classified the input first."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.311,
+                0.884,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Directions of future work comprise the development of a larger dataset, which, although not easily developed in our domain of focus, it could potentially allow us to draw more conclusions and lead to further findings. Furthermore, we plan to extend our dataset with calibrated evaluation subsets, in order to gain more perspective on the performance of systems for specific error types, including ones that occur naturally and are not meant to be corrected."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.469,
+                0.615,
+                0.483
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.494,
+                0.884,
+                0.591
+            ],
+            "angle": 0,
+            "content": "- Our study is focused on Byzantine Greek, from the 10th to the 16th century. Our hypothesis, that text classification could assist the post correction of HTRd output, should be studied for more languages and covering a longer period to draw general conclusions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.602,
+                0.885,
+                0.795
+            ],
+            "angle": 0,
+            "content": "- By contrast to OCR, where low CER values are reported, HTRd output can significantly differ from the human transcription. For example, correct text may be misplaced (e.g., due to mistaken word division) or by the introduction of an extra character (e.g., the last word of the fifth line in Table 4). Hence, the creation of ground truth at the character level is very challenging in this context, hindering experimentation at the token or the character level, as in OCR (Schulz and Kuhn, 2017; Schaefer and Neudecker, 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.807,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "- Sensitivity analysis has not been performed. Although we repeated our experiments to verify the validity of our findings regarding GreekBERT, we have not experimented with cross validation over the centuries, as in (Platanou et al., 2022). Training, however, an algorithm on texts from specific centuries, then as"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.532,
+                0.494,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7826"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.086,
+                0.489,
+                0.117
+            ],
+            "angle": 0,
+            "content": "sessing on ones from previous or future ones, could potentially reveal interesting patterns."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.128,
+                0.49,
+                0.336
+            ],
+            "angle": 0,
+            "content": "- Diplomatic transcriptions may lead to naturally-occurring errors (i.e., ones the editors would want to keep, to stay as close to the original as possible), which our proposed classifiers could and probably would classify as incorrect recognition output. This limitation could be bypassed if the proposed systems are used complementary with an expert, not trying to substitute one. For example, the expert could view only lines the classifier labelled as incorrect, bypassing or acknowledging errors which were meant to be there."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.376,
+                0.214,
+                0.391
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.399,
+                0.49,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Chantal Amrhein and Simon Clematide. 2018. SupervisedOCR error detection and correction using statistical and neural machine translation methods. Journal for Language Technology and Computational Linguistics (JLCL), 33(1):49-76."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.475,
+                0.489,
+                0.515
+            ],
+            "angle": 0,
+            "content": "William Chambers and Robert Chambers. 1891. Chambers' Encyclopaedia: A Dictionary of Universal Knowledge, volume 8. W. & R. Chambers."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.525,
+                0.49,
+                0.605
+            ],
+            "angle": 0,
+            "content": "Guillaume Chiron, Antoine Doucet, Mickaël Coustaty, and Jean-Philippe Moreux. 2017. Icdar2017 competition on post-ocr text correction. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 1, pages 1423-1428, Kyoto, Japan."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.614,
+                0.49,
+                0.668
+            ],
+            "angle": 0,
+            "content": "Dana Dannélls and Shafqat Virk. 2021. A supervised machine learning approach for post-ocr error detection for historical text. In Swedish Language Technology Conference and NLP4CALL, pages 13-20."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.677,
+                0.489,
+                0.73
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.74,
+                0.49,
+                0.807
+            ],
+            "angle": 0,
+            "content": "Tobias Mathias Hodel, David Selim Schoch, Christa Schneider, and Jake Purcell. 2021. General models for handwritten text recognition: feasibility and state-of-the-art. german current as an example. Journal of open humanities data, 7(13):1-10."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.816,
+                0.489,
+                0.843
+            ],
+            "angle": 0,
+            "content": "Geoffrey Horrocks. 2014. Greek: A History of the Language and its Speakers. John Wiley & Sons."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.853,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Adam Jatowt, Mickaël Coustaty, Nhu-Van Nguyen, Antoine Doucet, et al. 2019. Post-ocr error detection by generating plausible candidates. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 876-881, Sydney, Australia."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.399,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.513,
+                0.086,
+                0.885,
+                0.179
+            ],
+            "angle": 0,
+            "content": "Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.187,
+                0.885,
+                0.266
+            ],
+            "angle": 0,
+            "content": "Philip Kahle, Sebastian Colutto, Günter Hackl, and Günter Mühlberger. 2017. Transkribus - a service platform for transcription, recognition and retrieval of historical documents. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 04, pages 19-24."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.274,
+                0.885,
+                0.352
+            ],
+            "angle": 0,
+            "content": "Katerina Korre and John Pavlopoulos. 2022. Enriching grammatical error correction resources for Modern Greek. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4984-4991, Marseille, France. European Language Resources Association."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.362,
+                0.885,
+                0.415
+            ],
+            "angle": 0,
+            "content": "John Koutsikakis, Ilias Chalkidis, Prodromos Malakasiotis, and Ion Androutsopoulos. 2020. Greek-bert: The greeks visiting sesame street. In 11th Hellenic Conference on Artificial Intelligence, pages 110-117."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.423,
+                0.885,
+                0.489
+            ],
+            "angle": 0,
+            "content": "Qian Li, Hao Peng, Jianxin Li, Congying Xia, Renyu Yang, Lichao Sun, Philip S Yu, and Lifang He. 2022. A survey on text classification: From traditional to deep learning. ACM Transactions on Intelligent Systems and Technology (TIST), 13(2):1-41."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.498,
+                0.885,
+                0.525
+            ],
+            "angle": 0,
+            "content": "Anatole Lyovin. 1997. An Introduction to the Languages of the World. Oxford University Press, USA."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.533,
+                0.885,
+                0.587
+            ],
+            "angle": 0,
+            "content": "Lijun Lyu, Maria Koutraki, Martin Krickl, and Besnik Fetahu. 2021. NeuralOCR post-hoc correction of historical corpora. Transactions of the Association for Computational Linguistics, pages 479-493."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.594,
+                0.885,
+                0.634
+            ],
+            "angle": 0,
+            "content": "Arun S. Maiya. 2020. ktrain: A low-code library for augmented machine learning. arXiv preprint arXiv:2004.10703."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.643,
+                0.885,
+                0.683
+            ],
+            "angle": 0,
+            "content": "Barbara McGillivray, Thierry Poibeaun, and Pablo Ruiz Fabo. 2020. Digital humanities and natural language processing: \"je t'aime... moi non plus\"."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.691,
+                0.885,
+                0.757
+            ],
+            "angle": 0,
+            "content": "Shervin Minaee, Nal Kalchbrenner, Erik Cambria, Narjes Nikzad, Meysam Chenaghlu, and Jianfeng Gao. 2021. Deep learning-based text classification: a comprehensive review. ACM Computing Surveys (CSUR), 54(3):1-40."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.765,
+                0.885,
+                0.843
+            ],
+            "angle": 0,
+            "content": "Thi Tuyet Hai Nguyen, Adam Jatowt, Nhu-Van Nguyen, Mickael Coustaty, and Antoine Doucet. 2020. Neural machine translation with bert for post-ocr error detection and correction. In Proceedings of the ACM/IEEE joint conference on digital libraries in 2020, pages 333-336."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.853,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": "John Pavlopoulos, Vasiliki Kougia, Paraskevi Platanou, Stepan Shabalin, Konstantina Liagkou, Emmanouil Papadatos, Holger Essler, Jean-Baptiste Camps, and Franz Fischer. 2023. Error correcting htr'ed byzantine text."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7827"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Paraskevi Platanou, John Pavlopoulos, and Georgios Papaioannou. 2022. Handwritten paleographic greek text recognition: A century-based approach. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6585-6589."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.165,
+                0.488,
+                0.23
+            ],
+            "angle": 0,
+            "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.242,
+                0.488,
+                0.308
+            ],
+            "angle": 0,
+            "content": "Christophe Rigaud, Antoine Doucet, Mickaël Coustaty, and Jean-Philippe Moreux. 2019. Icdar 2019 competition on post-ocr text correction. In 2019 international conference on document analysis and recognition (ICDAR), pages 1588-1593, Sydney, Australia."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.32,
+                0.488,
+                0.398
+            ],
+            "angle": 0,
+            "content": "Robin Schaefer and Clemens Neudecker. 2020. A two-step approach for automaticOCR post-correction. In Proceedings of the The 4th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 52-57, held online."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.411,
+                0.488,
+                0.476
+            ],
+            "angle": 0,
+            "content": "Sarah Schulz and Jonas Kuhn. 2017. Multi-modular domain-tailoredOCR post-correction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2716-2726, Copenhagen, Denmark."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.489,
+                0.488,
+                0.58
+            ],
+            "angle": 0,
+            "content": "Pranaydeep Singh, Gorik Rutten, and Els Lefever. 2021. A pilot study for bert language modelling and morphological analysis for ancient and medieval greek. In The 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2021)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.593,
+                0.488,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Phillip Benjamin Ströbel, Simon Clematide, Martin Volk, Raphael Schwitter, Tobias Hodel, and David Schoch. 2022. Evaluation of htr models without ground truth material. arXiv preprint arXiv:2201.06170."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.671,
+                0.488,
+                0.75
+            ],
+            "angle": 0,
+            "content": "Shafqat Mumtaz Virk, Dana Dannélls, and Azam Sheikh Muhammad. 2021. A novel machine learning based approach for post-ocr error detection. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1463–1470, Varna, Bulgaria."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.762,
+                0.488,
+                0.841
+            ],
+            "angle": 0,
+            "content": "Sida Wang and Christopher Manning. 2012. Baselines and bigrams: Simple, good sentiment and topic classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 90-94, Jeju Island, Korea. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.853,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2020. mt5: A massively multilingual pre-trained text-to-text transformer. arXiv preprint arXiv:2010.11934."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.884,
+                0.191
+            ],
+            "angle": 0,
+            "content": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "7828"
+        }
+    ]
+]

2023/Detecting Erroneously Recognized Handwritten Byzantine Text/e23fb75f-6ceb-4c26-9972-d5c23aa51bcb_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6751e64f403b4d5e3632e22bbd41aaa2232baa250bee11f449c25dd60363501
+size 437113

2023/Detecting Erroneously Recognized Handwritten Byzantine Text/full.md

@@ -0,0 +1,275 @@
+# Detecting Erroneous Handwritten Byzantine Text Recognition
+
+John Pavlopoulos
+
+Department of Informatics
+
+Athens University of Economics and Business
+
+Greece
+
+annis@aueb.gr
+
+Vasiliki Kougia
+
+Faculty of Computer Science
+
+University of Vienna, Austria
+
+UniVie Doctoral School Computer Science
+
+vasiliki.kougia@univie.ac.at
+
+Paraskevi Platanou
+
+Department of Philology, School of Philosophy
+
+University of Athens, Greece
+
+pplatanou@phil.uoa.gr
+
+Holger Essler
+
+Department of Humanities
+
+Ca'Foscari University of Venice, Italy
+
+holger.essler@unive.it
+
+# Abstract
+
+Handwritten text recognition (HTR) yields textual output that comprises errors, which are considerably more compared to that of recognised printed (OCRed) text. Post-correcting methods can eliminate such errors but may also introduce errors. In this study, we investigate the issues arising from this reality in Byzantine Greek. We investigate the properties of the texts that lead post-correction systems to this adversarial behaviour and we experiment with text classification systems that learn to detect incorrect recognition output. A large masked language model, pre-trained in modern and fine-tuned in Byzantine Greek, achieves an Average Precision score of  $95\%$ . The score improves to  $97\%$  when using a model that is pretrained in modern and then in ancient Greek, the two language forms Byzantine Greek combines elements from. A century-based analysis shows that the advantage of the classifier that is further-pre-trained in ancient Greek concerns texts of older centuries. The application of this classifier before a neural post-corrector on HTRed text reduced significantly the post-correction mistakes.
+
+# 1 Introduction
+
+Handwritten Text Recognition (HTR) concerns the automated transcription of a handwritten text, extracting its computerised form from an image. Applying HTR on an old manuscript can lead to a high character error rate (CER), often due to the lack of training data (Pavlopoulos et al., 2023). In the real world, this erroneous HTR output is post-corrected by experts, who get involved, typically, in a tedious and time-consuming task. Delaying the delivery of manually curated data, hinders both: the preservation of digitised manuscripts that have not yet been
+
+![](images/1a6900ee3055eb9c40c6505bf00db057e7a0f2d9f507a3f97878c5d891264007.jpg)  
+Figure 1: Number of HTRed-lines according to their CER as presented by Pavlopoulos et al. (2023)
+
+transcribed; and the progress of HTR for historical manuscripts, which struggles with low resources. This is why researchers from Digital Humanities have started encouraging the development of Natural Language Processing (NLP) solutions, in an attempt to assist the experts toward this goal, since improving the accuracy of recognition will be exciting for the disciplines struggling with vast corpora and little man power (McGillivray et al., 2020).
+
+# 1.1 Background
+
+Although detecting erroneous recognised output is not a novel task (Chiron et al., 2017; Amrhein and Clematide, 2018; Rigaud et al., 2019; Jatowt et al., 2019; Nguyen et al., 2020; Schaefer and Neudecker, 2020), existing published studies focus on printed material (i.e., OCRed). Handwritten text, on the other hand, increases the challenge in recognition compared to printed material and hence, leads to more errors. The different hands, the greater time period, morphology and spelling
+
+<table><tr><td></td><td>Transcription</td><td>Recognition</td></tr><tr><td>a.</td><td>tov ποδων</td><td>τωη ποδων</td></tr><tr><td>b.</td><td>tov ποδων</td><td>αχοασδδεδ</td></tr><tr><td rowspan="2">a.</td><td>έγγινουμένα πάθη</td><td>έγγινουμένα πάθη</td></tr><tr><td>μή σβενύντε</td><td>μή σβενύντε</td></tr><tr><td rowspan="2">b.</td><td>έγγινουμένα πάθη</td><td>έγγινουμένδπάθη</td></tr><tr><td>μή σβενύντε</td><td>μαε σβανησοτε</td></tr><tr><td rowspan="3">a.</td><td>τες εμπυρίζουσι</td><td>τες εμπυρίζουσι</td></tr><tr><td>tόν ἀμπελώνα ἀλλὰ</td><td>tόν ἀμπελΟνα ἀλλὰ</td></tr><tr><td>xαλό διὰ τες</td><td>xαλό διὰ τες</td></tr><tr><td rowspan="3">b.</td><td>τες εμπυρίζουσι</td><td>τεσέμπυρίσουβοι</td></tr><tr><td>tόν ἀμπελώνα ἀλλὰ</td><td>xδα ἀμπελοαα xδλὰ</td></tr><tr><td>xαλό διὰ τες</td><td>xxxxαλ αδιοτες</td></tr></table>
+
+Table 1: Synthetic transcription (human-generated) and recognition (system-generated) output with few (a) and many (b) errors. Finding the differences between the transcription and the recognition is easier for the three examples shown in red (b).
+
+![](images/a4a3b7d2d163bed7f10abba2e1ed69e1a9eed264f270f5bc58f9693cb4b0948f.jpg)  
+Figure 2: Image of the handwritten text 'των ποδων' - in the recognised text, the third character was misspelled, written as eta ('η') instead of 'v' (Table 1, 1st row)
+
+(see Section 3), can explain the greater diversity of error rates (Fig. 1). The recognition quality, however, is related to the detection quality of erroneous output. To give an example, the detection of an erroneous recognition is easier when the recognition output is scrambled compared with a single incorrect character. This is shown in Table 1, where classifying recognised text (as erroneous/not) is easier for the human eye in (b) compared to that of (a). We hypothesise that this applies also to models.
+
+We argue that detecting erroneously recognised text can improve automatic error correction. Current approaches to post-correcting the often flawed recognised text, disregard the likelihood of a flawless recognition, assuming that the correct text can be simply copied (e.g., with an encoder-decoder-based correction). Consequently, they are faced with the possibility of introducing, instead of mitigating, errors that would be absent if post-correction hadn't been applied. Taking the case of a misspelled article (e.g., see Fig. 2), for instance, there is a variety of letters to choose from, for replacement, always accordingly to the declension and the case of the noun. Even if grammatical rules are set (e.g., in a rule-based system), an
+
+HTR system working at the character level is likely examining morphological aspects and often disregards semantic and pragmatic levels. A vowel replacement function addressing the morphological level, however, may omit important semantic aspects and relationships between sentence elements at the syntactic level and thus, misinterpret the case. For example, the line “λεντιον διεζωσεν εαυτόν” was HTRed as “λεντιον διεζωσενεαυτόν”. A neural corrector, then, which received as input the HTRed output, yielded “λεντιον διεζωσεν εαυτόν” (Pavlopoulos et al., 2023), correctly separating the two last words but incorrectly removing the accent. In this way and in this instance, post-correcting involves the risk of introducing errors by transforming probably correct output cases.
+
+# 1.2 The focus of the study
+
+This study quantifies the hazards of the above-mentioned adversarial post-correction, by focusing on the Byzantine Greek language. The development of this language entails a comprehensive examination of written language throughout the centuries under study. In essence, Byzantine Greek emerged as an academic style known as Atticised Greek, incorporating religious topics. As time progressed, elements of contemporary spoken Greek gradually became integrated into the language (Horrocks, 2014). As the results of our study also show, the combination of pre-trained data on Modern Greek, resembling what was considered as spoken language back then, and Ancient Greek, including the above-mentioned Atticised Greek, is beneficial for text classifiers detecting incorrect recognition output. Furthermore, by covering a period of more than seven centuries, the already challenging nature of recognition (Platanou et al., 2022) is increased, hence allowing a more diverse material for our investigation on post-correction. Errors in handwritten text recognition are often due to different hands (scribes) and scripts that evolved for many centuries. Printed text, on the other hand, exists for a much narrower period, there are no hands to cause similar errors, $^{1}$  and no clitics. As we show in this work, lower recognition error rate means more difficult classification of erroneous lines. By contrast, recognised handwritten text is often hard to parse even for humans, which makes an easier classification challenge and a much more difficult
+
+error-correction task, as shown in Table 1.
+
+# 1.3 The contributions
+
+We explore the extent to which classification to flawless/flawed texts can improve the post-correction of recognised output, by avoiding adversarial errors introduced when applied to flawless texts. We benchmark machine- and deep-learning-based text classifiers, showing that a Transformer that is pre-trained in modern Greek and then finetuned to detect erroneously recognised Byzantine Greek (10th-16th CE) text achieves an Average Precision of  $95\%$ . When the same Transformer is further pre-trained in ancient Greek prior to fine-tuning, the score increases to  $97\%$ . We quantify the benefits of employing our text classifier in the real world, by training an MT5 (Xue et al., 2021) encoder-decoder error-corrector, which altered  $80\%$  of the flawless texts. If our best-performing classifier had been applied prior to correction, the percentage would have been decreased to  $2\%$  with a trade-off of  $2\%$  texts being incorrectly classified as flawless.
+
+The better performance of a Transformer that is not pre-trained only in modern Greek can be explained by the fact that Byzantine Greek combines language elements from both, ancient and modern Greek. Since our study covers a broad historical period, we separated the evaluation dataset based on the century in order to investigate this hypothesis. Our results show that the detection of erroneous recognition depends on the chronology of the manuscript and that further pre-training in ancient Greek benefits the classification of lines from manuscripts in the following three centuries: 10thCE, 11thCE, and 13thCE. By following a similar segmentation-based evaluation method for the error rate level, instead of the century, we show that higher error rates make an easier classification goal. By contrast, the problem is harder for low error rates, when flawed and flawless texts look alike.
+
+# 2 Related work
+
+Our work focuses on the text classification of recognised handwritten text. In the absence of published studies for handwritten text, however, we summarise work focused on the recognition of printed text. Printed and handwritten materials differ, because the latter often yields scrambled recognition output (Table 1). This is probably what motivated Ströbel et al. (2022) in using the perplexity of lan
+
+guage models to detect the erroneous output in an unsupervised manner. On the other hand, although we consider that recognition errors are overall more for handwritten text compared to printed material, the quality of recognition can vary significantly for the former (Hodel et al., 2021), as is also shown in Fig. 1, and does not always come with a high error rate. Therefore, the detection of erroneously recognised handwritten text is also related to that for printed text, regarding their low error output.
+
+# Erroneous recognised printed text detection
+
+Initial work using rule-based approaches could not address all errors, and was followed by machine and deep learning approaches (Lyu et al., 2021). Standard approaches include support vector machines on top of n-gram-based features (Dannélls and Virk, 2021; Virk et al., 2021). Jatowt et al. (2019) used a gradient tree boosting binary classifier on top of various features, including character and word n-grams, part-of-speech, token frequency based on automatically created resources.
+
+Amrhein and Clematide (2018) performed detailed analyses and experiments of error detection and post-correction with various statistical and neural machine-translation methods. In general, they found that the former perform better in error correction, while the latter models are better in error detection. However, they concluded that there is not a single method that works best on all datasets while the results are highly affected by the data a model is trained on. Furthermore, they suggested that post-correction of recognised printed text should focus more on the improvement of error detection.
+
+Nguyen et al. (2020) used BERT pre-trained on named-entity recognition for token classification in order to perform error detection. After substring tokenization they obtain Glove or Fasttext word embeddings, which are combined with segment and positional embeddings and given as input to BERT (Devlin et al., 2018). The hidden states are fed to a dense layer on top that classifies each token as erroneous or not. Although interesting, token-level detection is not suited for high-error settings.
+
+# The ICDAR challenge
+
+ICDAR organised a competition focused on post-correcting text recognised from newspapers, shopping receipts, and other printed sources, including two subtasks: error detection, i.e., to detect the position and the length of the errors, and error correction (Chiron et al., 2017; Rigaud et al., 2019).
+
+In 2017, the best-performing method for the error detection subtask employed probabilistic character error models and weighted finite-state transducers while in 2019, the best was a pre-trained multilingual BERT with convolutional and fully-connected layers on top that classified each sub-token as erroneous or not. Schaefer and Neudecker (2020) followed a similar classification scheme, but focused on characters and not sub-tokens. Then, by excluding lines with at least one classified character from post-correction, they showed that false alteration (i.e., of correctly OCR red lines) by an encoder-decoder model could be avoided. Our work shows that casting this problem as a text classification task can lead to high accuracy for handwritten text, although similar benefits may apply to printed text.
+
+# 3 Data
+
+The written language of the manuscripts and papyri used in this study is Byzantine Greek.3 Within these texts, morphological categories such as the optative, the pluperfect, and the perfect have disappeared, while others such as the dative case have gradually decreased. Infinitives and participles are still there in the texts, as reminiscents of the classical tradition, encouraging one to treat the language as a unique variant, different from modern Greek. There are several spelling conventions that deviate from the older orthographic rules while the ancient punctuation signs are still in use, albeit not always with the same function. Therefore, following Classical and preceding Modern Greek, this language can be considered a combination of the two.
+
+The transcriptions and the HTR output were provided by the organisers of the recent HTREC challenge, which regarded the automated correction of HTR errors in Byzantine Greek (Pavlopoulos et al., 2023). Selected images of handwritten text in respective manuscripts (10th to 16th CE) were transcribed by a human expert. The training dataset of the challenge consists of 1,875 lines of transcribed text by a human expert and by an HTR model (examples shown in Table 2). Transkribus (Kahle et al., 2017) was used by the organisers to produce the system transcriptions, trained on seven images one per century. Our investigation of the CER of the opted HTR model reveals
+
+![](images/bdab47f54557d25fe59dce788ce6afe429b7b6ed1b8acf3a66b466b791bdf41b.jpg)  
+Figure 3: Violinplot of CER per century
+
+that a diversity of errors is produced, from very few to many per line (Fig. 1). This fact makes this dataset optimal for our training purposes, because we need a diversity of errors to train our classifiers. Furthermore, the errors are not uniformly distributed across centuries (Fig. 3), motivating a time/CER-based segmentation for evaluation purposes (discussed in Section 4.2).
+
+Our classification dataset consists of the 1,875 human-transcribed texts (lines), treated as the flawless negative (0) class, and the respective system-transcribed texts, treated as the flawed, positive (1) class. This resulted in total to an almost perfectly balanced dataset of 3,744 text lines, which we split to train, development and test subsets by using a 80/10/10 ratio. The average length of the transcription is 40.10 characters (6.83 words), and that of the HTR output is 37.09 characters (5.89 words).
+
+# 4 Experimental analysis
+
+We opted for machine and deep learning benchmarks, with different transfer learning settings.
+
+# Machine learning benchmarks
+
+Machine learning methods include a variant of Naive Bayes and Support Vector Machines (NB-SVM), which is often used as a baseline for text classification (Wang and Manning, 2012). Also, we experimented with a Random Forest and a multi-layer perceptron (MLP) using character n-grams
+
+<table><tr><td>Transcription by expert</td><td>HTR output</td></tr><tr><td>èγγινόμενα πάθη μή σβενύντες ἀλλὰ τή εχλύσει (the born-in passions not extinguishing but the release)</td><td>èγγενομεναπαδημησεννωτες ἀλλατήε κλησει</td></tr><tr><td>του βίου του καῦ ́ εαυτός πολλα γένεσθουσυ υμχωρουν (of the life of themselves many happening forgive)</td><td>του β ou του καλεαυτός πολλαγινεσθουσι συγχωρόν</td></tr><tr><td>τες ἄμυριζουσι του ὄμελώνα ἀλλα καὶ ὅ διὰ τες (- set on fire the vineyard but and the due to the)</td><td>εμμυριζου σιμαμπελώνα ἀλλακαι ὅδξα</td></tr></table>
+
+$(n \in 1,5)$ , term-frequency inverse document frequency, and default parameters otherwise.
+
+# Deep learning benchmarks
+
+Deep learning methods comprise a recurrent neural network using Gated Recurrent Units (GRU) and trainable word embeddings of 300 dimensions, originally pre-trained on (modern) Greek (Joulin et al., 2017). Also, we experimented with the state-of-the-art in text classification, which is BERT (Devlin et al., 2018) and its variants (Li et al., 2022; Minaee et al., 2021). In specific, we also fine-tuned two BERT models pre-trained in Greek. The first (GreekBERT) is pre-trained in modern Greek (Koutsikakis et al., 2020) while the second extends the first by being further pre-trained in ancient Greek (Singh et al., 2021).
+
+# Evaluation measures
+
+We opted for the  $F1$  score per class, the Area under the Receiver Operating Characteristic Curve (AUC), and Average Precision (AP) that is preferred in class-imbalanced settings. All evaluation measures were implemented in SCIKIT-LEARN. $^{12}$
+
+# 4.1 Text classification benchmark
+
+Table 3 shows the results of all our text classification benchmarks, including a random baseline (responding uniformly) that draws the lower performance limit. NBSVM is the worst machine learning model, followed by Forest and MLP. The latter achieves promising results, close to GRU (better in AP and  $F1^{-}$ ). GRU, on the other hand, is clearly
+
+Table 2: Human (on the left) and HTR-based (on the right) transcription (lines) from the 'Commentary on Isaiah' by Basil the Great (fol. 75r, Oxford, Bodleian Library MS. Barocci 102)  
+
+<table><tr><td></td><td>AP</td><td>AUC</td><td>F1+</td><td>F1-</td></tr><tr><td>Random</td><td>0.52</td><td>0.50</td><td>0.49</td><td>0.47</td></tr><tr><td>NBSVM</td><td>0.66</td><td>0.65</td><td>0.60</td><td>0.51</td></tr><tr><td>Forest</td><td>0.64</td><td>0.65</td><td>0.64</td><td>0.50</td></tr><tr><td>MLP</td><td>0.79</td><td>0.79</td><td>0.73</td><td>0.69</td></tr><tr><td>GRU</td><td>0.79</td><td>0.79</td><td>0.68</td><td>0.71</td></tr><tr><td>GreekBERT:M</td><td>0.95</td><td>0.94</td><td>0.88</td><td>0.88</td></tr><tr><td>GreekBERT:M+A</td><td>0.97</td><td>0.97</td><td>0.90</td><td>0.91</td></tr></table>
+
+Table 3: Evaluation with Average Precision (AP), AUC,  $F1$  for the classification of HTRed texts to flawed (+) or flawless (-) transcription. Random (uniform) classification was used as a baseline. GreekBERT was pre-trained in modern (:M) or in modern and ancient Greek (:A).
+
+outperformed in all metrics by GreekBERT. Two GreekBERT models were used, one pre-trained in modern Greek (:M) and one in modern and then in ancient Greek (:M+A). Byzantine Greek, as already discussed (Section 3), combines language elements found in both ancient and modern Greek. This is probably why GreekBERT:M+A, which was pretrained in modern and then in ancient Greek, was better compared to GreekBERT:M, which was pretrained only in modern Greek.[13]
+
+# 4.2 Evaluation per century
+
+We separated the lines of the test data based on the century of the manuscript that they came from. For each century, then, we evaluated the two Transformers, GreekBERT:M and GreekBERT:M+A. As can be seen in Fig. 4, $^{14}$  GreekBERT:M+A is better on the 10th (five percent units) and 11th century CE, with models performing similarly on the 12th and after the 13th century, when both models achieve a
+
+![](images/7f33b797aff5e59ccfcfed607a70a076263648627bbfa4090df6949c069b0456.jpg)  
+Figure 4: Average Precision per c. CE of GreekBERT:M, pretrained on modern Greek (solid), and GreekBERT:M+A, further pre-trained on ancient Greek (dashed) - error bands show the  $95\%$  bootstrap confidence intervals
+
+higher score. On the 13th century, the performance of both models drops, but considerably more for GreekBERT:M. The 14th, 15th and 16th century were more challenging in HTR (Platanou et al., 2022), which could explain the better performance of the classifiers. That is, more errors in the recognised output could potentially help the classifier distinguish between the two classes (correct/not).<sup>15</sup> Based on the previous observations, we report three findings:
+
+F1. The classification performance depends on the chronology of the manuscript.  
+F2. Older manuscripts are better handled by further pre-training on ancient Greek.  
+F3. Lines from recent manuscripts pose an easier classification challenge to both GreekBERT models.
+
+# 4.3 Evaluation per error zone
+
+Our dataset is balanced regarding whether the recognition output is erroneous (positive class) or not, yet it remains imbalanced regarding the CER level (see Fig. 1). In order to assess the two Transformers across different input types, we rolled a window of ten units, from low to high CER values, and we assessed the models per window (Fig. 5).<sup>16</sup>
+
+High-CER lines (i.e., lines with CER of thirty or more) appear to set an easy classification task to both models. Performance increases as the CER increases. This is reasonable considering that lines with a high CER comprise many errors and are
+
+hard to parse even by humans. This characteristic, however, makes them easier to distinguish from flawless lines, which comprise no errors from the recognition.
+
+Low-CER lines (i.e., lines with CER of zero to twenty) lead to a different picture, since Greek-BERT:M+A is consistently better, overall, than GreekBERT:M. Low CER means small differences between the two texts (flawless-flawed), which makes the classification task more difficult. In this zone, further pre-training on ancient Greek had a bigger impact, compared to zones with a higher error rate.
+
+# 5 Discussion
+
+In this section, first, we assess the possibly adversarial nature of post-correction. Then, we describe a synthetic error detection application, followed by a century-based error analysis we performed.
+
+# 5.1 Substantiality assessment
+
+Error-correction systems may introduce errors, even when applied to flawless HTRed texts. To assess the substantiality of this claim, we experimented with a neural encoder-decoder postcorrector. In specific, we employed a multilingual variant of the T5 Text-to-Text Transfer Transformer (Raffel et al., 2019), called MT5 (Xue et al., 2020), which was pre-trained on a dataset covering 101 languages.[17] MT5 has been used for Grammatical Error Correction in modern Greek and sets a strong baseline for HTR Error Correction (Korre and Pavlopoulos, 2022). We fine-tuned this model on the 1,875 training instances of the HTREC
+
+![](images/1b1fc96a65a97018d7f67907f01cd0a6068fe460d5c6b6f4d4c969915a3566d6.jpg)  
+Figure 5: Average Precision of GreekBERT:M, pretrained on modern Greek (blue, solid), and GreekBERT:M+A, further pre-trained on ancient Greek (orange, dashed), using a rolling window over CER (left edge shown horizontally) - error bands depict  $95\%$  bootstrap confidence intervals
+
+challenge, by encoding the system- and decoding the human-transcriptions. $^{18}$  This model is a neural error correction baseline that reduced the CER in 19 out of the 180 HTRed texts in the official HTREC2022 evaluation set. We applied this corrector on the respective 180 transcriptions of these data (the ground truth evaluation instances in HTREC2022) and found that 144  $(80\%)$  were mistakenly altered by our MT5 corrector. For example, the line 'αν τους μηδους οινες έτε αργυρτω παρα' was altered to 'αν τους μηδους οινες έτε αργυρτω παρα' (word division, 4th word). Similarly, 'χευσης μη φονευσης μη κλε' was altered to 'χλευσεως μη φονευσεως μη κλεπτ'. Most of these adversarial mistakes (all but the top three of Table 4) would have been avoided if our GreekBERT:M+A classifier had been applied prior to the application of this neural corrector. Conversely, only four (i.e., the rest in Table 4) would have been mistakenly found as correct from the system-transcribed instances.
+
+# 5.2 A synthetic application of error detection
+
+We experimented with the HTREC2022 evaluation dataset, which was used to assess the participating systems. This dataset comprises 180 lines that were transcribed from nine images. Out of these lines, one had zero CER achieved by the HTR model (perfect transcription), which was correctly classified by our GreekBERT:M+A model. We then performed two analyses, first by applying our model on the flawless human transcriptions (as if a transcription was the HTR output) and then by applying
+
+it on the flawed system recognitions (the same lines were used). The human transcriptions were correctly classified as flawless, all except three  $(1.7\%)$  error rate), which are shown in the top three rows of Table 4. The system recognitions were also correctly classified as flawed, all except four  $(2.3\%)$  error rate), shown in the last four rows of Table 4.
+
+# 5.3 Error analysis
+
+An error detection by experts in Table 4, $^{19}$  showed that no clue (indicating an error) was found for the first line (except from the rare ‘φιλαχόλουνός’ with double accent). As is shown in Figure 6 (a), which depicts the attention on the subwords of this line, this was indeed the network’s focus.
+
+On the second line, word 'iσμαηλήτι' was considered as the outcome of misspelling of the word 'iσμαηλιτη', matching the algorithm's decision. In this case, it was the ending of the word that attracted attention (Fig. 6), but the network focused more on the article 'των' preceding the word rather than on the word itself. The article is in plural form, not matching a word in singular. Apparently, 'iσμαηλήτι' is not a misspelling, but part of the word 'iσμαηλιτιχων' that was corrupted.
+
+In the last four lines, that were classified as correct, several flaws were raised, indicating a failure of the algorithm. For example, word divisions, such as in the case of  $\omega \sigma x\alpha$  instead of  $\omega \varsigma \chi \alpha \iota$ ; added characters, such as in the case of  $\varepsilon \dot{\upsilon}\upsilon \chi \eta \varsigma$  instead of  $\varepsilon \dot{\upsilon}\chi \eta \varsigma$ ; character replacements, such as in the case of  $\zeta \kappa \tau \varepsilon \iota \alpha \nu$  instead of  $\iota \kappa \varepsilon \tau \varepsilon \iota \alpha \nu$ .
+
+<table><tr><td>TRANSCRIPTIONS AND HTR OUTPUT</td><td>GREEKBERT:M+A</td><td>GT</td></tr><tr><td>èγώ δ ἀεί πως φιλαχόλουθός εἰμι(I am always readily following)</td><td>F</td><td>N</td></tr><tr><td>λα και γέδεων εκ τών σχύλων τών ομαηλητι(- and Gedeon from the dogs the - )</td><td>F</td><td>N</td></tr><tr><td>δραμων εἰς και γόνυπετη(having run to and falling on the knee)</td><td>F</td><td>N</td></tr><tr><td>τιθεις τας χειρας επαυτα(having put the hands on them)</td><td>F</td><td>N</td></tr><tr><td>ζετείαν χέδεσποτα, και ευμενώς Ἁκουσου τών τής προς εύυχης(- - - and favourably listen to the of the prayer)</td><td>N</td><td>F</td></tr><tr><td>èγενόμην icos. αραή τής μεό τιεγεζήθης ἐλειεμοήτιριος(became - - - - - )</td><td>N</td><td>F</td></tr><tr><td>ωσααι τής συνεχους διδασκαλιας τους βα(- of the continuous instruction the - )</td><td>N</td><td>F</td></tr></table>
+
+Table 4: HTRed lines and transcriptions that GreekBERT:M+A mistakenly predicted as flawed (F) or not (N) - the ground truth (GT) on the right shows whether there was an error or not
+
+# Error patterns
+
+Next, we proceed with an analysis of the recognition errors, using data from three centuries. We opt for 16th CE, the century on which both our models perform considerably well (Fig. 4). Next, we focus on the 14th CE, going back two centuries, but still observing good performance by both models. Last is shown the 13th CE, where GreekBERT:M performs much worse than GreekBERT:M+A.
+
+16th CE This is the data group with the highest average CER (see Figure 3). What seems to be challenging in this data group is the fact that there is a variety of punctuation marks which are sometimes read by the model as letters because of their shape. An illustrative example of this case is the word 'πλοῦς' which is read by the model as 'πλούου'. This is due to a comma following the word 'πλοῦς', which is very similar to the last letter of the word (ς). There are also many cases of ligatures in which the model fails to recognise both letters, especially the one written on the top of the ligature, such as the letter 'τ' in the combination 'τρ' as one can see in the following instance; 'ω πατρις' read as 'ως ωπαρις'.
+
+14th CE The data from the 14th century show the second highest CER on average (see Figure 3). This is partly because of a great number of ligatures appearing in this data group. A common ligature among this group is the one which represents the group of the letters  $\varepsilon$  and  $\rho$ . In this ligature it happens that the letter  $\varepsilon$  is written above the letter
+
+$\varphi^{\prime}$ . The way the model interprets this structure can be seen in the following case;  $\varphi \acute{\varepsilon} \rho^{\prime} \dot{\varepsilon} \pi^{\prime} \omega \tau \eta \nu$ $\iota \omega$ , where  $\varphi \acute{\varepsilon} \rho^{\prime}$  consists of one letter  $(\varphi)$  and one ligature  $(\varepsilon \rho)$  is read by the model as  $\varphi \varepsilon \pi \alpha \omega \tau \eta \nu$ $\iota \varsigma \omega$ . The model fails reading correctly the letter combination and returns only the letter  $\varphi^{\prime}$  and one of the two letters involved in the  $\varepsilon \rho^{\prime}$  combination  $(\rho)$ , which is actually the one written in the position next to the letter  $\varphi^{\prime}$ . The same instance shows clearly that there is a tendency towards adding characters because of different drawing representations of characters found in this data group. The letter  $\omega$  can be considered such a case. In the previous example, we can see that the word  $\iota \omega$  is read by the model as  $\iota \varsigma \omega$ . This is because, in this instance, the letter  $\omega$  has the shape of a laid eight. The same letter has the usual shape in other cases in the text of this century, which was quite confusing for the model and led the model to decide that it should be a combination of both  $\sigma^{\prime}$  and  $\omega^{\prime}$ , since the former is quite close to the half laid-eight shape.
+
+13th CE The data group that is easiest to read for the model seems to be the 13th century one. In this group, errors appear mainly because of special letter shapes and abbreviations. Such a letter shape is that of the letter  $\varepsilon$  which is written as the letter  $\delta$  with an added line pointing upwards. A characteristic example of the failure of the recognition task is this one;  $\tilde{\varepsilon}\omega\varsigma$  is read as  $\delta\omega\varsigma$ . Here the model has clearly classified the  $\tilde{\varepsilon}$  as a  $\delta$ . Regarding the abbreviations representing character groups, one can find among others the following
+
+![](images/b2b7a82b084099725c5d3a11f87f09630850cb3ae1b16ac4f0a3d88595c1dd97.jpg)  
+(c)  $\delta \rho \alpha \mu \omega \nu \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \varepsilon \vareeta$  
+Figure 6: The attention of GreekBERT:M+A (darker is more) for the three top (missclassified) lines of Table 4 - the symbol ## is used to concatenate a subword with the preceding one during tokenisation ( $\delta \rho \alpha \mu \omega \nu$  is  $\delta \rho \alpha \# \mu \omega \nu$ )
+
+one;  $\alpha v$  . In the case of ouxioov, where the last two letters are represented by the abbreviation, the model reads ouxoX, where  $\omega$  has replaced the abbreviation because of similarity in shape.
+
+# 5.4 Language and script generalisation
+
+Our work can potentially inspire analogous pretraining in other languages with a long documented history. To provide an example: Old, Middle, and Modern English and Japanese (Lyovin, 1997, pp. 2) can qualify for pre-training on both Old and Modern data, as in the Greek language in our case. The scripts our results are based on, however, are very common both in Greek and Latin manuscripts while common handwritten character shapes, e.g., epsilon (E) (Chambers and Chambers, 1891, pp. 703), may illuminate error patterns in Latin scripts.
+
+# 6 Conclusion
+
+This work showed that detecting erroneous recognised lines improves the accuracy of handwritten Byzantine text recognition, aiming to assist the disciplines struggling with vast corpora and little man power. By contrast to printed text, errors in handwritten text vary greatly, some of which are easier and others difficult to catch. Our experiments showed that two BERT models, one pre-trained in modern and the other pre-trained in modern then in ancient Greek perform well for the task of erroneous handwritten recognised text classification. The latter performs better, especially for manuscripts from older centuries and for lines with higher recognition error rate. By using a neural text-to-text Transformer for error correction along with our best performing classifier, we found that adversarial behaviour is not uncommon in the neural corrector's output which could have been significantly limited if a BERT classifier, pre-trained and
+
+fine-tuned, had classified the input first.
+
+Directions of future work comprise the development of a larger dataset, which, although not easily developed in our domain of focus, it could potentially allow us to draw more conclusions and lead to further findings. Furthermore, we plan to extend our dataset with calibrated evaluation subsets, in order to gain more perspective on the performance of systems for specific error types, including ones that occur naturally and are not meant to be corrected.
+
+# Limitations
+
+- Our study is focused on Byzantine Greek, from the 10th to the 16th century. Our hypothesis, that text classification could assist the post correction of HTRd output, should be studied for more languages and covering a longer period to draw general conclusions.  
+- By contrast to OCR, where low CER values are reported, HTRd output can significantly differ from the human transcription. For example, correct text may be misplaced (e.g., due to mistaken word division) or by the introduction of an extra character (e.g., the last word of the fifth line in Table 4). Hence, the creation of ground truth at the character level is very challenging in this context, hindering experimentation at the token or the character level, as in OCR (Schulz and Kuhn, 2017; Schaefer and Neudecker, 2020).  
+- Sensitivity analysis has not been performed. Although we repeated our experiments to verify the validity of our findings regarding GreekBERT, we have not experimented with cross validation over the centuries, as in (Platanou et al., 2022). Training, however, an algorithm on texts from specific centuries, then as
+
+sessing on ones from previous or future ones, could potentially reveal interesting patterns.
+
+- Diplomatic transcriptions may lead to naturally-occurring errors (i.e., ones the editors would want to keep, to stay as close to the original as possible), which our proposed classifiers could and probably would classify as incorrect recognition output. This limitation could be bypassed if the proposed systems are used complementary with an expert, not trying to substitute one. For example, the expert could view only lines the classifier labelled as incorrect, bypassing or acknowledging errors which were meant to be there.
+
+# References
+
+Chantal Amrhein and Simon Clematide. 2018. SupervisedOCR error detection and correction using statistical and neural machine translation methods. Journal for Language Technology and Computational Linguistics (JLCL), 33(1):49-76.  
+William Chambers and Robert Chambers. 1891. Chambers' Encyclopaedia: A Dictionary of Universal Knowledge, volume 8. W. & R. Chambers.  
+Guillaume Chiron, Antoine Doucet, Mickaël Coustaty, and Jean-Philippe Moreux. 2017. Icdar2017 competition on post-ocr text correction. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 1, pages 1423-1428, Kyoto, Japan.  
+Dana Dannélls and Shafqat Virk. 2021. A supervised machine learning approach for post-ocr error detection for historical text. In Swedish Language Technology Conference and NLP4CALL, pages 13-20.  
+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.  
+Tobias Mathias Hodel, David Selim Schoch, Christa Schneider, and Jake Purcell. 2021. General models for handwritten text recognition: feasibility and state-of-the-art. german current as an example. Journal of open humanities data, 7(13):1-10.  
+Geoffrey Horrocks. 2014. Greek: A History of the Language and its Speakers. John Wiley & Sons.  
+Adam Jatowt, Mickaël Coustaty, Nhu-Van Nguyen, Antoine Doucet, et al. 2019. Post-ocr error detection by generating plausible candidates. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 876-881, Sydney, Australia.
+
+Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431. Association for Computational Linguistics.  
+Philip Kahle, Sebastian Colutto, Günter Hackl, and Günter Mühlberger. 2017. Transkribus - a service platform for transcription, recognition and retrieval of historical documents. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 04, pages 19-24.  
+Katerina Korre and John Pavlopoulos. 2022. Enriching grammatical error correction resources for Modern Greek. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4984-4991, Marseille, France. European Language Resources Association.  
+John Koutsikakis, Ilias Chalkidis, Prodromos Malakasiotis, and Ion Androutsopoulos. 2020. Greek-bert: The greeks visiting sesame street. In 11th Hellenic Conference on Artificial Intelligence, pages 110-117.  
+Qian Li, Hao Peng, Jianxin Li, Congying Xia, Renyu Yang, Lichao Sun, Philip S Yu, and Lifang He. 2022. A survey on text classification: From traditional to deep learning. ACM Transactions on Intelligent Systems and Technology (TIST), 13(2):1-41.  
+Anatole Lyovin. 1997. An Introduction to the Languages of the World. Oxford University Press, USA.  
+Lijun Lyu, Maria Koutraki, Martin Krickl, and Besnik Fetahu. 2021. NeuralOCR post-hoc correction of historical corpora. Transactions of the Association for Computational Linguistics, pages 479-493.  
+Arun S. Maiya. 2020. ktrain: A low-code library for augmented machine learning. arXiv preprint arXiv:2004.10703.  
+Barbara McGillivray, Thierry Poibeaun, and Pablo Ruiz Fabo. 2020. Digital humanities and natural language processing: "je t'aime... moi non plus".  
+Shervin Minaee, Nal Kalchbrenner, Erik Cambria, Narjes Nikzad, Meysam Chenaghlu, and Jianfeng Gao. 2021. Deep learning-based text classification: a comprehensive review. ACM Computing Surveys (CSUR), 54(3):1-40.  
+Thi Tuyet Hai Nguyen, Adam Jatowt, Nhu-Van Nguyen, Mickael Coustaty, and Antoine Doucet. 2020. Neural machine translation with bert for post-ocr error detection and correction. In Proceedings of the ACM/IEEE joint conference on digital libraries in 2020, pages 333-336.  
+John Pavlopoulos, Vasiliki Kougia, Paraskevi Platanou, Stepan Shabalin, Konstantina Liagkou, Emmanouil Papadatos, Holger Essler, Jean-Baptiste Camps, and Franz Fischer. 2023. Error correcting htr'ed byzantine text.
+
+Paraskevi Platanou, John Pavlopoulos, and Georgios Papaioannou. 2022. Handwritten paleographic greek text recognition: A century-based approach. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6585-6589.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683.  
+Christophe Rigaud, Antoine Doucet, Mickaël Coustaty, and Jean-Philippe Moreux. 2019. Icdar 2019 competition on post-ocr text correction. In 2019 international conference on document analysis and recognition (ICDAR), pages 1588-1593, Sydney, Australia.  
+Robin Schaefer and Clemens Neudecker. 2020. A two-step approach for automaticOCR post-correction. In Proceedings of the The 4th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 52-57, held online.  
+Sarah Schulz and Jonas Kuhn. 2017. Multi-modular domain-tailoredOCR post-correction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2716-2726, Copenhagen, Denmark.  
+Pranaydeep Singh, Gorik Rutten, and Els Lefever. 2021. A pilot study for bert language modelling and morphological analysis for ancient and medieval greek. In The 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2021).  
+Phillip Benjamin Ströbel, Simon Clematide, Martin Volk, Raphael Schwitter, Tobias Hodel, and David Schoch. 2022. Evaluation of htr models without ground truth material. arXiv preprint arXiv:2201.06170.  
+Shafqat Mumtaz Virk, Dana Dannélls, and Azam Sheikh Muhammad. 2021. A novel machine learning based approach for post-ocr error detection. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1463–1470, Varna, Bulgaria.  
+Sida Wang and Christopher Manning. 2012. Baselines and bigrams: Simple, good sentiment and topic classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 90-94, Jeju Island, Korea. Association for Computational Linguistics.  
+Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2020. mt5: A massively multilingual pre-trained text-to-text transformer. arXiv preprint arXiv:2010.11934.
+
+Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics.

2023/Detecting Erroneously Recognized Handwritten Byzantine Text/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:782f850d4d1b1c2d9abfaa9793b7ee7f436cc57cefebc0a3e0f182b7f5e556bf
+size 343986

2023/Detecting Syntactic Change with Pre-trained Transformer Models/5f1aa19b-edb6-4425-a7bb-8580ef6a41b6_content_list.json

@@ -0,0 +1,1539 @@
+[
+    {
+        "type": "text",
+        "text": "Detecting Syntactic Change with Pre-trained Transformer Models",
+        "text_level": 1,
+        "bbox": [
+            156,
+            89,
+            842,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Liwen Hou  \nHarvard University  \nliwenhou@fas.harvard.edu",
+        "bbox": [
+            211,
+            142,
+            457,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "David A. Smith Northeastern University dasmith@ccs.neu.edu",
+        "bbox": [
+            566,
+            142,
+            764,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We investigate the ability of Transformer-based language models to find syntactic differences between the English of the early 1800s and that of the late 1900s. First, we show that a fine-tuned BERT model can distinguish between text from these two periods using syntactic information only; to show this, we employ a strategy to hide semantic information from the text. Second, we make further use of fine-tuned BERT models to identify specific instances of syntactic change and specific words for which a new part of speech was introduced. To do this, we employ a part-of-speech (POS) tagger and use it to train corpora-specific taggers based only on BERT representations pretrained on different corpora. Notably, our methods of identifying specific candidates for syntactic change avoid using any automatic POS tagger on old text, where its performance may be unreliable; instead, our methods only use untagged old text together with tagged modern text. We examine samples and distributional properties of the model output to validate automatically identified cases of syntactic change. Finally, we use our techniques to confirm the historical rise of the progressive construction, a known example of syntactic change.",
+        "bbox": [
+            142,
+            282,
+            460,
+            667
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            682,
+            260,
+            697
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Many researchers have taken advantage of the availability of pre-trained language models to study semantic change (e.g., Giulianelli et al., 2020; Hu et al., 2019; Fonteyn, 2020), but computational studies of syntactic change are more sporadic, relying on specially trained models (Merrill et al., 2019). In this paper, we investigate the extent to which BERT (Devlin et al., 2019) is able to distinguish between the syntax of English from the early 1800s and that of the late 1900s. To do so, we use the British Hansard corpus, which consists of a digitized version of the U.K. Parliamentary debates starting in the early 1800s and spanning",
+        "bbox": [
+            112,
+            709,
+            490,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "across two centuries. We isolate all of the parliamentary debates from the 1830s and the 1990s.<sup>1</sup> These debates can be found on the U.K. Parliament Website's Hansard archive<sup>2</sup>.",
+        "bbox": [
+            507,
+            253,
+            884,
+            315
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We fine-tune BERT-base-uncased, which has around 110 million parameters and was pre-trained on approximately 3.3 billion words. When training or testing BERT on a sentence, we first tokenize it using the original BERT-base-uncased tokenizer, as implemented by Hugging Face (Wolf et al., 2020).",
+        "bbox": [
+            507,
+            318,
+            884,
+            414
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We focus on two questions: First, to what extent can BERT distinguish text from different time periods on the basis of syntactic features alone? Second, can we use BERT to help identify words or sentences whose analysis reflects syntactic change in an unsupervised way?",
+        "bbox": [
+            507,
+            416,
+            882,
+            511
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In §2, we find evidence that BERT can be used to detect syntactic differences between the English of the 1830s and that of the 1990s. Specifically, even after we control for confounders, a fine-tuned BERT model can still distinguish between these two periods even when the semantic information is hidden.",
+        "bbox": [
+            507,
+            513,
+            882,
+            624
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In §3, we describe how we train classifiers that can assign POS tags to words based on a 1830s and a 1990s understanding of English, respectively. Our main contribution comes in §4, where we use these classifiers to identify words whose analysis reflects syntactic change. Our method can automatically find words whose part of speech has changed. Moreover, our method avoids running a POS tagger on old text, since automatic tagging may be unreliable in that scenario. We emphasize that our change-detection techniques work with no manual annotation and identify examples of language change without any manual filtering; see Appendix B for a list of the top automatically detected words whose part of speech has changed.",
+        "bbox": [
+            507,
+            626,
+            884,
+            868
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>We chose these decades because they were the earliest and latest decade with large amounts of data in every year. <sup>2</sup>https://www.hansard-archive.parliament.uk/",
+        "bbox": [
+            507,
+            879,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3564",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3564-3574 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            779,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To the best of our knowledge, such unsupervised syntactic change detection has not been previously done.",
+        "bbox": [
+            112,
+            84,
+            487,
+            130
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Finally, we confirm a known change (an increase in the frequency of progressives) using our classifiers. We show that a BERT model fine-tuned on text from the 1830s displays a poorer understanding of the progressive construction compared to a similar model trained on text from the 1990s.",
+        "bbox": [
+            112,
+            133,
+            489,
+            229
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Detecting Syntactic Change While Masking Semantics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            242,
+            443,
+            275
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we check the extent to which BERT is sensitive to changes in the syntax of English over time. This acts as a sanity check for later sections, in which we use BERT in a different way to identify specific instances of syntactic change.",
+        "bbox": [
+            112,
+            285,
+            487,
+            365
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We start by fine-tuning BERT-base-uncased on the classification task of predicting whether sentences come from the 1830s or the 1990s. To do so, we use the representation of the CLS token in the last layer of BERT and add a linear layer to it; this model is then fine-tuned using labeled sentences from the 1830s and the 1990s.",
+        "bbox": [
+            112,
+            366,
+            487,
+            476
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "After we fine-tune it on 100,000 sentences per time period, the BERT model achieves  $98.6\\%$  accuracy; however, it may use semantic information to distinguish between text from the two time periods, and we want to limit ourselves to syntactic information only in order to check the extent to which BERT can detect syntactic change over time.",
+        "bbox": [
+            112,
+            480,
+            487,
+            590
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To make BERT use syntactic information only, we mask semantic information by replacing all the open class words in each sentence with their POS tags, obtained using Stanford CoreNLP (Manning et al., 2014). In order to leverage BERT's pretrained knowledge of English, we then replace each POS tag with a common English word that usually has that tag (in a similar vein to Gulordava et al., 2018); this way, the input to BERT looks more like \"colorless green ideas\" and not \"JJ JJ NNS\".",
+        "bbox": [
+            112,
+            593,
+            487,
+            752
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "After replacing words (other than forms of \"to be\" and \"to have\", prepositions, and determiners) with POS tags and after replacing each POS tag with a representative word for that part of speech, we fine-tune BERT (with the added layer described above) on the resulting data. The BERT model performs surprisingly well on this classification task, achieving over  $90\\%$  accuracy for the task of pre",
+        "bbox": [
+            112,
+            753,
+            487,
+            882
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "dicting whether a sentence comes from the 1830s or the 1990s. However, there are still several ways for such a model to use non-syntactic information. First, the model may use sentence lengths, since the distribution of the sentence lengths is different for the 1830s than for the 1990s. Second, the BERT model may simply be using the counts of the POS tags rather than real syntactic information.",
+        "bbox": [
+            507,
+            84,
+            882,
+            212
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To mitigate the length confounder, we place restrictions on the lengths of our training sentences. We run experiments separately for the length bins 11-15 words, 21-25 words, 31-35 words, and 20 words. To check the extent to which BERT relies on syntactic information other than the counts of tags, we separately train a model on shuffled input data, in which the sequence of tags in the sentence appears out of order. Such shuffled data retains no information other than the tag counts; we then compare whether BERT trained on unshuffled data can outperform BERT trained on shuffled data. The results are summarized in Table 1.",
+        "bbox": [
+            507,
+            212,
+            882,
+            420
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We compare these results to a logistic regression model that takes in the frequencies for the 100 most common vocabulary items in the replaced corpus $^{4}$  (plus a catch-all category for the remaining items) and the sentence length as features. Such a logistic regression model achieves an accuracy of  $85.45\\%$ , which means it performs about as well as BERT fine-tuned on shuffled tags but loses to BERT when the latter is fine-tuned on unshuffled tags. Using sentence length as the only feature yields an accuracy of  $60.33\\%$ .",
+        "bbox": [
+            507,
+            423,
+            882,
+            598
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The conclusion of this experiment is that the fine-tuned BERT model must be using syntactic information to tell apart sentences from the 1830s and the 1990s. Indeed, even when the length of the sentence is fixed, and even when the words have been replaced by their parts of speech, the model gains advantage from seeing the true order of the tags as they appeared in the sentence, instead of seeing shuffled tags. This advantage must fundamentally come from differences in syntax that are present between the 1830s and 1990s sentences (and moreover, these differences must take a form that is not simply tag counts).",
+        "bbox": [
+            507,
+            599,
+            882,
+            808
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Having established that BERT is sensitive to changes in the syntax of English, we next train BERT in a different way to identify specific instances of such change. Importantly, from here on",
+        "bbox": [
+            507,
+            809,
+            882,
+            872
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "4 In addition to the typical words that represent different parts of speech, this replaced corpus also contains prepositions, determiners, and forms of \"to be\" and \"to have\".",
+        "bbox": [
+            507,
+            879,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "3The code for our experiments is available here: https://github.com/houliwen/detecting-syntactic-change",
+        "bbox": [
+            112,
+            892,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3565",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/4c1fefc37a810e339ed362df82ab658d41f83a63f796ef21152d75145f957300.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>sent.\nlength</td><td>shuffled\nplaced</td><td>unshuffled\nplaced</td><td>shuffled\nplain</td><td>unshuffled\nplain</td></tr><tr><td>11-15</td><td>83.37</td><td>87.74</td><td>94.70</td><td>96.69</td></tr><tr><td>21-25</td><td>86.44</td><td>90.54</td><td>97.39</td><td>98.63</td></tr><tr><td>31-35</td><td>89.37</td><td>93.12</td><td>98.20</td><td>98.90</td></tr><tr><td>20</td><td>85.20</td><td>89.30</td><td>96.60</td><td>97.40</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            485,
+            156
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Accuracy of BERT fine-tuned on different types of data for predicting time period. All models were trained on 20,000 sentences and tested on 2,000 (half from the 1830s and half from the 1990s); the exception was the last line, for which only 10,000 training sentences were used due to lack of data. Since test data contained an even mixture of 1830s and 1990s sentences, baseline accuracy is  $50\\%$ . These numbers are based on a single run; however, note that an accuracy of  $p$  will have standard deviation around  $\\sqrt{p(1 - p) / 2000}$ , which ranges from  $0.23\\%$  to  $0.83\\%$  for the above, with higher accuracies having lower standard deviations.",
+        "bbox": [
+            112,
+            167,
+            489,
+            338
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "out, we avoid using a POS tagger on older English.",
+        "bbox": [
+            112,
+            363,
+            487,
+            381
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Training Classifiers",
+        "text_level": 1,
+        "bbox": [
+            112,
+            393,
+            317,
+            409
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We use BERT to identify specific instances of change. To do so, we train two POS taggers, one based on text from the 1830s and the other based on text from the 1990s; in the subsequent sections, we will use these taggers and their errors to identify words or sentences that display syntactic change. In this section, we will focus on describing how these taggers are trained and their performance. Crucially, we will train these taggers without ever tagging text from the 1830s; all POS tags are only ever associated with text from the 1990s.",
+        "bbox": [
+            112,
+            419,
+            489,
+            594
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We start by fine-tuning two instances of BERT-base-uncased: one on text from the 1990s and the other on text from the 1830s. This is done via Masked-Language Modeling in PyTorch (Paszke et al., 2019), where we process one sentence at a time with  $15\\%$  of the tokens masked, and the model is trained to predict the masked tokens. We use each fine-tuned model to obtain BERT token representations for 1990s text and then average the token representations for all the WordPiece tokens within each word to obtain a representation of a word. We do this twice: once with the model fine-tuned on text from the 1830s and once with the model fine-tuned on text from the 1990s (yielding two different embeddings of each word). However, even though we have models trained on two time periods, we use both models only to embed words from the 1990s and not any words from the 1830s. At the end of this process, each word from the 1990s has two different BERT embeddings.",
+        "bbox": [
+            112,
+            596,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Next, using POS tags obtained from Stanford CoreNLP for the 1990s portion of the British Hansard, we train a 2-layer perceptron (with one hidden layer of size 50) to assign a Penn Treebank tag to each BERT representation of each word. An input to this 2-layer perceptron is a BERT embedding of a word (extracted as described above), and the output is a POS tag. This classification task ignores punctuation and has 21 classes: We merge the 'NN', 'NNP', 'NNS', and 'NNPS' tags into one single tag called 'Noun', after which we take the 20 most common tags ('Noun', 'IN', 'DT', 'JJ', 'VB', 'PRP', 'RB', 'VBN', 'VBZ', 'VBP', 'CC', 'MD', 'TO', 'VBD', 'CD', 'VBG', 'PRPS', 'WDT', 'WP', 'WRB') and then group the remaining POS tags into a catch-all class called 'other'. We actually train two such perceptrons separately: one for the 1830s BERT embeddings (of words from the 1990s only) and one for the 1990s BERT embeddings (also only of words from the 1990s).",
+        "bbox": [
+            505,
+            84,
+            884,
+            405
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Note that the training and test data for this procedure consist only of sentences from the 1990s; this is done because the POS tags obtained from Stanford CoreNLP are likely to be more accurate on our modern data. Using the representations from the final layer of BERT (when the classifier has access to 50,000 training sentences, containing 815,578 words, and when the test set has 50,000 sentences, containing 816,428 words), the accuracy obtained by the 1990s model is  $97.6\\%$ , and the accuracy of the 1830s model is  $97.2\\%$ .",
+        "bbox": [
+            507,
+            407,
+            884,
+            582
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We repeat the above analysis using POS tags from Trankit (Nguyen et al., 2021) instead of Stanford CoreNLP because Trankit is a state-of-the-art parser and tagger. By using Trankit tags to train and test the above classifiers based on BERT representations, the accuracy obtained by the 1990s model was  $98.4\\%$  (instead of  $97.6\\%$  using CoreNLP tags) and the accuracy obtained by the 1830s model was  $98.0\\%$  (instead of  $97.2\\%$  using CoreNLP tags). The higher accuracy means that the task became easier when predicting Trankit tags instead of CoreNLP tags; since Trankit tags are also more accurate, it suggests that our pipeline is working as expected.",
+        "bbox": [
+            507,
+            585,
+            884,
+            809
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In §4 and later, we will use these pipelines (and the differences between their predictions) to iden-",
+        "bbox": [
+            507,
+            810,
+            882,
+            843
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>5</sup>We tried hidden layer sizes 100 and 200; this hyperparameter did not make a large difference.",
+        "bbox": [
+            507,
+            854,
+            885,
+            879
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{6}$ www.cis.upenn.edu/~bies/manuals/tagguide.pdf",
+        "bbox": [
+            532,
+            879,
+            873,
+            892
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{7}$ https://trankit.readthedocs.io/en/latest/performance.html",
+        "bbox": [
+            510,
+            892,
+            852,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3566",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tify specific instances of syntactic change between the 1830s and the 1990s.",
+        "bbox": [
+            112,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Identifying Words that Underwent Change",
+        "text_level": 1,
+        "bbox": [
+            112,
+            130,
+            448,
+            165
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We now attempt to use the POS taggers trained above to identify words that underwent change. We employ two methods to do this: First, we find the instances for which the two trained models disagree on a POS tag (indicating that the model familiar with text from the 1830s would tag the word differently compared to the one familiar with text from the 1990s). The second method involves masking out the target word to see if the 1830s model predicts the tag correctly on the masked word. If the masked word is tagged correctly but the un-masked word is tagged incorrectly, it suggests that the 1830s model can correctly parse the rest of the sentence but is specifically confused by the target word; when this happens, it indicates that the target word may have undergone change. This latter strategy is depicted in Figure 1.",
+        "bbox": [
+            112,
+            175,
+            489,
+            450
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Tagging Disagreements",
+        "text_level": 1,
+        "bbox": [
+            112,
+            461,
+            344,
+            476
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After training, we have two new POS taggers, one based on an 1830s understanding of English and one based on a 1990s understanding. We now run these taggers on test data (from the 1990s) to find words on which they disagree. In particular, we find words for which the model trained on text from the 1990s agrees with the Trankit tag but where the 1830s model disagrees. The cases where such a disagreement occurs should be the cases in which an 1830s grammar leads the model to incorrectly parse the sentence (likely due to a change in English grammar for the relevant word or perhaps one of its neighbors). Moreover, we use our classifiers to extract probabilistic predictions for the tags, and we search for word instances where probability assigned to the correct tag by the 1830s model was at least 50 percentage points lower than that assigned by the 1990s model.",
+        "bbox": [
+            112,
+            483,
+            489,
+            772
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We do this for each word in each sentence of the 1990s test data. To help isolate interesting cases, we filter the words under consideration to those that occur at least 10 times in the 1830s text (indicating that they are not new words). We remove numbers, punctuation, and words written entirely in capital letters (likely acronyms). We also restrict to words for which such a tagging disagreement occurred at least twice. We then sort the resulting words",
+        "bbox": [
+            112,
+            774,
+            489,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "by the total number of tag disagreements for that word, normalized by the frequency of the word in the 1990s data, with additional weight given to words which were common in the 1830s (to reduce the number of rare words in the final list). We run this experiment on 100,000 test sentences, using classifiers trained on 25,000 sentences each.",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We then manually examine the top resulting words to check whether the identified words have truly undergone change; see §4.3. One example can be seen in the following sentence: \"My right hon. Friend last met M. Chevenement to discuss cooperative weapons development among other topics on 4 May.\" In this example, the model trained on text from the 1990s tags the word \"co-operative\" as an adjective whereas the model trained on the 1830s believes that \"co-operative\" is a noun; the 1990s model is correct in this case.",
+        "bbox": [
+            507,
+            199,
+            885,
+            375
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Masking and Confusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            393,
+            742,
+            407
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As before, we use both the 1830s model and the 1990s model to predict the POS tag of each word in held-out 1990s sentences. This time, we additionally repeat this process with one word in the sentence masked out: we ask both the 1990s model and the 1830s model to predict the POS tag of the masked word (without knowing what it is). We mask out each word in each sentence, one at a time; since the BERT embedding for a single word depends on the entire sentence, this requires passing the same sentence repeatedly through BERT with a different word masked each time. To speed up this process, we skip applying the mask to words that are too common (within the one thousand most common words in the 1990s), since these constitute the majority of word occurrences and yet are less likely to exhibit syntactic change.",
+        "bbox": [
+            507,
+            416,
+            884,
+            690
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After this process, each word in each sentence (except those we have deliberately skipped) will have four different contextual BERT embeddings: one by the 1830s model when the word was masked, one by the 1830s model when the word was unmasked, one by the 1990s model when the word was masked, and one by the 1990s model when the word was unmasked.",
+        "bbox": [
+            507,
+            692,
+            884,
+            820
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We now examine words for which the 1830s model gives a better prediction of the POS tag when the word is masked compared to when it is unmasked. When this happens, it means that the 1830s model correctly identified which tag to expect in the position of the masked word, and yet",
+        "bbox": [
+            507,
+            822,
+            884,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3567",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/8e1b734a3e8f9f745a0dab9b7fe9a79581b457fe1d249adea6f45db393b179fa.jpg",
+        "image_caption": [
+            "There is no example of sovereign countries which have given up their currencies. There is no example of [MASK] countries which have given up their currencies.",
+            "Figure 1: A depiction of the models' predictive process. The 1830s model predicts the correct tag when the word \"sovereign\" was masked (the tags assigned the highest likelihood after JJ were VBG with  $21\\%$  and DT with  $20\\%$  probability) but predicts it incorrectly when it was unmasked; this is the opposite of the usual pattern, in which the prediction is worse in the masked scenario. This type of tagging error indicates that the 1830s model expects \"sovereign\" to be a noun instead of an adjective."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            221,
+            112,
+            776,
+            260
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "after the mask is lifted (and the word itself was revealed) the 1830s model \"changed its mind\" and assigned the wrong tag instead. In such cases, we say that the 1830s model was \"confused\" by the token. We restrict to cases where the 1990s model correctly tagged the word, even when it was un-masked; this helps remove Trankit tagging errors as well as typographical errors in the text.",
+        "bbox": [
+            112,
+            366,
+            487,
+            495
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We then consider the remaining words for which the 1830s BERT model was confused more than once; specifically, we restrict to words for which there were at least two separate occurrences where the 1990s model tagged the word correctly (with a confidence of at least  $90\\%$ ) and the 1830s model assigned a probability that is at least 25 percentage points higher to the correct POS tag when seeing the masked version of the sentence than it did when seeing the unmasked version. We run this experiment on 100,000 test sentences; to train the 1990s and 1830s models, we use 25,000 sentences each (just as in the previous section). To help isolate interesting cases, we employ the same filters and the same sorting strategy as we did in §4.1.",
+        "bbox": [
+            112,
+            497,
+            487,
+            738
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "An example of a word that we found this way is \"material\", which was an adjective in the early 1800s but whose noun usage is modern. For instance, in the sentence \"the inquiry will be free to publish whatever material it wishes,\" Trankit correctly tags \"material\" as a noun, and so does the 1990s pipeline, but the 1830s pipeline only gives the correct tag when the word \"material\" is masked out in that sentence; when it is unmasked, an adjective is predicted.",
+        "bbox": [
+            112,
+            741,
+            487,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A list of the top resulting words (with exam-",
+        "bbox": [
+            132,
+            903,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ple sentences and tags provided by the models) is given in Appendix B. The words in this table were automatically generated with no manual curation.",
+        "bbox": [
+            507,
+            366,
+            880,
+            414
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            428,
+            640,
+            442
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the resulting words, we manually analyzed the top 20 candidates produced by §4.1 and the top 20 candidates produced by §4.2. For each candidate word, we examined 10 randomly sampled sentences containing that word from the 1990s and also 10 sentences from the 1830s, and we manually checked whether their syntax appeared to have changed.",
+        "bbox": [
+            505,
+            450,
+            882,
+            577
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Of the top 20 words from §4.1, seven (or  $35\\%$ ) were identified as having undergone syntactic change, with a part of speech in the 1990s that is either new or greatly increased in frequency. Of the top 20 words from §4.2, 11 (or  $55\\%$ ) were identified as having undergone syntactic change.",
+        "bbox": [
+            507,
+            579,
+            882,
+            675
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We note also that these two approaches identified different sets of words: Of the top 20 words on each list, there were only 4 words of overlap, while of the top 100 there were 15 words of overlap.",
+        "bbox": [
+            507,
+            676,
+            880,
+            740
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We also attempt an automatic evaluation by using Trankit to tag words from the 1830s and checking if these tags changed. Of course, our identification methods purposely avoided using a tagger on old text, since it may be unreliable, and a manual check reveals that the tagger made mistakes. Still, as a sanity check, we use it to compare the two methods for identifying syntactic change.",
+        "bbox": [
+            507,
+            741,
+            882,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For the purpose of automatic evaluation, we say that a word has changed and gained a new POS if the tag is at least twice as likely to appear for",
+        "bbox": [
+            507,
+            871,
+            880,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3568",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "1990s usages of the word than for 1830s usages of the word. Under this criterion for change, and using the Trankit tags for both 1830s and 1990s data, we get that the top 100 candidates from the simpler identification method (\\$4.1) contained 20 changed words, while the top 100 candidates from our masking strategy (\\$4.2) had 35 changed words. Note that of the manually evaluated words, the automatic evaluation agreed with  $75\\%$  of the human judgements. These should be interpreted with caution, but these results agree with our conclusion that the masking strategy is better at identifying changed words.",
+        "bbox": [
+            112,
+            84,
+            489,
+            294
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Confusion Scores at the Sentence Level",
+        "text_level": 1,
+        "bbox": [
+            112,
+            309,
+            485,
+            325
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this next experiment, we try to discern the extent to which masking out a word in a sentence makes the sentence more difficult for a fine-tuned BERT model to parse. The idea is that, for a typical word, once the word is masked out the model has less information about the sentence, and its syntactic understanding should therefore decrease; on the other hand, for a word that has an atypical or new usage, the model may end up being more confused after seeing the word than it is when the word is masked out. In the latter situation, we would expect the model's syntactic understanding of the sentence to improve if the word is masked out because, from the model's point of view, the word is misleading. A similar idea was used in §4.2 at the word level; in this section we will look at the accuracy of the tags for the whole sentence before and after a single word is masked.",
+        "bbox": [
+            112,
+            337,
+            489,
+            626
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For example, consider the following sentence: \"Opposition Members think that is how peace processes work.\" A model trained on English from the 1830s is likely to think that the word \"work\" is a noun rather than a verb. This could lead such a model to consider an incorrect way to parse this sentence, where the subject \"peace\" is doing the action of processing the object \"work\". In this case, if the word \"work\" is masked out, the sentence will be more likely to be parsed correctly.",
+        "bbox": [
+            110,
+            627,
+            489,
+            788
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As a proxy for the success of the parsing of the sentence, we feed in the sentence to the classifiers from §3 to get predicted tags for all the words in the sentence. We can use these classifiers to get probabilistic predictions, and in particular, we can get a probability assigned to the correct tag for each word in the sentence (using Trankit tags as ground truth). We then multiply together all",
+        "bbox": [
+            112,
+            790,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "these probabilities to get a measure of the overall probability the model assigns to the correct tags of the sentence. We can repeat this for both the 1830s and 1990s classifiers, and we can also repeat this when a specific word is either masked or unmasked.",
+        "bbox": [
+            507,
+            84,
+            884,
+            164
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We say a model is \"confused\" if the product of probabilities assigned to the correct (Trankit) tags is higher when the word is masked than when it is unmasked. For instance, in the above example sentence \"Opposition Members think that is how peace processes work,\" the 1830s model may assign a wrong tag to the word \"processes\" when the word \"work\" is unmasked, but it might correctly parse the sentence if the word \"work\" is masked (in the latter case, it will assume the masked word is an unknown verb, while in the former case it will try to parse the sentence with a noun in that position if it believes that \"work\" is a noun). In other words, if the 1830s model is confused by a word, it could be because this word underwent change from the 1830s to the 1990s. We create a \"confusion score\" by taking the logarithm of the ratio of the two aforementioned products (i.e. the product of probabilities assigned to the other words when the selected word is masked, and the corresponding product when the selected word is unmasked).",
+        "bbox": [
+            507,
+            167,
+            884,
+            502
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We can get such a confusion score for any word in any sentence, and for either the 1830s model or the 1990s model. We focus on sentences from the 1990s that were not part of the training data. Since calculating confusion scores is resource-intensive, we focus on a pre-selected set of word instances, such as those identified as likely having undergone syntactic change in §4.",
+        "bbox": [
+            507,
+            506,
+            885,
+            634
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Comparing Selected Words to Random Words",
+        "text_level": 1,
+        "bbox": [
+            507,
+            651,
+            865,
+            682
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We select 10,000 word occurrences for which the tag is predicted differently by the 1830s model and the 1990s model (as discussed in §4.1). For each word occurrence, we calculate a confusion score. We then compare these scores to the confusion scores of 10,000 randomly chosen word instances. Our goal is to check whether the confusion scores of the 10,000 selected words were higher than those of random words.",
+        "bbox": [
+            507,
+            692,
+            884,
+            835
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "One possible confounder is that the selected words may be rarer than the random words. To adjust for this, we choose the random words more carefully: for each non-random word in the first set, we pick a random word conditional on its overall",
+        "bbox": [
+            507,
+            839,
+            884,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3569",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/479a0c1717d54892947873d4695c0c65fa0259c0e4eb5ced8bb051b1633e021f.jpg",
+        "image_caption": [
+            "Figure 2: Distribution of confusion scores for selected words and random words."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            80,
+            484,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "frequency being close to that of the original word. We define close to be within 10 percentage points in the percentile of their frequencies. The resulting distribution of the confusion scores for the random words had lighter tails than the distribution of the selected words, especially on the right-hand side of the distribution. See Figure 2.",
+        "bbox": [
+            112,
+            316,
+            489,
+            429
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As the figure shows, the selected word instances (for which the 1990s model and 1830s model disagree on the tags) have a higher proportion of high confusion scores than random words. Therefore, calculating confusion scores for the selected word occurrences is a more resource-efficient way to find instances of high confusion scores.",
+        "bbox": [
+            112,
+            431,
+            489,
+            542
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Further Refinements",
+        "text_level": 1,
+        "bbox": [
+            112,
+            557,
+            326,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In addition to calculating confusion scores for a word in a given sentence using the 1830s model, we also calculate confusion scores in a similar way using the 1990s model. Recall that the confusion score tells us the extent to which the BERT-based tagger performs better when a word is masked out than when it is not. Since we have two such taggers (one fine-tuned on parliamentary speeches from the 1830s and one fine-tuned on speeches from the 1990s), we can get two confusion scores for a single word in a single sentence.",
+        "bbox": [
+            112,
+            580,
+            489,
+            755
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In order to zoom in on syntactic change, we restrict our attention to instances where the 1990s confusion score is less than or equal to zero. Since the confusion score is the logarithm of the ratio of masked to unmasked probabilities, this means we are restricting to word instances where the 1990s model gave a better prediction of the tags when the word was unmasked than it did when the word was masked. Applying this restriction helps to remove cases that are confusing for reasons other",
+        "bbox": [
+            112,
+            758,
+            490,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "than language change: for example, sentences that contain typographical errors, sentences that are incorrectly parsed by Trankit, sentences that were ungrammatical when originally spoken, or other special cases. By focusing on instances that the 1830s model found confusing but that the 1990s model did not find confusing, we increase the likelihood of finding syntactic change.",
+        "bbox": [
+            507,
+            84,
+            885,
+            214
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Manual Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            225,
+            709,
+            240
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We generate a list of top 100 candidates for syntactic change based on these confusion scores. Of those 100 sentences, our manual evaluation determined that 33 sentences were valid instances of syntactic change. We also computed the sum of the reciprocal ranks of the 33 manually selected sentences that exhibited syntactic change. This sum was 2.802, while the maximum possible sum of reciprocal ranks would be  $\\frac{1}{1} + \\frac{1}{2} + \\ldots + \\frac{1}{33} \\approx 4.089$ .",
+        "bbox": [
+            507,
+            248,
+            885,
+            395
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 Detecting the Rise of the Progressive",
+        "text_level": 1,
+        "bbox": [
+            507,
+            407,
+            858,
+            423
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we check the extent to which our techniques can identify syntactic change that was previously known to exist.",
+        "bbox": [
+            507,
+            434,
+            882,
+            481
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The frequency of the progressive has been increasing over time (e.g. Leech et al., 2009; Hou and Smith, 2021). An example of a progressive is \"he is reading\" (as opposed to \"he reads\"); the former has a form of \"to be\" followed by a verb that ends in \"-ing\". We identify progressive constructions using the dependency parse obtained from Trankit (Nguyen et al., 2021). We then mask out the present participle if it has the tag \"VBG\" and if the previous word is an inflection of the verb \"to be\". (This means that we mask out the tokens constituting the \"-ing\" word as well as the \"-ing\" suffix itself.)",
+        "bbox": [
+            505,
+            483,
+            885,
+            690
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Next, we use both the 1990s model and the 1830s model to predict the POS tags of all the words in these masked sentences. We compare these predictions to each other and also to the predictions that both models give for the unmasked versions of these sentences. Finally, we take randomly chosen words in randomly chosen sentences as a control group and repeat the process for this group.",
+        "bbox": [
+            507,
+            693,
+            885,
+            821
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We summarize the results in two figures. First, Figure 3 shows the confusion scores for the 1830s model on the progressive sentences compared to random sentences. As the figure shows, the 1830s model is negatively confused when we mask out progressives, which means that it gets the predic",
+        "bbox": [
+            507,
+            822,
+            885,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3570",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e4da15d8762aab18a0de5e880f8a1957059d101324d1d04ea001633ad1c4ea47.jpg",
+        "image_caption": [
+            "Figure 3: A histogram for the confusion scores of the 1830s model on progressives and on random sentences."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            80,
+            485,
+            253
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tion right more often on the unmasked version of these sentences than it does on the masked versions. This makes sense, as it corresponds to the insight that the 1830s model has a difficult time correctly tagging the masked progressives.",
+        "bbox": [
+            112,
+            317,
+            489,
+            398
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1610561e4206ac7e202022963eebc08063fe57eca2dc6d970b421f4d51c7bd97.jpg",
+        "image_caption": [
+            "Figure 4: A histogram for the logarithm of the ratios between the products of probabilities assigned to the true tags by the 1990s model and the 1830s model. We compare this for progressive sentences and random sentences. In all cases, both models were given the masked version of the sentences."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            131,
+            409,
+            470,
+            576
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 compares the predictions of the 1990s model to those of the 1830s model for the masked sentences. It shows that the 1990s model predicts the tags of the progressive sentences better than the 1830s model does, and the gap between their predictions is larger than it is for random sentences. This is evidence that the 1830s model is unfamiliar with progressive constructions.",
+        "bbox": [
+            112,
+            688,
+            489,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            828,
+            268,
+            844
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "It has been claimed that the performance of Transformers on downstream tasks is insensitive to word order (Hessel and Schofield, 2021; Sinha et al., 2021a; Sinha et al., 2021b; Pham et al., 2021;",
+        "bbox": [
+            112,
+            854,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Gupta et al., 2021), which is consistent with our observation that BERT fine-tuned on shuffled text can distinguish between two time periods with high accuracy; however, our experiments indicate BERT's accuracy still improves when it has access to word order, consistent with Papadimitriou et al. (2022), who argue that word order sometimes matters by considering non-prototypical cases where meanings and lexical expectations do not match.",
+        "bbox": [
+            507,
+            84,
+            884,
+            227
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A sizable literature considers syntactic probes for BERT embeddings (e.g. Tenney et al., 2019; Jawahar et al., 2019; Hewitt and Manning, 2019; Pimentel et al., 2020), but to our knowledge none of the works that probe syntax consider Transformers fine-tuned on different time periods.",
+        "bbox": [
+            507,
+            230,
+            884,
+            326
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Syntactic change detection has been understudied in the computational linguistics and NLP literature. One work on this topic is Merrill et al. (2019), in which an LSTM achieved higher accuracy compared to a model consisting of feedforward networks on the task of predicting the year of composition of sentences. However, in addition to predicting the time period of text, we also identify specific instances of syntactic change. Computational methods for the study of syntactic change also appear in Hou and Smith (2018); this work examined possible causes of the decline in the rate of passivization over time. Another related work is Lei and Wen (2020), which found dependency distance to be decreasing over time (between the years 1790 and 2017).",
+        "bbox": [
+            507,
+            328,
+            884,
+            583
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Many studies have used BERT to detect semantic change, including Giulianelli et al. (2020), Hu et al. (2019), and Fonteyn (2020). In addition, Beelen et al. (2021) fine-tune historical BERT models on a corpus of  $19^{\\mathrm{th}}$ -century English books in order to tackle the task of time-sensitive Targeted Sense Disambiguation and find that \"BERT works better when it is fine-tuned on data contemporaneous to the target period\".",
+        "bbox": [
+            507,
+            586,
+            882,
+            731
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "8 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            745,
+            640,
+            760
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We investigated the extent to which fine-tuned BERT models can distinguish between the syntax of 1830s English and 1990s English. Our approach used masking techniques, and we repeatedly compared BERT models on masked and unmasked text.",
+        "bbox": [
+            507,
+            772,
+            884,
+            852
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "First, in §2, we hid semantic information by replacing each word with its POS tag (and representing that tag as a common English word that exemplifies that tag). We fine-tuned BERT on such",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3571",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "syntax-only data to see if it could tell apart text from the 1830s and text from the 1990s. We found evidence that BERT can be used to detect syntactic differences between these two time periods.",
+        "bbox": [
+            112,
+            84,
+            487,
+            148
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In the later sections, we masked out individual words to see whether a POS tagger based on BERT embeddings (fine-tuned on either 1830s or 1990s) can still tag correctly when the word is masked. We filtered for a specific type of error, where the 1830s model predicts the correct tag when the word is masked but not when it is unmasked, and where the 1990s model is correct even when the word is unmasked. This procedure identified candidates for syntactic change, examples of which we provide in Appendix B. We were also able to confirm the rise of the progressive construction.",
+        "bbox": [
+            115,
+            149,
+            489,
+            341
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "An important point to highlight is that our identification of syntactic change relied on an automatic POS tagger only for the 1990s text; we avoided running any automatic tagging software on older English data. In fact, from §3 onwards, all we did with the older 1830s data was to fine-tune BERT on it in an unsupervised fashion. Despite this restriction, we were able to automatically identify words whose part of speech had undergone change between the two time periods.",
+        "bbox": [
+            112,
+            342,
+            489,
+            502
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            514,
+            218,
+            531
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "There are several limitations to this work. First, we worked only with English and only with text from the 1830s and 1990s. Our work also required the use of GPUs; see Appendix A for details of resources used. Although we were able to automatically identify specific instances of change (and we show automatically generated examples in Appendix B), our evaluation was limited to the manual judgement of one author, and we were unable to determine the recall of our detection system (because of the lack of objective criteria for what constitutes syntactic change). Additionally, our techniques are limited in the kinds of change we can detect: we focus on finding words whose typical part of speech has changed and on validating known syntactic changes such as the rise of the progressive. We hope that similar techniques can be extended to other types of syntactic change in future work.",
+        "bbox": [
+            112,
+            541,
+            489,
+            832
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            845,
+            278,
+            860
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We thank Jacob Eisenstein and the anonymous reviewers for their helpful feedback.",
+        "bbox": [
+            112,
+            871,
+            487,
+            903
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            83,
+            608,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Marc Alexander and Mark Davies. 2015. The Hansard Corpus 1803-2005. http://www.hansard-archive.parliament.uk/.",
+            "Kaspar Beelen, Federico Nanni, Mariona Coll Ardanuy, Kasra Hosseini, Giorgia Tolfo, and Barbara McGillivray. 2021. When time makes sense: A historically-aware approach to targeted sense disambiguation. In Findings of ACL-IJCNLP, pages 2751-2761.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In *NAACL*, pages 4171–4186.",
+            "Lauren Fonteyn. 2020. What about grammar? Using BERT embeddings to explore functional-semantic shifts of semi-lexical and grammatical constructions. Computational Humanities Research, pages 257-268.",
+            "Mario Giulianielli, Marco Del Tredici, and Raquel Fernandez. 2020. Analysing lexical semantic change with contextualised word representations. In ACL, pages 3960-3973.",
+            "Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In *NAACL*, pages 1195–1205.",
+            "Ashim Gupta, Giorgi Kvernadze, and Vivek Srikumar. 2021. BERT & family eat word salad: Experiments with text understanding. In AAAI, pages 12946-12954.",
+            "Jack Hessel and Alexandra Schofield. 2021. How effective is BERT without word ordering? implications for language understanding and data privacy. In ACLIJCNLP, pages 204-211.",
+            "John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In *NAACL*, pages 4129-4138.",
+            "Liwen Hou. 2022. Detecting and Modeling Syntactic Change. Ph.D. thesis, Northeastern University.",
+            "Liwen Hou and David A Smith. 2018. Modeling the decline in English passivization. In SCiL, pages 34-43.",
+            "Liwen Hou and David A. Smith. 2021. Drivers of English syntactic change in the Canadian parliament. In *SCiL*, pages 51–60.",
+            "Renfen Hu, Shen Li, and Shichen Liang. 2019. Diachronic sense modeling with deep contextualized word embeddings: An ecological view. In ACL, pages 3899-3908.",
+            "Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In ACL, pages 3651-3657."
+        ],
+        "bbox": [
+            509,
+            105,
+            885,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3572",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Geoffrey Leech, Marianne Hundt, Christian Mair, and Nicholas Smith. 2009. Change in Contemporary English: A Grammatical Study. Studies in English Language. Cambridge University Press.",
+            "Lei Lei and Ju Wen. 2020. Is dependency distance experiencing a process of minimization? A diachronic study based on the State of the Union addresses. *Lingua*, 239:102762.",
+            "Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In ACL System Demonstrations, pages 55-60.",
+            "William Merrill, Gigi Stark, and Robert Frank. 2019. Detecting syntactic change using a neural part-of-speech tagger. In International Workshop on Computational Approaches to Historical Language Change, pages 167-174.",
+            "Minh Van Nguyen, Viet Dac Lai, Amir Pouran Ben Veyseh, and Thien Huu Nguyen. 2021. Trankit: A lightweight transformer-based toolkit for multilingual natural language processing. In EACL System Demonstrations, pages 80-90.",
+            "Isabel Papadimitriou, Richard Futrell, and Kyle Mahowald. 2022. When classifying grammatical role, BERT doesn't care about word order... except when it matters. In ACL, pages 636-643.",
+            "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035.",
+            "Thang Pham, Trung Bui, Long Mai, and Anh Nguyen. 2021. Out of order: How important is the sequential order of words in a sentence in natural language understanding tasks? In Findings of ACL-IJCNLP, pages 1145-1160.",
+            "Tiago Pimentel, Josef Valvoda, Rowan Hall Maudslay, Ran Zmigrod, Adina Williams, and Ryan Cotterell. 2020. Information-theoretic probing for linguistic structure. In ACL, pages 4609-4622.",
+            "Koustuv Sinha, Robin Jia, Dieuwke Hupkes, Joelle Pineau, Adina Williams, and Douwe Kiela. 2021a. Masked language modeling and the distributional hypothesis: Order word matters pre-training for little. In EMNLP, pages 2888-2913.",
+            "Koustuv Sinha, Prasanna Parthasarathi, Joelle Pineau, and Adina Williams. 2021b. UnNatural Language Inference. In ACL-IJCNLP, pages 7329-7346.",
+            "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In ACL, pages 4593-4601."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "ref_text",
+        "text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In EMNLP System Demonstrations, pages 38-45.",
+        "bbox": [
+            510,
+            85,
+            882,
+            203
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Resources Used",
+        "text_level": 1,
+        "bbox": [
+            510,
+            214,
+            682,
+            228
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "In this work, we mentioned and used the following licensed resources: BERT (Devlin et al., 2019), which uses Apache 2.0 license; UK Parliament Hansard Corpus (Alexander and Davies, 2015), which uses Open Parliament Licence; Trankit large version 1.1.0 (Nguyen et al., 2021), which uses Apache 2.0 license; and Stanford CoreNLP version 4.2.2 (Manning et al., 2014), which uses the GNU General Public License v3.0. All of these licenses allow the type of research conducted in our work. Moreover, our work falls within the intended use of each of these resources.",
+        "bbox": [
+            507,
+            239,
+            882,
+            430
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Our work required GPU computation. We used NVIDIA Tesla V100 and P100 GPUs over several months while working on the results of this paper. Running Trankit was the slowest step and took around 72 GPU hours; the total GPU time to reproduce our final results would be under a week.",
+        "bbox": [
+            507,
+            432,
+            882,
+            527
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "B Automatically Identified Words",
+        "text_level": 1,
+        "bbox": [
+            509,
+            539,
+            816,
+            555
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 2 shows the top of the list of automatically identified words whose part-of-speech usage may have changed; a longer list can be found in Appendix A of the first author's thesis (Hou, 2022). For each word, we give an example sentence from the 1990s on which the 1830s model was \"confused\". For each example sentence, we give the Trankit tag and two sets of probabilistic tag predictions given by the 1830s model; one set corresponds to the tags predicted by the 1830s model for the highlighted word when the word was masked, and the other set corresponds to the tags predicted by the 1830s model for the highlighted word when the word was unmasked. (For brevity, we only list the tag predictions to which the model assigns at least  $30\\%$  probability.) An arrow such as 'JJ  $\\rightarrow$  Noun' indicates that the correct tag in the 1990s was 'Noun' but that the unmasked 1830s model put a high weight on 'JJ' instead; this suggests the word changed from 'JJ' in the 1830s to 'Noun' in the 1990s. (Recall that all POS tags for nouns were merged into a single 'Noun' category.)",
+        "bbox": [
+            507,
+            565,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3573",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/458e61b411ddf01828f69463869fa73eba1c8a66bbeb4709ffab62ba8b18addc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">opposite\nJJ→Noun</td><td colspan=\"3\">“The mandatory life sentence does the opposite of what it sets out to achieve, and it should go.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 84% Noun</td><td>unmasked tag: 93% JJ</td></tr><tr><td rowspan=\"2\">presume\nVB→VBP</td><td colspan=\"3\">“I presume that that shows they will be on message until they die.”</td></tr><tr><td>Trankit: VBP</td><td>masked tag: 98% VBP</td><td>unmasked tag: 38% VB, 32% VBP</td></tr><tr><td rowspan=\"2\">corporal\nNoun→JJ</td><td colspan=\"3\">“The school board wrote to me on 29 October advocating the reintroduction of corporal punishment in schools as a deterrent of last resort.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 74% JJ</td><td>unmasked tag: 95% Noun</td></tr><tr><td rowspan=\"2\">intrepid\nJJ→Noun</td><td colspan=\"3\">“The second category are the Royal Marine&#x27;s principal transport and landing ships, Fearless and Intrepid.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 64% Noun</td><td>unmasked tag: 97% JJ</td></tr><tr><td rowspan=\"2\">evil\nNoun→JJ</td><td colspan=\"3\">“The gunmen, in the evil atrocity that they committed, symbolised the past.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 82% JJ</td><td>unmasked tag: 99% Noun</td></tr><tr><td rowspan=\"2\">remembered\nJJ→VBN</td><td colspan=\"3\">“It should be remembered that, during the period 1974 – 79, when he was a senior Cabinet Minister, the steel industry was inefficient and unprofitable.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 86% VBN</td><td>unmasked tag: 64% JJ, 36% VBN</td></tr><tr><td rowspan=\"2\">shut\nVBD→VBN</td><td colspan=\"3\">“My second charge is that the Prime Minister has consistently talked about the incentive society with open opportunities for all, yet her policies have effectively shut out a quarter of the nation.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 97% VBN</td><td>unmasked tag: 86% VBD</td></tr><tr><td rowspan=\"2\">offending\nJJ→Noun</td><td colspan=\"3\">“I can vouch for the fact that juvenile offending is a major problem in my constituency.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 98% Noun</td><td>unmasked tag: 58% JJ</td></tr><tr><td rowspan=\"2\">granting\nVBG→Noun</td><td colspan=\"3\">“In most normal cases there will be the usual opportunity for consultation and granting of opinion from the local community.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 99% Noun</td><td>unmasked tag: 98% VBG</td></tr><tr><td rowspan=\"2\">disposed\nNoun→VBN</td><td colspan=\"3\">“Mr. Streeter: To ask the Secretary of State for Defence what Ministry of Defence land and buildings in Plymouth he expects to be surplus to MOD requirement in the next five years and to be disposed of to the private sector.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 87% VBN</td><td>unmasked tag: 48% Noun, 31% VBN</td></tr><tr><td rowspan=\"2\">exciting\nVBG→JJ</td><td colspan=\"3\">“We should also encourage housing associations to feel the exciting and exhilarating wind of competition.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 99% JJ</td><td>unmasked tag: 94% VBG</td></tr><tr><td rowspan=\"2\">laying\nVBG→Noun</td><td colspan=\"3\">“More significant orders, especially those exercising the so-called Henry VIII powers, are to be subject to laying in draft and affirmative resolution.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 98% Noun</td><td>unmasked tag: 98% VBG</td></tr><tr><td rowspan=\"2\">notwithstanding\nVBG→IN</td><td colspan=\"3\">&quot;Notwithstanding signal failures and operational difficulties, I look forward to using that fast service.”</td></tr><tr><td>Trankit: IN</td><td>masked tag: 84% IN</td><td>unmasked tag: 86% VBG</td></tr></table>",
+        "bbox": [
+            117,
+            166,
+            878,
+            804
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Automatically identified words whose part-of-speech usage may have changed",
+        "bbox": [
+            201,
+            814,
+            794,
+            829
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "3574",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2023/Detecting Syntactic Change with Pre-trained Transformer Models/5f1aa19b-edb6-4425-a7bb-8580ef6a41b6_model.json

@@ -0,0 +1,1839 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.157,
+                0.09,
+                0.843,
+                0.112
+            ],
+            "angle": 0,
+            "content": "Detecting Syntactic Change with Pre-trained Transformer Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.212,
+                0.143,
+                0.458,
+                0.193
+            ],
+            "angle": 0,
+            "content": "Liwen Hou  \nHarvard University  \nliwenhou@fas.harvard.edu"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.567,
+                0.143,
+                0.765,
+                0.193
+            ],
+            "angle": 0,
+            "content": "David A. Smith Northeastern University dasmith@ccs.neu.edu"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.144,
+                0.284,
+                0.461,
+                0.668
+            ],
+            "angle": 0,
+            "content": "We investigate the ability of Transformer-based language models to find syntactic differences between the English of the early 1800s and that of the late 1900s. First, we show that a fine-tuned BERT model can distinguish between text from these two periods using syntactic information only; to show this, we employ a strategy to hide semantic information from the text. Second, we make further use of fine-tuned BERT models to identify specific instances of syntactic change and specific words for which a new part of speech was introduced. To do this, we employ a part-of-speech (POS) tagger and use it to train corpora-specific taggers based only on BERT representations pretrained on different corpora. Notably, our methods of identifying specific candidates for syntactic change avoid using any automatic POS tagger on old text, where its performance may be unreliable; instead, our methods only use untagged old text together with tagged modern text. We examine samples and distributional properties of the model output to validate automatically identified cases of syntactic change. Finally, we use our techniques to confirm the historical rise of the progressive construction, a known example of syntactic change."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.683,
+                0.262,
+                0.699
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.71,
+                0.491,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Many researchers have taken advantage of the availability of pre-trained language models to study semantic change (e.g., Giulianelli et al., 2020; Hu et al., 2019; Fonteyn, 2020), but computational studies of syntactic change are more sporadic, relying on specially trained models (Merrill et al., 2019). In this paper, we investigate the extent to which BERT (Devlin et al., 2019) is able to distinguish between the syntax of English from the early 1800s and that of the late 1900s. To do so, we use the British Hansard corpus, which consists of a digitized version of the U.K. Parliamentary debates starting in the early 1800s and spanning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.254,
+                0.885,
+                0.316
+            ],
+            "angle": 0,
+            "content": "across two centuries. We isolate all of the parliamentary debates from the 1830s and the 1990s.<sup>1</sup> These debates can be found on the U.K. Parliament Website's Hansard archive<sup>2</sup>."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.319,
+                0.885,
+                0.415
+            ],
+            "angle": 0,
+            "content": "We fine-tune BERT-base-uncased, which has around 110 million parameters and was pre-trained on approximately 3.3 billion words. When training or testing BERT on a sentence, we first tokenize it using the original BERT-base-uncased tokenizer, as implemented by Hugging Face (Wolf et al., 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.417,
+                0.884,
+                0.512
+            ],
+            "angle": 0,
+            "content": "We focus on two questions: First, to what extent can BERT distinguish text from different time periods on the basis of syntactic features alone? Second, can we use BERT to help identify words or sentences whose analysis reflects syntactic change in an unsupervised way?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.514,
+                0.884,
+                0.625
+            ],
+            "angle": 0,
+            "content": "In §2, we find evidence that BERT can be used to detect syntactic differences between the English of the 1830s and that of the 1990s. Specifically, even after we control for confounders, a fine-tuned BERT model can still distinguish between these two periods even when the semantic information is hidden."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.627,
+                0.885,
+                0.869
+            ],
+            "angle": 0,
+            "content": "In §3, we describe how we train classifiers that can assign POS tags to words based on a 1830s and a 1990s understanding of English, respectively. Our main contribution comes in §4, where we use these classifiers to identify words whose analysis reflects syntactic change. Our method can automatically find words whose part of speech has changed. Moreover, our method avoids running a POS tagger on old text, since automatic tagging may be unreliable in that scenario. We emphasize that our change-detection techniques work with no manual annotation and identify examples of language change without any manual filtering; see Appendix B for a list of the top automatically detected words whose part of speech has changed."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.88,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>We chose these decades because they were the earliest and latest decade with large amounts of data in every year. <sup>2</sup>https://www.hansard-archive.parliament.uk/"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3564"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.78,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3564-3574 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.131
+            ],
+            "angle": 0,
+            "content": "To the best of our knowledge, such unsupervised syntactic change detection has not been previously done."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.134,
+                0.49,
+                0.23
+            ],
+            "angle": 0,
+            "content": "Finally, we confirm a known change (an increase in the frequency of progressives) using our classifiers. We show that a BERT model fine-tuned on text from the 1830s displays a poorer understanding of the progressive construction compared to a similar model trained on text from the 1990s."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.243,
+                0.444,
+                0.276
+            ],
+            "angle": 0,
+            "content": "2 Detecting Syntactic Change While Masking Semantics"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.286,
+                0.489,
+                0.366
+            ],
+            "angle": 0,
+            "content": "In this section, we check the extent to which BERT is sensitive to changes in the syntax of English over time. This acts as a sanity check for later sections, in which we use BERT in a different way to identify specific instances of syntactic change."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.367,
+                0.489,
+                0.478
+            ],
+            "angle": 0,
+            "content": "We start by fine-tuning BERT-base-uncased on the classification task of predicting whether sentences come from the 1830s or the 1990s. To do so, we use the representation of the CLS token in the last layer of BERT and add a linear layer to it; this model is then fine-tuned using labeled sentences from the 1830s and the 1990s."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.481,
+                0.489,
+                0.592
+            ],
+            "angle": 0,
+            "content": "After we fine-tune it on 100,000 sentences per time period, the BERT model achieves \\(98.6\\%\\) accuracy; however, it may use semantic information to distinguish between text from the two time periods, and we want to limit ourselves to syntactic information only in order to check the extent to which BERT can detect syntactic change over time."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.594,
+                0.489,
+                0.753
+            ],
+            "angle": 0,
+            "content": "To make BERT use syntactic information only, we mask semantic information by replacing all the open class words in each sentence with their POS tags, obtained using Stanford CoreNLP (Manning et al., 2014). In order to leverage BERT's pretrained knowledge of English, we then replace each POS tag with a common English word that usually has that tag (in a similar vein to Gulordava et al., 2018); this way, the input to BERT looks more like \"colorless green ideas\" and not \"JJ JJ NNS\"."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.755,
+                0.489,
+                0.883
+            ],
+            "angle": 0,
+            "content": "After replacing words (other than forms of \"to be\" and \"to have\", prepositions, and determiners) with POS tags and after replacing each POS tag with a representative word for that part of speech, we fine-tune BERT (with the added layer described above) on the resulting data. The BERT model performs surprisingly well on this classification task, achieving over \\(90\\%\\) accuracy for the task of pre"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.883,
+                0.213
+            ],
+            "angle": 0,
+            "content": "dicting whether a sentence comes from the 1830s or the 1990s. However, there are still several ways for such a model to use non-syntactic information. First, the model may use sentence lengths, since the distribution of the sentence lengths is different for the 1830s than for the 1990s. Second, the BERT model may simply be using the counts of the POS tags rather than real syntactic information."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.214,
+                0.884,
+                0.421
+            ],
+            "angle": 0,
+            "content": "To mitigate the length confounder, we place restrictions on the lengths of our training sentences. We run experiments separately for the length bins 11-15 words, 21-25 words, 31-35 words, and 20 words. To check the extent to which BERT relies on syntactic information other than the counts of tags, we separately train a model on shuffled input data, in which the sequence of tags in the sentence appears out of order. Such shuffled data retains no information other than the tag counts; we then compare whether BERT trained on unshuffled data can outperform BERT trained on shuffled data. The results are summarized in Table 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.424,
+                0.883,
+                0.599
+            ],
+            "angle": 0,
+            "content": "We compare these results to a logistic regression model that takes in the frequencies for the 100 most common vocabulary items in the replaced corpus\\(^{4}\\) (plus a catch-all category for the remaining items) and the sentence length as features. Such a logistic regression model achieves an accuracy of \\(85.45\\%\\), which means it performs about as well as BERT fine-tuned on shuffled tags but loses to BERT when the latter is fine-tuned on unshuffled tags. Using sentence length as the only feature yields an accuracy of \\(60.33\\%\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.6,
+                0.883,
+                0.809
+            ],
+            "angle": 0,
+            "content": "The conclusion of this experiment is that the fine-tuned BERT model must be using syntactic information to tell apart sentences from the 1830s and the 1990s. Indeed, even when the length of the sentence is fixed, and even when the words have been replaced by their parts of speech, the model gains advantage from seeing the true order of the tags as they appeared in the sentence, instead of seeing shuffled tags. This advantage must fundamentally come from differences in syntax that are present between the 1830s and 1990s sentences (and moreover, these differences must take a form that is not simply tag counts)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.81,
+                0.883,
+                0.873
+            ],
+            "angle": 0,
+            "content": "Having established that BERT is sensitive to changes in the syntax of English, we next train BERT in a different way to identify specific instances of such change. Importantly, from here on"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.881,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": "4 In addition to the typical words that represent different parts of speech, this replaced corpus also contains prepositions, determiners, and forms of \"to be\" and \"to have\"."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.893,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": "3The code for our experiments is available here: https://github.com/houliwen/detecting-syntactic-change"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3565"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.082,
+                0.486,
+                0.158
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>sent.\nlength</td><td>shuffled\nplaced</td><td>unshuffled\nplaced</td><td>shuffled\nplain</td><td>unshuffled\nplain</td></tr><tr><td>11-15</td><td>83.37</td><td>87.74</td><td>94.70</td><td>96.69</td></tr><tr><td>21-25</td><td>86.44</td><td>90.54</td><td>97.39</td><td>98.63</td></tr><tr><td>31-35</td><td>89.37</td><td>93.12</td><td>98.20</td><td>98.90</td></tr><tr><td>20</td><td>85.20</td><td>89.30</td><td>96.60</td><td>97.40</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.168,
+                0.49,
+                0.34
+            ],
+            "angle": 0,
+            "content": "Table 1: Accuracy of BERT fine-tuned on different types of data for predicting time period. All models were trained on 20,000 sentences and tested on 2,000 (half from the 1830s and half from the 1990s); the exception was the last line, for which only 10,000 training sentences were used due to lack of data. Since test data contained an even mixture of 1830s and 1990s sentences, baseline accuracy is \\(50\\%\\). These numbers are based on a single run; however, note that an accuracy of \\(p\\) will have standard deviation around \\(\\sqrt{p(1 - p) / 2000}\\), which ranges from \\(0.23\\%\\) to \\(0.83\\%\\) for the above, with higher accuracies having lower standard deviations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.365,
+                0.489,
+                0.382
+            ],
+            "angle": 0,
+            "content": "out, we avoid using a POS tagger on older English."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.394,
+                0.319,
+                0.41
+            ],
+            "angle": 0,
+            "content": "3 Training Classifiers"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.42,
+                0.49,
+                0.595
+            ],
+            "angle": 0,
+            "content": "We use BERT to identify specific instances of change. To do so, we train two POS taggers, one based on text from the 1830s and the other based on text from the 1990s; in the subsequent sections, we will use these taggers and their errors to identify words or sentences that display syntactic change. In this section, we will focus on describing how these taggers are trained and their performance. Crucially, we will train these taggers without ever tagging text from the 1830s; all POS tags are only ever associated with text from the 1990s."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.598,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We start by fine-tuning two instances of BERT-base-uncased: one on text from the 1990s and the other on text from the 1830s. This is done via Masked-Language Modeling in PyTorch (Paszke et al., 2019), where we process one sentence at a time with \\(15\\%\\) of the tokens masked, and the model is trained to predict the masked tokens. We use each fine-tuned model to obtain BERT token representations for 1990s text and then average the token representations for all the WordPiece tokens within each word to obtain a representation of a word. We do this twice: once with the model fine-tuned on text from the 1830s and once with the model fine-tuned on text from the 1990s (yielding two different embeddings of each word). However, even though we have models trained on two time periods, we use both models only to embed words from the 1990s and not any words from the 1830s. At the end of this process, each word from the 1990s has two different BERT embeddings."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.085,
+                0.885,
+                0.406
+            ],
+            "angle": 0,
+            "content": "Next, using POS tags obtained from Stanford CoreNLP for the 1990s portion of the British Hansard, we train a 2-layer perceptron (with one hidden layer of size 50) to assign a Penn Treebank tag to each BERT representation of each word. An input to this 2-layer perceptron is a BERT embedding of a word (extracted as described above), and the output is a POS tag. This classification task ignores punctuation and has 21 classes: We merge the 'NN', 'NNP', 'NNS', and 'NNPS' tags into one single tag called 'Noun', after which we take the 20 most common tags ('Noun', 'IN', 'DT', 'JJ', 'VB', 'PRP', 'RB', 'VBN', 'VBZ', 'VBP', 'CC', 'MD', 'TO', 'VBD', 'CD', 'VBG', 'PRPS', 'WDT', 'WP', 'WRB') and then group the remaining POS tags into a catch-all class called 'other'. We actually train two such perceptrons separately: one for the 1830s BERT embeddings (of words from the 1990s only) and one for the 1990s BERT embeddings (also only of words from the 1990s)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.408,
+                0.885,
+                0.583
+            ],
+            "angle": 0,
+            "content": "Note that the training and test data for this procedure consist only of sentences from the 1990s; this is done because the POS tags obtained from Stanford CoreNLP are likely to be more accurate on our modern data. Using the representations from the final layer of BERT (when the classifier has access to 50,000 training sentences, containing 815,578 words, and when the test set has 50,000 sentences, containing 816,428 words), the accuracy obtained by the 1990s model is \\(97.6\\%\\), and the accuracy of the 1830s model is \\(97.2\\%\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.586,
+                0.885,
+                0.81
+            ],
+            "angle": 0,
+            "content": "We repeat the above analysis using POS tags from Trankit (Nguyen et al., 2021) instead of Stanford CoreNLP because Trankit is a state-of-the-art parser and tagger. By using Trankit tags to train and test the above classifiers based on BERT representations, the accuracy obtained by the 1990s model was \\(98.4\\%\\) (instead of \\(97.6\\%\\) using CoreNLP tags) and the accuracy obtained by the 1830s model was \\(98.0\\%\\) (instead of \\(97.2\\%\\) using CoreNLP tags). The higher accuracy means that the task became easier when predicting Trankit tags instead of CoreNLP tags; since Trankit tags are also more accurate, it suggests that our pipeline is working as expected."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.812,
+                0.884,
+                0.844
+            ],
+            "angle": 0,
+            "content": "In §4 and later, we will use these pipelines (and the differences between their predictions) to iden-"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.855,
+                0.887,
+                0.88
+            ],
+            "angle": 0,
+            "content": "<sup>5</sup>We tried hidden layer sizes 100 and 200; this hyperparameter did not make a large difference."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.534,
+                0.881,
+                0.874,
+                0.894
+            ],
+            "angle": 0,
+            "content": "\\(^{6}\\)www.cis.upenn.edu/~bies/manuals/tagguide.pdf"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.511,
+                0.894,
+                0.853,
+                0.918
+            ],
+            "angle": 0,
+            "content": "\\(^{7}\\)https://trankit.readthedocs.io/en/latest/performance.html"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.508,
+                0.855,
+                0.887,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3566"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.085,
+                0.488,
+                0.117
+            ],
+            "angle": 0,
+            "content": "tify specific instances of syntactic change between the 1830s and the 1990s."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.131,
+                0.449,
+                0.166
+            ],
+            "angle": 0,
+            "content": "4 Identifying Words that Underwent Change"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.176,
+                0.49,
+                0.451
+            ],
+            "angle": 0,
+            "content": "We now attempt to use the POS taggers trained above to identify words that underwent change. We employ two methods to do this: First, we find the instances for which the two trained models disagree on a POS tag (indicating that the model familiar with text from the 1830s would tag the word differently compared to the one familiar with text from the 1990s). The second method involves masking out the target word to see if the 1830s model predicts the tag correctly on the masked word. If the masked word is tagged correctly but the un-masked word is tagged incorrectly, it suggests that the 1830s model can correctly parse the rest of the sentence but is specifically confused by the target word; when this happens, it indicates that the target word may have undergone change. This latter strategy is depicted in Figure 1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.462,
+                0.345,
+                0.478
+            ],
+            "angle": 0,
+            "content": "4.1 Tagging Disagreements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.485,
+                0.49,
+                0.773
+            ],
+            "angle": 0,
+            "content": "After training, we have two new POS taggers, one based on an 1830s understanding of English and one based on a 1990s understanding. We now run these taggers on test data (from the 1990s) to find words on which they disagree. In particular, we find words for which the model trained on text from the 1990s agrees with the Trankit tag but where the 1830s model disagrees. The cases where such a disagreement occurs should be the cases in which an 1830s grammar leads the model to incorrectly parse the sentence (likely due to a change in English grammar for the relevant word or perhaps one of its neighbors). Moreover, we use our classifiers to extract probabilistic predictions for the tags, and we search for word instances where probability assigned to the correct tag by the 1830s model was at least 50 percentage points lower than that assigned by the 1990s model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.775,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "We do this for each word in each sentence of the 1990s test data. To help isolate interesting cases, we filter the words under consideration to those that occur at least 10 times in the 1830s text (indicating that they are not new words). We remove numbers, punctuation, and words written entirely in capital letters (likely acronyms). We also restrict to words for which such a tagging disagreement occurred at least twice. We then sort the resulting words"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.198
+            ],
+            "angle": 0,
+            "content": "by the total number of tag disagreements for that word, normalized by the frequency of the word in the 1990s data, with additional weight given to words which were common in the 1830s (to reduce the number of rare words in the final list). We run this experiment on 100,000 test sentences, using classifiers trained on 25,000 sentences each."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.2,
+                0.887,
+                0.376
+            ],
+            "angle": 0,
+            "content": "We then manually examine the top resulting words to check whether the identified words have truly undergone change; see §4.3. One example can be seen in the following sentence: \"My right hon. Friend last met M. Chevenement to discuss cooperative weapons development among other topics on 4 May.\" In this example, the model trained on text from the 1990s tags the word \"co-operative\" as an adjective whereas the model trained on the 1830s believes that \"co-operative\" is a noun; the 1990s model is correct in this case."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.394,
+                0.744,
+                0.409
+            ],
+            "angle": 0,
+            "content": "4.2 Masking and Confusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.417,
+                0.885,
+                0.691
+            ],
+            "angle": 0,
+            "content": "As before, we use both the 1830s model and the 1990s model to predict the POS tag of each word in held-out 1990s sentences. This time, we additionally repeat this process with one word in the sentence masked out: we ask both the 1990s model and the 1830s model to predict the POS tag of the masked word (without knowing what it is). We mask out each word in each sentence, one at a time; since the BERT embedding for a single word depends on the entire sentence, this requires passing the same sentence repeatedly through BERT with a different word masked each time. To speed up this process, we skip applying the mask to words that are too common (within the one thousand most common words in the 1990s), since these constitute the majority of word occurrences and yet are less likely to exhibit syntactic change."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.693,
+                0.885,
+                0.821
+            ],
+            "angle": 0,
+            "content": "After this process, each word in each sentence (except those we have deliberately skipped) will have four different contextual BERT embeddings: one by the 1830s model when the word was masked, one by the 1830s model when the word was unmasked, one by the 1990s model when the word was masked, and one by the 1990s model when the word was unmasked."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.885,
+                0.921
+            ],
+            "angle": 0,
+            "content": "We now examine words for which the 1830s model gives a better prediction of the POS tag when the word is masked compared to when it is unmasked. When this happens, it means that the 1830s model correctly identified which tag to expect in the position of the masked word, and yet"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3567"
+        }
+    ],
+    [
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.219,
+                0.082,
+                0.747,
+                0.112
+            ],
+            "angle": 0,
+            "content": "There is no example of sovereign countries which have given up their currencies. There is no example of [MASK] countries which have given up their currencies."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.222,
+                0.113,
+                0.778,
+                0.261
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.271,
+                0.882,
+                0.342
+            ],
+            "angle": 0,
+            "content": "Figure 1: A depiction of the models' predictive process. The 1830s model predicts the correct tag when the word \"sovereign\" was masked (the tags assigned the highest likelihood after JJ were VBG with \\(21\\%\\) and DT with \\(20\\%\\) probability) but predicts it incorrectly when it was unmasked; this is the opposite of the usual pattern, in which the prediction is worse in the masked scenario. This type of tagging error indicates that the 1830s model expects \"sovereign\" to be a noun instead of an adjective."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.367,
+                0.489,
+                0.496
+            ],
+            "angle": 0,
+            "content": "after the mask is lifted (and the word itself was revealed) the 1830s model \"changed its mind\" and assigned the wrong tag instead. In such cases, we say that the 1830s model was \"confused\" by the token. We restrict to cases where the 1990s model correctly tagged the word, even when it was un-masked; this helps remove Trankit tagging errors as well as typographical errors in the text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.498,
+                0.489,
+                0.739
+            ],
+            "angle": 0,
+            "content": "We then consider the remaining words for which the 1830s BERT model was confused more than once; specifically, we restrict to words for which there were at least two separate occurrences where the 1990s model tagged the word correctly (with a confidence of at least \\(90\\%\\)) and the 1830s model assigned a probability that is at least 25 percentage points higher to the correct POS tag when seeing the masked version of the sentence than it did when seeing the unmasked version. We run this experiment on 100,000 test sentences; to train the 1990s and 1830s models, we use 25,000 sentences each (just as in the previous section). To help isolate interesting cases, we employ the same filters and the same sorting strategy as we did in §4.1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.742,
+                0.489,
+                0.902
+            ],
+            "angle": 0,
+            "content": "An example of a word that we found this way is \"material\", which was an adjective in the early 1800s but whose noun usage is modern. For instance, in the sentence \"the inquiry will be free to publish whatever material it wishes,\" Trankit correctly tags \"material\" as a noun, and so does the 1990s pipeline, but the 1830s pipeline only gives the correct tag when the word \"material\" is masked out in that sentence; when it is unmasked, an adjective is predicted."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.904,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "A list of the top resulting words (with exam-"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.367,
+                0.882,
+                0.415
+            ],
+            "angle": 0,
+            "content": "ple sentences and tags provided by the models) is given in Appendix B. The words in this table were automatically generated with no manual curation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.429,
+                0.642,
+                0.443
+            ],
+            "angle": 0,
+            "content": "4.3 Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.451,
+                0.884,
+                0.579
+            ],
+            "angle": 0,
+            "content": "To evaluate the resulting words, we manually analyzed the top 20 candidates produced by §4.1 and the top 20 candidates produced by §4.2. For each candidate word, we examined 10 randomly sampled sentences containing that word from the 1990s and also 10 sentences from the 1830s, and we manually checked whether their syntax appeared to have changed."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.58,
+                0.884,
+                0.676
+            ],
+            "angle": 0,
+            "content": "Of the top 20 words from §4.1, seven (or \\(35\\%\\)) were identified as having undergone syntactic change, with a part of speech in the 1990s that is either new or greatly increased in frequency. Of the top 20 words from §4.2, 11 (or \\(55\\%\\)) were identified as having undergone syntactic change."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.677,
+                0.882,
+                0.741
+            ],
+            "angle": 0,
+            "content": "We note also that these two approaches identified different sets of words: Of the top 20 words on each list, there were only 4 words of overlap, while of the top 100 there were 15 words of overlap."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.742,
+                0.884,
+                0.87
+            ],
+            "angle": 0,
+            "content": "We also attempt an automatic evaluation by using Trankit to tag words from the 1830s and checking if these tags changed. Of course, our identification methods purposely avoided using a tagger on old text, since it may be unreliable, and a manual check reveals that the tagger made mistakes. Still, as a sanity check, we use it to compare the two methods for identifying syntactic change."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.872,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "For the purpose of automatic evaluation, we say that a word has changed and gained a new POS if the tag is at least twice as likely to appear for"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3568"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.295
+            ],
+            "angle": 0,
+            "content": "1990s usages of the word than for 1830s usages of the word. Under this criterion for change, and using the Trankit tags for both 1830s and 1990s data, we get that the top 100 candidates from the simpler identification method (\\$4.1) contained 20 changed words, while the top 100 candidates from our masking strategy (\\$4.2) had 35 changed words. Note that of the manually evaluated words, the automatic evaluation agreed with \\(75\\%\\) of the human judgements. These should be interpreted with caution, but these results agree with our conclusion that the masking strategy is better at identifying changed words."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.31,
+                0.486,
+                0.326
+            ],
+            "angle": 0,
+            "content": "5 Confusion Scores at the Sentence Level"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.338,
+                0.49,
+                0.627
+            ],
+            "angle": 0,
+            "content": "In this next experiment, we try to discern the extent to which masking out a word in a sentence makes the sentence more difficult for a fine-tuned BERT model to parse. The idea is that, for a typical word, once the word is masked out the model has less information about the sentence, and its syntactic understanding should therefore decrease; on the other hand, for a word that has an atypical or new usage, the model may end up being more confused after seeing the word than it is when the word is masked out. In the latter situation, we would expect the model's syntactic understanding of the sentence to improve if the word is masked out because, from the model's point of view, the word is misleading. A similar idea was used in §4.2 at the word level; in this section we will look at the accuracy of the tags for the whole sentence before and after a single word is masked."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.629,
+                0.49,
+                0.789
+            ],
+            "angle": 0,
+            "content": "For example, consider the following sentence: \"Opposition Members think that is how peace processes work.\" A model trained on English from the 1830s is likely to think that the word \"work\" is a noun rather than a verb. This could lead such a model to consider an incorrect way to parse this sentence, where the subject \"peace\" is doing the action of processing the object \"work\". In this case, if the word \"work\" is masked out, the sentence will be more likely to be parsed correctly."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.791,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "As a proxy for the success of the parsing of the sentence, we feed in the sentence to the classifiers from §3 to get predicted tags for all the words in the sentence. We can use these classifiers to get probabilistic predictions, and in particular, we can get a probability assigned to the correct tag for each word in the sentence (using Trankit tags as ground truth). We then multiply together all"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.165
+            ],
+            "angle": 0,
+            "content": "these probabilities to get a measure of the overall probability the model assigns to the correct tags of the sentence. We can repeat this for both the 1830s and 1990s classifiers, and we can also repeat this when a specific word is either masked or unmasked."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.168,
+                0.885,
+                0.504
+            ],
+            "angle": 0,
+            "content": "We say a model is \"confused\" if the product of probabilities assigned to the correct (Trankit) tags is higher when the word is masked than when it is unmasked. For instance, in the above example sentence \"Opposition Members think that is how peace processes work,\" the 1830s model may assign a wrong tag to the word \"processes\" when the word \"work\" is unmasked, but it might correctly parse the sentence if the word \"work\" is masked (in the latter case, it will assume the masked word is an unknown verb, while in the former case it will try to parse the sentence with a noun in that position if it believes that \"work\" is a noun). In other words, if the 1830s model is confused by a word, it could be because this word underwent change from the 1830s to the 1990s. We create a \"confusion score\" by taking the logarithm of the ratio of the two aforementioned products (i.e. the product of probabilities assigned to the other words when the selected word is masked, and the corresponding product when the selected word is unmasked)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.507,
+                0.887,
+                0.636
+            ],
+            "angle": 0,
+            "content": "We can get such a confusion score for any word in any sentence, and for either the 1830s model or the 1990s model. We focus on sentences from the 1990s that were not part of the training data. Since calculating confusion scores is resource-intensive, we focus on a pre-selected set of word instances, such as those identified as likely having undergone syntactic change in §4."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.652,
+                0.866,
+                0.683
+            ],
+            "angle": 0,
+            "content": "5.1 Comparing Selected Words to Random Words"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.693,
+                0.885,
+                0.837
+            ],
+            "angle": 0,
+            "content": "We select 10,000 word occurrences for which the tag is predicted differently by the 1830s model and the 1990s model (as discussed in §4.1). For each word occurrence, we calculate a confusion score. We then compare these scores to the confusion scores of 10,000 randomly chosen word instances. Our goal is to check whether the confusion scores of the 10,000 selected words were higher than those of random words."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "One possible confounder is that the selected words may be rarer than the random words. To adjust for this, we choose the random words more carefully: for each non-random word in the first set, we pick a random word conditional on its overall"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3569"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.081,
+                0.485,
+                0.252
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.261,
+                0.489,
+                0.292
+            ],
+            "angle": 0,
+            "content": "Figure 2: Distribution of confusion scores for selected words and random words."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.317,
+                0.49,
+                0.43
+            ],
+            "angle": 0,
+            "content": "frequency being close to that of the original word. We define close to be within 10 percentage points in the percentile of their frequencies. The resulting distribution of the confusion scores for the random words had lighter tails than the distribution of the selected words, especially on the right-hand side of the distribution. See Figure 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.432,
+                0.49,
+                0.543
+            ],
+            "angle": 0,
+            "content": "As the figure shows, the selected word instances (for which the 1990s model and 1830s model disagree on the tags) have a higher proportion of high confusion scores than random words. Therefore, calculating confusion scores for the selected word occurrences is a more resource-efficient way to find instances of high confusion scores."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.558,
+                0.327,
+                0.572
+            ],
+            "angle": 0,
+            "content": "5.2 Further Refinements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.581,
+                0.49,
+                0.756
+            ],
+            "angle": 0,
+            "content": "In addition to calculating confusion scores for a word in a given sentence using the 1830s model, we also calculate confusion scores in a similar way using the 1990s model. Recall that the confusion score tells us the extent to which the BERT-based tagger performs better when a word is masked out than when it is not. Since we have two such taggers (one fine-tuned on parliamentary speeches from the 1830s and one fine-tuned on speeches from the 1990s), we can get two confusion scores for a single word in a single sentence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.491,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In order to zoom in on syntactic change, we restrict our attention to instances where the 1990s confusion score is less than or equal to zero. Since the confusion score is the logarithm of the ratio of masked to unmasked probabilities, this means we are restricting to word instances where the 1990s model gave a better prediction of the tags when the word was unmasked than it did when the word was masked. Applying this restriction helps to remove cases that are confusing for reasons other"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.886,
+                0.215
+            ],
+            "angle": 0,
+            "content": "than language change: for example, sentences that contain typographical errors, sentences that are incorrectly parsed by Trankit, sentences that were ungrammatical when originally spoken, or other special cases. By focusing on instances that the 1830s model found confusing but that the 1990s model did not find confusing, we increase the likelihood of finding syntactic change."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.227,
+                0.71,
+                0.241
+            ],
+            "angle": 0,
+            "content": "5.3 Manual Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.249,
+                0.886,
+                0.396
+            ],
+            "angle": 0,
+            "content": "We generate a list of top 100 candidates for syntactic change based on these confusion scores. Of those 100 sentences, our manual evaluation determined that 33 sentences were valid instances of syntactic change. We also computed the sum of the reciprocal ranks of the 33 manually selected sentences that exhibited syntactic change. This sum was 2.802, while the maximum possible sum of reciprocal ranks would be \\( \\frac{1}{1} + \\frac{1}{2} + \\ldots + \\frac{1}{33} \\approx 4.089 \\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.508,
+                0.408,
+                0.859,
+                0.424
+            ],
+            "angle": 0,
+            "content": "6 Detecting the Rise of the Progressive"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.435,
+                0.884,
+                0.482
+            ],
+            "angle": 0,
+            "content": "In this section, we check the extent to which our techniques can identify syntactic change that was previously known to exist."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.484,
+                0.886,
+                0.692
+            ],
+            "angle": 0,
+            "content": "The frequency of the progressive has been increasing over time (e.g. Leech et al., 2009; Hou and Smith, 2021). An example of a progressive is \"he is reading\" (as opposed to \"he reads\"); the former has a form of \"to be\" followed by a verb that ends in \"-ing\". We identify progressive constructions using the dependency parse obtained from Trankit (Nguyen et al., 2021). We then mask out the present participle if it has the tag \"VBG\" and if the previous word is an inflection of the verb \"to be\". (This means that we mask out the tokens constituting the \"-ing\" word as well as the \"-ing\" suffix itself.)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.694,
+                0.886,
+                0.822
+            ],
+            "angle": 0,
+            "content": "Next, we use both the 1990s model and the 1830s model to predict the POS tags of all the words in these masked sentences. We compare these predictions to each other and also to the predictions that both models give for the unmasked versions of these sentences. Finally, we take randomly chosen words in randomly chosen sentences as a control group and repeat the process for this group."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.824,
+                0.886,
+                0.92
+            ],
+            "angle": 0,
+            "content": "We summarize the results in two figures. First, Figure 3 shows the confusion scores for the 1830s model on the progressive sentences compared to random sentences. As the figure shows, the 1830s model is negatively confused when we mask out progressives, which means that it gets the predic"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3570"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.082,
+                0.486,
+                0.254
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.263,
+                0.488,
+                0.293
+            ],
+            "angle": 0,
+            "content": "Figure 3: A histogram for the confusion scores of the 1830s model on progressives and on random sentences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.318,
+                0.49,
+                0.399
+            ],
+            "angle": 0,
+            "content": "tion right more often on the unmasked version of these sentences than it does on the masked versions. This makes sense, as it corresponds to the insight that the 1830s model has a difficult time correctly tagging the masked progressives."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.132,
+                0.41,
+                0.472,
+                0.577
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.587,
+                0.49,
+                0.673
+            ],
+            "angle": 0,
+            "content": "Figure 4: A histogram for the logarithm of the ratios between the products of probabilities assigned to the true tags by the 1990s model and the 1830s model. We compare this for progressive sentences and random sentences. In all cases, both models were given the masked version of the sentences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.689,
+                0.49,
+                0.817
+            ],
+            "angle": 0,
+            "content": "Figure 4 compares the predictions of the 1990s model to those of the 1830s model for the masked sentences. It shows that the 1990s model predicts the tags of the progressive sentences better than the 1830s model does, and the gap between their predictions is larger than it is for random sentences. This is evidence that the 1830s model is unfamiliar with progressive constructions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.829,
+                0.27,
+                0.845
+            ],
+            "angle": 0,
+            "content": "7 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.855,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "It has been claimed that the performance of Transformers on downstream tasks is insensitive to word order (Hessel and Schofield, 2021; Sinha et al., 2021a; Sinha et al., 2021b; Pham et al., 2021;"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.228
+            ],
+            "angle": 0,
+            "content": "Gupta et al., 2021), which is consistent with our observation that BERT fine-tuned on shuffled text can distinguish between two time periods with high accuracy; however, our experiments indicate BERT's accuracy still improves when it has access to word order, consistent with Papadimitriou et al. (2022), who argue that word order sometimes matters by considering non-prototypical cases where meanings and lexical expectations do not match."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.231,
+                0.885,
+                0.327
+            ],
+            "angle": 0,
+            "content": "A sizable literature considers syntactic probes for BERT embeddings (e.g. Tenney et al., 2019; Jawahar et al., 2019; Hewitt and Manning, 2019; Pimentel et al., 2020), but to our knowledge none of the works that probe syntax consider Transformers fine-tuned on different time periods."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.329,
+                0.885,
+                0.585
+            ],
+            "angle": 0,
+            "content": "Syntactic change detection has been understudied in the computational linguistics and NLP literature. One work on this topic is Merrill et al. (2019), in which an LSTM achieved higher accuracy compared to a model consisting of feedforward networks on the task of predicting the year of composition of sentences. However, in addition to predicting the time period of text, we also identify specific instances of syntactic change. Computational methods for the study of syntactic change also appear in Hou and Smith (2018); this work examined possible causes of the decline in the rate of passivization over time. Another related work is Lei and Wen (2020), which found dependency distance to be decreasing over time (between the years 1790 and 2017)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.587,
+                0.884,
+                0.732
+            ],
+            "angle": 0,
+            "content": "Many studies have used BERT to detect semantic change, including Giulianelli et al. (2020), Hu et al. (2019), and Fonteyn (2020). In addition, Beelen et al. (2021) fine-tune historical BERT models on a corpus of \\(19^{\\mathrm{th}}\\)-century English books in order to tackle the task of time-sensitive Targeted Sense Disambiguation and find that \"BERT works better when it is fine-tuned on data contemporaneous to the target period\"."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.746,
+                0.642,
+                0.761
+            ],
+            "angle": 0,
+            "content": "8 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.774,
+                0.885,
+                0.853
+            ],
+            "angle": 0,
+            "content": "We investigated the extent to which fine-tuned BERT models can distinguish between the syntax of 1830s English and 1990s English. Our approach used masking techniques, and we repeatedly compared BERT models on masked and unmasked text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.855,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "First, in §2, we hid semantic information by replacing each word with its POS tag (and representing that tag as a common English word that exemplifies that tag). We fine-tuned BERT on such"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3571"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.489,
+                0.149
+            ],
+            "angle": 0,
+            "content": "syntax-only data to see if it could tell apart text from the 1830s and text from the 1990s. We found evidence that BERT can be used to detect syntactic differences between these two time periods."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.151,
+                0.49,
+                0.342
+            ],
+            "angle": 0,
+            "content": "In the later sections, we masked out individual words to see whether a POS tagger based on BERT embeddings (fine-tuned on either 1830s or 1990s) can still tag correctly when the word is masked. We filtered for a specific type of error, where the 1830s model predicts the correct tag when the word is masked but not when it is unmasked, and where the 1990s model is correct even when the word is unmasked. This procedure identified candidates for syntactic change, examples of which we provide in Appendix B. We were also able to confirm the rise of the progressive construction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.343,
+                0.49,
+                0.504
+            ],
+            "angle": 0,
+            "content": "An important point to highlight is that our identification of syntactic change relied on an automatic POS tagger only for the 1990s text; we avoided running any automatic tagging software on older English data. In fact, from §3 onwards, all we did with the older 1830s data was to fine-tune BERT on it in an unsupervised fashion. Despite this restriction, we were able to automatically identify words whose part of speech had undergone change between the two time periods."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.516,
+                0.22,
+                0.532
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.542,
+                0.49,
+                0.833
+            ],
+            "angle": 0,
+            "content": "There are several limitations to this work. First, we worked only with English and only with text from the 1830s and 1990s. Our work also required the use of GPUs; see Appendix A for details of resources used. Although we were able to automatically identify specific instances of change (and we show automatically generated examples in Appendix B), our evaluation was limited to the manual judgement of one author, and we were unable to determine the recall of our detection system (because of the lack of objective criteria for what constitutes syntactic change). Additionally, our techniques are limited in the kinds of change we can detect: we focus on finding words whose typical part of speech has changed and on validating known syntactic changes such as the rise of the progressive. We hope that similar techniques can be extended to other types of syntactic change in future work."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.846,
+                0.279,
+                0.862
+            ],
+            "angle": 0,
+            "content": "Acknowledgments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.872,
+                0.489,
+                0.904
+            ],
+            "angle": 0,
+            "content": "We thank Jacob Eisenstein and the anonymous reviewers for their helpful feedback."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.084,
+                0.61,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.107,
+                0.886,
+                0.147
+            ],
+            "angle": 0,
+            "content": "Marc Alexander and Mark Davies. 2015. The Hansard Corpus 1803-2005. http://www.hansard-archive.parliament.uk/."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.156,
+                0.886,
+                0.236
+            ],
+            "angle": 0,
+            "content": "Kaspar Beelen, Federico Nanni, Mariona Coll Ardanuy, Kasra Hosseini, Giorgia Tolfo, and Barbara McGillivray. 2021. When time makes sense: A historically-aware approach to targeted sense disambiguation. In Findings of ACL-IJCNLP, pages 2751-2761."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.244,
+                0.885,
+                0.299
+            ],
+            "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 *NAACL*, pages 4171–4186."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.307,
+                0.886,
+                0.373
+            ],
+            "angle": 0,
+            "content": "Lauren Fonteyn. 2020. What about grammar? Using BERT embeddings to explore functional-semantic shifts of semi-lexical and grammatical constructions. Computational Humanities Research, pages 257-268."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.382,
+                0.885,
+                0.436
+            ],
+            "angle": 0,
+            "content": "Mario Giulianielli, Marco Del Tredici, and Raquel Fernandez. 2020. Analysing lexical semantic change with contextualised word representations. In ACL, pages 3960-3973."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.445,
+                0.885,
+                0.499
+            ],
+            "angle": 0,
+            "content": "Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In *NAACL*, pages 1195–1205."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.507,
+                0.886,
+                0.56
+            ],
+            "angle": 0,
+            "content": "Ashim Gupta, Giorgi Kvernadze, and Vivek Srikumar. 2021. BERT & family eat word salad: Experiments with text understanding. In AAAI, pages 12946-12954."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.569,
+                0.885,
+                0.624
+            ],
+            "angle": 0,
+            "content": "Jack Hessel and Alexandra Schofield. 2021. How effective is BERT without word ordering? implications for language understanding and data privacy. In ACLIJCNLP, pages 204-211."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.632,
+                0.885,
+                0.673
+            ],
+            "angle": 0,
+            "content": "John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In *NAACL*, pages 4129-4138."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.681,
+                0.885,
+                0.71
+            ],
+            "angle": 0,
+            "content": "Liwen Hou. 2022. Detecting and Modeling Syntactic Change. Ph.D. thesis, Northeastern University."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.717,
+                0.885,
+                0.758
+            ],
+            "angle": 0,
+            "content": "Liwen Hou and David A Smith. 2018. Modeling the decline in English passivization. In SCiL, pages 34-43."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.766,
+                0.885,
+                0.808
+            ],
+            "angle": 0,
+            "content": "Liwen Hou and David A. Smith. 2021. Drivers of English syntactic change in the Canadian parliament. In *SCiL*, pages 51–60."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.816,
+                0.885,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Renfen Hu, Shen Li, and Shichen Liang. 2019. Diachronic sense modeling with deep contextualized word embeddings: An ecological view. In ACL, pages 3899-3908."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.878,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In ACL, pages 3651-3657."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.51,
+                0.107,
+                0.886,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3572"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.086,
+                0.487,
+                0.139
+            ],
+            "angle": 0,
+            "content": "Geoffrey Leech, Marianne Hundt, Christian Mair, and Nicholas Smith. 2009. Change in Contemporary English: A Grammatical Study. Studies in English Language. Cambridge University Press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.147,
+                0.488,
+                0.2
+            ],
+            "angle": 0,
+            "content": "Lei Lei and Ju Wen. 2020. Is dependency distance experiencing a process of minimization? A diachronic study based on the State of the Union addresses. *Lingua*, 239:102762."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.209,
+                0.488,
+                0.275
+            ],
+            "angle": 0,
+            "content": "Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In ACL System Demonstrations, pages 55-60."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.283,
+                0.488,
+                0.349
+            ],
+            "angle": 0,
+            "content": "William Merrill, Gigi Stark, and Robert Frank. 2019. Detecting syntactic change using a neural part-of-speech tagger. In International Workshop on Computational Approaches to Historical Language Change, pages 167-174."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.358,
+                0.488,
+                0.423
+            ],
+            "angle": 0,
+            "content": "Minh Van Nguyen, Viet Dac Lai, Amir Pouran Ben Veyseh, and Thien Huu Nguyen. 2021. Trankit: A lightweight transformer-based toolkit for multilingual natural language processing. In EACL System Demonstrations, pages 80-90."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.432,
+                0.488,
+                0.485
+            ],
+            "angle": 0,
+            "content": "Isabel Papadimitriou, Richard Futrell, and Kyle Mahowald. 2022. When classifying grammatical role, BERT doesn't care about word order... except when it matters. In ACL, pages 636-643."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.493,
+                0.488,
+                0.611
+            ],
+            "angle": 0,
+            "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.62,
+                0.488,
+                0.686
+            ],
+            "angle": 0,
+            "content": "Thang Pham, Trung Bui, Long Mai, and Anh Nguyen. 2021. Out of order: How important is the sequential order of words in a sentence in natural language understanding tasks? In Findings of ACL-IJCNLP, pages 1145-1160."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.694,
+                0.488,
+                0.747
+            ],
+            "angle": 0,
+            "content": "Tiago Pimentel, Josef Valvoda, Rowan Hall Maudslay, Ran Zmigrod, Adina Williams, and Ryan Cotterell. 2020. Information-theoretic probing for linguistic structure. In ACL, pages 4609-4622."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.755,
+                0.488,
+                0.821
+            ],
+            "angle": 0,
+            "content": "Koustuv Sinha, Robin Jia, Dieuwke Hupkes, Joelle Pineau, Adina Williams, and Douwe Kiela. 2021a. Masked language modeling and the distributional hypothesis: Order word matters pre-training for little. In EMNLP, pages 2888-2913."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.83,
+                0.488,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Koustuv Sinha, Prasanna Parthasarathi, Joelle Pineau, and Adina Williams. 2021b. UnNatural Language Inference. In ACL-IJCNLP, pages 7329-7346."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.878,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In ACL, pages 4593-4601."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.086,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.204
+            ],
+            "angle": 0,
+            "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In EMNLP System Demonstrations, pages 38-45."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.215,
+                0.684,
+                0.229
+            ],
+            "angle": 0,
+            "content": "A Resources Used"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.24,
+                0.884,
+                0.431
+            ],
+            "angle": 0,
+            "content": "In this work, we mentioned and used the following licensed resources: BERT (Devlin et al., 2019), which uses Apache 2.0 license; UK Parliament Hansard Corpus (Alexander and Davies, 2015), which uses Open Parliament Licence; Trankit large version 1.1.0 (Nguyen et al., 2021), which uses Apache 2.0 license; and Stanford CoreNLP version 4.2.2 (Manning et al., 2014), which uses the GNU General Public License v3.0. All of these licenses allow the type of research conducted in our work. Moreover, our work falls within the intended use of each of these resources."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.433,
+                0.884,
+                0.528
+            ],
+            "angle": 0,
+            "content": "Our work required GPU computation. We used NVIDIA Tesla V100 and P100 GPUs over several months while working on the results of this paper. Running Trankit was the slowest step and took around 72 GPU hours; the total GPU time to reproduce our final results would be under a week."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.541,
+                0.818,
+                0.556
+            ],
+            "angle": 0,
+            "content": "B Automatically Identified Words"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.566,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Table 2 shows the top of the list of automatically identified words whose part-of-speech usage may have changed; a longer list can be found in Appendix A of the first author's thesis (Hou, 2022). For each word, we give an example sentence from the 1990s on which the 1830s model was \"confused\". For each example sentence, we give the Trankit tag and two sets of probabilistic tag predictions given by the 1830s model; one set corresponds to the tags predicted by the 1830s model for the highlighted word when the word was masked, and the other set corresponds to the tags predicted by the 1830s model for the highlighted word when the word was unmasked. (For brevity, we only list the tag predictions to which the model assigns at least \\(30\\%\\) probability.) An arrow such as 'JJ \\(\\rightarrow\\) Noun' indicates that the correct tag in the 1990s was 'Noun' but that the unmasked 1830s model put a high weight on 'JJ' instead; this suggests the word changed from 'JJ' in the 1830s to 'Noun' in the 1990s. (Recall that all POS tags for nouns were merged into a single 'Noun' category.)"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3573"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.118,
+                0.167,
+                0.88,
+                0.805
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">opposite\nJJ→Noun</td><td colspan=\"3\">“The mandatory life sentence does the opposite of what it sets out to achieve, and it should go.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 84% Noun</td><td>unmasked tag: 93% JJ</td></tr><tr><td rowspan=\"2\">presume\nVB→VBP</td><td colspan=\"3\">“I presume that that shows they will be on message until they die.”</td></tr><tr><td>Trankit: VBP</td><td>masked tag: 98% VBP</td><td>unmasked tag: 38% VB, 32% VBP</td></tr><tr><td rowspan=\"2\">corporal\nNoun→JJ</td><td colspan=\"3\">“The school board wrote to me on 29 October advocating the reintroduction of corporal punishment in schools as a deterrent of last resort.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 74% JJ</td><td>unmasked tag: 95% Noun</td></tr><tr><td rowspan=\"2\">intrepid\nJJ→Noun</td><td colspan=\"3\">“The second category are the Royal Marine&#x27;s principal transport and landing ships, Fearless and Intrepid.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 64% Noun</td><td>unmasked tag: 97% JJ</td></tr><tr><td rowspan=\"2\">evil\nNoun→JJ</td><td colspan=\"3\">“The gunmen, in the evil atrocity that they committed, symbolised the past.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 82% JJ</td><td>unmasked tag: 99% Noun</td></tr><tr><td rowspan=\"2\">remembered\nJJ→VBN</td><td colspan=\"3\">“It should be remembered that, during the period 1974 – 79, when he was a senior Cabinet Minister, the steel industry was inefficient and unprofitable.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 86% VBN</td><td>unmasked tag: 64% JJ, 36% VBN</td></tr><tr><td rowspan=\"2\">shut\nVBD→VBN</td><td colspan=\"3\">“My second charge is that the Prime Minister has consistently talked about the incentive society with open opportunities for all, yet her policies have effectively shut out a quarter of the nation.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 97% VBN</td><td>unmasked tag: 86% VBD</td></tr><tr><td rowspan=\"2\">offending\nJJ→Noun</td><td colspan=\"3\">“I can vouch for the fact that juvenile offending is a major problem in my constituency.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 98% Noun</td><td>unmasked tag: 58% JJ</td></tr><tr><td rowspan=\"2\">granting\nVBG→Noun</td><td colspan=\"3\">“In most normal cases there will be the usual opportunity for consultation and granting of opinion from the local community.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 99% Noun</td><td>unmasked tag: 98% VBG</td></tr><tr><td rowspan=\"2\">disposed\nNoun→VBN</td><td colspan=\"3\">“Mr. Streeter: To ask the Secretary of State for Defence what Ministry of Defence land and buildings in Plymouth he expects to be surplus to MOD requirement in the next five years and to be disposed of to the private sector.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 87% VBN</td><td>unmasked tag: 48% Noun, 31% VBN</td></tr><tr><td rowspan=\"2\">exciting\nVBG→JJ</td><td colspan=\"3\">“We should also encourage housing associations to feel the exciting and exhilarating wind of competition.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 99% JJ</td><td>unmasked tag: 94% VBG</td></tr><tr><td rowspan=\"2\">laying\nVBG→Noun</td><td colspan=\"3\">“More significant orders, especially those exercising the so-called Henry VIII powers, are to be subject to laying in draft and affirmative resolution.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 98% Noun</td><td>unmasked tag: 98% VBG</td></tr><tr><td rowspan=\"2\">notwithstanding\nVBG→IN</td><td colspan=\"3\">&quot;Notwithstanding signal failures and operational difficulties, I look forward to using that fast service.”</td></tr><tr><td>Trankit: IN</td><td>masked tag: 84% IN</td><td>unmasked tag: 86% VBG</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.202,
+                0.815,
+                0.795,
+                0.83
+            ],
+            "angle": 0,
+            "content": "Table 2: Automatically identified words whose part-of-speech usage may have changed"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "3574"
+        }
+    ]
+]

2023/Detecting Syntactic Change with Pre-trained Transformer Models/5f1aa19b-edb6-4425-a7bb-8580ef6a41b6_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00c79584a7d7fa2eef79ff00a3d2793328a6a6a1fc7c907a56fe61840d5a48a1
+size 472916

2023/Detecting Syntactic Change with Pre-trained Transformer Models/full.md

@@ -0,0 +1,278 @@
+# Detecting Syntactic Change with Pre-trained Transformer Models
+
+Liwen Hou  
+Harvard University  
+liwenhou@fas.harvard.edu
+
+David A. Smith Northeastern University dasmith@ccs.neu.edu
+
+# Abstract
+
+We investigate the ability of Transformer-based language models to find syntactic differences between the English of the early 1800s and that of the late 1900s. First, we show that a fine-tuned BERT model can distinguish between text from these two periods using syntactic information only; to show this, we employ a strategy to hide semantic information from the text. Second, we make further use of fine-tuned BERT models to identify specific instances of syntactic change and specific words for which a new part of speech was introduced. To do this, we employ a part-of-speech (POS) tagger and use it to train corpora-specific taggers based only on BERT representations pretrained on different corpora. Notably, our methods of identifying specific candidates for syntactic change avoid using any automatic POS tagger on old text, where its performance may be unreliable; instead, our methods only use untagged old text together with tagged modern text. We examine samples and distributional properties of the model output to validate automatically identified cases of syntactic change. Finally, we use our techniques to confirm the historical rise of the progressive construction, a known example of syntactic change.
+
+# 1 Introduction
+
+Many researchers have taken advantage of the availability of pre-trained language models to study semantic change (e.g., Giulianelli et al., 2020; Hu et al., 2019; Fonteyn, 2020), but computational studies of syntactic change are more sporadic, relying on specially trained models (Merrill et al., 2019). In this paper, we investigate the extent to which BERT (Devlin et al., 2019) is able to distinguish between the syntax of English from the early 1800s and that of the late 1900s. To do so, we use the British Hansard corpus, which consists of a digitized version of the U.K. Parliamentary debates starting in the early 1800s and spanning
+
+across two centuries. We isolate all of the parliamentary debates from the 1830s and the 1990s.<sup>1</sup> These debates can be found on the U.K. Parliament Website's Hansard archive<sup>2</sup>.
+
+We fine-tune BERT-base-uncased, which has around 110 million parameters and was pre-trained on approximately 3.3 billion words. When training or testing BERT on a sentence, we first tokenize it using the original BERT-base-uncased tokenizer, as implemented by Hugging Face (Wolf et al., 2020).
+
+We focus on two questions: First, to what extent can BERT distinguish text from different time periods on the basis of syntactic features alone? Second, can we use BERT to help identify words or sentences whose analysis reflects syntactic change in an unsupervised way?
+
+In §2, we find evidence that BERT can be used to detect syntactic differences between the English of the 1830s and that of the 1990s. Specifically, even after we control for confounders, a fine-tuned BERT model can still distinguish between these two periods even when the semantic information is hidden.
+
+In §3, we describe how we train classifiers that can assign POS tags to words based on a 1830s and a 1990s understanding of English, respectively. Our main contribution comes in §4, where we use these classifiers to identify words whose analysis reflects syntactic change. Our method can automatically find words whose part of speech has changed. Moreover, our method avoids running a POS tagger on old text, since automatic tagging may be unreliable in that scenario. We emphasize that our change-detection techniques work with no manual annotation and identify examples of language change without any manual filtering; see Appendix B for a list of the top automatically detected words whose part of speech has changed.
+
+To the best of our knowledge, such unsupervised syntactic change detection has not been previously done.
+
+Finally, we confirm a known change (an increase in the frequency of progressives) using our classifiers. We show that a BERT model fine-tuned on text from the 1830s displays a poorer understanding of the progressive construction compared to a similar model trained on text from the 1990s.
+
+# 2 Detecting Syntactic Change While Masking Semantics
+
+In this section, we check the extent to which BERT is sensitive to changes in the syntax of English over time. This acts as a sanity check for later sections, in which we use BERT in a different way to identify specific instances of syntactic change.
+
+We start by fine-tuning BERT-base-uncased on the classification task of predicting whether sentences come from the 1830s or the 1990s. To do so, we use the representation of the CLS token in the last layer of BERT and add a linear layer to it; this model is then fine-tuned using labeled sentences from the 1830s and the 1990s.
+
+After we fine-tune it on 100,000 sentences per time period, the BERT model achieves  $98.6\%$  accuracy; however, it may use semantic information to distinguish between text from the two time periods, and we want to limit ourselves to syntactic information only in order to check the extent to which BERT can detect syntactic change over time.
+
+To make BERT use syntactic information only, we mask semantic information by replacing all the open class words in each sentence with their POS tags, obtained using Stanford CoreNLP (Manning et al., 2014). In order to leverage BERT's pretrained knowledge of English, we then replace each POS tag with a common English word that usually has that tag (in a similar vein to Gulordava et al., 2018); this way, the input to BERT looks more like "colorless green ideas" and not "JJ JJ NNS".
+
+After replacing words (other than forms of "to be" and "to have", prepositions, and determiners) with POS tags and after replacing each POS tag with a representative word for that part of speech, we fine-tune BERT (with the added layer described above) on the resulting data. The BERT model performs surprisingly well on this classification task, achieving over  $90\%$  accuracy for the task of pre
+
+dicting whether a sentence comes from the 1830s or the 1990s. However, there are still several ways for such a model to use non-syntactic information. First, the model may use sentence lengths, since the distribution of the sentence lengths is different for the 1830s than for the 1990s. Second, the BERT model may simply be using the counts of the POS tags rather than real syntactic information.
+
+To mitigate the length confounder, we place restrictions on the lengths of our training sentences. We run experiments separately for the length bins 11-15 words, 21-25 words, 31-35 words, and 20 words. To check the extent to which BERT relies on syntactic information other than the counts of tags, we separately train a model on shuffled input data, in which the sequence of tags in the sentence appears out of order. Such shuffled data retains no information other than the tag counts; we then compare whether BERT trained on unshuffled data can outperform BERT trained on shuffled data. The results are summarized in Table 1.
+
+We compare these results to a logistic regression model that takes in the frequencies for the 100 most common vocabulary items in the replaced corpus $^{4}$  (plus a catch-all category for the remaining items) and the sentence length as features. Such a logistic regression model achieves an accuracy of  $85.45\%$ , which means it performs about as well as BERT fine-tuned on shuffled tags but loses to BERT when the latter is fine-tuned on unshuffled tags. Using sentence length as the only feature yields an accuracy of  $60.33\%$ .
+
+The conclusion of this experiment is that the fine-tuned BERT model must be using syntactic information to tell apart sentences from the 1830s and the 1990s. Indeed, even when the length of the sentence is fixed, and even when the words have been replaced by their parts of speech, the model gains advantage from seeing the true order of the tags as they appeared in the sentence, instead of seeing shuffled tags. This advantage must fundamentally come from differences in syntax that are present between the 1830s and 1990s sentences (and moreover, these differences must take a form that is not simply tag counts).
+
+Having established that BERT is sensitive to changes in the syntax of English, we next train BERT in a different way to identify specific instances of such change. Importantly, from here on
+
+<table><tr><td>sent.
+length</td><td>shuffled
+placed</td><td>unshuffled
+placed</td><td>shuffled
+plain</td><td>unshuffled
+plain</td></tr><tr><td>11-15</td><td>83.37</td><td>87.74</td><td>94.70</td><td>96.69</td></tr><tr><td>21-25</td><td>86.44</td><td>90.54</td><td>97.39</td><td>98.63</td></tr><tr><td>31-35</td><td>89.37</td><td>93.12</td><td>98.20</td><td>98.90</td></tr><tr><td>20</td><td>85.20</td><td>89.30</td><td>96.60</td><td>97.40</td></tr></table>
+
+Table 1: Accuracy of BERT fine-tuned on different types of data for predicting time period. All models were trained on 20,000 sentences and tested on 2,000 (half from the 1830s and half from the 1990s); the exception was the last line, for which only 10,000 training sentences were used due to lack of data. Since test data contained an even mixture of 1830s and 1990s sentences, baseline accuracy is  $50\%$ . These numbers are based on a single run; however, note that an accuracy of  $p$  will have standard deviation around  $\sqrt{p(1 - p) / 2000}$ , which ranges from  $0.23\%$  to  $0.83\%$  for the above, with higher accuracies having lower standard deviations.
+
+out, we avoid using a POS tagger on older English.
+
+# 3 Training Classifiers
+
+We use BERT to identify specific instances of change. To do so, we train two POS taggers, one based on text from the 1830s and the other based on text from the 1990s; in the subsequent sections, we will use these taggers and their errors to identify words or sentences that display syntactic change. In this section, we will focus on describing how these taggers are trained and their performance. Crucially, we will train these taggers without ever tagging text from the 1830s; all POS tags are only ever associated with text from the 1990s.
+
+We start by fine-tuning two instances of BERT-base-uncased: one on text from the 1990s and the other on text from the 1830s. This is done via Masked-Language Modeling in PyTorch (Paszke et al., 2019), where we process one sentence at a time with  $15\%$  of the tokens masked, and the model is trained to predict the masked tokens. We use each fine-tuned model to obtain BERT token representations for 1990s text and then average the token representations for all the WordPiece tokens within each word to obtain a representation of a word. We do this twice: once with the model fine-tuned on text from the 1830s and once with the model fine-tuned on text from the 1990s (yielding two different embeddings of each word). However, even though we have models trained on two time periods, we use both models only to embed words from the 1990s and not any words from the 1830s. At the end of this process, each word from the 1990s has two different BERT embeddings.
+
+Next, using POS tags obtained from Stanford CoreNLP for the 1990s portion of the British Hansard, we train a 2-layer perceptron (with one hidden layer of size 50) to assign a Penn Treebank tag to each BERT representation of each word. An input to this 2-layer perceptron is a BERT embedding of a word (extracted as described above), and the output is a POS tag. This classification task ignores punctuation and has 21 classes: We merge the 'NN', 'NNP', 'NNS', and 'NNPS' tags into one single tag called 'Noun', after which we take the 20 most common tags ('Noun', 'IN', 'DT', 'JJ', 'VB', 'PRP', 'RB', 'VBN', 'VBZ', 'VBP', 'CC', 'MD', 'TO', 'VBD', 'CD', 'VBG', 'PRPS', 'WDT', 'WP', 'WRB') and then group the remaining POS tags into a catch-all class called 'other'. We actually train two such perceptrons separately: one for the 1830s BERT embeddings (of words from the 1990s only) and one for the 1990s BERT embeddings (also only of words from the 1990s).
+
+Note that the training and test data for this procedure consist only of sentences from the 1990s; this is done because the POS tags obtained from Stanford CoreNLP are likely to be more accurate on our modern data. Using the representations from the final layer of BERT (when the classifier has access to 50,000 training sentences, containing 815,578 words, and when the test set has 50,000 sentences, containing 816,428 words), the accuracy obtained by the 1990s model is  $97.6\%$ , and the accuracy of the 1830s model is  $97.2\%$ .
+
+We repeat the above analysis using POS tags from Trankit (Nguyen et al., 2021) instead of Stanford CoreNLP because Trankit is a state-of-the-art parser and tagger. By using Trankit tags to train and test the above classifiers based on BERT representations, the accuracy obtained by the 1990s model was  $98.4\%$  (instead of  $97.6\%$  using CoreNLP tags) and the accuracy obtained by the 1830s model was  $98.0\%$  (instead of  $97.2\%$  using CoreNLP tags). The higher accuracy means that the task became easier when predicting Trankit tags instead of CoreNLP tags; since Trankit tags are also more accurate, it suggests that our pipeline is working as expected.
+
+In §4 and later, we will use these pipelines (and the differences between their predictions) to iden-
+
+tify specific instances of syntactic change between the 1830s and the 1990s.
+
+# 4 Identifying Words that Underwent Change
+
+We now attempt to use the POS taggers trained above to identify words that underwent change. We employ two methods to do this: First, we find the instances for which the two trained models disagree on a POS tag (indicating that the model familiar with text from the 1830s would tag the word differently compared to the one familiar with text from the 1990s). The second method involves masking out the target word to see if the 1830s model predicts the tag correctly on the masked word. If the masked word is tagged correctly but the un-masked word is tagged incorrectly, it suggests that the 1830s model can correctly parse the rest of the sentence but is specifically confused by the target word; when this happens, it indicates that the target word may have undergone change. This latter strategy is depicted in Figure 1.
+
+# 4.1 Tagging Disagreements
+
+After training, we have two new POS taggers, one based on an 1830s understanding of English and one based on a 1990s understanding. We now run these taggers on test data (from the 1990s) to find words on which they disagree. In particular, we find words for which the model trained on text from the 1990s agrees with the Trankit tag but where the 1830s model disagrees. The cases where such a disagreement occurs should be the cases in which an 1830s grammar leads the model to incorrectly parse the sentence (likely due to a change in English grammar for the relevant word or perhaps one of its neighbors). Moreover, we use our classifiers to extract probabilistic predictions for the tags, and we search for word instances where probability assigned to the correct tag by the 1830s model was at least 50 percentage points lower than that assigned by the 1990s model.
+
+We do this for each word in each sentence of the 1990s test data. To help isolate interesting cases, we filter the words under consideration to those that occur at least 10 times in the 1830s text (indicating that they are not new words). We remove numbers, punctuation, and words written entirely in capital letters (likely acronyms). We also restrict to words for which such a tagging disagreement occurred at least twice. We then sort the resulting words
+
+by the total number of tag disagreements for that word, normalized by the frequency of the word in the 1990s data, with additional weight given to words which were common in the 1830s (to reduce the number of rare words in the final list). We run this experiment on 100,000 test sentences, using classifiers trained on 25,000 sentences each.
+
+We then manually examine the top resulting words to check whether the identified words have truly undergone change; see §4.3. One example can be seen in the following sentence: "My right hon. Friend last met M. Chevenement to discuss cooperative weapons development among other topics on 4 May." In this example, the model trained on text from the 1990s tags the word "co-operative" as an adjective whereas the model trained on the 1830s believes that "co-operative" is a noun; the 1990s model is correct in this case.
+
+# 4.2 Masking and Confusion
+
+As before, we use both the 1830s model and the 1990s model to predict the POS tag of each word in held-out 1990s sentences. This time, we additionally repeat this process with one word in the sentence masked out: we ask both the 1990s model and the 1830s model to predict the POS tag of the masked word (without knowing what it is). We mask out each word in each sentence, one at a time; since the BERT embedding for a single word depends on the entire sentence, this requires passing the same sentence repeatedly through BERT with a different word masked each time. To speed up this process, we skip applying the mask to words that are too common (within the one thousand most common words in the 1990s), since these constitute the majority of word occurrences and yet are less likely to exhibit syntactic change.
+
+After this process, each word in each sentence (except those we have deliberately skipped) will have four different contextual BERT embeddings: one by the 1830s model when the word was masked, one by the 1830s model when the word was unmasked, one by the 1990s model when the word was masked, and one by the 1990s model when the word was unmasked.
+
+We now examine words for which the 1830s model gives a better prediction of the POS tag when the word is masked compared to when it is unmasked. When this happens, it means that the 1830s model correctly identified which tag to expect in the position of the masked word, and yet
+
+![](images/8e1b734a3e8f9f745a0dab9b7fe9a79581b457fe1d249adea6f45db393b179fa.jpg)  
+There is no example of sovereign countries which have given up their currencies. There is no example of [MASK] countries which have given up their currencies.  
+Figure 1: A depiction of the models' predictive process. The 1830s model predicts the correct tag when the word "sovereign" was masked (the tags assigned the highest likelihood after JJ were VBG with  $21\%$  and DT with  $20\%$  probability) but predicts it incorrectly when it was unmasked; this is the opposite of the usual pattern, in which the prediction is worse in the masked scenario. This type of tagging error indicates that the 1830s model expects "sovereign" to be a noun instead of an adjective.
+
+after the mask is lifted (and the word itself was revealed) the 1830s model "changed its mind" and assigned the wrong tag instead. In such cases, we say that the 1830s model was "confused" by the token. We restrict to cases where the 1990s model correctly tagged the word, even when it was un-masked; this helps remove Trankit tagging errors as well as typographical errors in the text.
+
+We then consider the remaining words for which the 1830s BERT model was confused more than once; specifically, we restrict to words for which there were at least two separate occurrences where the 1990s model tagged the word correctly (with a confidence of at least  $90\%$ ) and the 1830s model assigned a probability that is at least 25 percentage points higher to the correct POS tag when seeing the masked version of the sentence than it did when seeing the unmasked version. We run this experiment on 100,000 test sentences; to train the 1990s and 1830s models, we use 25,000 sentences each (just as in the previous section). To help isolate interesting cases, we employ the same filters and the same sorting strategy as we did in §4.1.
+
+An example of a word that we found this way is "material", which was an adjective in the early 1800s but whose noun usage is modern. For instance, in the sentence "the inquiry will be free to publish whatever material it wishes," Trankit correctly tags "material" as a noun, and so does the 1990s pipeline, but the 1830s pipeline only gives the correct tag when the word "material" is masked out in that sentence; when it is unmasked, an adjective is predicted.
+
+A list of the top resulting words (with exam-
+
+ple sentences and tags provided by the models) is given in Appendix B. The words in this table were automatically generated with no manual curation.
+
+# 4.3 Evaluation
+
+To evaluate the resulting words, we manually analyzed the top 20 candidates produced by §4.1 and the top 20 candidates produced by §4.2. For each candidate word, we examined 10 randomly sampled sentences containing that word from the 1990s and also 10 sentences from the 1830s, and we manually checked whether their syntax appeared to have changed.
+
+Of the top 20 words from §4.1, seven (or  $35\%$ ) were identified as having undergone syntactic change, with a part of speech in the 1990s that is either new or greatly increased in frequency. Of the top 20 words from §4.2, 11 (or  $55\%$ ) were identified as having undergone syntactic change.
+
+We note also that these two approaches identified different sets of words: Of the top 20 words on each list, there were only 4 words of overlap, while of the top 100 there were 15 words of overlap.
+
+We also attempt an automatic evaluation by using Trankit to tag words from the 1830s and checking if these tags changed. Of course, our identification methods purposely avoided using a tagger on old text, since it may be unreliable, and a manual check reveals that the tagger made mistakes. Still, as a sanity check, we use it to compare the two methods for identifying syntactic change.
+
+For the purpose of automatic evaluation, we say that a word has changed and gained a new POS if the tag is at least twice as likely to appear for
+
+1990s usages of the word than for 1830s usages of the word. Under this criterion for change, and using the Trankit tags for both 1830s and 1990s data, we get that the top 100 candidates from the simpler identification method (\$4.1) contained 20 changed words, while the top 100 candidates from our masking strategy (\$4.2) had 35 changed words. Note that of the manually evaluated words, the automatic evaluation agreed with  $75\%$  of the human judgements. These should be interpreted with caution, but these results agree with our conclusion that the masking strategy is better at identifying changed words.
+
+# 5 Confusion Scores at the Sentence Level
+
+In this next experiment, we try to discern the extent to which masking out a word in a sentence makes the sentence more difficult for a fine-tuned BERT model to parse. The idea is that, for a typical word, once the word is masked out the model has less information about the sentence, and its syntactic understanding should therefore decrease; on the other hand, for a word that has an atypical or new usage, the model may end up being more confused after seeing the word than it is when the word is masked out. In the latter situation, we would expect the model's syntactic understanding of the sentence to improve if the word is masked out because, from the model's point of view, the word is misleading. A similar idea was used in §4.2 at the word level; in this section we will look at the accuracy of the tags for the whole sentence before and after a single word is masked.
+
+For example, consider the following sentence: "Opposition Members think that is how peace processes work." A model trained on English from the 1830s is likely to think that the word "work" is a noun rather than a verb. This could lead such a model to consider an incorrect way to parse this sentence, where the subject "peace" is doing the action of processing the object "work". In this case, if the word "work" is masked out, the sentence will be more likely to be parsed correctly.
+
+As a proxy for the success of the parsing of the sentence, we feed in the sentence to the classifiers from §3 to get predicted tags for all the words in the sentence. We can use these classifiers to get probabilistic predictions, and in particular, we can get a probability assigned to the correct tag for each word in the sentence (using Trankit tags as ground truth). We then multiply together all
+
+these probabilities to get a measure of the overall probability the model assigns to the correct tags of the sentence. We can repeat this for both the 1830s and 1990s classifiers, and we can also repeat this when a specific word is either masked or unmasked.
+
+We say a model is "confused" if the product of probabilities assigned to the correct (Trankit) tags is higher when the word is masked than when it is unmasked. For instance, in the above example sentence "Opposition Members think that is how peace processes work," the 1830s model may assign a wrong tag to the word "processes" when the word "work" is unmasked, but it might correctly parse the sentence if the word "work" is masked (in the latter case, it will assume the masked word is an unknown verb, while in the former case it will try to parse the sentence with a noun in that position if it believes that "work" is a noun). In other words, if the 1830s model is confused by a word, it could be because this word underwent change from the 1830s to the 1990s. We create a "confusion score" by taking the logarithm of the ratio of the two aforementioned products (i.e. the product of probabilities assigned to the other words when the selected word is masked, and the corresponding product when the selected word is unmasked).
+
+We can get such a confusion score for any word in any sentence, and for either the 1830s model or the 1990s model. We focus on sentences from the 1990s that were not part of the training data. Since calculating confusion scores is resource-intensive, we focus on a pre-selected set of word instances, such as those identified as likely having undergone syntactic change in §4.
+
+# 5.1 Comparing Selected Words to Random Words
+
+We select 10,000 word occurrences for which the tag is predicted differently by the 1830s model and the 1990s model (as discussed in §4.1). For each word occurrence, we calculate a confusion score. We then compare these scores to the confusion scores of 10,000 randomly chosen word instances. Our goal is to check whether the confusion scores of the 10,000 selected words were higher than those of random words.
+
+One possible confounder is that the selected words may be rarer than the random words. To adjust for this, we choose the random words more carefully: for each non-random word in the first set, we pick a random word conditional on its overall
+
+![](images/479a0c1717d54892947873d4695c0c65fa0259c0e4eb5ced8bb051b1633e021f.jpg)  
+Figure 2: Distribution of confusion scores for selected words and random words.
+
+frequency being close to that of the original word. We define close to be within 10 percentage points in the percentile of their frequencies. The resulting distribution of the confusion scores for the random words had lighter tails than the distribution of the selected words, especially on the right-hand side of the distribution. See Figure 2.
+
+As the figure shows, the selected word instances (for which the 1990s model and 1830s model disagree on the tags) have a higher proportion of high confusion scores than random words. Therefore, calculating confusion scores for the selected word occurrences is a more resource-efficient way to find instances of high confusion scores.
+
+# 5.2 Further Refinements
+
+In addition to calculating confusion scores for a word in a given sentence using the 1830s model, we also calculate confusion scores in a similar way using the 1990s model. Recall that the confusion score tells us the extent to which the BERT-based tagger performs better when a word is masked out than when it is not. Since we have two such taggers (one fine-tuned on parliamentary speeches from the 1830s and one fine-tuned on speeches from the 1990s), we can get two confusion scores for a single word in a single sentence.
+
+In order to zoom in on syntactic change, we restrict our attention to instances where the 1990s confusion score is less than or equal to zero. Since the confusion score is the logarithm of the ratio of masked to unmasked probabilities, this means we are restricting to word instances where the 1990s model gave a better prediction of the tags when the word was unmasked than it did when the word was masked. Applying this restriction helps to remove cases that are confusing for reasons other
+
+than language change: for example, sentences that contain typographical errors, sentences that are incorrectly parsed by Trankit, sentences that were ungrammatical when originally spoken, or other special cases. By focusing on instances that the 1830s model found confusing but that the 1990s model did not find confusing, we increase the likelihood of finding syntactic change.
+
+# 5.3 Manual Evaluation
+
+We generate a list of top 100 candidates for syntactic change based on these confusion scores. Of those 100 sentences, our manual evaluation determined that 33 sentences were valid instances of syntactic change. We also computed the sum of the reciprocal ranks of the 33 manually selected sentences that exhibited syntactic change. This sum was 2.802, while the maximum possible sum of reciprocal ranks would be  $\frac{1}{1} + \frac{1}{2} + \ldots + \frac{1}{33} \approx 4.089$ .
+
+# 6 Detecting the Rise of the Progressive
+
+In this section, we check the extent to which our techniques can identify syntactic change that was previously known to exist.
+
+The frequency of the progressive has been increasing over time (e.g. Leech et al., 2009; Hou and Smith, 2021). An example of a progressive is "he is reading" (as opposed to "he reads"); the former has a form of "to be" followed by a verb that ends in "-ing". We identify progressive constructions using the dependency parse obtained from Trankit (Nguyen et al., 2021). We then mask out the present participle if it has the tag "VBG" and if the previous word is an inflection of the verb "to be". (This means that we mask out the tokens constituting the "-ing" word as well as the "-ing" suffix itself.)
+
+Next, we use both the 1990s model and the 1830s model to predict the POS tags of all the words in these masked sentences. We compare these predictions to each other and also to the predictions that both models give for the unmasked versions of these sentences. Finally, we take randomly chosen words in randomly chosen sentences as a control group and repeat the process for this group.
+
+We summarize the results in two figures. First, Figure 3 shows the confusion scores for the 1830s model on the progressive sentences compared to random sentences. As the figure shows, the 1830s model is negatively confused when we mask out progressives, which means that it gets the predic
+
+![](images/e4da15d8762aab18a0de5e880f8a1957059d101324d1d04ea001633ad1c4ea47.jpg)  
+Figure 3: A histogram for the confusion scores of the 1830s model on progressives and on random sentences.
+
+tion right more often on the unmasked version of these sentences than it does on the masked versions. This makes sense, as it corresponds to the insight that the 1830s model has a difficult time correctly tagging the masked progressives.
+
+![](images/1610561e4206ac7e202022963eebc08063fe57eca2dc6d970b421f4d51c7bd97.jpg)  
+Figure 4: A histogram for the logarithm of the ratios between the products of probabilities assigned to the true tags by the 1990s model and the 1830s model. We compare this for progressive sentences and random sentences. In all cases, both models were given the masked version of the sentences.
+
+Figure 4 compares the predictions of the 1990s model to those of the 1830s model for the masked sentences. It shows that the 1990s model predicts the tags of the progressive sentences better than the 1830s model does, and the gap between their predictions is larger than it is for random sentences. This is evidence that the 1830s model is unfamiliar with progressive constructions.
+
+# 7 Related Work
+
+It has been claimed that the performance of Transformers on downstream tasks is insensitive to word order (Hessel and Schofield, 2021; Sinha et al., 2021a; Sinha et al., 2021b; Pham et al., 2021;
+
+Gupta et al., 2021), which is consistent with our observation that BERT fine-tuned on shuffled text can distinguish between two time periods with high accuracy; however, our experiments indicate BERT's accuracy still improves when it has access to word order, consistent with Papadimitriou et al. (2022), who argue that word order sometimes matters by considering non-prototypical cases where meanings and lexical expectations do not match.
+
+A sizable literature considers syntactic probes for BERT embeddings (e.g. Tenney et al., 2019; Jawahar et al., 2019; Hewitt and Manning, 2019; Pimentel et al., 2020), but to our knowledge none of the works that probe syntax consider Transformers fine-tuned on different time periods.
+
+Syntactic change detection has been understudied in the computational linguistics and NLP literature. One work on this topic is Merrill et al. (2019), in which an LSTM achieved higher accuracy compared to a model consisting of feedforward networks on the task of predicting the year of composition of sentences. However, in addition to predicting the time period of text, we also identify specific instances of syntactic change. Computational methods for the study of syntactic change also appear in Hou and Smith (2018); this work examined possible causes of the decline in the rate of passivization over time. Another related work is Lei and Wen (2020), which found dependency distance to be decreasing over time (between the years 1790 and 2017).
+
+Many studies have used BERT to detect semantic change, including Giulianelli et al. (2020), Hu et al. (2019), and Fonteyn (2020). In addition, Beelen et al. (2021) fine-tune historical BERT models on a corpus of  $19^{\mathrm{th}}$ -century English books in order to tackle the task of time-sensitive Targeted Sense Disambiguation and find that "BERT works better when it is fine-tuned on data contemporaneous to the target period".
+
+# 8 Conclusion
+
+We investigated the extent to which fine-tuned BERT models can distinguish between the syntax of 1830s English and 1990s English. Our approach used masking techniques, and we repeatedly compared BERT models on masked and unmasked text.
+
+First, in §2, we hid semantic information by replacing each word with its POS tag (and representing that tag as a common English word that exemplifies that tag). We fine-tuned BERT on such
+
+syntax-only data to see if it could tell apart text from the 1830s and text from the 1990s. We found evidence that BERT can be used to detect syntactic differences between these two time periods.
+
+In the later sections, we masked out individual words to see whether a POS tagger based on BERT embeddings (fine-tuned on either 1830s or 1990s) can still tag correctly when the word is masked. We filtered for a specific type of error, where the 1830s model predicts the correct tag when the word is masked but not when it is unmasked, and where the 1990s model is correct even when the word is unmasked. This procedure identified candidates for syntactic change, examples of which we provide in Appendix B. We were also able to confirm the rise of the progressive construction.
+
+An important point to highlight is that our identification of syntactic change relied on an automatic POS tagger only for the 1990s text; we avoided running any automatic tagging software on older English data. In fact, from §3 onwards, all we did with the older 1830s data was to fine-tune BERT on it in an unsupervised fashion. Despite this restriction, we were able to automatically identify words whose part of speech had undergone change between the two time periods.
+
+# Limitations
+
+There are several limitations to this work. First, we worked only with English and only with text from the 1830s and 1990s. Our work also required the use of GPUs; see Appendix A for details of resources used. Although we were able to automatically identify specific instances of change (and we show automatically generated examples in Appendix B), our evaluation was limited to the manual judgement of one author, and we were unable to determine the recall of our detection system (because of the lack of objective criteria for what constitutes syntactic change). Additionally, our techniques are limited in the kinds of change we can detect: we focus on finding words whose typical part of speech has changed and on validating known syntactic changes such as the rise of the progressive. We hope that similar techniques can be extended to other types of syntactic change in future work.
+
+# Acknowledgments
+
+We thank Jacob Eisenstein and the anonymous reviewers for their helpful feedback.
+
+# References
+
+Marc Alexander and Mark Davies. 2015. The Hansard Corpus 1803-2005. http://www.hansard-archive.parliament.uk/.  
+Kaspar Beelen, Federico Nanni, Mariona Coll Ardanuy, Kasra Hosseini, Giorgia Tolfo, and Barbara McGillivray. 2021. When time makes sense: A historically-aware approach to targeted sense disambiguation. In Findings of ACL-IJCNLP, pages 2751-2761.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In *NAACL*, pages 4171–4186.  
+Lauren Fonteyn. 2020. What about grammar? Using BERT embeddings to explore functional-semantic shifts of semi-lexical and grammatical constructions. Computational Humanities Research, pages 257-268.  
+Mario Giulianielli, Marco Del Tredici, and Raquel Fernandez. 2020. Analysing lexical semantic change with contextualised word representations. In ACL, pages 3960-3973.  
+Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In *NAACL*, pages 1195–1205.  
+Ashim Gupta, Giorgi Kvernadze, and Vivek Srikumar. 2021. BERT & family eat word salad: Experiments with text understanding. In AAAI, pages 12946-12954.  
+Jack Hessel and Alexandra Schofield. 2021. How effective is BERT without word ordering? implications for language understanding and data privacy. In ACLIJCNLP, pages 204-211.  
+John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In *NAACL*, pages 4129-4138.  
+Liwen Hou. 2022. Detecting and Modeling Syntactic Change. Ph.D. thesis, Northeastern University.  
+Liwen Hou and David A Smith. 2018. Modeling the decline in English passivization. In SCiL, pages 34-43.  
+Liwen Hou and David A. Smith. 2021. Drivers of English syntactic change in the Canadian parliament. In *SCiL*, pages 51–60.  
+Renfen Hu, Shen Li, and Shichen Liang. 2019. Diachronic sense modeling with deep contextualized word embeddings: An ecological view. In ACL, pages 3899-3908.  
+Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In ACL, pages 3651-3657.
+
+Geoffrey Leech, Marianne Hundt, Christian Mair, and Nicholas Smith. 2009. Change in Contemporary English: A Grammatical Study. Studies in English Language. Cambridge University Press.  
+Lei Lei and Ju Wen. 2020. Is dependency distance experiencing a process of minimization? A diachronic study based on the State of the Union addresses. *Lingua*, 239:102762.  
+Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In ACL System Demonstrations, pages 55-60.  
+William Merrill, Gigi Stark, and Robert Frank. 2019. Detecting syntactic change using a neural part-of-speech tagger. In International Workshop on Computational Approaches to Historical Language Change, pages 167-174.  
+Minh Van Nguyen, Viet Dac Lai, Amir Pouran Ben Veyseh, and Thien Huu Nguyen. 2021. Trankit: A lightweight transformer-based toolkit for multilingual natural language processing. In EACL System Demonstrations, pages 80-90.  
+Isabel Papadimitriou, Richard Futrell, and Kyle Mahowald. 2022. When classifying grammatical role, BERT doesn't care about word order... except when it matters. In ACL, pages 636-643.  
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8024-8035.  
+Thang Pham, Trung Bui, Long Mai, and Anh Nguyen. 2021. Out of order: How important is the sequential order of words in a sentence in natural language understanding tasks? In Findings of ACL-IJCNLP, pages 1145-1160.  
+Tiago Pimentel, Josef Valvoda, Rowan Hall Maudslay, Ran Zmigrod, Adina Williams, and Ryan Cotterell. 2020. Information-theoretic probing for linguistic structure. In ACL, pages 4609-4622.  
+Koustuv Sinha, Robin Jia, Dieuwke Hupkes, Joelle Pineau, Adina Williams, and Douwe Kiela. 2021a. Masked language modeling and the distributional hypothesis: Order word matters pre-training for little. In EMNLP, pages 2888-2913.  
+Koustuv Sinha, Prasanna Parthasarathi, Joelle Pineau, and Adina Williams. 2021b. UnNatural Language Inference. In ACL-IJCNLP, pages 7329-7346.  
+Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In ACL, pages 4593-4601.
+
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In EMNLP System Demonstrations, pages 38-45.
+
+# A Resources Used
+
+In this work, we mentioned and used the following licensed resources: BERT (Devlin et al., 2019), which uses Apache 2.0 license; UK Parliament Hansard Corpus (Alexander and Davies, 2015), which uses Open Parliament Licence; Trankit large version 1.1.0 (Nguyen et al., 2021), which uses Apache 2.0 license; and Stanford CoreNLP version 4.2.2 (Manning et al., 2014), which uses the GNU General Public License v3.0. All of these licenses allow the type of research conducted in our work. Moreover, our work falls within the intended use of each of these resources.
+
+Our work required GPU computation. We used NVIDIA Tesla V100 and P100 GPUs over several months while working on the results of this paper. Running Trankit was the slowest step and took around 72 GPU hours; the total GPU time to reproduce our final results would be under a week.
+
+# B Automatically Identified Words
+
+Table 2 shows the top of the list of automatically identified words whose part-of-speech usage may have changed; a longer list can be found in Appendix A of the first author's thesis (Hou, 2022). For each word, we give an example sentence from the 1990s on which the 1830s model was "confused". For each example sentence, we give the Trankit tag and two sets of probabilistic tag predictions given by the 1830s model; one set corresponds to the tags predicted by the 1830s model for the highlighted word when the word was masked, and the other set corresponds to the tags predicted by the 1830s model for the highlighted word when the word was unmasked. (For brevity, we only list the tag predictions to which the model assigns at least  $30\%$  probability.) An arrow such as 'JJ  $\rightarrow$  Noun' indicates that the correct tag in the 1990s was 'Noun' but that the unmasked 1830s model put a high weight on 'JJ' instead; this suggests the word changed from 'JJ' in the 1830s to 'Noun' in the 1990s. (Recall that all POS tags for nouns were merged into a single 'Noun' category.)
+
+<table><tr><td rowspan="2">opposite
+JJ→Noun</td><td colspan="3">“The mandatory life sentence does the opposite of what it sets out to achieve, and it should go.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 84% Noun</td><td>unmasked tag: 93% JJ</td></tr><tr><td rowspan="2">presume
+VB→VBP</td><td colspan="3">“I presume that that shows they will be on message until they die.”</td></tr><tr><td>Trankit: VBP</td><td>masked tag: 98% VBP</td><td>unmasked tag: 38% VB, 32% VBP</td></tr><tr><td rowspan="2">corporal
+Noun→JJ</td><td colspan="3">“The school board wrote to me on 29 October advocating the reintroduction of corporal punishment in schools as a deterrent of last resort.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 74% JJ</td><td>unmasked tag: 95% Noun</td></tr><tr><td rowspan="2">intrepid
+JJ→Noun</td><td colspan="3">“The second category are the Royal Marine&#x27;s principal transport and landing ships, Fearless and Intrepid.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 64% Noun</td><td>unmasked tag: 97% JJ</td></tr><tr><td rowspan="2">evil
+Noun→JJ</td><td colspan="3">“The gunmen, in the evil atrocity that they committed, symbolised the past.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 82% JJ</td><td>unmasked tag: 99% Noun</td></tr><tr><td rowspan="2">remembered
+JJ→VBN</td><td colspan="3">“It should be remembered that, during the period 1974 – 79, when he was a senior Cabinet Minister, the steel industry was inefficient and unprofitable.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 86% VBN</td><td>unmasked tag: 64% JJ, 36% VBN</td></tr><tr><td rowspan="2">shut
+VBD→VBN</td><td colspan="3">“My second charge is that the Prime Minister has consistently talked about the incentive society with open opportunities for all, yet her policies have effectively shut out a quarter of the nation.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 97% VBN</td><td>unmasked tag: 86% VBD</td></tr><tr><td rowspan="2">offending
+JJ→Noun</td><td colspan="3">“I can vouch for the fact that juvenile offending is a major problem in my constituency.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 98% Noun</td><td>unmasked tag: 58% JJ</td></tr><tr><td rowspan="2">granting
+VBG→Noun</td><td colspan="3">“In most normal cases there will be the usual opportunity for consultation and granting of opinion from the local community.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 99% Noun</td><td>unmasked tag: 98% VBG</td></tr><tr><td rowspan="2">disposed
+Noun→VBN</td><td colspan="3">“Mr. Streeter: To ask the Secretary of State for Defence what Ministry of Defence land and buildings in Plymouth he expects to be surplus to MOD requirement in the next five years and to be disposed of to the private sector.”</td></tr><tr><td>Trankit: VBN</td><td>masked tag: 87% VBN</td><td>unmasked tag: 48% Noun, 31% VBN</td></tr><tr><td rowspan="2">exciting
+VBG→JJ</td><td colspan="3">“We should also encourage housing associations to feel the exciting and exhilarating wind of competition.”</td></tr><tr><td>Trankit: JJ</td><td>masked tag: 99% JJ</td><td>unmasked tag: 94% VBG</td></tr><tr><td rowspan="2">laying
+VBG→Noun</td><td colspan="3">“More significant orders, especially those exercising the so-called Henry VIII powers, are to be subject to laying in draft and affirmative resolution.”</td></tr><tr><td>Trankit: Noun</td><td>masked tag: 98% Noun</td><td>unmasked tag: 98% VBG</td></tr><tr><td rowspan="2">notwithstanding
+VBG→IN</td><td colspan="3">&quot;Notwithstanding signal failures and operational difficulties, I look forward to using that fast service.”</td></tr><tr><td>Trankit: IN</td><td>masked tag: 84% IN</td><td>unmasked tag: 86% VBG</td></tr></table>
+
+Table 2: Automatically identified words whose part-of-speech usage may have changed

2023/Detecting Syntactic Change with Pre-trained Transformer Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9504779080062c63524025b74ae9e8c5c4af4136fe8e9b69a4859cac0c9edba6
+size 489339